Rolling bearing fault diagnosis based on multiscale ResNet and LSTM

Aiming at the problems of weak fault characteristics and complex and variable operating conditions of rolling bearings in strong noise environments, a fault diagnosis method based on the combination of multiscale residual neural network (MResNet) and long short-term memory (LSTM) neural network is proposed. Firstly, the vibration signal is decomposed using Complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), and components with autocorrelation coefficients greater than 0.2 are reconstructed to achieve signal denoising. Secondly, a dropout layer is introduced between multiscale residual blocks to prevent network overfitting, and the powerful time series information capturing ability of LSTM is combined to improve diagnostic accuracy. Finally, the bearing data from the mechanical comprehensive simulation experimental platform was used for experimental verification, and the results showed that the proposed method had high diagnostic accuracy. In order to simulate the real working environment of rotating machinery equipment, Gaussian white noise was added to the bearing diagnosis signal. The proposed method achieved diagnostic accuracies of 91.13% and 89.72% at −2 dB and −4 dB signal-to-noise ratios, respectively, higher than comparison methods such as WKCNN , WDCNN and ResNet-18. Experiments showed that the proposed method still has enhanced diagnostic ability even when the signal is heavily polluted by noise.