Deep Feature Learning Network for Fault Detection and Isolation

Abstract

Prognostics and Health Management (PHM) approaches typically involve several signal processing and feature engineering steps. The state of the art on feature engineering, comprising feature extraction and feature dimensionality reduction, often only provides specific solutions for specific problems, but rarely supports transferability or generalization: it often requires expert knowledge and extensive intervention. In this paper, we propose a new integrated feature learning approach for jointly achieving fault detection and fault isolation in high-dimensional condition monitoring data. The proposed approach, based on Hierarchical Extreme Learning Machines (HELM) demonstrates a good ability to detect and isolate faults in large datasets comprising signals of different natures, non-informative signals, non-linear relationships and noise. The method includes stacked auto-encoders that are able to learn the underlying high-level features, and a one-class classifier to combine the learned features in an indicator that represents the deviation from the normal system behavior. Once a deviation is identified, features are used to isolate the most deviating signal components. Two case studies highlight the benefits of the approach: First, a synthetic dataset with the typical characteristics of condition monitoring data and different types of faults is applied to evaluate the performance with objective metrics. Second, the approach is tested on data stemming from a power plant generator interturn failure. In both cases, the results are compared to other commonly applied approaches for fault isolation.