Recently health management techniques for mechanical equipment based on vibration signals have been widely studied. However, the working environment of many mechanical equipment has variable excitation sources, and the motion process of the equipment itself is also very complex. The useful information in the sensor vibration signal is scattered in different time scales and is difficult to be extracted directly. To solve the above problems, Stacked Residual Multi-Attention Network (SAMANet) is proposed for feature extraction of vibration signals and provide a study in detail. Stacking using one-dimensional convolution modules allows SAMANet to adapt to arbitrary data lengths. Designing Squeeze-excitation residual blocks (SE-Res blocks) to obtain additive features with low redundancy. The attention fusion unit is proposed to ensure the interpretability of the model and ultimately to obtain representative features. Finally, the interpretability, identification accuracy and adaptability of the model to different operating conditions are verified on 12 different fault tasks in planetary gearboxes. As a result of the study, SAMANet performs better in fault diagnosis tasks compared to other deep learning benchmark models.

Machine fault diagnosis;Universal feature extractor;Convolutional neural network;Attention fusion;Interpretability