Leveraging Feedback and Causality-Enriched Multimodal Context for Predictive Maintenance

We propose an anomaly detector-agnostic framework to exploit heterogeneous and multidimensional streams in industrial predictive maintenance, with the main objective of detecting early data anomalies preceding asset failures. Our novelty lies in fusing multiple data streams, combining both discrete and continuous sources with potentially different sample rates, exploiting causality graphs, and leveraging (automated) feedback on past false positive alarms. This forms the contextual information, which, combined with feedback on produced alarms, aims to prune new false positive alarms of similar context before they are raised, in an inherently explainable manner. The framework is applied in two predictive maintenance case studies with different characteristics, and the results show that it can significantly enhance standalone anomaly detectors; e.g., we have observed decreases in false positive rates up to 7 times. Our implementation is provided as a python library enabling experiment repeatability and application to arbitrary other cases.