In view of the complex environment and frequent faults in the actual operation of mine hoist, a fault diagnosis method based on Convolution Attention Autoencoder (CAAE) is proposed through theoretical analysis and experimental verification to improve the diagnostic stability of mine hoist under strong noise. First, a CAAE is constructed, which uses a combination of a convolutional neural network (CNN) and a channel attention module (CAM) to compress and encode the input signal, and then the input signal is reconstructed by a decoder to train the CAAE to extract the original signal fault features. Then, a fault diagnosis classifier is constructed to classify different fault patterns. Finally, experimental validation is performed with the Case Western Reserve University bearing dataset. The results show that the method has a strong feature extraction capability and a high classification accuracy for bearing failure modes compared with existing methods. And the experiments on the application effect of the proposed method in noisy environment are conducted to verify that the method is highly effective and challenging.