A Novel Automated Feature Extraction Method for Fault Diagnosis of Rotating Mechanical Systems

Abstract

A novel approach to feature extraction, capable of generating a number of robust features in an automated way, is introduced. Although the proposed method focuses on features on the frequency domain for vibration data related to rotating mechanical systems, it can be extended on different types of features. The method comprises of two simple models for the feature generation and a Particle Swarm Optimization system for establishing optimum or near optimum parameters for these models. The generated features are evaluated with a number of metrics, before they are used for diagnosis purposes. The features derive from real-world data related to a case of corroded bearings in a helicopter system. The extracted features of the proposed method are compared with some which were manually generated, and the former are found to be of superior quality. A series of diagnosis experiments based on the best extracted features was carried out. The results of these experiments appear to validate the performance of the automatically generated features.