Feature Selecting Hierarchical Neural Network for Industrial System Health Monitoring

Industrial System Health Monitoring relies usually on the monitoring of well-designed features. This requires both, the engineering of reliable features and a good methodology for their analysis. If traditionally, features were engineered based on the physics of the system, recent advances in machine learning demonstrated that features could be automatically learned and monitored. In particular, using Hierarchical Extreme Learning Machines (HELM), based on random features, very good results have already been achieved for health monitoring with training on healthy data only. Yet, although very useful and mathematically sound, random features have little popularity as they contradict the intuition and seem to rely on luck. This tends to increase the “blackbox” effect often associated with Machine Learning. To mitigate this, in this paper, we propose to modify the traditional HELM architecture such that, while still relying on random features, only the most useful features among a large population will be selected. Traditional HELM are made of stacked contractive autoencoders with `1- or `2-regularisation and of a classifier as last layer. To achieve our objective, we propose to opt for expanding auto-encoders instead, but trained with a strong Group-LASSO regularization. This Group-LASSO regularisation fosters the selection of as few features as possible, making the auto-encoder in reality (or in testing condition) contractive. This deterministic selection provides useful features for health monitoring, without the need of learning or manually engineering them. The proposed approach demonstrates a better performance for fault detection and isolation on case studies developed for HELM evaluation.