Unsupervised Probabilistic Anomaly Detection Over Nominal Subsystem Events Through a Hierarchical Variational Autoencoder

Abstract

This work develops a versatile approach to discover anomalies in operational data for nominal (i.e., non-parametric) subsystem event signals using unsupervised Deep Learning techniques. Firstly, it builds a neural convolutional frameworkto extract both intrasubsystem and intersubsystem patterns. This is done by applying banks of voxel filters on the charteddata. Secondly, it generalizes the learned embedded regularity of a Variational Autoencoder manifold by merging latentspace-overlapping deviations with non-overlapping synthetic irregularities. Contingencies like novel data, modeldrift, etc., are therefore seamlessly managed by the proposed data-augmented approach. Finally, it creates a smooth diagnosis probabilistic function on the ensuing low-dimensional distributed representation. The resulting enhanced solution warrants analytically strong tools for a critical industrial environment. It also facilitates its hierarchical integrability, and provides visually interpretable insights of the degraded condition hazard to increase the confidence in its predictions. This strategy has been validated with eight pairwise-interrelated subsystems from high-speed trains. Its outcome also leads to further reliable explainability from a causal perspective.