Fault Feature Extraction of Rolling Bearing with Sparse Representation Auto-Encoder Driven by Impact Response Mechanism

Due to the interference of slipping, signal loss in non-bearing area, noise and other factors, it is hard to accurately extract the fault feature of rolling bearings. Additionally, the abstract features extracted by intelligent diagnosis methods are not interpretable. Firstly, combined the impact response mechanism and sparse representation theory, an interpretable sparse representation sparse representation Auto-Encoder network is designed, which regards the coding and decoding layer of Auto-Encoder as the solution of sparse vector and the learning of over complete dictionary respectively. Secondly, an adaptively optimization algorithm is designed based on the loss function characteristics of impact response parameters, which effectively realizing the fast global optimization of characteristic parameters. Combined with the designed sparse representation Auto-Encoder network and the rolling bearing signal features, a two-layer neural network is built to perform high-precision feature reconstruction of bearing fault signals. Finally, simulation analysis shows that the proposed method can extract impact fault feature parameters with clear physical meaning, which has high feature extraction accuracy and good anti-noise performance. Moreover, experimental signals further verify the effectiveness of the proposed method. © 2023 Editorial Office of Chinese Journal of Mechanical Engineering. All rights reserved.