Hydrogenerator early fault detection: Sparse Dictionary Learning jointly with the Variational Autoencoder

Abstract

Monitoring the continuous health status of a Hydraulic Turbine Generator Unit (HTGU) is a strategic task to prevent any unexpected downtime. In addition to the loss of energy production following a prolonged shutdown, the various maintenance costs represent an undesirable effect due to unanticipated failure. One of the main functions of a condition-based maintenance (CBM), is the early detection of any unexpected changes in a machine behavior. Indeed, being able to detect any drift or change in behavior compared to a reference behavior represents a major challenge in the monitoring of complex machines like a HTGU. In this paper, an early anomaly detection model for HTGU using Variational Autoencoders (VAEs) and Sparse Dictionary Learning (SDL) is proposed. The combined reconstruction error thus obtained from the VAE-SDL model is used as an early fault detection. Experimental tests on vibratory signals show that the two models thus joined increase the sensitivity as well as the robustness of the anomaly detection.