Abstract: A fault recognition model combining signal optimization preprocessing and deep learning is proposed to address
the problems of high noise in vibration signals and difficult feature extraction of rolling bearings in mining belt conveyors.
The model first utilizes the Whale Optimization Algorithm (WOA) optimized Variational Mode Decomposition (VMD) method
to adaptively denoise and reconstruct the original vibration signal to accurately extract fault features. Then,convert the reconstructed
signal into a two-dimensional grayscale image as input to the model. Finally,an improved Extreme Inception (Xception)
and Convolutional Neural Network (Xception-CNN) model was constructed during the recognition and classification
stage. This network integrates the deep separable convolution advantages of Xception architecture to more efficiently utilize
computing resources,while also introducing channel attention mechanism to enhance attention to key fault features,and embedding
residual learning module to alleviate the gradient vanishing problem of deep networks,ultimately achieving end-to-end intelligent
classification of fault states. The results showed that the proposed Xception-CNN fault recognition model achieved the
highest recognition accuracy of 98. 61% on the test set,with an F1 score of 0. 985. Under strong noise,i. e. with a signal-tonoise
ratio of 10 dB interference,the accuracy of the model still remains at 98. 61%,significantly better than the comparison
method,and has good robustness. At the same time,the model parameter size is only 42. 7 MB,and the time-consumed for the single-sample inference is only 12. 3 ms,which ensures high accuracy while having good engineering application efficiency.
This provides new ideas and methods for fault diagnosis and predictive maintenance of mining equipment,which is of great significance
for promoting the progress of mining equipment maintenance technology.

Key words: rolling bearing,fault recognition,signal processing,whale optimization algorithm,variational mode decomposition,
convolutional neural network