Incremental novelty detection and fault identification scheme applied to a kinematic chain under non-stationary operation

Abstract

Classical methods for monitoring electromechanical systems lack two critical functions for effective industrial application: management of unexpected events and the incorporation of new patterns into the knowledge database. This study presents a novel, high-performance condition-monitoring method based on a four-stage incremental learning approach. First, non-stationary operation is characterised using normalised time-frequency maps. Second, operating novelties are detected using multivariate kernel density estimators. Third, the operating novelties are characterised and labelled to increase the knowledge available for subsequent diagnosis. Fourth, operating faults are diagnosed and classified using neural networks. The proposed method is validated experimentally with an industrial camshaft-based machine under a variety of operating conditions.

Keywords: Condition monitoring; Data-driven modelling; Fault diagnosis; Non-stationary operation; Novelty detection.