Integrating adaptive input length selection strategy and unsupervised transfer learning for bearing fault diagnosis under noisy conditions

Transfer learning (TL) has made great progress in intelligent fault diagnosis of bearing. However, due to the harsh working conditions of bearings in engineering practice, in addition to the bearing vibration signals, sensors also inevitably collect noise signals, which leads to the performance degradation of TL. In this paper, a novel model based on TL is proposed to solve the problem. Different from previous studies, an adaptive input length model instead of a fixed input length is established to diagnose bearings with different parameters. A signal processing method combining wide kernels in the first convolutional layer and pooling layer is used to feature denoise the input. Group convolutional and instance normalization based on 1D-CNN are constructed for feature extraction, health conditions classification, and denoise. An optimization objective function based on maximum mean discrepancy is introduced to align the feature distribution discrepancy. The model does not rely on any bearing data label information and realizes unsupervised TL. In the meanwhile, seven experiments including TL model and anti-noise model are conducted to compare and analyze the performance of intelligent fault diagnosis. Experimented on the axlebox bearing dataset of High-speed train, the accuracy is 85.98%, 93.62%, and 99.82% under SNR equal -6 dB, -4 dB, and -2 dB respectively. Experimented between bearing in electromechanical drive systems and high-speed train axlebox bearing dataset, the accuracy of the proposed method is 77.56%, 79.33%, and 96.22% under SNR equal -6 dB, -4 dB, and -2 dB respectively. In the meanwhile, the accuracy of the three advanced TL models and four state-of-the-art anti-noise models are below 50 %. The results indicate that the proposed model has good bearing fault diagnosis ability for TL different datasets under noisy conditions.