An intelligent bearing fault diagnosis framework: one-dimensional improved self-attention-enhanced CNN and empirical wavelet transform

The complexity of the internal structure of rolling bearings and the harshness of their operating environment result in strong non-stationarity and nonlinearity of the vibration signals. It remains a challenging and attractive task to accomplish more accurate classification through signal processing techniques and pattern recognition methods. To realize this aim, a novel one-dimensional improved self-attention-enhanced convolutional neural network (1D-ISACNN) with empirical wavelet transform (EWT) is proposed for rolling bearing fault classification. Firstly, the EWT algorithm is employed to decompose the raw signal into three frequency components, allowing for further extraction of multi-frequency components to enhance signal characteristics. Subsequently, a creative1D-ISACNN leverages the merits of a newly developed attention mechanism and an optimized meta-activation concatenation function in feature learning to better capture and map crucial information within the signal. Furthermore, label smoothing regularization is designed as the loss function of the 1D-ISACNN, which takes into account not only the loss of correctly labeled positions in the training samples but also the loss of other mislabeled positions. Finally, the adaptive moment projection estimation is designed to ensure a more robust gradient update strategy for updating the parameters of the proposed model. The developed model tested on three different sets of bearing data, has achieved a classification accuracy of 100%. In ablative experiments and other comparative experiments, the proposed method demonstrates higher recognition accuracy and more robust generalization capabilities compared to other excellent approaches.