Operator Feedback-Compatible Fault Detection Framework for Industrial Time-series Data Streams

While Time-Series Anomaly Detection is an evergrowing field, there is still a gap between its commonly adopted offline learning, update-less behavior and the requirements of real-time predictive maintenance that represents its main industrial goal. The Incremental Learning paradigm attempts to address this issue conjugating an initial modeling with a continuous local enhancing ability to tackle the frequent normality drifts of a data stream as labelled through an operator's feedback. In this paper, a streaming-compatible extension of a recently proposed model-free, explainable algorithm is proposed, which is compatible with both drift adaptation and anomalous instance learning despite an unsupervised initial setting. Moreover, as a general industrial algorithm, the feature engineering process involved in the proposed framework is heavily inspired by the implicit physical laws that should govern machinery-sourced time-series. Two examples are explored in order to assess the online-learning abilities along a streaming scenario.