An Unsupervised Multiview Sparse Filtering Approach for Current-Based Wind Turbine Gearbox Fault Diagnosis

Abstract

Gearboxes are critical components in wind turbines, and their fault diagnosis has gained increasing and considerable attention. Compared to traditional vibration-based methods, current-based fault diagnosis has significant advantages in terms of cost, implementation, and reliability. However, it is quite challenging to extract informative fault-related features from raw current signals due to the presence of dominant current fundamental component and harmonic component as well as electrical noise. In order to address this challenge, this article presents a novel unsupervised feature learning approach based on a two-layer sparse filtering algorithm for current-based gearbox fault diagnosis. Specifically, a multiview sparse filtering (MVSF) method is proposed to automatically extract useful and complementary features under different views from raw current signals. The proposed method can fuse multiview feature representations learned concurrently to improve the fault diagnosis performance. The effectiveness of the proposed MVSF method is verified through experiments on a wind turbine gearbox test rig. Experimental results demonstrate that the proposed approach can effectively recognize the health state of the gearbox and exhibits superior performance in feature learning and diagnosis compared with traditional feature extraction approaches.