An unsupervised transfer network with adaptive input and dynamic channel pruning for train axle bearing fault diagnosis

The field of bearing fault diagnosis has witnessed remarkable advancements with cross-domain fault diagnosis techniques. Nonetheless, these existing methods suffer from two main drawbacks. First, the input length of these methods is fixed, such as 2048 sample points, irrespective of the diverse sampling frequencies, bearing structure parameters, and rotational speeds observed among transfer objects. Additionally, the transfer learning methods currently employed are not robust to noise, rendering them incapable of functioning optimally in contaminated target domains. To address the aforementioned challenges, this study presents an unsupervised transfer network for train axle bearing fault diagnosis. First, an adaptive input module is proposed, which enables the input length of the proposed network to be adaptively selected based on parameters such as sampling frequency and bearing structure. Then, an enhanced feature learning block with sharing parameters is designed to enhance the transfer learning feature extraction capability under noise condition. Next, a dynamic channel pruning module is proposed to optimize of the proposed network. Finally, the transferability of the proposed network is demonstrated through experiments involving two types of transfer learning tasks. The proposed network exhibits robustness to noise and outperforms existing methods by achieving higher diagnostic accuracy and stability.