Abstract
The fault diagnosis of marine propulsion motors has been an important focus of attention in the marine industry. Various types and numbers of sensors have been used to monitor and diagnose faults in permanent magnet propulsion motors, and the comprehensive application of multi-sensor signals has played a key role in improving fault diagnosis performance, but the issue of how to efficiently exploit this multi-source information remains a difficult problem. In this paper, we propose a multi-sensor-signal feature-level fusion method based on a multi-input convolutional neural network and a CBAM attention mechanism, which fully utilises the end-to-end learning capability of deep learning and the interpretability and domain expert knowledge of traditional methods. A synchrosqueezing wavelet transform is used to extract the high-resolution feature information of the current and vibration signals; the multi-input neural network extracts the high-level abstract features in the current and vibration signals; the CBAM attention mechanism is introduced to make the network more targeted, to deal with the key feature information; and a Bayesian optimisation algorithm is used to automatically determine suitable combinations of hyperparameters for training of the network. A fault test of a permanent magnet synchronous motor shows that the diagnostic accuracy of the proposed method reaches 99.08%, a value 1.29% higher than a scheme without the CBAM attention mechanism, with an increase in the detection time for each sample of only 2.33 ms. Our approach also has better anti-noise interference ability and generalisation performance.