A time-frequency images driven multi-component intelligent fault diagnosis method of mechanical transmission system

Abstract In order to realize the multi-component fault diagnosis of bearings, rotors and gears in a mechanical transmission system with weak faults under low-speed, strong noise and few measurement points, an intelligent diagnosis method combining wavelet time-frequency image with lightweight multi-scale kernel residual network (LMSKRN) is proposed. Firstly, the vibration signals are transformed by continuous wavelet transform (CWT) to obtain a high-quality 2D time-frequency image. Then, based on the standard convolutional neural network, an improved LMSKRN is established by designing the multi-scale kernel residual network, extracting the multi-scale features and model lightweight design. Aiming at the multi-component fault diagnosis of a mechanical transmission system, the diagnosis experiment with a single measurement point at a low speed of 700 r min −1 is designed, and seven single faults of bearings, rotors and gears and three mixed faults among components are fabricated. The investigation shows that the parameters of the proposed diagnosis method are 0.86 M, the floating point operations are 417.88 M, the model size is 3.31 MB, and achieves an average accuracy of 99.76% on the original data set, and an average accuracy of 98.71% on the 0 dB Gaussian noisy data set. Finally, the pretrained model is transferred to the actual noisy data set and achieves an average accuracy of more than 97% on multiple noisy data sets, which provides a new way for the data driven deep learning based multi-component fault diagnosis of a mechanical transmission system.