A deep multi-signal fusion adversarial model based transfer learning and residual network for axial piston pump fault diagnosis

Deep learning has made remarkable achievements in fault diagnosis. However, the working conditions of the axial piston pump are diverse, and the distribution of the data is not the same, which causes most of the deep learning models to invalid. A deep multi-signal fusion adversarial model based transfer learning (MFAN) is presented to solve this problem. A multi-signal fusion module is designed to assigns weights to vibration signals and acoustic signals, which improves the dynamic adjustment ability of the method. Moreover, the residual network is embedded in the shared feature generation module to obtain abundant feature information. According to the different working loads of the axial piston pump, nine transfer scenarios are designed, and the proposed method is compared with five typical diagnosis methods. The average accuracy of MFAN on all scenarios reaches 98.5%, indicating this method has excellent performance in cross-domain fault detection of axial piston pumps.

Automated summary

A deep multi-signal fusion adversarial model based transfer learning method is proposed to address the diverse working conditions of axial piston pumps in fault diagnosis. The method improves performance through a multi-signal fusion module and embedding a residual network.