Fault diagnosis methods for rotating machine parts include traditional methods and deep learning, and the former often requires a lot of expert experience and the diagnosis accuracy is poor. A multi-scale attention deep convolutional neural network (MADCNN) is proposed to improve the fault diagnosis method. The MADCNN method provides three convolutional channels, and the principle of differential kernel size of each channel effectively widens the network to achieve multi-scale feature extraction of the original time-domain data. At the same time, CBAM further assigns weights to the extracted features to enhance the differentiation of the model for different types of faults. The accuracy of the validation set was improved by 7. 76% compared with the traditional deep convolutional model by using the bearing failure data from Case Western Reserve University (CWRU) and the planetary gearbox test bench failure data. The experimental results show that the method has high diagnostic accuracy and good generalization performance.