├── .gitignore ├── .pre-commit-config.yaml ├── LICENSE ├── README.md ├── core ├── Conv_AE.py ├── Isolation_Forest.py ├── LSTM_AE.py ├── LSTM_VAE.py ├── MSCRED.py ├── MSET.py ├── Vanilla_AE.py ├── Vanilla_LSTM.py ├── __init.py__ ├── metrics.py ├── t2.py └── utils.py ├── data ├── README.md ├── anomaly-free │ └── anomaly-free.csv ├── other │ ├── 1.csv │ ├── 10.csv │ ├── 11.csv │ ├── 12.csv │ ├── 13.csv │ ├── 14.csv │ ├── 2.csv │ ├── 3.csv │ ├── 4.csv │ ├── 5.csv │ ├── 6.csv │ ├── 7.csv │ ├── 8.csv │ └── 9.csv ├── valve1 │ ├── 0.csv │ ├── 1.csv │ ├── 10.csv │ ├── 11.csv │ ├── 12.csv │ ├── 13.csv │ ├── 14.csv │ ├── 15.csv │ ├── 2.csv │ ├── 3.csv │ ├── 4.csv │ ├── 5.csv │ ├── 6.csv │ ├── 7.csv │ ├── 8.csv │ └── 9.csv └── valve2 │ ├── 0.csv │ ├── 1.csv │ ├── 2.csv │ └── 3.csv ├── docs ├── contributing.md └── pictures │ ├── nab-metric.jpg │ ├── skab.png │ └── testbed.png ├── notebooks ├── ArimaFD.ipynb ├── Conv_AE.ipynb ├── LSTM_AE.ipynb ├── MSET.ipynb ├── README.md ├── Vanilla_AE.ipynb ├── Vanilla_LSTM.ipynb ├── isolation_forest.ipynb ├── mscred.ipynb ├── t2_SKAB.ipynb └── t2_with_q_SKAB.ipynb ├── poetry.lock ├── pyproject.toml └── results ├── results-Arima_anomaly_detection.pkl ├── results-Conv_AE.pkl ├── results-Isolation_Forest.pkl ├── results-LSTM_AE.pkl ├── results-MSCRED.pkl ├── results-MSET.pkl ├── results-T2-q.pkl ├── results-T2.pkl ├── results-Vanilla_AE.pkl └── results-Vanilla_LSTM.pkl /.gitignore: -------------------------------------------------------------------------------- 1 | .DS_Store 2 | *.pickle 3 | 4 | # Byte-compiled / optimized / DLL files 5 | __pycache__/ 6 | algorithms/__pycache__/ 7 | notebooks/__pycache__/ 8 | *.py[cod] 9 | *$py.class 10 | 11 | # Distribution / packaging 12 | .Python 13 | build/ 14 | develop-eggs/ 15 | dist/ 16 | downloads/ 17 | eggs/ 18 | .eggs/ 19 | lib/ 20 | lib64/ 21 | parts/ 22 | sdist/ 23 | var/ 24 | wheels/ 25 | pip-wheel-metadata/ 26 | share/python-wheels/ 27 | *.egg-info/ 28 | .installed.cfg 29 | *.egg 30 | MANIFEST 31 | 32 | # PyInstaller 33 | # Usually these files are written by a python script from a template 34 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 35 | *.manifest 36 | *.spec 37 | 38 | # Installer logs 39 | pip-log.txt 40 | pip-delete-this-directory.txt 41 | 42 | # Unit test / coverage reports 43 | htmlcov/ 44 | .tox/ 45 | .nox/ 46 | .coverage 47 | .coverage.* 48 | .cache 49 | nosetests.xml 50 | coverage.xml 51 | *.cover 52 | *.py,cover 53 | .hypothesis/ 54 | .pytest_cache/ 55 | 56 | # Jupyter Notebook 57 | .ipynb_checkpoints 58 | notebooks/.ipynb_checkpoints 59 | algorithms/.ipynb_checkpoints 60 | 61 | # IPython 62 | profile_default/ 63 | ipython_config.py 64 | 65 | # pyenv 66 | .python-version 67 | 68 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow 69 | __pypackages__/ 70 | 71 | # Environments 72 | .env 73 | .venv 74 | env/ 75 | venv/ 76 | ENV/ 77 | env.bak/ 78 | venv.bak/ 79 | -------------------------------------------------------------------------------- /.pre-commit-config.yaml: -------------------------------------------------------------------------------- 1 | # https://pre-commit.com 2 | exclude: 'examples|reports' 3 | repos: 4 | - repo: https://github.com/pre-commit/pre-commit-hooks 5 | rev: v4.5.0 6 | hooks: 7 | - id: debug-statements #Check for debugger imports and breakpoint() in python files 8 | - id: check-ast #Simply check whether files parse as valid python 9 | - id: fix-byte-order-marker #removes UTF-8 byte order marker 10 | - id: check-json 11 | - id: detect-private-key # detect-private-key is not in repo 12 | - id: check-yaml 13 | - id: check-added-large-files 14 | - id: check-shebang-scripts-are-executable 15 | - id: check-case-conflict #Check for files with names that would conflict on a case-insensitive filesystem like MacOS HFS+ or Windows FAT 16 | - id: end-of-file-fixer #Makes sure files end in a newline and only a newline 17 | - id: trailing-whitespace 18 | - id: mixed-line-ending 19 | - repo: https://github.com/astral-sh/ruff-pre-commit 20 | rev: v0.3.4 21 | hooks: 22 | - id: ruff 23 | args: [--fix, --exit-non-zero-on-fix, --line-length=79] 24 | types_or: [python, pyi, jupyter] 25 | - id: ruff-format 26 | args: [--line-length=79] 27 | types_or: [python, pyi, jupyter] 28 | - repo: https://github.com/pycqa/isort 29 | rev: 5.13.2 30 | hooks: 31 | - id: isort #isort is a pre-commit hook that runs to check for issues in imports and docstrings 32 | args: [ 33 | "--profile", "black", "--filter-files", 34 | "-l", "79" 35 | ] 36 | - repo: https://github.com/asottile/blacken-docs 37 | rev: 1.16.0 38 | hooks: 39 | - id: blacken-docs #blacken-docs is a pre-commit hook that runs to check for issues in the docs 40 | additional_dependencies: [black] 41 | - repo: https://github.com/asottile/pyupgrade 42 | rev: v3.15.2 43 | hooks: 44 | - id: pyupgrade #pyupgrade is a pre-commit hook that runs to check for issues in the code 45 | args: [--py36-plus] 46 | - repo: local 47 | hooks: 48 | - id: mypy # mypy is a pre-commit hook that runs as a linter to check for type errors 49 | name: mypy 50 | entry: mypy --implicit-optional 51 | language: system 52 | types: [python] 53 | args: [ 54 | "--ignore-missing-imports", 55 | "--explicit-package-bases", 56 | "--check-untyped-defs" 57 | ] 58 | stages: 59 | - "pre-push" 60 | - "pre-merge-commit" 61 | - repo: local 62 | hooks: 63 | - id: pytest-check 64 | name: pytest-check 65 | language: python 66 | types: [python] 67 | entry: pytest 68 | pass_filenames: false 69 | always_run: true 70 | args: [ 71 | --doctest-modules, 72 | -o, addopts="" 73 | ] 74 | - repo: https://github.com/roy-ht/pre-commit-jupyter 75 | rev: v1.2.1 76 | hooks: 77 | - id: jupyter-notebook-cleanup 78 | args: 79 | - --remove-kernel-metadata 80 | - --pin-patterns 81 | - "[pin];[donotremove]" 82 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # ![skab](docs/pictures/skab.png) 2 | 3 | 🛠🛠🛠**The testbed is under repair right now. Unfortunately, we can't tell exactly when it will be ready and we be able to continue data collection. Information about it will be in the repository. Sorry for the delay.** 4 | 5 | ❗️❗️❗️The current version of SKAB (v0.9) contains 34 datasets with collective anomalies. But the update to v1.0 will contain 300+ additional files with point and collective anomalies. It will make SKAB one of the largest changepoint-containing benchmarks, especially in the technical field. 6 | 7 | ## About SKAB [![Maintenance](https://img.shields.io/badge/Maintained%3F-yes-green.svg)](https://github.com/waico/SKAB/graphs/commit-activity) [![DOI](https://img.shields.io/badge/DOI-10.34740/kaggle/dsv/1693952-blue.svg)](https://doi.org/10.34740/KAGGLE/DSV/1693952) [![License: GPL v3.0](https://img.shields.io/badge/License-GPL%20v3.0-green.svg)](https://www.gnu.org/licenses/gpl-3.0.html) 8 | 9 | We propose the [Skoltech](https://www.skoltech.ru/en) Anomaly Benchmark (SKAB) designed for evaluating the anomaly detection core. SKAB allows working with two main problems (there are two markups for anomalies): 10 | 11 | 1. Outlier detection (anomalies considered and marked up as single-point anomalies) 12 | 2. Changepoint detection (anomalies considered and marked up as collective anomalies) 13 | 14 | SKAB consists of the following artifacts: 15 | 16 | 1. [Datasets](#datasets) 17 | 2. [Proposed Leaderboard](#proposed-leaderboard) for outlier detection and changepoint detection problems 18 | 3. Python modules for algorithms’ evaluation (now evaluation modules are being imported from [TSAD](https://github.com/waico/tsad) framework, while the details regarding the evaluation process are presented [here](https://github.com/waico/tsad/blob/main/examples/Evaluating.ipynb)) 19 | 4. Python [core](core/) with algorithms’ implementation 20 | 5. Python [notebooks](#notebooks) with anomaly detection pipeline implementation for various algorithms 21 | 22 | All the details about SKAB are presented in the following artifacts: 23 | 24 | - Position paper (*currently submitted for publication*) 25 | - Talk about the project: [English](https://youtu.be/hjzuKeNYUho) version and [Russian](https://www.youtube.com/watch?v=VLmmYGc4v2c) version 26 | - Slides about the project: [English](https://drive.google.com/open?id=1dHUevwPp6ftQCEKnRgB4KMp9oLBMSiDM) version and [Russian](https://drive.google.com/file/d/1gThPCNbEaIxhENLm-WTFGO_9PU1Wdwjq/view?usp=share_link) version 27 | 28 | ## Datasets 29 | 30 | The SKAB v0.9 corpus contains 35 individual data files in .csv format (datasets). The [data](data/) folder contains datasets from the benchmark. The structure of the data folder is presented in the [structure](./data/README.md) file. Each dataset represents a single experiment and contains a single anomaly. The datasets represent a multivariate time series collected from the sensors installed on the testbed. Columns in each data file are following: 31 | 32 | - `datetime` - Represents dates and times of the moment when the value is written to the database (YYYY-MM-DD hh:mm:ss) 33 | - `Accelerometer1RMS` - Shows a vibration acceleration (Amount of g units) 34 | - `Accelerometer2RMS` - Shows a vibration acceleration (Amount of g units) 35 | - `Current` - Shows the amperage on the electric motor (Ampere) 36 | - `Pressure` - Represents the pressure in the loop after the water pump (Bar) 37 | - `Temperature` - Shows the temperature of the engine body (The degree Celsius) 38 | - `Thermocouple` - Represents the temperature of the fluid in the circulation loop (The degree Celsius) 39 | - `Voltage` - Shows the voltage on the electric motor (Volt) 40 | - `RateRMS` - Represents the circulation flow rate of the fluid inside the loop (Liter per minute) 41 | - `anomaly` - Shows if the point is anomalous (0 or 1) 42 | - `changepoint` - Shows if the point is a changepoint for collective anomalies (0 or 1) 43 | 44 | Exploratory Data Analysis (EDA) for SKAB is presented [here (tbd)]. Russian version of EDA is available on [kaggle](https://www.kaggle.com/newintown/eda-example). 45 | 46 | ℹ️We have also made a *SKAB teaser* that is a small dataset collected separately but from the same testbed. SKAB teaser is made just for learning/teaching purposes and contains only 4 collective anomalies. All the information is available on [kaggle](https://www.kaggle.com/datasets/yuriykatser/skoltech-anomaly-benchmark-skab-teaser). 47 | 48 | ## Proposed Leaderboard 49 | 50 | This leaderboard shows performance of algorithms on test set, unlike leaderboard for SKAB v0.9 which evaluates both training and testing data all together. Moreover, the evaluated window of change points is to the right side of actual change point occurence which is in accordance with fact, that it should be impossible to capture event before it occurs. Lastly, the window size for the NAB detection algorithm is set to 60 seconds to reflect the dynamics of the transition as presented in the slides to enable detection of the start of the transition phase which is also marked as change-point. 51 | 52 | You can present and evaluate your algorithm using SKAB on [kaggle](https://www.kaggle.com/yuriykatser/skoltech-anomaly-benchmark-skab). Leaderboards are also available at paperswithcode.com: [CPD problem](https://paperswithcode.com/sota/change-point-detection-on-skab). 53 | 54 | Information about the metrics for anomaly detection and intuition behind the metrics selection can be found in [this](https://medium.com/@katser/a-review-of-anomaly-detection-metrics-with-a-lot-of-related-information-736d88774712) medium article. 55 | 56 | ### Outlier detection problem 57 | 58 | *Sorted by F1; for F1 bigger is better; both for FAR (False Alarm Rate) and MAR (Missing Alarm Rate) less is better* 59 | *Evaluated as binary classification problem.* 60 | 61 | | Algorithm | F1 | FAR, % | MAR, % 62 | |---|---|---|--- 63 | |Perfect detector | 1 | 0 | 0 64 | |Conv-AE |0.78 | 13.55 | 28.02 65 | |MSET |0.78 | 39.73 | 14.13 66 | |T-squared+Q (PCA-based) | 0.76 | 26.62 | 24.92 67 | |LSTM-AE |0.74 | 29.96 | 25.92 68 | |T-squared | 0.66 | 19.21 | 42.6 69 | |LSTM-VAE | 0.56 | 9.13 | 55.03 70 | |Vanilla LSTM | 0.54 | 12.54 | 59.53 71 | |MSCRED | 0.36 | 49.94 | 69.88 72 | |Vanilla AE | 0.39 | 2.59 | 75.15 73 | |Isolation forest | 0.29 | 2.56 | 82.89 74 | |Null detector | 0 | 0 | 100 75 | 76 | ### Changepoint detection problem 77 | 78 | *Sorted by NAB (standard); for NAB (standard), NAB (LowFP), NAB (LowFN) bigger is better, for Number of Missed CPs, Number of FPs lower is better* 79 | *The current leaderboard is obtained with the window size for the NAB detection algorithm equal to 60 sec and to the right side of true change point.* 80 | 81 | | Algorithm | NAB (standard) | NAB (LowFP) | NAB (LowFN) | Number of Missed CPs | Number of FPs 82 | |---|---|---|---|---|--- 83 | |Perfect detector | 100 | 100 | 100 | 0 | 0 84 | |MSCRED | 32.42 | 16.53 | 40.28 | 55 | 342 85 | |Isolation forest | 26.16 | 19.5 | 30.82 | 76 | 135 86 | |T-squared+Q (PCA-based) | 25.35 | 14.51 | 31.33 | 72 | 232 87 | |Conv-AE | 23.61 | 21.54 | 27.55 | 82 | 23 88 | |LSTM-AE | 23.51 | 20.11 | 25.91 | 88 | 69 89 | |T-squared | 19.54 | 10.2 | 24.31 | 70 | 106 90 | |MSET | 13.84 | 10.22 | 17.37 | 96 | 66 91 | |Vanilla AE | 11.41 | 6.53 | 13.91 | 103 | 106 92 | |Vanilla LSTM | 11.31 | -3.8 | 17.25 | 90 | 342 93 | |ArimaFD | -0.09 | -0.17 | -0.06 | 127 | 2 94 | |Null detector | 0 | 0 | 0 | - | - 95 | 96 | ## Notebooks 97 | 98 | The [notebooks](notebooks/) folder contains jupyter notebooks with the code for the proposed leaderboard results reproducing. We have calculated the results for following commonly known anomaly detection algorithms: 99 | 100 | - Isolation forest - *Outlier detection algorithm based on Random forest concept* 101 | - Vanilla LSTM - *NN with LSTM layer* 102 | - Vanilla AE - *Feed-Forward Autoencoder* 103 | - LSTM-AE - *LSTM Autoencoder* 104 | - LSTM-VAE - *LSTM Variational Autoencoder* 105 | - Conv-AE - *Convolutional Autoencoder* 106 | - MSCRED - *Multi-Scale Convolutional Recurrent Encoder-Decoder* 107 | - MSET - *Multivariate State Estimation Technique* 108 | 109 | Additionally on the leaderboard were shown the externally calculated results of the following algorithms: 110 | 111 | - [ArimaFD](https://github.com/waico/arimafd) - *ARIMA-based fault detection algorithm* 112 | - [T-squared](http://github.com/YKatser/ControlCharts/tree/main/examples) - *Hotelling's T-squared statistics* 113 | - [T-squared+Q (PCA-based)](http://github.com/YKatser/ControlCharts/tree/main/examples) - *Hotelling's T-squared statistics + Q statistics based on PCA* 114 | - [ruptures](https://github.com/deepcharles/ruptures) - *Changepoint detection (CPD) algorithms from ruptures package* 115 | - [CPDE](https://github.com/YKatser/CPDE) - *Ruptures-based changepoint detection ensemble (CPDE) algorithms* 116 | 117 | Details regarding the algorithms, including short description, references to scientific papers and code of the initial implementation is available in [this readme](https://github.com/waico/SKAB/tree/master/notebooks#anomaly-detection-algorithms). 118 | 119 | ## Installation 120 | 121 | 1. install Python 3.10+ (tested on 3.10.13) 122 | 123 | 1. install [poetry](https://python-poetry.org/docs/) package manager 124 | - `brew install poetry` 125 | > Poetry installs dependencies and locks versions for deterministic installs. Poetry uses [Python's built-in `venv` module](https://docs.python.org/3/library/venv.html) to create virtual environments. It also uses PEP [517](https://peps.python.org/pep-0517) & [518](https://peps.python.org/pep-0518) specifications to build packages without requiring `setup.py` or `requirements.txt` files. 126 | 127 | 1. LightGBM base install 128 | - `brew install lightgbm` 129 | 130 | 1. install SKAB dependencies, see [pyproject.toml](pyproject.toml) for details 131 | - `poetry install` 132 | 133 | 1. confirm installation 134 | - `poetry show --tree` - shows all dependencies installed 135 | - `poetry env info` - displays information about the current environment (Python version, path, etc) 136 | - `poetry list` - lists all cli commands 137 | 138 | ## Citation 139 | 140 | Please cite our project in your publications if it helps your research. 141 | 142 | ```bibtex 143 | @misc{skab, 144 | author = {Katser, Iurii D. and Kozitsin, Vyacheslav O.}, 145 | title = {Skoltech Anomaly Benchmark (SKAB)}, 146 | year = {2020}, 147 | publisher = {Kaggle}, 148 | howpublished = {\url{https://www.kaggle.com/dsv/1693952}}, 149 | DOI = {10.34740/KAGGLE/DSV/1693952} 150 | } 151 | ``` 152 | 153 | ## Notable mentions 154 | 155 | SKAB is acknowledged by some ML resources. 156 | 157 | - [Anomaly Detection Learning Resources](https://github.com/yzhao062/anomaly-detection-resources#34-datasets) 158 | - [awesome-TS-anomaly-detection](https://github.com/rob-med/awesome-TS-anomaly-detection#benchmark-datasets) 159 | - [List of datasets for machine-learning research](https://en.wikipedia.org/wiki/List_of_datasets_for_machine-learning_research#Anomaly_data) 160 | - [paperswithcode.com](https://paperswithcode.com/dataset/skab) 161 | - [Google datasets](https://datasetsearch.research.google.com/search?query=skoltech%20anomaly%20benchmark&docid=IIIE4VWbqUKszygyAAAAAA%3D%3D) 162 | - [Industrial ML Datasets](https://github.com/nicolasj92/industrial-ml-datasets) 163 | - etc. 164 | -------------------------------------------------------------------------------- /core/Conv_AE.py: -------------------------------------------------------------------------------- 1 | from tensorflow.keras.callbacks import EarlyStopping 2 | from tensorflow.keras.layers import Conv1D, Conv1DTranspose, Dropout, Input 3 | from tensorflow.keras.models import Sequential 4 | from tensorflow.keras.optimizers import Adam 5 | 6 | 7 | class Conv_AE: 8 | """ 9 | A reconstruction convolutional autoencoder model to detect anomalies in timeseries data using reconstruction error as an anomaly score. 10 | 11 | Parameters 12 | ---------- 13 | No parameters are required for initializing the class. 14 | 15 | Attributes 16 | ---------- 17 | model : Sequential 18 | The trained convolutional autoencoder model. 19 | 20 | Examples 21 | -------- 22 | >>> from Conv_AE import Conv_AE 23 | >>> CAutoencoder = Conv_AE() 24 | >>> CAutoencoder.fit(train_data) 25 | >>> prediction = CAutoencoder.predict(test_data) 26 | """ 27 | 28 | def __init__(self): 29 | self._Random(0) 30 | 31 | def _Random(self, seed_value): 32 | import os 33 | 34 | os.environ["PYTHONHASHSEED"] = str(seed_value) 35 | 36 | import random 37 | 38 | random.seed(seed_value) 39 | 40 | import numpy as np 41 | 42 | np.random.seed(seed_value) 43 | 44 | import tensorflow as tf 45 | 46 | tf.random.set_seed(seed_value) 47 | 48 | def _build_model(self): 49 | model = Sequential( 50 | [ 51 | Input(shape=(self.shape[1], self.shape[2])), 52 | Conv1D( 53 | filters=32, 54 | kernel_size=7, 55 | padding="same", 56 | strides=2, 57 | activation="relu", 58 | ), 59 | Dropout(rate=0.2), 60 | Conv1D( 61 | filters=16, 62 | kernel_size=7, 63 | padding="same", 64 | strides=2, 65 | activation="relu", 66 | ), 67 | Conv1DTranspose( 68 | filters=16, 69 | kernel_size=7, 70 | padding="same", 71 | strides=2, 72 | activation="relu", 73 | ), 74 | Dropout(rate=0.2), 75 | Conv1DTranspose( 76 | filters=32, 77 | kernel_size=7, 78 | padding="same", 79 | strides=2, 80 | activation="relu", 81 | ), 82 | Conv1DTranspose(filters=1, kernel_size=7, padding="same"), 83 | ] 84 | ) 85 | model.compile(optimizer=Adam(learning_rate=0.001), loss="mse") 86 | 87 | return model 88 | 89 | def fit(self, data): 90 | """ 91 | Train the convolutional autoencoder model on the provided data. 92 | 93 | Parameters 94 | ---------- 95 | data : numpy.ndarray 96 | Input data for training the autoencoder model. 97 | """ 98 | 99 | self.shape = data.shape 100 | self.model = self._build_model() 101 | 102 | self.model.fit( 103 | data, 104 | data, 105 | epochs=100, 106 | batch_size=32, 107 | validation_split=0.1, 108 | verbose=0, 109 | callbacks=[ 110 | EarlyStopping( 111 | monitor="val_loss", patience=5, mode="min", verbose=0 112 | ) 113 | ], 114 | ) 115 | 116 | def predict(self, data): 117 | """ 118 | Generate predictions using the trained convolutional autoencoder model. 119 | 120 | Parameters 121 | ---------- 122 | data : numpy.ndarray 123 | Input data for generating predictions. 124 | 125 | Returns 126 | ------- 127 | numpy.ndarray 128 | Predicted output data. 129 | """ 130 | 131 | return self.model.predict(data) 132 | -------------------------------------------------------------------------------- /core/Isolation_Forest.py: -------------------------------------------------------------------------------- 1 | from sklearn.ensemble import IsolationForest 2 | 3 | 4 | class Isolation_Forest: 5 | """ 6 | Isolation Forest or iForest builds an ensemble of iTrees for a given data set, then anomalies are those instances which have short average path lengths on the iTrees. 7 | 8 | Parameters 9 | ---------- 10 | params : list 11 | A list containing three parameters: random_state, n_jobs, and contamination. 12 | 13 | Attributes 14 | ---------- 15 | random_state : int 16 | The random seed used for reproducibility. 17 | n_jobs : int 18 | The number of CPU cores to use for parallelism. 19 | contamination : float 20 | The expected proportion of anomalies in the dataset. 21 | 22 | Examples 23 | -------- 24 | >>> from Isolation_Forest import Isolation_Forest 25 | >>> PARAMS = [random_state, n_jobs, contamination] 26 | >>> model = Isolation_Forest(PARAMS) 27 | >>> model.fit(X_train) 28 | >>> predictions = model.predict(test_data) 29 | """ 30 | 31 | def __init__(self, params): 32 | self.params = params 33 | self.random_state = self.params[0] 34 | self.n_jobs = self.params[1] 35 | self.contamination = self.params[2] 36 | 37 | def _Random(self, seed_value): 38 | import os 39 | 40 | os.environ["PYTHONHASHSEED"] = str(seed_value) 41 | 42 | import random 43 | 44 | random.seed(seed_value) 45 | 46 | import numpy as np 47 | 48 | np.random.seed(seed_value) 49 | 50 | import tensorflow as tf 51 | 52 | tf.random.set_seed(seed_value) 53 | 54 | def _build_model(self): 55 | self._Random(0) 56 | 57 | model = IsolationForest( 58 | random_state=self.random_state, 59 | n_jobs=self.n_jobs, 60 | contamination=self.contamination, 61 | ) 62 | return model 63 | 64 | def fit(self, X): 65 | """ 66 | Train the Isolation Forest model on the provided data. 67 | 68 | Parameters 69 | ---------- 70 | X : numpy.ndarray 71 | Input data for training the model. 72 | """ 73 | 74 | self.model = self._build_model() 75 | 76 | self.model.fit(X) 77 | 78 | def predict(self, data): 79 | """ 80 | Generate predictions using the trained Isolation Forest model. 81 | 82 | Parameters 83 | ---------- 84 | data : numpy.ndarray 85 | Input data for generating predictions. 86 | 87 | Returns 88 | ------- 89 | numpy.ndarray 90 | Predicted output data. 91 | """ 92 | 93 | return self.model.predict(data) 94 | -------------------------------------------------------------------------------- /core/LSTM_AE.py: -------------------------------------------------------------------------------- 1 | from tensorflow.keras import Model 2 | from tensorflow.keras.callbacks import EarlyStopping 3 | from tensorflow.keras.layers import ( 4 | LSTM, 5 | Dense, 6 | Input, 7 | RepeatVector, 8 | TimeDistributed, 9 | ) 10 | 11 | 12 | class LSTM_AE: 13 | """ 14 | A reconstruction sequence-to-sequence (LSTM-based) autoencoder model to detect anomalies in timeseries data using reconstruction error as an anomaly score. 15 | 16 | Parameters 17 | ---------- 18 | params : list 19 | A list of hyperparameters for the model, containing the following elements: 20 | - EPOCHS : int 21 | The number of training epochs. 22 | - BATCH_SIZE : int 23 | The batch size for training. 24 | - VAL_SPLIT : float 25 | The validation split ratio during training. 26 | 27 | Attributes 28 | ---------- 29 | params : list 30 | The hyperparameters for the model. 31 | 32 | Examples 33 | -------- 34 | >>> from LSTM_AE import LSTM_AE 35 | >>> PARAMS = [EPOCHS, BATCH_SIZE, VAL_SPLIT] 36 | >>> model = LSTM_AE(PARAMS) 37 | >>> model.fit(train_data) 38 | >>> predictions = model.predict(test_data) 39 | """ 40 | 41 | def __init__(self, params): 42 | self.params = params 43 | 44 | def _Random(self, seed_value): 45 | import os 46 | 47 | os.environ["PYTHONHASHSEED"] = str(seed_value) 48 | 49 | import random 50 | 51 | random.seed(seed_value) 52 | 53 | import numpy as np 54 | 55 | np.random.seed(seed_value) 56 | 57 | import tensorflow as tf 58 | 59 | tf.random.set_seed(seed_value) 60 | 61 | def _build_model(self): 62 | self._Random(0) 63 | 64 | inputs = Input(shape=(self.shape[1], self.shape[2])) 65 | encoded = LSTM(100, activation="relu")(inputs) 66 | 67 | decoded = RepeatVector(self.shape[1])(encoded) 68 | decoded = LSTM(100, activation="relu", return_sequences=True)(decoded) 69 | decoded = TimeDistributed(Dense(self.shape[2]))(decoded) 70 | 71 | model = Model(inputs, decoded) 72 | _ = Model(inputs, encoded) 73 | 74 | model.compile(optimizer="adam", loss="mae", metrics=["mse"]) 75 | 76 | return model 77 | 78 | def fit(self, X): 79 | """ 80 | Train the sequence-to-sequence (LSTM-based) autoencoder model on the provided data. 81 | 82 | Parameters 83 | ---------- 84 | X : numpy.ndarray 85 | Input data for training the model. 86 | """ 87 | 88 | self.shape = X.shape 89 | self.model = self._build_model() 90 | 91 | early_stopping = EarlyStopping(patience=5, verbose=0) 92 | 93 | self.model.fit( 94 | X, 95 | X, 96 | validation_split=self.params[2], 97 | epochs=self.params[0], 98 | batch_size=self.params[1], 99 | verbose=0, 100 | shuffle=False, 101 | callbacks=[early_stopping], 102 | ) 103 | 104 | def predict(self, data): 105 | """ 106 | Generate predictions using the trained sequence-to-sequence (LSTM-based) autoencoder model. 107 | 108 | Parameters 109 | ---------- 110 | data : numpy.ndarray 111 | Input data for generating predictions. 112 | 113 | Returns 114 | ------- 115 | numpy.ndarray 116 | Predicted output data. 117 | """ 118 | 119 | return self.model.predict(data) 120 | -------------------------------------------------------------------------------- /core/LSTM_VAE.py: -------------------------------------------------------------------------------- 1 | from tensorflow.keras import backend as K 2 | from tensorflow.keras import losses 3 | from tensorflow.keras.callbacks import EarlyStopping 4 | from tensorflow.keras.layers import LSTM, Dense, Input, Layer, RepeatVector 5 | from tensorflow.keras.models import Model 6 | 7 | 8 | class KLDivergenceLayer(Layer): 9 | def __init__(self, **kwargs): 10 | super().__init__(**kwargs) 11 | 12 | def call(self, inputs): 13 | z_mean, z_log_sigma = inputs 14 | kl_loss = -0.5 * K.mean( 15 | 1 + z_log_sigma - K.square(z_mean) - K.exp(z_log_sigma), axis=-1 16 | ) 17 | self.add_loss(kl_loss) 18 | return kl_loss # Return KL loss value 19 | 20 | 21 | class Sampling(Layer): 22 | def __init__(self, latent_dim, epsilon_std=1.0, **kwargs): 23 | super().__init__(**kwargs) 24 | self.latent_dim = latent_dim 25 | self.epsilon_std = epsilon_std 26 | 27 | def call(self, inputs): 28 | z_mean, z_log_sigma = inputs 29 | batch = K.shape(z_mean)[0] 30 | dim = K.shape(z_mean)[1] 31 | epsilon = K.random_normal( 32 | shape=(batch, dim), mean=0.0, stddev=self.epsilon_std 33 | ) 34 | return z_mean + z_log_sigma * epsilon 35 | 36 | def compute_output_shape(self, input_shape): 37 | return input_shape[0] # Same shape as z_mean and z_log_sigma 38 | 39 | 40 | class LSTM_VAE: 41 | """ 42 | A reconstruction LSTM variational autoencoder model to detect anomalies in timeseries data using reconstruction error as an anomaly score. 43 | 44 | Parameters 45 | ---------- 46 | TenserFlow_backend : bool, optional 47 | Flag to specify whether to use TensorFlow backend (default is False). 48 | 49 | Attributes 50 | ---------- 51 | None 52 | 53 | Examples 54 | ------- 55 | >>> from LSTM_VAE import LSTM_VAE 56 | >>> model = LSTM_VAE() 57 | >>> model.fit(train_data) 58 | >>> predictions = model.predict(test_data) 59 | """ 60 | 61 | def __init__(self, params): 62 | self.params = params 63 | 64 | def _build_model(self, input_dim, timesteps, intermediate_dim, latent_dim): 65 | self._Random(0) 66 | 67 | x = Input( 68 | shape=( 69 | timesteps, 70 | input_dim, 71 | ) 72 | ) 73 | 74 | h = LSTM(intermediate_dim)(x) 75 | 76 | self.z_mean = Dense(latent_dim)(h) 77 | self.z_log_sigma = Dense(latent_dim)(h) 78 | 79 | z = Sampling(latent_dim)([self.z_mean, self.z_log_sigma]) 80 | 81 | h_decoded = RepeatVector(timesteps)(z) 82 | decoder_h = LSTM(intermediate_dim, return_sequences=True)(h_decoded) 83 | decoder_mean = LSTM(input_dim, return_sequences=True)(decoder_h) 84 | 85 | vae = Model(x, decoder_mean) 86 | 87 | _ = Model(x, self.z_mean) 88 | 89 | decoder_input = Input(shape=(latent_dim,)) 90 | 91 | _h_decoded = RepeatVector(timesteps)(decoder_input) 92 | _h_decoded = LSTM(intermediate_dim, return_sequences=True)(_h_decoded) 93 | 94 | _x_decoded_mean = LSTM(input_dim, return_sequences=True)(_h_decoded) 95 | _ = Model(decoder_input, _x_decoded_mean) 96 | 97 | vae.compile(optimizer="rmsprop", loss=self.vae_loss) 98 | 99 | return vae 100 | 101 | def _Random(self, seed_value): 102 | import os 103 | 104 | os.environ["PYTHONHASHSEED"] = str(seed_value) 105 | 106 | import random 107 | 108 | random.seed(seed_value) 109 | 110 | import numpy as np 111 | 112 | np.random.seed(seed_value) 113 | 114 | import tensorflow as tf 115 | 116 | tf.random.set_seed(seed_value) 117 | 118 | def vae_loss(self, x, x_decoded_mean): 119 | """ 120 | Calculate the VAE loss. 121 | 122 | Parameters 123 | ---------- 124 | x : tensorflow.Tensor 125 | Input data. 126 | x_decoded_mean : tensorflow.Tensor 127 | Decoded output data. 128 | 129 | Returns 130 | ------- 131 | loss : tensorflow.Tensor 132 | VAE loss value. 133 | """ 134 | mse = losses.MeanSquaredError() 135 | xent_loss = mse(x, x_decoded_mean) 136 | kl_loss = KLDivergenceLayer()([self.z_mean, self.z_log_sigma]) 137 | loss = xent_loss + kl_loss 138 | return loss 139 | 140 | def fit(self, X): 141 | """ 142 | Train the LSTM variational autoencoder model on the provided data. 143 | 144 | Parameters 145 | ---------- 146 | data : numpy.ndarray 147 | Input data for training. 148 | epochs : int, optional 149 | Number of training epochs (default is 20). 150 | validation_split : float, optional 151 | Fraction of the training data to be used as validation data (default is 0.1). 152 | BATCH_SIZE : int, optional 153 | Batch size for training (default is 1). 154 | early_stopping : bool, optional 155 | Whether to use early stopping during training (default is True). 156 | """ 157 | 158 | self.shape = X.shape 159 | self.input_dim = self.shape[-1] 160 | self.timesteps = self.shape[1] 161 | self.latent_dim = 100 162 | self.epsilon_std = 1.0 163 | self.intermediate_dim = 32 164 | 165 | self.model = self._build_model( 166 | self.input_dim, 167 | timesteps=self.timesteps, 168 | intermediate_dim=self.intermediate_dim, 169 | latent_dim=self.latent_dim, 170 | ) 171 | 172 | early_stopping = EarlyStopping(patience=5, verbose=0) 173 | 174 | self.model.fit( 175 | X, 176 | X, 177 | validation_split=self.params[2], 178 | epochs=self.params[0], 179 | batch_size=self.params[1], 180 | verbose=0, 181 | shuffle=False, 182 | callbacks=[early_stopping], 183 | ) 184 | 185 | def predict(self, data): 186 | """ 187 | Generate predictions using the trained LSTM variational autoencoder model. 188 | 189 | Parameters 190 | ---------- 191 | data : numpy.ndarray 192 | Input data for making predictions. 193 | 194 | Returns 195 | ------- 196 | predictions : numpy.ndarray 197 | The reconstructed output predictions. 198 | """ 199 | 200 | return self.model.predict(data) 201 | -------------------------------------------------------------------------------- /core/MSCRED.py: -------------------------------------------------------------------------------- 1 | import math 2 | 3 | import tensorflow as tf 4 | from tensorflow.keras import Model 5 | from tensorflow.keras.callbacks import ReduceLROnPlateau 6 | from tensorflow.keras.layers import ( 7 | Conv2D, 8 | Conv2DTranspose, 9 | ConvLSTM2D, 10 | Input, 11 | Layer, 12 | TimeDistributed, 13 | ) 14 | from tensorflow.keras.optimizers import Adam 15 | 16 | 17 | class MSCRED: 18 | """ 19 | MSCRED - Multi-Scale Convolutional Recurrent Encoder-Decoder first constructs multi-scale (resolution) signature matrices to characterize multiple levels of the system statuses across different time steps. In particular, different levels of the system statuses are used to indicate the severity of different abnormal incidents. Subsequently, given the signature matrices, a convolutional encoder is employed to encode the inter-sensor (time series) correlations patterns and an attention based Convolutional Long-Short Term Memory (ConvLSTM) network is developed to capture the temporal patterns. Finally, with the feature maps which encode the inter-sensor correlations and temporal information, a convolutional decoder is used to reconstruct the signature matrices and the residual signature matrices are further utilized to detect and diagnose anomalies. The intuition is that MSCRED may not reconstruct the signature matrices well if it never observes similar system statuses before. 20 | 21 | Parameters 22 | ---------- 23 | params : list 24 | A list containing configuration parameters for the MSCRED model. 25 | 26 | Attributes 27 | ---------- 28 | model : Model 29 | The trained MSCRED model. 30 | 31 | Examples 32 | -------- 33 | >>> from MSCRED import MSCRED 34 | >>> PARAMS = [sensor_n, scale_n, step_max] 35 | >>> model = MSCRED(PARAMS) 36 | >>> model.fit(X_train, Y_train, X_test, Y_test) 37 | >>> prediction = model.predict(test_data) 38 | """ 39 | 40 | def __init__(self, params): 41 | self.params = params 42 | 43 | def _build_model(self): 44 | self._Random(0) 45 | 46 | class MyPadLayer(Layer): 47 | def __init__(self, paddings, **kwargs): 48 | super().__init__(**kwargs) 49 | self.paddings = paddings 50 | 51 | def call(self, inputs): 52 | return tf.pad(inputs, self.paddings) 53 | 54 | class MyAttentionLayer(Layer): 55 | def __init__(self, attention_fun, **kwargs): 56 | super().__init__(**kwargs) 57 | self.attention = attention_fun 58 | 59 | def call(self, inputs, **kwargs): 60 | # Your attention mechanism implementation here 61 | return self.attention(inputs, **kwargs) 62 | 63 | class MyConcatLayer(Layer): 64 | def __init__(self, axis, **kwargs): 65 | super().__init__(**kwargs) 66 | self.axis = axis 67 | 68 | def call(self, inputs): 69 | return tf.concat(inputs, axis=self.axis) 70 | 71 | input_size = ( 72 | self.params[2], 73 | self.params[0], 74 | self.params[0], 75 | self.params[1], 76 | ) 77 | inputs = Input(input_size) 78 | 79 | if self.params[0] % 8 != 0: 80 | self.sensor_n_pad = (self.params[0] // 8) * 8 + 8 81 | else: 82 | self.sensor_n_pad = self.params[0] 83 | 84 | paddings = tf.constant( 85 | [ 86 | [0, 0], 87 | [0, 0], 88 | [0, self.sensor_n_pad - self.params[0]], 89 | [0, self.sensor_n_pad - self.params[0]], 90 | [0, 0], 91 | ] 92 | ) 93 | 94 | inputs_pad = MyPadLayer(paddings)(inputs) 95 | 96 | conv1 = TimeDistributed( 97 | Conv2D( 98 | filters=32, 99 | kernel_size=3, 100 | strides=1, 101 | kernel_initializer="glorot_uniform", 102 | padding="same", 103 | activation="selu", 104 | name="conv1", 105 | ) 106 | )(inputs_pad) 107 | 108 | conv2 = TimeDistributed( 109 | Conv2D( 110 | filters=64, 111 | kernel_size=3, 112 | strides=2, 113 | kernel_initializer="glorot_uniform", 114 | padding="same", 115 | activation="selu", 116 | name="conv2", 117 | ) 118 | )(conv1) 119 | 120 | conv3 = TimeDistributed( 121 | Conv2D( 122 | filters=128, 123 | kernel_size=2, 124 | strides=2, 125 | kernel_initializer="glorot_uniform", 126 | padding="same", 127 | activation="selu", 128 | name="conv3", 129 | ) 130 | )(conv2) 131 | 132 | conv4 = TimeDistributed( 133 | Conv2D( 134 | filters=256, 135 | kernel_size=2, 136 | strides=2, 137 | kernel_initializer="glorot_uniform", 138 | padding="same", 139 | activation="selu", 140 | name="conv4", 141 | ) 142 | )(conv3) 143 | 144 | convLSTM1 = ConvLSTM2D( 145 | filters=32, 146 | kernel_size=2, 147 | padding="same", 148 | return_sequences=True, 149 | name="convLSTM1", 150 | )(conv1) 151 | convLSTM1_out = MyAttentionLayer(self.attention)( 152 | convLSTM1, **{"koef": 1} 153 | ) 154 | 155 | convLSTM2 = ConvLSTM2D( 156 | filters=64, 157 | kernel_size=2, 158 | padding="same", 159 | return_sequences=True, 160 | name="convLSTM2", 161 | )(conv2) 162 | convLSTM2_out = MyAttentionLayer(self.attention)( 163 | convLSTM2, **{"koef": 2} 164 | ) 165 | 166 | convLSTM3 = ConvLSTM2D( 167 | filters=128, 168 | kernel_size=2, 169 | padding="same", 170 | return_sequences=True, 171 | name="convLSTM3", 172 | )(conv3) 173 | convLSTM3_out = MyAttentionLayer(self.attention)( 174 | convLSTM3, **{"koef": 4} 175 | ) 176 | 177 | convLSTM4 = ConvLSTM2D( 178 | filters=256, 179 | kernel_size=2, 180 | padding="same", 181 | return_sequences=True, 182 | name="convLSTM4", 183 | )(conv4) 184 | convLSTM4_out = MyAttentionLayer(self.attention)( 185 | convLSTM4, **{"koef": 8} 186 | ) 187 | 188 | deconv4 = Conv2DTranspose( 189 | filters=128, 190 | kernel_size=2, 191 | strides=2, 192 | kernel_initializer="glorot_uniform", 193 | padding="same", 194 | activation="selu", 195 | name="deconv4", 196 | )(convLSTM4_out) 197 | deconv4_out = MyConcatLayer(axis=3)([deconv4, convLSTM3_out]) 198 | 199 | deconv3 = Conv2DTranspose( 200 | filters=64, 201 | kernel_size=2, 202 | strides=2, 203 | kernel_initializer="glorot_uniform", 204 | padding="same", 205 | activation="selu", 206 | name="deconv3", 207 | )(deconv4_out) 208 | deconv3_out = MyConcatLayer(axis=3)([deconv3, convLSTM2_out]) 209 | 210 | deconv2 = Conv2DTranspose( 211 | filters=32, 212 | kernel_size=3, 213 | strides=2, 214 | kernel_initializer="glorot_uniform", 215 | padding="same", 216 | activation="selu", 217 | name="deconv2", 218 | )(deconv3_out) 219 | deconv2_out = MyConcatLayer(axis=3)([deconv2, convLSTM1_out]) 220 | 221 | deconv1 = Conv2DTranspose( 222 | filters=self.params[1], 223 | kernel_size=3, 224 | strides=1, 225 | kernel_initializer="glorot_uniform", 226 | padding="same", 227 | activation="selu", 228 | name="deconv1", 229 | )(deconv2_out) 230 | 231 | model = Model( 232 | inputs=inputs, 233 | outputs=deconv1[:, : self.params[0], : self.params[0], :], 234 | ) 235 | 236 | return model 237 | 238 | def attention(self, outputs, koef): 239 | """ 240 | Attention mechanism to weigh the importance of each step in the sequence. 241 | 242 | Parameters 243 | ---------- 244 | outputs : tf.Tensor 245 | The output tensor from ConvLSTM layers. 246 | koef : int 247 | A coefficient to scale the attention mechanism. 248 | 249 | Returns 250 | ------- 251 | tf.Tensor 252 | Weighted output tensor. 253 | """ 254 | 255 | attention_w = [] 256 | for k in range(self.params[2]): 257 | attention_w.append( 258 | tf.reduce_sum( 259 | tf.multiply(outputs[:, k], outputs[:, -1]), axis=(1, 2, 3) 260 | ) 261 | / self.params[2] 262 | ) 263 | attention_w = tf.reshape( 264 | tf.nn.softmax(tf.stack(attention_w, axis=1)), 265 | [-1, 1, self.params[2]], 266 | ) 267 | outputs = tf.reshape( 268 | outputs, 269 | [-1, self.params[2], tf.reduce_prod(outputs.shape.as_list()[2:])], 270 | ) 271 | outputs = tf.matmul(attention_w, outputs) 272 | outputs = tf.reshape( 273 | outputs, 274 | [ 275 | -1, 276 | math.ceil(self.sensor_n_pad / koef), 277 | math.ceil(self.sensor_n_pad / koef), 278 | 32 * koef, 279 | ], 280 | ) 281 | return outputs 282 | 283 | def _Random(self, seed_value): 284 | import os 285 | 286 | os.environ["PYTHONHASHSEED"] = str(seed_value) 287 | 288 | import random 289 | 290 | random.seed(seed_value) 291 | 292 | import numpy as np 293 | 294 | np.random.seed(seed_value) 295 | 296 | import tensorflow as tf 297 | 298 | tf.random.set_seed(seed_value) 299 | 300 | def _loss_fn(self, y_true, y_pred): 301 | return tf.reduce_mean(tf.square(y_true - y_pred)) 302 | 303 | def fit(self, X_train, Y_train, batch_size=200, epochs=25): 304 | """ 305 | Train the MSCRED model on the provided data. 306 | 307 | Parameters 308 | ---------- 309 | X_train : numpy.ndarray 310 | The training input data. 311 | Y_train : numpy.ndarray 312 | The training target data. 313 | X_test : numpy.ndarray 314 | The testing input data. 315 | Y_test : numpy.ndarray 316 | The testing target data. 317 | batch_size : int, optional 318 | The batch size for training, by default 200. 319 | epochs : int, optional 320 | The number of training epochs, by default 25. 321 | """ 322 | 323 | self.model = self._build_model() 324 | 325 | self.model.compile( 326 | optimizer=Adam(learning_rate=1e-3), 327 | loss=self._loss_fn, 328 | ) 329 | reduce_lr = ReduceLROnPlateau( 330 | monitor="loss", factor=0.8, patience=6, min_lr=0.000001, verbose=1 331 | ) 332 | self.model.fit( 333 | X_train, 334 | Y_train, 335 | batch_size=batch_size, 336 | epochs=epochs, 337 | # validation_data=(X_test, Y_test), 338 | callbacks=reduce_lr, 339 | ) 340 | 341 | def predict(self, data): 342 | """ 343 | Generate predictions using the trained MSCRED model. 344 | 345 | Parameters 346 | ---------- 347 | data : numpy.ndarray 348 | Input data for generating predictions. 349 | 350 | Returns 351 | ------- 352 | numpy.ndarray 353 | Predicted output data. 354 | """ 355 | 356 | return self.model.predict(data) 357 | -------------------------------------------------------------------------------- /core/MSET.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import pandas as pd 3 | from scipy import linalg as spla 4 | from sklearn.preprocessing import StandardScaler 5 | 6 | 7 | class MSET: 8 | """ 9 | MSET - multivariate state estimation technique is a non-parametric and statistical modeling method, which calculates the estimated values based on the weighted average of historical data. In terms of procedure, MSET is similar to some nonparametric regression methods, such as, auto-associative kernel regression. 10 | 11 | Parameters 12 | ---------- 13 | None 14 | 15 | Attributes 16 | ---------- 17 | None 18 | 19 | Examples 20 | -------- 21 | >>> from MSET import MSET 22 | >>> model = MSET() 23 | >>> model.fit(data) 24 | >>> prediction = model.predict(test_data) 25 | """ 26 | 27 | def __init__(self): 28 | self._Random(0) 29 | 30 | def _build_model(self): 31 | self.SS = StandardScaler() 32 | 33 | def _Random(self, seed_value): 34 | import os 35 | 36 | os.environ["PYTHONHASHSEED"] = str(seed_value) 37 | 38 | import random 39 | 40 | random.seed(seed_value) 41 | 42 | import numpy as np 43 | 44 | np.random.seed(seed_value) 45 | 46 | import tensorflow as tf 47 | 48 | tf.random.set_seed(seed_value) 49 | 50 | def calc_W(self, X_obs): 51 | """ 52 | Calculate the weight matrix W. 53 | 54 | Parameters 55 | ---------- 56 | X_obs : numpy.ndarray 57 | Observations for which to calculate the weight matrix. 58 | 59 | Returns 60 | ------- 61 | numpy.ndarray 62 | Weight matrix W. 63 | """ 64 | 65 | DxX_obs = self.otimes(self.D, X_obs) 66 | # try: 67 | W = spla.lu_solve(self.LU_factors, DxX_obs) 68 | # except: 69 | # W = np.linalg.solve(self.DxD, DxX_obs) 70 | 71 | return W 72 | 73 | def otimes(self, X, Y): 74 | """ 75 | Compute the outer product of two matrices X and Y. 76 | 77 | Parameters 78 | ---------- 79 | X : numpy.ndarray 80 | First matrix. 81 | Y : numpy.ndarray 82 | Second matrix. 83 | 84 | Returns 85 | ------- 86 | numpy.ndarray 87 | Outer product of X and Y. 88 | """ 89 | 90 | m1, n = np.shape(X) 91 | m2, p = np.shape(Y) 92 | 93 | if m1 != m2: 94 | raise Exception("dimensionality mismatch between X and Y.") 95 | 96 | Z = np.zeros((n, p)) 97 | 98 | if n != p: 99 | for i in range(n): 100 | for j in range(p): 101 | Z[i, j] = self.kernel(X[:, i], Y[:, j]) 102 | else: 103 | for i in range(n): 104 | for j in range(i, p): 105 | Z[i, j] = self.kernel(X[:, i], Y[:, j]) 106 | Z[j, i] = Z[i, j] 107 | 108 | return Z 109 | 110 | def kernel(self, x, y): 111 | """ 112 | Compute the kernel function value. 113 | 114 | Parameters 115 | ---------- 116 | x : numpy.ndarray 117 | First vector. 118 | y : numpy.ndarray 119 | Second vector. 120 | 121 | Returns 122 | ------- 123 | float 124 | Kernel function s(x,y) = 1 - ||x-y||/(||x|| + ||y||) value. 125 | """ 126 | 127 | if all(x == y): 128 | return 1.0 129 | else: 130 | return 1.0 - np.linalg.norm(x - y) / ( 131 | np.linalg.norm(x) + np.linalg.norm(y) 132 | ) 133 | 134 | def fit(self, df, train_start=None, train_stop=None): 135 | """ 136 | Train the MSET model on the provided data. 137 | 138 | Parameters 139 | ---------- 140 | df : pandas.DataFrame 141 | Input data for training the model. 142 | train_start : int, optional 143 | Index to start training, by default None. 144 | train_stop : int, optional 145 | Index to stop training, by default None. 146 | 147 | Returns 148 | ------- 149 | None 150 | """ 151 | 152 | self.model = self._build_model() 153 | 154 | self.D = df[train_start:train_stop].values.T.copy() 155 | self.D = self.SS.fit_transform(self.D.T).T 156 | 157 | self.DxD = self.otimes(self.D, self.D) 158 | self.LU_factors = spla.lu_factor(self.DxD) 159 | 160 | def predict(self, data): 161 | """ 162 | Generate predictions using the trained MSET model. 163 | 164 | Parameters 165 | ---------- 166 | data : pandas.DataFrame 167 | Input data for generating predictions. 168 | 169 | Returns 170 | ------- 171 | pandas.DataFrame 172 | Predicted output data. 173 | """ 174 | 175 | X_obs = data.values.T.copy() 176 | X_obs = self.SS.transform(X_obs.T).T 177 | 178 | pred = np.zeros(X_obs.T.shape) 179 | 180 | for i in range(X_obs.shape[1]): 181 | pred[[i], :] = ( 182 | self.D @ self.calc_W(X_obs[:, i].reshape([-1, 1])) 183 | ).T 184 | 185 | return pd.DataFrame( 186 | self.SS.inverse_transform(pred), 187 | index=data.index, 188 | columns=data.columns, 189 | ) 190 | -------------------------------------------------------------------------------- /core/Vanilla_AE.py: -------------------------------------------------------------------------------- 1 | from tensorflow.keras.callbacks import EarlyStopping 2 | from tensorflow.keras.layers import ( 3 | Activation, 4 | BatchNormalization, 5 | Dense, 6 | Input, 7 | ) 8 | from tensorflow.keras.models import Model 9 | from tensorflow.keras.optimizers import Adam 10 | 11 | 12 | class Vanilla_AE: 13 | """ 14 | Feed-forward neural network with autoencoder architecture for anomaly detection using reconstruction error as an anomaly score. 15 | 16 | Parameters 17 | ---------- 18 | params : list 19 | List containing the following hyperparameters in order: 20 | - Number of neurons in the first encoder layer 21 | - Number of neurons in the bottleneck layer (latent representation) 22 | - Number of neurons in the first decoder layer 23 | - Learning rate for the optimizer 24 | - Batch size for training 25 | 26 | Attributes 27 | ---------- 28 | model : tensorflow.keras.models.Model 29 | The autoencoder model. 30 | 31 | Examples 32 | ------- 33 | >>> from Vanilla_AE import AutoEncoder 34 | >>> autoencoder = AutoEncoder(param=[5, 4, 2, 0.005, 32]) 35 | >>> autoencoder.fit(train_data) 36 | >>> predictions = autoencoder.predict(test_data) 37 | """ 38 | 39 | def __init__(self, params): 40 | self.param = params 41 | 42 | def _build_model(self): 43 | self._Random(0) 44 | 45 | input_dots = Input(shape=(self.shape,)) 46 | x = Dense(self.param[0])(input_dots) 47 | x = BatchNormalization()(x) 48 | x = Activation("relu")(x) 49 | 50 | x = Dense(self.param[1])(x) 51 | x = BatchNormalization()(x) 52 | x = Activation("relu")(x) 53 | 54 | bottleneck = Dense(self.param[2], activation="linear")(x) 55 | 56 | x = Dense(self.param[1])(bottleneck) 57 | x = BatchNormalization()(x) 58 | x = Activation("relu")(x) 59 | 60 | x = Dense(self.param[0])(x) 61 | x = BatchNormalization()(x) 62 | x = Activation("relu")(x) 63 | 64 | out = Dense(self.shape, activation="linear")(x) 65 | 66 | model = Model(input_dots, out) 67 | model.compile( 68 | optimizer=Adam(self.param[3]), loss="mae", metrics=["mse"] 69 | ) 70 | self.model = model 71 | 72 | return model 73 | 74 | def _Random(self, seed_value): 75 | import os 76 | 77 | os.environ["PYTHONHASHSEED"] = str(seed_value) 78 | 79 | import random 80 | 81 | random.seed(seed_value) 82 | 83 | import numpy as np 84 | 85 | np.random.seed(seed_value) 86 | 87 | import tensorflow as tf 88 | 89 | tf.random.set_seed(seed_value) 90 | 91 | def fit( 92 | self, 93 | data, 94 | early_stopping=True, 95 | validation_split=0.2, 96 | epochs=40, 97 | verbose=0, 98 | shuffle=True, 99 | ): 100 | """ 101 | Train the autoencoder model on the provided data. 102 | 103 | Parameters 104 | ---------- 105 | data : numpy.ndarray 106 | Input data for training. 107 | early_stopping : bool, optional 108 | Whether to use early stopping during training. 109 | validation_split : float, optional 110 | Fraction of the training data to be used as validation data. 111 | epochs : int, optional 112 | Number of training epochs. 113 | verbose : int, optional 114 | Verbosity mode (0 = silent, 1 = progress bar, 2 = current epoch and losses, 3 = each training iteration). 115 | shuffle : bool, optional 116 | Whether to shuffle the training data before each epoch. 117 | """ 118 | 119 | self.shape = data.shape[1] 120 | self.model = self._build_model() 121 | callbacks = [] 122 | if early_stopping: 123 | callbacks.append(EarlyStopping(patience=3, verbose=0)) 124 | self.model.fit( 125 | data, 126 | data, 127 | validation_split=validation_split, 128 | epochs=epochs, 129 | batch_size=self.param[4], 130 | verbose=verbose, 131 | shuffle=shuffle, 132 | callbacks=callbacks, 133 | ) 134 | 135 | def predict(self, data): 136 | """ 137 | Generate predictions using the trained autoencoder model. 138 | 139 | Parameters 140 | ---------- 141 | data : numpy.ndarray 142 | Input data for making predictions. 143 | 144 | Returns 145 | ------- 146 | numpy.ndarray 147 | The reconstructed output predictions. 148 | """ 149 | 150 | return self.model.predict(data) 151 | -------------------------------------------------------------------------------- /core/Vanilla_LSTM.py: -------------------------------------------------------------------------------- 1 | from tensorflow.keras.callbacks import EarlyStopping, ReduceLROnPlateau 2 | from tensorflow.keras.layers import LSTM, Dense 3 | from tensorflow.keras.models import Sequential 4 | 5 | 6 | class Vanilla_LSTM: 7 | """ 8 | LSTM-based neural network for anomaly detection using reconstruction error as an anomaly score. 9 | 10 | Parameters 11 | ---------- 12 | params : list 13 | A list containing various parameters for configuring the LSTM model. 14 | 15 | Attributes 16 | ---------- 17 | model : Sequential 18 | The trained LSTM model. 19 | 20 | Examples 21 | -------- 22 | >>> from Vanilla_LSTM import Vanilla_LSTM 23 | >>> PARAMS = [N_STEPS, EPOCHS, BATCH_SIZE, VAL_SPLIT] 24 | >>> lstm_model = Vanilla_LSTM(PARAMS) 25 | >>> lstm_model.fit(train_data, train_labels) 26 | >>> predictions = lstm_model.predict(test_data) 27 | """ 28 | 29 | def __init__(self, params): 30 | self.params = params 31 | 32 | def _Random(self, seed_value): 33 | import os 34 | 35 | os.environ["PYTHONHASHSEED"] = str(seed_value) 36 | 37 | import random 38 | 39 | random.seed(seed_value) 40 | 41 | import numpy as np 42 | 43 | np.random.seed(seed_value) 44 | 45 | import tensorflow as tf 46 | 47 | tf.random.set_seed(seed_value) 48 | 49 | def _build_model(self): 50 | self._Random(0) 51 | 52 | model = Sequential() 53 | model.add( 54 | LSTM( 55 | 100, 56 | activation="relu", 57 | return_sequences=True, 58 | input_shape=(self.params[0], self.n_features), 59 | ) 60 | ) 61 | model.add(LSTM(100, activation="relu")) 62 | model.add(Dense(self.n_features)) 63 | model.compile(optimizer="adam", loss="mae", metrics=["mse"]) 64 | return model 65 | 66 | def fit(self, X, y): 67 | """ 68 | Train the LSTM model on the provided data. 69 | 70 | Parameters 71 | ---------- 72 | X : numpy.ndarray 73 | Input data for training the model. 74 | y : numpy.ndarray 75 | Target data for training the model. 76 | """ 77 | self.n_features = X.shape[2] 78 | self.model = self._build_model() 79 | 80 | early_stopping = EarlyStopping(patience=10, verbose=0) 81 | 82 | reduce_lr = ReduceLROnPlateau( 83 | factor=0.1, patience=5, min_lr=0.0001, verbose=0 84 | ) 85 | 86 | self.model.fit( 87 | X, 88 | y, 89 | validation_split=self.params[3], 90 | epochs=self.params[1], 91 | batch_size=self.params[2], 92 | verbose=0, 93 | shuffle=False, 94 | callbacks=[early_stopping, reduce_lr], 95 | ) 96 | 97 | def predict(self, data): 98 | """ 99 | Generate predictions using the trained LSTM model. 100 | 101 | Parameters 102 | ---------- 103 | data : numpy.ndarray 104 | Input data for generating predictions. 105 | 106 | Returns 107 | ------- 108 | numpy.ndarray 109 | Predicted output data. 110 | """ 111 | 112 | return self.model.predict(data) 113 | -------------------------------------------------------------------------------- /core/__init.py__: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/waico/SKAB/b2c0d46c2971dcbfe71e26087b6d231998bb91c2/core/__init.py__ -------------------------------------------------------------------------------- /core/metrics.py: -------------------------------------------------------------------------------- 1 | """ 2 | This module is part of library (tsad)[https://github.com/waico/tsad] 3 | """ 4 | 5 | import matplotlib.gridspec as gridspec 6 | import matplotlib.pyplot as plt 7 | import numpy as np 8 | import pandas as pd 9 | 10 | 11 | def filter_detecting_boundaries(detecting_boundaries): 12 | """ 13 | [[t1,t2],[],[t1,t2]] -> [[t1,t2],[t1,t2]] 14 | [[],[]] -> [] 15 | """ 16 | _detecting_boundaries = [] 17 | for couple in detecting_boundaries.copy(): 18 | if len(couple) != 0: 19 | _detecting_boundaries.append(couple) 20 | detecting_boundaries = _detecting_boundaries 21 | return detecting_boundaries 22 | 23 | 24 | def single_detecting_boundaries( 25 | true_series, 26 | true_list_ts, 27 | prediction, 28 | portion, 29 | window_width, 30 | anomaly_window_destination, 31 | intersection_mode, 32 | ): 33 | """ 34 | Extract detecting_boundaries from series or list of timestamps 35 | """ 36 | 37 | if (true_series is not None) and (true_list_ts is not None): 38 | raise Exception("Choose the ONE type") 39 | elif true_series is not None: 40 | true_timestamps = true_series[true_series == 1].index 41 | elif true_list_ts is not None: 42 | if len(true_list_ts) == 0: 43 | return [[]] 44 | else: 45 | true_timestamps = true_list_ts 46 | else: 47 | raise Exception("Choose the type") 48 | # 49 | detecting_boundaries = [] 50 | td = ( 51 | pd.Timedelta(window_width) 52 | if window_width is not None 53 | else pd.Timedelta( 54 | (prediction.index[-1] - prediction.index[0]) 55 | / (len(true_timestamps) + 1) 56 | * portion 57 | ) 58 | ) 59 | for val in true_timestamps: 60 | if anomaly_window_destination == "lefter": 61 | detecting_boundaries.append([val - td, val]) 62 | elif anomaly_window_destination == "righter": 63 | detecting_boundaries.append([val, val + td]) 64 | elif anomaly_window_destination == "center": 65 | detecting_boundaries.append([val - td / 2, val + td / 2]) 66 | else: 67 | raise RuntimeError("choose anomaly_window_destination") 68 | 69 | # block for resolving intersection problem: 70 | # important to watch right boundary to be never included to avoid windows intersection 71 | if len(detecting_boundaries) == 0: 72 | return detecting_boundaries 73 | 74 | new_detecting_boundaries = detecting_boundaries.copy() 75 | intersection_count = 0 76 | for i in range(len(new_detecting_boundaries) - 1): 77 | if ( 78 | new_detecting_boundaries[i][1] 79 | >= new_detecting_boundaries[i + 1][0] 80 | ): 81 | # transform print to list of intersections 82 | # print(f'Intersection of scoring windows {new_detecting_boundaries[i][1], new_detecting_boundaries[i+1][0]}') 83 | intersection_count += 1 84 | if intersection_mode == "cut left window": 85 | new_detecting_boundaries[i][1] = new_detecting_boundaries[ 86 | i + 1 87 | ][0] 88 | elif intersection_mode == "cut right window": 89 | new_detecting_boundaries[i + 1][0] = new_detecting_boundaries[ 90 | i 91 | ][1] 92 | elif intersection_mode == "cut both": 93 | _a = new_detecting_boundaries[i][1] 94 | new_detecting_boundaries[i][1] = new_detecting_boundaries[ 95 | i + 1 96 | ][0] 97 | new_detecting_boundaries[i + 1][0] = _a 98 | else: 99 | raise Exception("choose the intersection_mode") 100 | # print(f'There are {intersection_count} intersections of scoring windows') 101 | detecting_boundaries = new_detecting_boundaries.copy() 102 | return detecting_boundaries 103 | 104 | 105 | def check_errors(my_list): 106 | """ 107 | Check format of input true data 108 | 109 | Parameters 110 | ---------- 111 | my_list - uniform format of true (See evaluate.evaluate) 112 | 113 | Returns 114 | ---------- 115 | mx : depth of list, or variant of processing 116 | """ 117 | assert isinstance(my_list, list) 118 | mx = 1 119 | # ravel = [] 120 | level_list = {} 121 | 122 | def check_error(my_list): 123 | return not ( 124 | (all(isinstance(my_el, list) for my_el in my_list)) 125 | or (all(isinstance(my_el, pd.Series) for my_el in my_list)) 126 | or (all(isinstance(my_el, pd.Timestamp) for my_el in my_list)) 127 | ) 128 | 129 | def recurse(my_list, level=1): 130 | nonlocal mx 131 | nonlocal level_list 132 | 133 | if check_error(my_list): 134 | raise Exception( 135 | f"Non uniform data format in level {level}: {my_list}" 136 | ) 137 | 138 | if level not in level_list.keys(): 139 | level_list[level] = [] # for checking format 140 | 141 | for my_el in my_list: 142 | level_list[level].append(my_el) 143 | if isinstance(my_el, list): 144 | mx = max([mx, level + 1]) 145 | recurse(my_el, level + 1) 146 | 147 | recurse(my_list) 148 | for level in level_list: 149 | if check_error(level_list[level]): 150 | raise Exception( 151 | f"Non uniform data format in level {level}: {my_list}" 152 | ) 153 | 154 | if 3 in level_list: 155 | for el in level_list[2]: 156 | if not ((len(el) == 2) or (len(el) == 0)): 157 | raise Exception( 158 | f"Non uniform data format in level {2}: {my_list}" 159 | ) 160 | return mx 161 | 162 | 163 | def extract_cp_confusion_matrix( 164 | detecting_boundaries, prediction, point=0, binary=False 165 | ): 166 | """ 167 | prediction: pd.Series 168 | 169 | point=None for binary case 170 | Returns 171 | ---------- 172 | dict: TPs: dict of numer window of [t1,t_cp,t2] 173 | FPs: list of timestamps 174 | FNs: list of numer window 175 | """ 176 | _detecting_boundaries = [] 177 | for couple in detecting_boundaries.copy(): 178 | if len(couple) != 0: 179 | _detecting_boundaries.append(couple) 180 | detecting_boundaries = _detecting_boundaries 181 | 182 | times_pred = prediction[prediction.dropna() == 1].sort_index().index 183 | 184 | my_dict = {} 185 | my_dict["TPs"] = {} 186 | my_dict["FPs"] = [] 187 | my_dict["FNs"] = [] 188 | 189 | if len(detecting_boundaries) != 0: 190 | my_dict["FPs"].append( 191 | times_pred[times_pred < detecting_boundaries[0][0]] 192 | ) # left 193 | for i in range(len(detecting_boundaries)): 194 | times_pred_window = times_pred[ 195 | (times_pred >= detecting_boundaries[i][0]) 196 | & (times_pred <= detecting_boundaries[i][1]) 197 | ] 198 | times_prediction_in_window = prediction[ 199 | detecting_boundaries[i][0] : detecting_boundaries[i][1] 200 | ].index 201 | if len(times_pred_window) == 0: 202 | if not binary: 203 | my_dict["FNs"].append(i) 204 | else: 205 | my_dict["FNs"].append(times_prediction_in_window) 206 | else: 207 | my_dict["TPs"][i] = [ 208 | detecting_boundaries[i][0], 209 | times_pred_window[point] 210 | if not binary 211 | else times_pred_window, # attention 212 | detecting_boundaries[i][1], 213 | ] 214 | if binary: 215 | my_dict["FNs"].append( 216 | times_prediction_in_window[ 217 | ~times_prediction_in_window.isin(times_pred_window) 218 | ] 219 | ) 220 | if len(detecting_boundaries) > i + 1: 221 | my_dict["FPs"].append( 222 | times_pred[ 223 | (times_pred > detecting_boundaries[i][1]) 224 | & (times_pred < detecting_boundaries[i + 1][0]) 225 | ] 226 | ) 227 | 228 | my_dict["FPs"].append( 229 | times_pred[times_pred > detecting_boundaries[i][1]] 230 | ) # right 231 | else: 232 | my_dict["FPs"].append(times_pred) 233 | 234 | if len(my_dict["FPs"]) > 1: 235 | my_dict["FPs"] = np.concatenate(my_dict["FPs"]) 236 | elif len(my_dict["FPs"]) == 1: 237 | my_dict["FPs"] = my_dict["FPs"][0] 238 | if len(my_dict["FPs"]) == 0: # not elif on purpose 239 | my_dict["FPs"] = [] 240 | 241 | if binary: 242 | if len(my_dict["FNs"]) > 1: 243 | my_dict["FNs"] = np.concatenate(my_dict["FNs"]) 244 | elif len(my_dict["FNs"]) == 1: 245 | my_dict["FNs"] = my_dict["FNs"][0] 246 | if len(my_dict["FNs"]) == 0: # not elif on purpose 247 | my_dict["FNs"] = [] 248 | return my_dict 249 | 250 | 251 | def confusion_matrix(true, prediction): 252 | true_ = true == 1 253 | prediction_ = prediction == 1 254 | TP = (true_ & prediction_).sum() 255 | TN = (~true_ & ~prediction_).sum() 256 | FP = (~true_ & prediction_).sum() 257 | FN = (true_ & ~prediction_).sum() 258 | return TP, TN, FP, FN 259 | 260 | 261 | def single_average_delay( 262 | detecting_boundaries, 263 | prediction, 264 | anomaly_window_destination, 265 | clear_anomalies_mode, 266 | ): 267 | """ 268 | anomaly_window_destination: 'lefter', 'righter', 'center'. Default='right' 269 | """ 270 | detecting_boundaries = filter_detecting_boundaries(detecting_boundaries) 271 | point = 0 if clear_anomalies_mode else -1 272 | dict_cp_confusion = extract_cp_confusion_matrix( 273 | detecting_boundaries, prediction, point=point 274 | ) 275 | 276 | missing = 0 277 | detectHistory = [] 278 | all_true_anom = 0 279 | FP = 0 280 | 281 | FP += len(dict_cp_confusion["FPs"]) 282 | missing += len(dict_cp_confusion["FNs"]) 283 | all_true_anom += len(dict_cp_confusion["TPs"]) + len( 284 | dict_cp_confusion["FNs"] 285 | ) 286 | 287 | if anomaly_window_destination == "lefter": 288 | 289 | def average_time(output_cp_cm_tp): 290 | return output_cp_cm_tp[2] - output_cp_cm_tp[1] 291 | elif anomaly_window_destination == "righter": 292 | 293 | def average_time(output_cp_cm_tp): 294 | return output_cp_cm_tp[1] - output_cp_cm_tp[0] 295 | elif anomaly_window_destination == "center": 296 | 297 | def average_time(output_cp_cm_tp): 298 | return output_cp_cm_tp[1] - ( 299 | output_cp_cm_tp[0] 300 | + (output_cp_cm_tp[2] - output_cp_cm_tp[0]) / 2 301 | ) 302 | else: 303 | raise Exception("Choose anomaly_window_destination") 304 | 305 | for fp_case_window in dict_cp_confusion["TPs"]: 306 | detectHistory.append( 307 | average_time(dict_cp_confusion["TPs"][fp_case_window]) 308 | ) 309 | return missing, detectHistory, FP, all_true_anom 310 | 311 | 312 | def my_scale( 313 | fp_case_window=None, 314 | A_tp=1, 315 | A_fp=0, 316 | koef=1, 317 | detalization=1000, 318 | clear_anomalies_mode=True, 319 | plot_figure=False, 320 | ): 321 | """ 322 | ts - segment on which the window is applied 323 | """ 324 | x = np.linspace(-np.pi / 2, np.pi / 2, detalization) 325 | x = x if clear_anomalies_mode else x[::-1] 326 | y = ( 327 | (A_tp - A_fp) 328 | / 2 329 | * -1 330 | * np.tanh(koef * x) 331 | / (np.tanh(np.pi * koef / 2)) 332 | + (A_tp - A_fp) / 2 333 | + A_fp 334 | ) 335 | if not plot_figure and fp_case_window is not None: 336 | event = int( 337 | (fp_case_window[1] - fp_case_window[0]) 338 | / (fp_case_window[-1] - fp_case_window[0]) 339 | * detalization 340 | ) 341 | if event >= len(x): 342 | event = len(x) - 1 343 | score = y[event] 344 | return score 345 | else: 346 | return y 347 | 348 | 349 | def single_evaluate_nab( 350 | detecting_boundaries, 351 | prediction, 352 | table_of_coef=None, 353 | clear_anomalies_mode=True, 354 | scale_func="improved", 355 | scale_koef=1, 356 | ): 357 | """ 358 | 359 | detecting_boundaries: list of list of two float values 360 | The list of lists of left and right boundary indices 361 | for scoring results of labeling if empty. Can be [[]], or [[],[t1,t2],[]] 362 | table_of_coef: pandas array (3x4) of float values 363 | Table of coefficients for NAB score function 364 | indices: 'Standard','LowFP','LowFN' 365 | columns:'A_tp','A_fp','A_tn','A_fn' 366 | 367 | scale_func {default}, improved 368 | недостатки scale_func default - 369 | 1 - зависит от относительного шага, а это значит, что если 370 | слишком много точек в scoring window то перепад будет слишком 371 | жестким в середение. 372 | 2- то самая левая точка не равно Atp, а права не равна Afp 373 | (особенно если пррименять расплывающую множитель) 374 | 375 | clear_anomalies_mode тогда слева от границы Atp срправа Afp, 376 | иначе fault mode, когда слева от границы Afp срправа Atp 377 | """ 378 | if scale_func == "improved": 379 | scale_func = my_scale 380 | else: 381 | raise Exception("choose the scale_func") 382 | 383 | # filter 384 | detecting_boundaries = filter_detecting_boundaries(detecting_boundaries) 385 | 386 | if table_of_coef is None: 387 | table_of_coef = pd.DataFrame( 388 | [ 389 | [1.0, -0.11, 1.0, -1.0], 390 | [1.0, -0.22, 1.0, -1.0], 391 | [1.0, -0.11, 1.0, -2.0], 392 | ] 393 | ) 394 | table_of_coef.index = pd.Index(["Standard", "LowFP", "LowFN"]) 395 | table_of_coef.index.name = "Metric" 396 | table_of_coef.columns = ["A_tp", "A_fp", "A_tn", "A_fn"] 397 | 398 | # GO 399 | point = 0 if clear_anomalies_mode else -1 400 | dict_cp_confusion = extract_cp_confusion_matrix( 401 | detecting_boundaries, prediction, point=point 402 | ) 403 | 404 | Scores, Scores_perfect, Scores_null = [], [], [] 405 | for profile in ["Standard", "LowFP", "LowFN"]: 406 | A_tp = table_of_coef["A_tp"][profile] 407 | A_fp = table_of_coef["A_fp"][profile] 408 | A_fn = table_of_coef["A_fn"][profile] 409 | 410 | score = 0 411 | score += A_fp * len(dict_cp_confusion["FPs"]) 412 | score += A_fn * len(dict_cp_confusion["FNs"]) 413 | for fp_case_window in dict_cp_confusion["TPs"]: 414 | set_times = dict_cp_confusion["TPs"][fp_case_window] 415 | score += scale_func(set_times, A_tp, A_fp, koef=scale_koef) 416 | 417 | Scores.append(score) 418 | Scores_perfect.append(len(detecting_boundaries) * A_tp) 419 | Scores_null.append(len(detecting_boundaries) * A_fn) 420 | 421 | return np.array( 422 | [np.array(Scores), np.array(Scores_null), np.array(Scores_perfect)] 423 | ) 424 | 425 | 426 | def chp_score( 427 | true, 428 | prediction, 429 | metric="nab", 430 | window_width=None, 431 | portion=0.1, 432 | anomaly_window_destination="lefter", 433 | clear_anomalies_mode=True, 434 | intersection_mode="cut right window", 435 | table_of_coef=None, 436 | scale_func="improved", 437 | scale_koef=1, 438 | plot_figure=False, 439 | verbose=True, 440 | ): 441 | """ 442 | Parameters 443 | ---------- 444 | true: variants: 445 | or: if one dataset : pd.Series with binary int labels (1 is 446 | anomaly, 0 is not anomaly); 447 | 448 | or: if one dataset : list of pd.Timestamp of true labels, or [] 449 | if haven't labels ; 450 | 451 | or: if one dataset : list of list of t1,t2: left and right 452 | detection, boundaries of pd.Timestamp or [[]] if haven't labels 453 | 454 | or: if many datasets: list (len of number of datasets) of pd.Series 455 | with binary int labels; 456 | 457 | or: if many datasets: list of list of pd.Timestamp of true labels, or 458 | true = [ts,[]] if haven't labels for specific dataset; 459 | 460 | or: if many datasets: list of list of list of t1,t2: left and right 461 | detection boundaries of pd.Timestamp; 462 | If we haven't true labels for specific dataset then we must insert 463 | empty list of labels: true = [[[]],[[t1,t2],[t1,t2]]]. 464 | 465 | __True labels of anomalies or changepoints. 466 | It is important to have appropriate labels (CP or 467 | anomaly) for corresponding metric (See later "metric") 468 | 469 | prediction: variants: 470 | or: if one dataset : pd.Series with binary int labels 471 | (1 is anomaly, 0 is not anomaly); 472 | 473 | or: if many datasets: list (len of number of datasets) 474 | of pd.Series with binary int labels. 475 | 476 | __Predicted labels of anomalies or changepoints. 477 | It is important to have appropriate labels (CP or 478 | anomaly) for corresponding metric (See later "metric") 479 | 480 | metric: {'nab', 'binary', 'average_time', 'confusion_matrix'}. 481 | Default='nab' 482 | Affects to output (see later: Returns) 483 | Changepoint problem: {'nab', 'average_time'}. 484 | Standard AD problem: {'binary', 'confusion_matrix'}. 485 | 'nab' is Numenta Anomaly Benchmark metric 486 | 487 | 'average_time' is both average delay or time to failure 488 | depend on situation. 489 | 490 | 'binary': FAR, MAR, F1. 491 | 492 | 'confusion_matrix' standard confusion_matrix for any point. 493 | 494 | window_width: 'str' for pd.Timedelta 495 | Width of detection window. Default=None. 496 | 497 | portion : float, default=0.1 498 | The portion is needed if window_width = None. 499 | The width of the detection window in this case is equal 500 | to a portion of the width of the length of prediction divided 501 | by the number of real CPs in this dataset. Default=0.1. 502 | 503 | anomaly_window_destination: {'lefter', 'righter', 'center'}. Default='right' 504 | The parameter of the location of the detection window relative to the anomaly. 505 | 'lefter' : the detection window will be on the left side of the anomaly 506 | 'righter' : the detection window will be on the right side of the anomaly 507 | 'center' : the scoring window will be positioned relative to the center of anom. 508 | 509 | clear_anomalies_mode : boolean, default=True. 510 | True : then the `left value of a Scoring function is Atp and the 511 | `right is Afp. Only the `first value inside the detection window is taken. 512 | False: then the `right value of a Scoring function is Atp and the 513 | `left is Afp. Only the `last value inside the detection window is taken. 514 | 515 | intersection_mode: {'cut left window', 'cut right window', 'both'}. 516 | Default='cut right window' 517 | The parameter will be used if the detection windows overlap for 518 | true changepoints, which is generally undesirable and requires a 519 | different approach than simply cropping the scoring window using 520 | this parameter. 521 | 'cut left window' : will cut the overlapping part of the left window 522 | 'cut right window': will cut the intersecting part of the right window 523 | 'both' : will crop the intersecting portion of both the left 524 | and right windows 525 | 526 | verbose: boolean, default=True. 527 | If True, then output useful information 528 | 529 | plot_figure : boolean, default=False. 530 | If True, then drawing the score fuctions, detection windows and predictions 531 | It is used for example, for calibration the scale_koef. 532 | 533 | table_of_coef (metric='nab'): pd.DataFrame of specific form. See bellow. 534 | Application profiles of NAB metric.If Default is None: 535 | table_of_coef = pd.DataFrame([[1.0,-0.11,1.0,-1.0], 536 | [1.0,-0.22,1.0,-1.0], 537 | [1.0,-0.11,1.0,-2.0]]) 538 | table_of_coef.index = ['Standard','LowFP','LowFN'] 539 | table_of_coef.index.name = "Metric" 540 | table_of_coef.columns = ['A_tp','A_fp','A_tn','A_fn'] 541 | 542 | scale_func (metric='nab'): "default" of "improved". Default="improved". 543 | Scoring function in NAB metric. 544 | 'default' : standard NAB scoring function 545 | 'improved' : Our function for resolving disadvantages 546 | of standard NAB scoring function 547 | 548 | scale_koef : float > 0. Default=1.0. 549 | Smoothing factor. The smaller it is, 550 | the smoother the scoring function is. 551 | 552 | Returns 553 | ---------- 554 | metrics : value of metrics, depend on metric 555 | 'nab': tuple 556 | - Standard profile, float 557 | - Low FP profile, float 558 | - Low FN profile 559 | 'average_time': tuple 560 | - Average time (average delay, or time to failure) 561 | - Missing changepoints, int 562 | - FPs, int 563 | - Number of true changepoints, int 564 | 'binary': tuple 565 | - F1 metric, float 566 | - False alarm rate, %, float 567 | - Missing Alarm Rate, %, float 568 | 'binary': tuple 569 | - TPs, int 570 | - TNs, int 571 | - FPs, int 572 | - FNS, int 573 | 574 | """ 575 | 576 | assert isinstance(true, pd.Series) or isinstance(true, list) 577 | # checking prediction 578 | if isinstance(prediction, pd.Series): 579 | true = [true] 580 | prediction = [prediction] 581 | elif isinstance(prediction, list): 582 | if not all(isinstance(my_el, pd.Series) for my_el in prediction): 583 | raise Exception("Incorrect format for prediction") 584 | else: 585 | raise Exception("Incorrect format for prediction") 586 | 587 | # checking dataset length: Number of dataset unequal 588 | assert len(true) == len(prediction) 589 | 590 | # final check 591 | input_variant = check_errors(true) 592 | 593 | def check_sort(my_list, input_variant): 594 | for dataset in my_list: 595 | if input_variant == 2: 596 | assert all(np.sort(dataset) == np.array(dataset)) 597 | elif input_variant == 3: 598 | assert all( 599 | np.sort(np.concatenate(dataset)) == np.concatenate(dataset) 600 | ) 601 | elif input_variant == 1: 602 | assert all( 603 | dataset.index.values == dataset.sort_index().index.values 604 | ) 605 | 606 | check_sort(true, input_variant) 607 | check_sort(prediction, 1) 608 | 609 | # part 2. To detected boundaries 610 | if ( 611 | ((metric == "nab") or (metric == "average_time")) 612 | and (window_width is None) 613 | and (input_variant != 3) 614 | ): 615 | print( 616 | f"Since you didn't choose window_width and portion, portion will be default ({portion})" 617 | ) 618 | 619 | if input_variant == 1: 620 | detecting_boundaries = [ 621 | single_detecting_boundaries( 622 | true_series=true[i], 623 | true_list_ts=None, 624 | prediction=prediction[i], 625 | window_width=window_width, 626 | portion=portion, 627 | anomaly_window_destination=anomaly_window_destination, 628 | intersection_mode=intersection_mode, 629 | ) 630 | for i in range(len(true)) 631 | ] 632 | 633 | elif input_variant == 2: 634 | detecting_boundaries = [ 635 | single_detecting_boundaries( 636 | true_series=None, 637 | true_list_ts=true[i], 638 | prediction=prediction[i], 639 | window_width=window_width, 640 | portion=portion, 641 | anomaly_window_destination=anomaly_window_destination, 642 | intersection_mode=intersection_mode, 643 | ) 644 | for i in range(len(true)) 645 | ] 646 | 647 | elif input_variant == 3: 648 | detecting_boundaries = true.copy() 649 | # Next anti fool system [[[t1,t2]],[]] -> [[[t1,t2]],[[]]] 650 | for i in range(len(detecting_boundaries)): 651 | if len(detecting_boundaries[i]) == 0: 652 | detecting_boundaries[i] = [[]] 653 | else: 654 | raise Exception("Unknown format for true data") 655 | 656 | # part 3. To compute metric 657 | if plot_figure: 658 | num_datasets = len(true) 659 | if ((metric == "binary") or (metric == "confusion_matrix")) and ( 660 | input_variant == 1 661 | ): 662 | f = plt.figure(figsize=(16, 5 * num_datasets)) 663 | grid = gridspec.GridSpec(num_datasets, 1) 664 | for i in range(num_datasets): 665 | globals()["ax" + str(i)] = f.add_subplot(grid[i]) 666 | prediction[i].plot( 667 | ax=globals()["ax" + str(i)], label="pred", marker="o" 668 | ) 669 | true[i].plot( # type: ignore 670 | ax=globals()["ax" + str(i)], label="true", marker="o" 671 | ) 672 | globals()["ax" + str(i)].legend() 673 | plt.show() 674 | else: 675 | f = plt.figure(figsize=(16, 5 * num_datasets)) 676 | grid = gridspec.GridSpec(num_datasets, 1) 677 | detalization = 100 678 | for i in range(num_datasets): 679 | globals()["ax" + str(i)] = f.add_subplot(grid[i]) 680 | print_legend_boundary = True 681 | 682 | def plot_cp(couple, anomaly_window_destination, ax, label): 683 | if anomaly_window_destination == "lefter": 684 | ax.axvline(couple[1], c="r", label=label) 685 | elif anomaly_window_destination == "righter": 686 | ax.axvline(couple[0], c="r", label=label) 687 | elif anomaly_window_destination == "center": 688 | ax.axvline( 689 | couple[0] + ((couple[1] - couple[0]) / 2), 690 | c="r", 691 | label=label, 692 | ) 693 | 694 | for couple in detecting_boundaries[i]: 695 | if len(couple) > 0: 696 | globals()["ax" + str(i)].axvspan( 697 | couple[0], 698 | couple[1], 699 | alpha=0.5, 700 | color="green", 701 | label="detection \nboundary" 702 | if print_legend_boundary 703 | else None, 704 | ) 705 | nab = pd.Series( 706 | my_scale( 707 | plot_figure=True, detalization=detalization 708 | ), 709 | index=pd.date_range( 710 | couple[0], couple[1], periods=detalization 711 | ), 712 | ) 713 | nab.plot( 714 | ax=globals()["ax" + str(i)], 715 | linewidth=0.4, 716 | color="brown", 717 | label="nab scoring func" 718 | if print_legend_boundary 719 | else None, 720 | ) 721 | plot_cp( 722 | couple, 723 | anomaly_window_destination, 724 | globals()["ax" + str(i)], 725 | label="Changepoint" 726 | if print_legend_boundary 727 | else None, 728 | ) 729 | print_legend_boundary = False 730 | else: 731 | pass 732 | prediction[i].plot( 733 | ax=globals()["ax" + str(i)], label="pred", marker="o" 734 | ) 735 | globals()["ax" + str(i)].legend() 736 | plt.show() 737 | 738 | if metric == "nab": 739 | matrix = np.zeros((3, 3)) 740 | for i in range(len(prediction)): 741 | matrix_ = single_evaluate_nab( 742 | detecting_boundaries[i], 743 | prediction[i], 744 | table_of_coef=table_of_coef, 745 | clear_anomalies_mode=clear_anomalies_mode, 746 | scale_func=scale_func, 747 | scale_koef=scale_koef, 748 | # plot_figure=plot_figure, 749 | ) 750 | matrix = matrix + matrix_ 751 | 752 | results = {} 753 | desc = ["Standard", "LowFP", "LowFN"] 754 | for t, profile_name in enumerate(desc): 755 | results[profile_name] = round( 756 | 100 757 | * (matrix[0, t] - matrix[1, t]) 758 | / (matrix[2, t] - matrix[1, t]), 759 | 2, 760 | ) 761 | if verbose: 762 | print(profile_name, " - ", results[profile_name]) 763 | return results 764 | 765 | elif metric == "average_time": 766 | missing, detectHistory, FP, all_true_anom = 0, [], 0, 0 767 | for i in range(len(prediction)): 768 | missing_, detectHistory_, FP_, all_true_anom_ = ( 769 | single_average_delay( 770 | detecting_boundaries[i], 771 | prediction[i], 772 | anomaly_window_destination=anomaly_window_destination, 773 | clear_anomalies_mode=clear_anomalies_mode, 774 | ) 775 | ) 776 | missing, detectHistory, FP, all_true_anom = ( 777 | missing + missing_, 778 | detectHistory + detectHistory_, 779 | FP + FP_, 780 | all_true_anom + all_true_anom_, 781 | ) 782 | add = np.mean(detectHistory) 783 | if verbose: 784 | print("Amount of true anomalies", all_true_anom) 785 | print(f"A number of missed CPs = {missing}") 786 | print(f"A number of FPs = {int(FP)}") 787 | print("Average time", add) 788 | return add, missing, int(FP), all_true_anom 789 | 790 | elif (metric == "binary") or (metric == "confusion_matrix"): 791 | if all(isinstance(my_el, pd.Series) for my_el in true): 792 | TP, TN, FP, FN = 0, 0, 0, 0 793 | for i in range(len(prediction)): 794 | TP_, TN_, FP_, FN_ = confusion_matrix(true[i], prediction[i]) 795 | TP, TN, FP, FN = TP + TP_, TN + TN_, FP + FP_, FN + FN_ 796 | else: 797 | print( 798 | "For this metric it is better if you use pd.Series format for true \nwith common index of true and prediction" 799 | ) 800 | TP, TN, FP, FN = 0, 0, 0, 0 801 | for i in range(len(prediction)): 802 | dict_cp_confusion = extract_cp_confusion_matrix( 803 | detecting_boundaries[i], prediction[i], binary=True 804 | ) 805 | TP += np.sum( 806 | [ 807 | len(dict_cp_confusion["TPs"][window][1]) 808 | for window in dict_cp_confusion["TPs"] 809 | ] 810 | ) 811 | FP += len(dict_cp_confusion["FPs"]) 812 | FN += len(dict_cp_confusion["FNs"]) 813 | TN += len(prediction[i]) - TP - FP - FN 814 | 815 | if metric == "binary": 816 | f1 = round(TP / (TP + (FN + FP) / 2), 2) 817 | far = round(FP / (FP + TN) * 100, 2) 818 | mar = round(FN / (FN + TP) * 100, 2) 819 | if verbose: 820 | print(f"False Alarm Rate {far} %") 821 | print(f"Missing Alarm Rate {mar} %") 822 | print(f"F1 metric {f1}") 823 | return f1, far, mar 824 | 825 | elif metric == "confusion_matrix": 826 | if verbose: 827 | print("TP", TP) 828 | print("TN", TN) 829 | print("FP", FP) 830 | print("FN", FN) 831 | return TP, TN, FP, FN 832 | else: 833 | raise Exception("Choose the performance metric") 834 | -------------------------------------------------------------------------------- /core/t2.py: -------------------------------------------------------------------------------- 1 | # Author: Iurii Katser 2 | 3 | import os 4 | from math import sqrt 5 | 6 | import numpy as np 7 | import scipy.stats as SS 8 | from matplotlib import pyplot as plt 9 | from numpy import linalg as LA 10 | from pandas import DataFrame 11 | from sklearn.decomposition import PCA 12 | from sklearn.preprocessing import StandardScaler 13 | 14 | 15 | class T2: 16 | """Calculation of the Hotelling's 1-dimensional T-squared 17 | statistic or T-squared statistic+Q-statistic based on PCA for 18 | anomaly detection in multivariate data. 19 | 20 | Based on the following papers: 21 | [1] - Q-statistic and T2-statistic PCA-based measures for damage 22 | assessment in structures / LE Mujica, J. Rodellar, A. Ferna ́ndez, 23 | A. Gu ̈emes // Structural Health Monitoring: An International 24 | Journal. — 2010. — nov. — Vol. 10, no. 5. — Pp. 539–553. 25 | [2] - Zhao Chunhui, Gao Furong. Online fault prognosis with 26 | relative deviation analysis and vector autoregressive modeling // 27 | Chemical Engineering Science. — 2015. — dec. — Vol. 138. — Pp. 28 | 531–543. 29 | [3] - Li Wei, Peng Minjun, Wang Qingzhong. False alarm reducing in 30 | PCA method for sensor fault detection in a nuclear power plant // 31 | Annals of Nuclear Energy. — 2018. — aug. — Vol. 118. — Pp. 131–139. 32 | 33 | Parameters 34 | ---------- 35 | scaling : boolean, default = False 36 | If True StandartScaler is used in the pipeline. 37 | If False no scaling procedures are used. 38 | 39 | using_pca : boolean, default = True 40 | If True T2+Q based on PCA is used as anomaly detection method. 41 | If False T2 without PCA is used as anomaly detection method. 42 | 43 | explained_variance : object, default = 0.85 44 | Proportion of the explained variance for principal components 45 | selection. Relevant only if using_pca=True. 46 | 47 | p_value : object, default = 0.999 48 | P value for upper control limits selection. Shows the proportion 49 | of the number of points in train set perceived as normal. 50 | 51 | Examples 52 | -------- 53 | T2+Q based on PCA: 54 | 55 | from ControlCharts import T2 56 | import pandas as pd 57 | import numpy as np 58 | df = pd.DataFrame(np.random.randn(100, 4), columns=list('ABCD')) 59 | t2 = T2() 60 | t2.fit(df.iloc[:20]) 61 | t2.predict(df) 62 | 63 | T2 without PCA: 64 | 65 | from ControlCharts import T2 66 | import pandas as pd 67 | import numpy as np 68 | df = pd.DataFrame(np.random.randn(100, 4), columns=list('ABCD')) 69 | t2 = T2(using_pca=False) 70 | t2.fit(df.iloc[:20]) 71 | t2.predict(df) 72 | 73 | More examples at: 74 | https://github.com/YKatser/control-charts/tree/main/examples 75 | """ 76 | 77 | def __init__( 78 | self, 79 | scaling=False, 80 | using_pca=True, 81 | explained_variance=0.85, 82 | p_value=0.999, 83 | ): 84 | self.scaling = scaling 85 | self.using_pca = using_pca 86 | self.explained_variance = explained_variance 87 | self.p_value = p_value 88 | 89 | # T2 and Q statistics calculations 90 | def _t2_calculation(self, x): 91 | t2 = [] 92 | for i in range(len(x)): 93 | t2.append(x[i] @ self.inv_cov @ x[i].T) 94 | return t2 95 | 96 | def _q_calculation(self, x): 97 | q = [] 98 | for i in range(len(x)): 99 | q.append(x[i] @ self.transform_rc @ x[i].T) 100 | return q 101 | 102 | # CALCULATING UPPER CONTROL LIMITS 103 | def _t2_ucl(self, x): 104 | if self.using_pca: 105 | m = self.n_components 106 | else: 107 | m = x.shape[1] 108 | 109 | n = len(x) 110 | linspace = np.linspace(0, 15, 10000) 111 | c_alpha = linspace[SS.f.cdf(linspace, m, n - m) < self.p_value][-1] 112 | # koef = m * (n-1) / (n-m) 113 | koef = m * (n - 1) * (n + 1) / (n * (n - m)) 114 | 115 | self.t2_ucl = koef * c_alpha 116 | 117 | def _q_ucl(self, x): 118 | w, v = LA.eig(np.cov(x.T)) 119 | sum_ = 0 120 | for i in range(self.n_components, len(w)): 121 | sum_ += w[i] 122 | 123 | tetta = [] 124 | for i in [1, 2, 3]: 125 | tetta.append(sum_**i) 126 | h0 = 1 - 2 * tetta[0] * tetta[2] / (3 * tetta[1] ** 2) 127 | linspace = np.linspace(0, 15, 10000) 128 | c_alpha = linspace[SS.norm.cdf(linspace) < self.p_value][-1] 129 | 130 | self.q_ucl = tetta[0] * ( 131 | 1 132 | + (c_alpha * h0 * sqrt(2 * tetta[1]) / tetta[0]) 133 | + tetta[1] * h0 * (h0 - 1) / tetta[0] ** 2 134 | ) ** (1 / h0) 135 | 136 | # applying pca 137 | def _pca_applying(self, x): 138 | self.pca = PCA(n_components=self.explained_variance).fit(x) 139 | self.n_components = self.pca.n_components_ 140 | self._EV = self.pca.components_.T 141 | return self.pca.transform(x) 142 | 143 | # PLOTTING AND SAVING RESULTS 144 | def plot_t2(self, t2=None, t2_ucl=None, save_fig=False, fig_name="T2"): 145 | """Plotting results of T2-statistic calculation with matplotlib 146 | 147 | Parameters 148 | ---------- 149 | t2 : pandas.DataFrame(), default = None 150 | Results of T2-statistic calculation. 151 | 152 | t2_ucl : float or int, default = None 153 | Upper control limit for T2. 154 | 155 | save_fig : boolean, default = False 156 | If True there will be saved T2 chart as .png to the 157 | current folder. 158 | 159 | fig_name : str, default = 'T2' 160 | Name of the saved figure. 161 | 162 | Returns 163 | ------- 164 | self : object. 165 | """ 166 | 167 | if t2 is None: 168 | t2 = self.t2 169 | if t2_ucl is None: 170 | t2_ucl = self.t2_ucl 171 | plt.figure(figsize=(12, 4)) 172 | plt.plot(t2, label="$T^2$-statistic") 173 | # for i in self.final_list: 174 | # plt.axvspan(i[0], i[1], facecolor='green', alpha=0.2, zorder=0, 175 | # label='Train set') 176 | plt.grid(True) 177 | plt.axhline(t2_ucl, zorder=10, color="r", label="UCL") 178 | plt.ylim(0, 3 * max(t2.min().values, t2_ucl)) 179 | plt.xlim(t2.index.values[0], t2.index.values[-1]) 180 | plt.title("$T^2$-statistic chart") 181 | plt.xlabel("Time") 182 | plt.ylabel("$T^2$-statistic value") 183 | plt.legend(["$T^2$-statistic", "UCL", "Train set"]) 184 | plt.tight_layout() 185 | if save_fig: 186 | self._save(name=fig_name) 187 | 188 | def plot_q(self, q=None, q_ucl=None, save_fig=False, fig_name="Q"): 189 | """Plotting results of Q-statistic calculation with matplotlib 190 | 191 | Parameters 192 | ---------- 193 | q : pandas.DataFrame(), default = None 194 | Results of Q-statistic calculation. 195 | 196 | q_ucl : float or int, default = None 197 | Upper control limit for Q. 198 | 199 | save_fig : boolean, default = False 200 | If True there will be saved Q chart as .png to the 201 | current folder. 202 | 203 | fig_name : str, default = 'Q' 204 | Name of the saved figure. 205 | 206 | Returns 207 | ------- 208 | self : object. 209 | """ 210 | 211 | if q is None: 212 | q = self.q 213 | if q_ucl is None: 214 | q_ucl = self.q_ucl 215 | plt.figure(figsize=(12, 4)) 216 | plt.plot(q, label="$Q$-statistic") 217 | # for i in self.final_list: 218 | # plt.axvspan(i[0], i[1], facecolor='green', alpha=0.2, zorder=0, 219 | # label='Train set') 220 | plt.grid(True) 221 | plt.axhline(q_ucl, zorder=10, color="r", label="UCL") 222 | plt.ylim(0, 3 * max(q.min().values, q_ucl)) 223 | plt.xlim(q.index.values[0], q.index.values[-1]) 224 | plt.title("$Q$-statistic chart") 225 | plt.xlabel("Time") 226 | plt.ylabel("$Q$-statistic value") 227 | plt.legend(["$Q$-statistic", "UCL", "Train set"]) 228 | plt.tight_layout() 229 | if save_fig: 230 | self._save(name=fig_name) 231 | 232 | @staticmethod 233 | def _save(name="", fmt="png"): 234 | pwd = os.getcwd() 235 | iPath = pwd + "/pictures/" 236 | if not os.path.exists(iPath): 237 | os.mkdir(iPath) 238 | os.chdir(iPath) 239 | plt.savefig(f"{name}.{fmt}", fmt="png", dpi=150, bbox_inches="tight") 240 | os.chdir(pwd) 241 | 242 | def fit(self, x): 243 | """Computation of the inversed covariance matrix, matrix of 244 | transformation to the residual space (in case of 245 | using_pca=True) and standart scaler fitting (in case of using 246 | scaling=True). 247 | 248 | Parameters 249 | ---------- 250 | x : pandas.DataFrame() 251 | Training set. 252 | 253 | Returns 254 | ------- 255 | self : object. 256 | """ 257 | 258 | x = x.copy() 259 | 260 | # removing constant columns 261 | initial_cols_number = len(x.columns) 262 | x = x.loc[:, (x != x.iloc[0]).any()] 263 | self._feature_names_in = x.columns 264 | if initial_cols_number > len(x.columns): 265 | print("Constant columns removed") 266 | 267 | if self.scaling: 268 | # fitting PCA and calculation of scaler, EV 269 | self.scaler = StandardScaler() 270 | self.scaler.fit(x) 271 | x_ = self.scaler.transform(x) 272 | else: 273 | x_ = x.values 274 | 275 | if self.using_pca: 276 | x_pc = self._pca_applying(x_) 277 | else: 278 | self.n_components = x.shape[1] 279 | 280 | if self.n_components == x.shape[1]: 281 | # preparing inv_cov for T2 282 | self.inv_cov = LA.inv(np.cov(x_.T)) 283 | 284 | # calculating T2_ucl 285 | self._t2_ucl(x_) 286 | if self.using_pca: 287 | print("""Number of principal components is equal to dataset \ 288 | shape. Q-statistics is unavailable.""") 289 | else: 290 | # preparing inv_cov for T2 (principal space) 291 | self.inv_cov = LA.inv(np.cov(x_pc.T)) 292 | 293 | # preparing transform matrix for Q (to residual space) 294 | self.transform_rc = np.eye(len(self._EV)) - np.dot( 295 | self._EV, self._EV.T 296 | ) 297 | 298 | # calculating t2_ucl and q_ucl 299 | self._t2_ucl(x_) 300 | self._q_ucl(x_) 301 | 302 | # calculating train indices 303 | # indices = x.index.tolist() 304 | # diff = x.index.to_series().diff() 305 | # list_of_ind = diff[diff > diff.mean() + 3 * diff.std()].index.tolist() 306 | 307 | def predict( 308 | self, 309 | x, 310 | plot_fig=True, 311 | save_fig=False, 312 | fig_name=["T2", "Q"], 313 | window_size=1, 314 | ): 315 | """Computation of T2-statistic or T2-statistic+Q-statistic for 316 | the test dataset. 317 | 318 | Parameters 319 | ---------- 320 | x : pandas.DataFrame() 321 | Testing dataset. 322 | 323 | plot_fig : boolean, default = True 324 | If True there will be plotted T2-statistics or 325 | T2-statistics+Q-statistics chart. 326 | 327 | save_fig : boolean, default = False 328 | If True there will be saved T2 and Q charts as .png to the 329 | current folder. 330 | 331 | fig_name : list of one or two str, default = ['T2','Q'] 332 | Names of the saved figures. 333 | 334 | window_size : int, default = 1 335 | Size of the window for median filter as a postprocessing. 336 | 337 | Returns 338 | ------- 339 | self : object 340 | Plotting and saving T2 or T2+Q charts. To get DataFrames 341 | with T2 or Q values call self.t2 or self.q. 342 | """ 343 | 344 | x = x.copy() 345 | x = x.loc[:, self._feature_names_in] 346 | if self.scaling: 347 | x_ = self.scaler.transform(x) 348 | else: 349 | x_ = x.values 350 | 351 | if self.n_components != x.shape[1]: 352 | # calculating T2 353 | self.t2 = ( 354 | DataFrame( 355 | self._t2_calculation(self.pca.transform(x_)), 356 | index=x.index, 357 | columns=["T2"], 358 | ) 359 | .rolling(window_size) 360 | .median() 361 | ) 362 | 363 | # calculating Q 364 | self.q = ( 365 | DataFrame( 366 | self._q_calculation(x_), index=x.index, columns=["Q"] 367 | ) 368 | .rolling(window_size) 369 | .median() 370 | ) 371 | 372 | # plotting 373 | if plot_fig: 374 | self.plot_t2( 375 | t2=self.t2, 376 | t2_ucl=self.t2_ucl, 377 | save_fig=save_fig, 378 | fig_name=fig_name[0], 379 | ) 380 | self.plot_q( 381 | q=self.q, 382 | q_ucl=self.q_ucl, 383 | save_fig=save_fig, 384 | fig_name=fig_name[1], 385 | ) 386 | 387 | else: 388 | # calculating T2 389 | self.t2 = ( 390 | DataFrame( 391 | self._t2_calculation(x_), index=x.index, columns=["T2"] 392 | ) 393 | .rolling(window_size) 394 | .median() 395 | ) 396 | 397 | # plotting 398 | if plot_fig: 399 | self.plot_t2(self.t2) 400 | if save_fig: 401 | self._save(name=fig_name[0]) 402 | -------------------------------------------------------------------------------- /core/utils.py: -------------------------------------------------------------------------------- 1 | import os 2 | 3 | import matplotlib.pyplot as plt 4 | import numpy as np 5 | import pandas as pd 6 | from sklearn.model_selection import train_test_split 7 | 8 | from .metrics import chp_score 9 | 10 | 11 | def load_skab(): 12 | path_to_data = "../data/" 13 | # benchmark files checking 14 | all_files = [] 15 | 16 | for root, dirs, files in os.walk(path_to_data): 17 | for file in files: 18 | if file.endswith(".csv"): 19 | all_files.append(os.path.join(root, file)) 20 | 21 | # datasets with anomalies loading 22 | list_of_df = [ 23 | pd.read_csv(file, sep=";", index_col="datetime", parse_dates=True) 24 | for file in all_files 25 | if "anomaly-free" not in file 26 | ] 27 | # anomaly-free df loading 28 | anomaly_free_df = pd.read_csv( 29 | [file for file in all_files if "anomaly-free" in file][0], 30 | sep=";", 31 | index_col="datetime", 32 | parse_dates=True, 33 | ) 34 | 35 | return list_of_df, anomaly_free_df 36 | 37 | 38 | def preprocess_skab(list_of_df): 39 | Xy_traintest_list: list[list] = [] 40 | for df in list_of_df: 41 | Xy_traintest_list.append( 42 | train_test_split( 43 | df.drop(["anomaly", "changepoint"], axis=1), 44 | df[["anomaly", "changepoint"]], 45 | train_size=400, 46 | shuffle=False, 47 | random_state=0, 48 | ) 49 | ) 50 | return Xy_traintest_list 51 | 52 | 53 | def load_preprocess_skab(): 54 | list_of_df, _ = load_skab() 55 | Xy_traintest_list = preprocess_skab(list_of_df) 56 | return Xy_traintest_list 57 | 58 | 59 | # Generated training sequences for use in the model. 60 | def create_sequences(values, time_steps): 61 | output = [] 62 | for i in range(len(values) - time_steps + 1): 63 | output.append(values[i : (i + time_steps)]) 64 | return np.stack(output) 65 | 66 | 67 | def plot_results(*true_pred_pairs: tuple[pd.Series, pd.Series]): 68 | n = len(true_pred_pairs) 69 | fig, axs = plt.subplots(n, 1, figsize=(12, 3 * n), sharex=True) 70 | if not isinstance(axs, (list | np.ndarray)): 71 | axs = [axs] 72 | for ax, (true, pred) in zip(axs, true_pred_pairs): 73 | ax.plot(true, label="True", marker="o", markersize=5) 74 | ax.plot(pred, label="Predicted", marker="x", markersize=5) 75 | ax.set_title(f"{true.name} detection") 76 | ax.legend() 77 | fig.show() 78 | 79 | 80 | def print_results( 81 | y_true, 82 | y_pred, 83 | score_kwargs: list[dict], 84 | ): 85 | for kwargs in score_kwargs: 86 | print(kwargs) 87 | chp_score(y_true, y_pred, **kwargs) 88 | print() 89 | -------------------------------------------------------------------------------- /data/README.md: -------------------------------------------------------------------------------- 1 | ``` 2 | └── data # Data files and processing Jupyter Notebook 3 | ├── Load data.ipynb # Jupyter Notebook to load all data 4 | ├── anomaly-free 5 | │ └── anomaly-free.csv # Data obtained from the experiments with normal mode 6 | ├── valve2 # Data obtained from the experiments with closing the valve at the outlet of the flow from the pump. 7 | │ ├── 1.csv 8 | │ ├── 2.csv 9 | │ ├── 3.csv 10 | │ └── 4.csv 11 | ├── valve1 # Data obtained from the experiments with closing the valve at the flow inlet to the pump. 12 | │ ├── 1.csv 13 | │ ├── 2.csv 14 | │ ├── 3.csv 15 | │ ├── 4.csv 16 | │ ├── 5.csv 17 | │ ├── 6.csv 18 | │ ├── 7.csv 19 | │ ├── 8.csv 20 | │ ├── 9.csv 21 | │ ├── 10.csv 22 | │ ├── 11.csv 23 | │ ├── 12.csv 24 | │ ├── 12.csv 25 | │ ├── 13.csv 26 | │ ├── 14.csv 27 | │ ├── 15.csv 28 | │ └── 16.csv 29 | └── other # Data obtained from the other experiments 30 | ├── 1.csv # Simulation of fluid leaks and fluid additions 31 | ├── 2.csv # Simulation of fluid leaks and fluid additions 32 | ├── 3.csv # Simulation of fluid leaks and fluid additions 33 | ├── 4.csv # Simulation of fluid leaks and fluid additions 34 | ├── 5.csv # Sharply behavior of rotor imbalance 35 | ├── 6.csv # Linear behavior of rotor imbalance 36 | ├── 7.csv # Step behavior of rotor imbalance 37 | ├── 8.csv # Dirac delta function behavior of rotor imbalance 38 | ├── 9.csv # Exponential behavior of rotor imbalance 39 | ├── 10.csv # The slow increase in the amount of water in the circuit 40 | ├── 11.csv # The sudden increase in the amount of water in the circuit 41 | ├── 12.csv # Draining water from the tank until cavitation 42 | ├── 13.csv # Two-phase flow supply to the pump inlet (cavitation) 43 | └── 14.csv # Water supply of increased temperature 44 | ``` 45 | -------------------------------------------------------------------------------- /docs/contributing.md: -------------------------------------------------------------------------------- 1 | # Contributing to SKAB repository 2 | 3 | We are glad you are reading this because work on the SKAB benchmark is still in progress, and contribution is welcome. 4 | 5 | ## How to Contribute 6 | 7 | 1. Make sure you have a GitHub account. 8 | 2. Fork the repository for the relevant book. 9 | 3. Create a new branch on which to make your change, e.g. 10 | `git checkout -b my_contribution` 11 | 4. Commit your change. Include a commit message describing the correction. Please note that if your commit message is not clear, the correction will not be accepted. 12 | 5. Submit a [pull request](https://github.com/waico/skab/compare?expand=1). 13 | 14 | Thank you for your contribution! 15 | -------------------------------------------------------------------------------- /docs/pictures/nab-metric.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/waico/SKAB/b2c0d46c2971dcbfe71e26087b6d231998bb91c2/docs/pictures/nab-metric.jpg -------------------------------------------------------------------------------- /docs/pictures/skab.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/waico/SKAB/b2c0d46c2971dcbfe71e26087b6d231998bb91c2/docs/pictures/skab.png -------------------------------------------------------------------------------- /docs/pictures/testbed.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/waico/SKAB/b2c0d46c2971dcbfe71e26087b6d231998bb91c2/docs/pictures/testbed.png -------------------------------------------------------------------------------- /notebooks/ArimaFD.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "cells": [ 3 | { 4 | "cell_type": "markdown", 5 | "metadata": {}, 6 | "source": [ 7 | "# Pipeline for the anomaly detection on the SKAB using ARIMA fault detection algorithm" 8 | ] 9 | }, 10 | { 11 | "cell_type": "markdown", 12 | "metadata": {}, 13 | "source": [ 14 | "Details regarding the SKAB one can find in the [SKAB repository](https://github.com/waico/SKAB)." 15 | ] 16 | }, 17 | { 18 | "cell_type": "markdown", 19 | "metadata": {}, 20 | "source": [ 21 | "The idea behind this algorithm is to use ARIMA weights as features for the anomaly detection algorithm. Using discrete differences of weight coefficients for different heuristic methods for obtaining function, which characterized the state (anomaly, not anomaly) using a threshold. \n", 22 | "\n", 23 | "Links at [PyPi](https://pypi.org/project/arimafd/), [GitHub](https://github.com/waico/arimafd) and [paper](https://waico.ru)" 24 | ] 25 | }, 26 | { 27 | "cell_type": "code", 28 | "execution_count": null, 29 | "metadata": {}, 30 | "outputs": [], 31 | "source": [ 32 | "# libraries importing\n", 33 | "import sys\n", 34 | "import warnings\n", 35 | "\n", 36 | "import pandas as pd\n", 37 | "from arimafd import Arima_anomaly_detection\n", 38 | "\n", 39 | "sys.path.append(\"..\")\n", 40 | "from core.metrics import chp_score\n", 41 | "from core.utils import load_preprocess_skab, plot_results\n", 42 | "\n", 43 | "warnings.filterwarnings(\"ignore\", category=UserWarning)" 44 | ] 45 | }, 46 | { 47 | "cell_type": "markdown", 48 | "metadata": {}, 49 | "source": [ 50 | "## Data" 51 | ] 52 | }, 53 | { 54 | "cell_type": "code", 55 | "execution_count": null, 56 | "metadata": {}, 57 | "outputs": [], 58 | "source": [ 59 | "Xy_traintest_list = load_preprocess_skab()" 60 | ] 61 | }, 62 | { 63 | "cell_type": "markdown", 64 | "metadata": {}, 65 | "source": [ 66 | "## Method" 67 | ] 68 | }, 69 | { 70 | "cell_type": "code", 71 | "execution_count": null, 72 | "metadata": {}, 73 | "outputs": [], 74 | "source": [ 75 | "predicted_outlier, predicted_cp = [], []\n", 76 | "true_outlier, true_cp = [], []\n", 77 | "for X_train, X_test, y_train, y_test in Xy_traintest_list:\n", 78 | " model = Arima_anomaly_detection()\n", 79 | " model.fit(X_train)\n", 80 | " prediction = pd.Series(\n", 81 | " model.predict(X_test),\n", 82 | " index=X_test.index,\n", 83 | " )\n", 84 | "\n", 85 | " # predicted outliers saving\n", 86 | " predicted_outlier.append(prediction)\n", 87 | "\n", 88 | " # predicted CPs saving\n", 89 | " prediction_cp = prediction.rolling(30).max().fillna(0).diff().abs()\n", 90 | " prediction_cp[0] = prediction[0]\n", 91 | " predicted_cp.append(prediction_cp)\n", 92 | "\n", 93 | " true_outlier.append(y_test[\"anomaly\"])\n", 94 | " true_cp.append(y_test[\"changepoint\"])" 95 | ] 96 | }, 97 | { 98 | "cell_type": "markdown", 99 | "metadata": {}, 100 | "source": [ 101 | "### Results visualization" 102 | ] 103 | }, 104 | { 105 | "cell_type": "code", 106 | "execution_count": null, 107 | "metadata": {}, 108 | "outputs": [ 109 | { 110 | "data": { 111 | "image/png": "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", 112 | "text/plain": [ 113 | "
" 114 | ] 115 | }, 116 | "metadata": {}, 117 | "output_type": "display_data" 118 | } 119 | ], 120 | "source": [ 121 | "# [donotremove]\n", 122 | "plot_results(\n", 123 | " (true_cp[1], predicted_cp[1]),\n", 124 | ")" 125 | ] 126 | }, 127 | { 128 | "cell_type": "markdown", 129 | "metadata": {}, 130 | "source": [ 131 | "## Evaluation (metrics calculation)" 132 | ] 133 | }, 134 | { 135 | "cell_type": "code", 136 | "execution_count": null, 137 | "metadata": {}, 138 | "outputs": [], 139 | "source": [ 140 | "import pickle\n", 141 | "\n", 142 | "pickle.dump(\n", 143 | " predicted_outlier,\n", 144 | " open(f\"../results/results-{model.__class__.__name__}.pkl\", \"wb\"),\n", 145 | ")" 146 | ] 147 | }, 148 | { 149 | "cell_type": "markdown", 150 | "metadata": {}, 151 | "source": [ 152 | "### Binary classification (outlier detection) metrics" 153 | ] 154 | }, 155 | { 156 | "cell_type": "code", 157 | "execution_count": null, 158 | "metadata": {}, 159 | "outputs": [ 160 | { 161 | "name": "stdout", 162 | "output_type": "stream", 163 | "text": [ 164 | "False Alarm Rate 0.01 %\n", 165 | "Missing Alarm Rate 100.0 %\n", 166 | "F1 metric 0.0\n" 167 | ] 168 | } 169 | ], 170 | "source": [ 171 | "# [donotremove]\n", 172 | "# binary classification metrics calculation\n", 173 | "binary = chp_score(true_outlier, predicted_outlier, metric=\"binary\")" 174 | ] 175 | }, 176 | { 177 | "cell_type": "markdown", 178 | "metadata": {}, 179 | "source": [ 180 | "not implemented" 181 | ] 182 | }, 183 | { 184 | "cell_type": "markdown", 185 | "metadata": {}, 186 | "source": [ 187 | "### Changepoint detection metrics" 188 | ] 189 | }, 190 | { 191 | "cell_type": "code", 192 | "execution_count": null, 193 | "metadata": {}, 194 | "outputs": [ 195 | { 196 | "name": "stdout", 197 | "output_type": "stream", 198 | "text": [ 199 | "Amount of true anomalies 127\n", 200 | "A number of missed CPs = 127\n", 201 | "A number of FPs = 2\n", 202 | "Average time nan\n" 203 | ] 204 | }, 205 | { 206 | "name": "stderr", 207 | "output_type": "stream", 208 | "text": [ 209 | "/Users/mw/pyprojects/SKAB/.venv/lib/python3.10/site-packages/numpy/core/fromnumeric.py:3504: RuntimeWarning: Mean of empty slice.\n", 210 | " return _methods._mean(a, axis=axis, dtype=dtype,\n", 211 | "/Users/mw/pyprojects/SKAB/.venv/lib/python3.10/site-packages/numpy/core/_methods.py:129: RuntimeWarning: invalid value encountered in scalar divide\n", 212 | " ret = ret.dtype.type(ret / rcount)\n" 213 | ] 214 | } 215 | ], 216 | "source": [ 217 | "# [donotremove]\n", 218 | "# average detection delay metric calculation\n", 219 | "add = chp_score(\n", 220 | " true_cp,\n", 221 | " predicted_cp,\n", 222 | " metric=\"average_time\",\n", 223 | " window_width=\"60s\",\n", 224 | " anomaly_window_destination=\"righter\",\n", 225 | ")" 226 | ] 227 | }, 228 | { 229 | "cell_type": "code", 230 | "execution_count": null, 231 | "metadata": {}, 232 | "outputs": [ 233 | { 234 | "name": "stdout", 235 | "output_type": "stream", 236 | "text": [ 237 | "Standard - -0.09\n", 238 | "LowFP - -0.17\n", 239 | "LowFN - -0.06\n" 240 | ] 241 | } 242 | ], 243 | "source": [ 244 | "# [donotremove]\n", 245 | "# nab metric calculation\n", 246 | "nab = chp_score(\n", 247 | " true_cp,\n", 248 | " predicted_cp,\n", 249 | " metric=\"nab\",\n", 250 | " window_width=\"60s\",\n", 251 | " anomaly_window_destination=\"righter\",\n", 252 | ")" 253 | ] 254 | } 255 | ], 256 | "metadata": { 257 | "kernelspec": { 258 | "display_name": "", 259 | "language": "python", 260 | "name": "" 261 | }, 262 | "language_info": { 263 | "codemirror_mode": { 264 | "name": "ipython", 265 | "version": 3 266 | }, 267 | "file_extension": ".py", 268 | "mimetype": "text/x-python", 269 | "name": "python", 270 | "nbconvert_exporter": "python", 271 | "pygments_lexer": "ipython3", 272 | "version": "3.10.14" 273 | }, 274 | "toc": { 275 | "base_numbering": 1, 276 | "nav_menu": { 277 | "height": "282.997px", 278 | "width": "471.989px" 279 | }, 280 | "number_sections": true, 281 | "sideBar": true, 282 | "skip_h1_title": false, 283 | "title_cell": "Table of Contents", 284 | "title_sidebar": "Contents", 285 | "toc_cell": false, 286 | "toc_position": {}, 287 | "toc_section_display": true, 288 | "toc_window_display": false 289 | } 290 | }, 291 | "nbformat": 4, 292 | "nbformat_minor": 4 293 | } 294 | -------------------------------------------------------------------------------- /notebooks/README.md: -------------------------------------------------------------------------------- 1 | # Anomaly Detection Algorithms 2 | 3 | ### Hotelling's T-squared statistic 4 | Hotelling's statistic is one of the most popular statistical process control techniques. It is based on the Mahalanobis distance. 5 | Generally, it measures the distance between the new vector of values and the previously defined vector of normal values additionally using variances. 6 | 7 | [[notebook]](https://github.com/YKatser/ControlCharts/blob/main/examples/t2_SKAB.ipynb) [[paper]](https://www.semanticscholar.org/paper/Multivariate-Quality-Control-illustrated-by-the-air-Hotelling/529ba6c1a80b684d2f704a7565da305bb84f14e8) 8 | 9 | ### Hotelling's T-squared statistic + Q statistic (SPE index) based on PCA 10 | The combined index is based on PCA. 11 | Hotelling’s T-squared statistic measures variations in the principal component subspace. 12 | Q statistic measures the projection of the sample vector on the residual subspace. 13 | To avoid using two separated indicators (Hotelling's T-squared and Q statistics) for the process monitoring, we use a combined one based on logical or. 14 | 15 | [[notebook]](https://github.com/YKatser/ControlCharts/blob/main/examples/t2_with_q_SKAB.ipynb) [[paper]](https://analyticalsciencejournals.onlinelibrary.wiley.com/doi/abs/10.1002/cem.800) 16 | 17 | ### Isolation Forest 18 | Isolation Forest or iForest builds an ensemble of iTrees for a given data set, then anomalies are those instances which have short average path lengths on the iTrees. 19 | 20 | [[notebook]](https://github.com/waico/SKAB/blob/master/notebooks/isolation_forest.ipynb) [[paper]](https://ieeexplore.ieee.org/abstract/document/4781136?casa_token=kiHmrqDyGL4AAAAA:O4yM7O2WCXdQH2sQbpKUXAHiepBxUhc5odzbydmgTiz5f7ZEDYgkXltodCahlgIzArxUldce5LB9mg) 21 | 22 | ### LSTM-based NN (LSTM) 23 | LSTM-based neural network for anomaly detection using reconstruction error as an anomaly score. 24 | 25 | [[notebook]](https://github.com/waico/SKAB/blob/master/notebooks/LSTM.ipynb) [[paper]](https://arxiv.org/abs/1612.06676) 26 | 27 | ### Feed-Forward Autoencoder 28 | Feed-forward neural network with autoencoder architecture for anomaly detection using reconstruction error as an anomaly score. 29 | 30 | [[notebook]](https://github.com/waico/SKAB/blob/master/notebooks/autoencoder.ipynb) [[paper]](https://epubs.siam.org/doi/abs/10.1137/1.9781611974973.11) 31 | 32 | ### Convolutional Autoencoder (Conv-AE) 33 | A reconstruction convolutional autoencoder model to detect anomalies in timeseries data using reconstruction error as an anomaly score. 34 | 35 | [[notebook]](https://github.com/waico/SKAB/blob/master/notebooks/CAE.ipynb) [[paper]](https://keras.io/examples/timeseries/timeseries_anomaly_detection/) 36 | 37 | ### LSTM Autoencoder (LSTM-AE) 38 | If you inputs are sequences, rather than vectors or 2D images, then you may want to use as encoder and decoder a type of model that can capture temporal structure, such as a LSTM. To build a LSTM-based autoencoder, first use a LSTM encoder to turn your input sequences into a single vector that contains information about the entire sequence, then repeat this vector n times (where n is the number of timesteps in the output sequence), and run a LSTM decoder to turn this constant sequence into the target sequence. 39 | 40 | A reconstruction sequence-to-sequence (LSTM-based) autoencoder model to detect anomalies in timeseries data using reconstruction error as an anomaly score. 41 | 42 | [[notebook]](https://github.com/waico/SKAB/blob/master/notebooks/LSTM-AE.ipynb) [[paper]](https://machinelearningmastery.com/lstm-autoencoders/) [[paper]](https://blog.keras.io/building-autoencoders-in-keras.html) 43 | 44 | ### LSTM Variational Autoencoder (LSTM-VAE) 45 | A reconstruction LSTM variational autoencoder model to detect anomalies in timeseries data using reconstruction error as an anomaly score. 46 | 47 | [[notebook]](https://github.com/waico/SKAB/blob/master/notebooks/LSTM-VAE.ipynb) [[paper]](https://arxiv.org/pdf/1511.06349.pdf) [[code]](https://github.com/twairball/keras_lstm_vae) 48 | 49 | ### Variational Autoencoder (VAE) 50 | A reconstruction variational autoencoder (VAE) model to detect anomalies in timeseries data using reconstruction error as an anomaly score. VAE is an autoencoder that learns a latent variable model for its input data. So instead of letting your neural network learn an arbitrary function, you are learning the parameters of a probability distribution modeling your data. If you sample points from this distribution, you can generate new input data samples: a VAE is a "generative model". 51 | 52 | [[notebook]](https://github.com/waico/SKAB/blob/master/notebooks/VAE.ipynb) [[paper1]](https://arxiv.org/pdf/1312.6114.pdf) [[paper2]](https://dl.acm.org/doi/pdf/10.1145/3178876.3185996?casa_token=HVY_9X3NxToAAAAA%3AZzZNSpmDdI9bEbTCqC1R3fPLiP4SDHyH9l9VyHxZ9zsL_3UXblc7Fe-ZdMPI7gkyVN9orRYQ5j9C) [[code]](https://blog.keras.io/building-autoencoders-in-keras.html) 53 | 54 | 55 | ### MSCRED 56 | MSCRED - Multi-Scale Convolutional Recurrent Encoder-Decoder first constructs multi-scale (resolution) signature matrices to characterize multiple levels of the system statuses across different time steps. 57 | In particular, different levels of the system statuses are used to indicate the severity of different abnormal incidents. 58 | Subsequently, given the signature matrices, a convolutional encoder is employed to encode the inter-sensor (time series) correlations patterns and an attention based Convolutional Long-Short Term Memory (ConvLSTM) network is developed to capture the temporal patterns. 59 | Finally, with the feature maps which encode the inter-sensor correlations and temporal information, a convolutional decoder is used to reconstruct the signature matrices and the residual signature matrices are further utilized to detect and diagnose anomalies. 60 | The intuition is that MSCRED may not reconstruct the signature matrices well if it never observes similar system statuses before. 61 | 62 | [[notebook]](https://github.com/waico/SKAB/blob/master/notebooks/MSCRED.ipynb) [[paper]](https://ojs.aaai.org/index.php/AAAI/article/view/3942) 63 | 64 | ### MSET 65 | MSET - multivariate state estimation technique is a non-parametric and statistical modeling method, which calculates the estimated values based on the weighted average of historical data. In terms of procedure, MSET is similar to some nonparametric regression methods, such as, auto-associative kernel regression. 66 | 67 | [[notebook]](https://github.com/waico/SKAB/blob/master/notebooks/MSET.ipynb) [[paper]](https://inis.iaea.org/collection/NCLCollectionStore/_Public/32/025/32025817.pdf) 68 | -------------------------------------------------------------------------------- /pyproject.toml: -------------------------------------------------------------------------------- 1 | [tool.poetry] 2 | name = "notebooks" 3 | version = "0.1.0" 4 | description = "" 5 | authors = ["Your Name "] 6 | readme = "README.md" 7 | 8 | [tool.poetry.dependencies] 9 | python = ">=3.10,<3.11" 10 | arimafd = {git = "https://github.com/InSuperposition/arimafd.git", branch = "master"} 11 | tensorflow = "^2.15.0" 12 | sympy = "^1.12" 13 | lightgbm = "^4.3.0" 14 | matplotlib = "^3.8.4" 15 | ipykernel = "^6.29.4" 16 | ipython = "^8.24.0" 17 | 18 | [tool.poetry.group.dev.dependencies] 19 | pre-commit = "^3.7.0" 20 | ruff = "^0.4.2" 21 | ipykernel = "^6.29.4" 22 | pytest = "^8.2.0" 23 | 24 | [build-system] 25 | requires = ["poetry-core"] 26 | build-backend = "poetry.core.masonry.api" 27 | -------------------------------------------------------------------------------- /results/results-Arima_anomaly_detection.pkl: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/waico/SKAB/b2c0d46c2971dcbfe71e26087b6d231998bb91c2/results/results-Arima_anomaly_detection.pkl -------------------------------------------------------------------------------- /results/results-Conv_AE.pkl: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/waico/SKAB/b2c0d46c2971dcbfe71e26087b6d231998bb91c2/results/results-Conv_AE.pkl -------------------------------------------------------------------------------- /results/results-Isolation_Forest.pkl: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/waico/SKAB/b2c0d46c2971dcbfe71e26087b6d231998bb91c2/results/results-Isolation_Forest.pkl -------------------------------------------------------------------------------- /results/results-LSTM_AE.pkl: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/waico/SKAB/b2c0d46c2971dcbfe71e26087b6d231998bb91c2/results/results-LSTM_AE.pkl -------------------------------------------------------------------------------- /results/results-MSCRED.pkl: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/waico/SKAB/b2c0d46c2971dcbfe71e26087b6d231998bb91c2/results/results-MSCRED.pkl -------------------------------------------------------------------------------- /results/results-MSET.pkl: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/waico/SKAB/b2c0d46c2971dcbfe71e26087b6d231998bb91c2/results/results-MSET.pkl -------------------------------------------------------------------------------- /results/results-T2-q.pkl: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/waico/SKAB/b2c0d46c2971dcbfe71e26087b6d231998bb91c2/results/results-T2-q.pkl -------------------------------------------------------------------------------- /results/results-T2.pkl: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/waico/SKAB/b2c0d46c2971dcbfe71e26087b6d231998bb91c2/results/results-T2.pkl -------------------------------------------------------------------------------- /results/results-Vanilla_AE.pkl: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/waico/SKAB/b2c0d46c2971dcbfe71e26087b6d231998bb91c2/results/results-Vanilla_AE.pkl -------------------------------------------------------------------------------- /results/results-Vanilla_LSTM.pkl: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/waico/SKAB/b2c0d46c2971dcbfe71e26087b6d231998bb91c2/results/results-Vanilla_LSTM.pkl --------------------------------------------------------------------------------