Anomaly Detection

A suite of tools for performing anomaly detection and classification on time series.

  • Anomaly Scorers are at the core of the anomaly detection module. They produce anomaly scores time series, either for single series (score()), or for series accompanied by some predictions (score_from_prediction()). Scorers can be trainable (e.g., KMeansScorer) or not (e.g., NormScorer).

  • Anomaly Models offer a convenient way to produce anomaly scores from any of Darts forecasting models (ForecastingAnomalyModel) or filtering models (FilteringAnomalyModel), by comparing models’ predictions with actual observations. These classes take as parameters one Darts model, and one or multiple scorers, and can be readily used to produce anomaly scores with the score() method.

  • Anomaly Detectors: transform raw time series (such as anaomly scores) into binary anomaly time series.

  • Anomaly Aggregators: combine multiple binary anomaly time series (in the form of multivariate time series) into a single binary anomaly time series applying boolean logic.