Anomaly-Detection 🩺

The service is responsible for detecting anomalies in the (multivariate) time-series data. It provides different algorithms for the detection.

Requirements

• Python ≥ 3.10
• All packages from requirements.txt

Available models

• Isolation Forest
• One-Class-SVM
• DAGMM
• LSTM-Autoencoder

Development

Local

Install dependencies from requirements.txt

Start the service:

uvicorn main:app --reload

Docker

We provide a docker-compose in the root directory of ADEPT to start all services bundled together.

Adding functionality

New algorithms can be easily added by following the instructions below. Additional features such as configuration options and on-demand training are also explained.

Directory structure

\-Anomaly-Detection
    ├── algorithms                              # Contains all files for the anomaly detection
    │   ├── interface                           # Contains all interfaces
    │   │   ├── algorithm_config.py
    │   │   ├── algorithm_information.py
    │   │   └── algorithm_interface.py
    │   ├── models                              # Contains trained models and training code
    │   │   ├── DAGMM                           # Contains all files for the DAGMM model
    │   │   │   ├── train                       # Contains the training code for DAGMM
    │   │   │   └── universal                   # Contains a generic pre-trained DAGMM model
    │   │   └── [...]
    │   ├── util                                # Contains useful functions for anomaly detection
    │   │   ├── window_util.py
    │   │   └── [...]
    │   ├── dagmm.py                            # DAGMM Implementation
    │   ├── ocsvm.py                            # OCSVM Implementation
    │   └── [...]
    ├── src                                     # Python source files for base functions
    │   ├── dynamic_algorithm_loading.py
    │   └── [...]
    ├── Dockerfile
    ├── main.py                                 # Main module with all API definitions
    ├── requirements.txt                        # Required python dependencies
    └── [...]

Adding anomaly detection algorithms

1. Create a python module in the algorithms package
2. Create a class named Algorithm that implements the algorithm interface
3. Implement the anomaly detection algorithms and expose its functions via the interface

Configuring the anomaly detection algorithm

Metadata

The algorithm metadata can be adjusted through the AlgorithmInformation object.

• name is the display name in ADEPT
• deep is used for grouping the algorithms based on their type
• explainable is used to denote the support for advanced explainability techniques

Configuration Options

The configuration options that are exposed to the user can be adjusted through the AlgorithmConfig object. ADEPT supports multiple different types of settings:

• Numeric (Integer, Float)
• Slider
• Toggle
• Dropdown

There are a few rules for using algorithm configs:

• The id's of the settings need to be unique
• Dropdown settings need to contain at least one Option
• Option names must be unique
• Options can include further settings (Numeric, Slider, Toggle) but no further dropdowns
• The default values have to be in the specified range
• The step value has to be greater than zero

An example for no exposed settings:

config = AlgorithmConfig()

An example for a simple toggle setting:

toggle_setting = ToggleSetting(id="toggle",
                               name="A simple toggle setting",
                               description="This is a description.",
                               default=False)

config = AlgorithmConfig([toggle_setting])

An example for multiple settings:

toggle_setting = ToggleSetting(id="toggle",
                               name="A simple toggle setting",
                               description="This is a description.",
                               default=False)

float_setting = FloatSetting(id="float",
                             name="A float setting",
                             description="This is a description.",
                             default=3.14)

slider_setting = SliderSetting(id="slider",
                               name="A slider setting",
                               description="This is a description.",
                               default=42,
                               step=0.5,
                               lowBound=0,
                               highBound=100)

config = AlgorithmConfig([toggle_setting, float_setting, slider_setting])

An example for a simple dropdown:

dropdown = OptionSetting(id="dropdown",
                         name="Language",
                         description="Select the desired language.",
                         default="English",
                         options=[Option("English"), Option("German")])

config = AlgorithmConfig([dropdown])

An example for a dropdown with settings for options:

toggle_setting = ToggleSetting(id="toggle",
                               name="A simple toggle setting",
                               description="This is a description.",
                               default=False)

float_setting = FloatSetting(id="float",
                             name="A float setting",
                             description="This is a description.",
                             default=3.14)

option_a = Option(name="Toggle Option", settings=[toggle_setting])

option_b = Option(name="Float Option", settings=[float_setting])

dropdown = OptionSetting(id="dropdown",
                         name="Dropdown",
                         description="This is a description.",
                         default="Toggle Option",
                         options=[option_a, option_b])

config = AlgorithmConfig([dropdown])

Copyright © ADEPT ML, TU Dortmund 2023