An Unsupervised Bayesian OC-SVM Approach for Early Degradation Detection, Thresholding, and Fault Prediction in Machinery Monitoring
Stanley Fong, Sriram Narasimhan
Abstract
In the literature pertaining to condition-based maintenance using machine learning, the class of unsupervised approaches has yet to be fully explored. In particular, the topics of fault prediction and unsupervised degradation modeling, which have significant importance to practice, have not been widely studied. In this article, we propose a novel unsupervised approach for early degradation detection and fault prediction, termed the Bayesian one-class support vector machine (B-OCSVM), which is based on OC-SVM and the hierarchical Bayesian framework. The proposed approach aims to address key gaps in the literature by proposing a new unsupervised OC-SVM hyperparameter estimation method for early degradation detection and new degradation threshold, which is used to predict when a detected degradation process will manifest as a fault. Unlike traditional prognostic approaches, the proposed degradation threshold can be defined without prior knowledge pertaining to the fault. Fault prediction is achieved using a novel prognostic in conjunction with a hierarchical Bayesian degradation framework. We validate and benchmark the proposed approach against the state of the art using run-to-failure data from three datasets to demonstrate the earlier detection point of the proposed approach compared to existing approaches and to showcase how the proposed degradation threshold and novel prognostic can be used effectively to predict when faults will occur. In addition, the results show that B-OCSVM generalizes well to a number of different operating conditions and fault types, which signifies that the early detection and prognostic elements of the proposed approach have strong practical significance in real-world applications.