<p>In the coal mining industry, the smooth operation of idlers is crucial. In order to solve the challenge of fault classification and prediction of roller bearings, an ICCEMDAN-WTD-CNN-BiLSTM diagnostic model was proposed. The model first decomposes the vibration signal into noise IMF and effective IMF by ICCEMDAN technology, and then uses sample entropy evaluation to deal with noise IMF by wavelet threshold denoising. Then, the convolutional neural network (CNN) was used to extract and fuse the time-frequency features, and finally the fault sequence pattern was learned through the bidirectional long short-term memory network (BiLSTM) to achieve classification prediction. The experimental results show that the model has excellent performance in bearing fault identification, with an average recognition accuracy of 99.63 %, which is better than the traditional model and has a short running time. The t-SNE visualization further verifies the fault diagnosis performance, generalization ability and computational efficiency of the model. This study shows that the ICCEMDAN-WTD-CNN-BiLSTM model has high efficiency in the fault identification of roller bearings, and has practical value for intelligent online monitoring and intelligent mine construction.</p>