Abstract:For the problem of low diagnostic accuracy and robustness due to the difficulty in extracting the signal features of rolling bearings during fault diagnosis, a new rolling bearing fault diagnosis method is proposed based on Squeeze-Excitation-ResNeXt(SE-ResNeXt). The collected one-dimensional bearing vibration signals were taken as input, the sliding window sampling and standardization were conducted, the feature re-calibration was carried out through compression and excitation operation, the model receptive field was enlarged and the fault signal characteristics were extracted adaptively by cascading aggregate residual transformation network. In the process of model training, the optimal compression rate was selected as 1/8 and 8 sets of convolution, Relu function was introduced to accelerate the convergence of the network, global average pooling was used to replace the full connection layer to avoid overfitting, and an optimal fault diagnosis model capable of independent characterization learning was constructed. Simulation experiments show that compared with the current deep learning algorithm, the SE-ResNeXt network can accurately realize bearing fault diagnosis and still has good robustness under high noise environment.
