Fault diagnosis of rotating machinery components with deep ELM ensemble induced by real-valued output-based diversity metric

Abstract

Abstract Fault diagnosis of rotating machinery components under different working conditions or noisy environment has been a major challenge. The domain shift caused by fluctuation of working condition or noise makes existing deep learning methods’ diagnosis performance degrade significantly. To solve this problem, in this paper, we develop an ensemble learning diagnosis method. The method employs denoising multi-layer extreme learning machine (DMELM) as base classifier because it can learn features with high-level of abstraction and also has fast training speed. The ensemble method is characterized by two ingredients: 1) Candidate classifiers are generated by setting different activation functions and denoising criteria simultaneously. 2) A novel real-valued output-based diversity metric-RM is first proposed and RM induced selection scheme is designed to select accurate as well as diverse classifiers more efficiently from candidates to form a compact ensemble. The proposed ensemble method is applied to detect the faults of gearbox and an engine rolling bearing under various working conditions and noise levels. Results show our method can effectively solve the domain shift problem caused by condition variation and noise, thus outperforms not only single deep learning or deep ELM classifier, but also other state-of-the-art ensemble methods and domain adaptation methods.