A lightweighted planetary gearbox fault diagnosis method based on multisensor fusion

Abstract Most research methods for planetary gearbox fault diagnosis with multisensors have some shortcomings, such as large numbers of parameters and unsuitability for deployment. To address these issues, a lightweighted planetary gearbox fault diagnosis method based on multisensor fusion is proposed. Firstly, the collected planetary gearbox signals are preprocessed by a tunable- Q wavelet transform and Gramian angular field image encoding method. Secondly, a multibranch feature fusion module is constructed for multisensor fusion and a multiscale residual attention network model is built to extract fault features at multi-scales. The mixture-of-depths is adopted to lightweight the model. The results show that the number of parameters and floating point operations of the proposed fault diagnosis model are 3.5 million and 452.6 million, respectively. The method’s diagnostic accuracy reaches 97.2%, and it has good anti-noise performance and stability.