Rolling bearing fault diagnosis method based on DSICNN under strong noise background

To address the issues of inadequate fault diagnosis accuracy and suboptimal generalisation performance of rolling bearings in the presence of significant noise and varying operational conditions, a fault diagnosis approach utilising a dual-stream interactive convolutional neural network (DSICNN) is presented. To fully leverage the fault characteristics in vibration signals, both time-domain and frequency-domain signals are concurrently employed as inputs to the neural network. Then, dual attention mechanisms are presented, among which the Dynamic Weighted Channel Attention Mechanism (DWCAM) dynamically calibrates the channel weights of different inputs based on the importance of different channels, and the Aggregated Spatial Self Attention Mechanism (ASSAM) assigns greater weights to important regions while enhancing feature expression ability. Meanwhile, an arctangent linear unit function (AT-LU) is constructed to improve the problem of information loss contained in the signal when the linear rectifier unit has negative input. Finally, the signals are input into DSICNN, and the features are extracted more fully by interacting with two information streams. The model is trained and evaluated against other fault diagnosis models. The experimental results indicate that the suggested method exhibits superior classification performance, generalisation capability and robustness in the presence of significant noise and varying operating conditions.