Joint Semitensor Compressive Sensing With Multiscale Pyramidal Residual Network for Rolling Bearing Fault Diagnosis

Fault diagnosis is an indispensable part of mechanical systems. However, due to the severe working conditions, the collected bearing vibration signals are usually nonlinear and nonstationary and may contain multiple time scales and more noise, which results in low accuracy in fault identification. To address these issues, this article introduces a multiscale pyramidal residual network (MPRNet) for bearing fault diagnosis. First, based on semitensor product compression sensing (STP-CS), we put forward parallel compressive sampling matching pursuit (PCSMP). Compressing and reconstructing vibration signals helps solve data storage difficulties and slow transmission speeds and then we established MPRNet. The pyramid network architecture combines the residual module and the multiscale network module. It avoids model degradation in deeper networks and allows for the expression of more feature information on the output side. Finally, establish an end-to-end fault diagnosis model and test it on two different laboratory data. The results demonstrate an average diagnostic accuracy of 99.47% and 99.56%, respectively. Compared to other methods, this method has a better performance and provides theoretical support for joint method fault diagnosis.