Anomaly Detection and Factor Estimation by Graph Deep Learning in Storage Batteries

Abstract

Predictive maintenance is a technique to perform maintenance before failures happen by finding their Indications in advance and is a key to streamlining Maintenance operations and reducing the downtime. Methods for predictive maintenance based on anomaly detection using deep learning have been actively studied, but the identification of anomalous sensors remains a challenging task. As sensors corresponding to the cause of anomaly do not necessarily indicate large anomaly scores, it is important to watch how a model computes the scores. In this work, we use a graph neural network for anomaly detection and isolation. The vertices of the graph that appears in the network correspond to the sensors, so we can interpret the relevant weights as the relationship between the sensors. We specifically used a sparse variant of graph attention network for anomaly detection and isolation. We applied it to real-world storage battery data and confirmed the effectiveness of the method.