Abstract:

Permanent Magnet Synchronous Motor （PMSM） is widely used in electric vehicle drive systems due to its wide speed range， large output torque， fast speed response， small size and lightweight. PMSMs are susceptible to inter turn short-circuit faults， demagnetization faults， bearing wear faults and other faults due to harsh climate， abnormal vibration and frequent start-brake conditions. In this paper， to address the problems of difficulty in distinguishing PMSM similar faults with single-dimension signals and poor robustness of traditional diagnostic methods when the operating conditions change， a fault diagnosis method is proposed for automotive PMSMs based on the combination of Empirical Mode Decomposition-Symmetric Dot Pattern （EMD-SDP） image features and improved DenseNet convolutional neural network. Firstly， the vibration signals of PMSMs in different states under multiple operating conditions are experimentally obtained， and the pre-processed vibration signals are subjected to EMD to solve the Intrinsic Mode Function （IMF） at different levels. Secondly， the original vibration signals are transformed to the SDP images， and the IMFs at different levels are transformed into RGB color features to be displayed on the SDP images. Then， a classification network model is constructed by improving DenseNet learning image dataset through the fusion of scSE attention mechanism. Finally， the motor state to be measured is evaluated and diagnosed through the signal-image-network process. The diagnostic results show that the proposed method performs well under both steady state and variable speed transient conditions. Under constant speed and load conditions， the proposed method achieves the highest fault diagnosis accuracy （99.72%）， which is 1.66% higher than the accuracy of the DenseNet （98.06%）. The ROC curves of the improved DenseNet model and the DenseNet model are closest to the upper left corner， with the mean AUC values of 0.997 4 and 0.974 5. Under acceleration and deceleration with constant load conditions， the improved DenseNet model also achieves the highest diagnostic accuracies of 96.88% and 97.08%， with the mean AUC values of 0.987 7 and 0.986 9. The overall performance of the proposed method is better than conventional methods and can be effectively used for fault diagnosis during speed changes.

Key words: permanent magnet synchronous motors, fault diagnosis, empirical mode decomposition, symmetric dot pattern, scSE, DenseNet