Sparsity Based Anomaly Detection Framework

Unsupervised variables selection This submodule selects a subset of variables among an initial set while limiting the loss of information, independently of any a priori predictive target, by pruning a mutual information graph.  Counterfactual analysis based unsupervised anomaly detection and diagnosis  This submodule computes a multivariate time series that is as close as possible to the input time series, while lowering the global anomaly score.  The anomaly score is computed based on points anomaly detection in an embedding space derived from a use-case specific pretext-task model.  The sparse residual between the original and the counterfactual time series enables anomalies diagnostic.  This approach suffers from several methodological limitations. Developments will not be pursued in this direction for the time being.  Collaborative anomaly detection  This submodule performs anomalous segments detection in univariate time series, leveraging their structural similarities.  A multiscale representation of the signal is derived to enable the detection of anomalous segments of unprescribed lengths.  As an important by-product, a pattern-based time series dissimilarity metric is provided.

Documentation

Methodological Guidelines

• Methodological Guideline for Anomaly Detection Models
• Methodological Guideline for Time Series Anomaly Detection

Benchmarks

• Unsupervised Anomaly Detection and Explainability tools for Time Series

Applications over use-cases

• Anomaly Detection Models Use Case level document

Demonstrator

• A restricted demonstrator is available here. It is based on an anomaly-detection use case.

Support