An optimized recurrence plot method for intelligent diagnosis of early weak faults

Due to strong background noise and complex operating conditions, deep learning-based fault diagnosis encounters excessive data dependency, prolonged training durations, and suboptimal prediction accuracy. In this article, using an improved residual neural network (ResNet) model, an optimized recurrence plot (RP) (OptRP) method is developed to detect early weak faults in bearings. The technique comprises a multiscale RP, a maximum information entropy RP, and a signed RP. First, in order to address the data redundancy and single-scale limitations of RPs, an asymmetric multiscale RP is constructed. Second, using the maximum information entropy as the optimization criterion, an optimal threshold function is designed to overcome the conventional RP’s inability to automatically select the optimal threshold. Finally, a direction vector is defined to resolve the serious trend confusion problem inherent in the conventional RP algorithm during the phase space reconstruction process, and the OptRP image is generated by multiplying the direction vector matrix with the multiscale, maximum information entropy RP. Furthermore, an improved ResNet model, called as improved residual neural network (IResNet), is proposed by incorporating the Inception module. IResNet effectively enhances the multiscale feature extraction capability of the existing ResNet. The effectiveness of the OptRP-IResNet method is validated on a small sample dataset of early weak faults in high-speed train bearings, achieving a 5.61% improvement in a long-time series fault diagnosis experiment. The superiority of the method is further proved by gradient-weighted class activation mapping visualization results.