The measurement error of traditional high-resistance connection (HRC) fault diagnosis method for permanent magnet synchronous motor (PMSM) is typically high, and it is difficult to evaluate HRC fault comprehensively and quantitatively. Modern diagnosticmethods of HRC fault are realized by the symmetry monitoring system of injection signal method and zero-sequence component detection method. However, these methods based on signal injection strategy are invasive diagnostics, which may have an impact on the original drive system. The method based on zero sequence component must have neutral points and additional resistance network, and the application scenario is limited. To solve the above problems, this paper proposes a data-driven method to realize the HRC fault diagnosis of PMSM, and establishes the fault mode of the system through a large number of existing historical data, without the need for a prior known model or signal mode. Firstly, a simulated PMSM model is established to obtain the three-phase current signal as an effective fault signal. Secondly, the recursive graph method is innovatively introduced to map the three-phase current signals into two-dimensional recursive images and perform multi-channel fusion. The fused images contain more feature information and avoid the influence of artificial feature extraction during fault signal processing. Then, ConvNeXt is introduced to effectively solve the gradient dispersion problem of the existing classification model, and a new ECA-ConvNeXt classification model is obtained by integrating the attention mechanism, which improves the adaptability of the network in the channel dimension and enhances the generalization ability of the model. The simulation results show that compared with the basic ConvNeXt, the accuracy of ECA-ConvNeXt on the test set sample is improved from 98.05% to 99.18%. Compared with models such as GoogLeNet, ResNet and Swin Transformer, the ECA-ConvNeXt proposed in this paper also has higher performance indexes and more advantages in identifying early HRC faults. To verify the robustness of the model in the presence of input data noise, Gaussian noise is added to the original input data. The results show that the effect of these noises is not significant. Finally, through the establishment of prototype experimental platform verification, it is proved that the identification accuracy of the new method proposed in this paper is as high as 97.35%, and it can accurately identify the location and severity of high resistance contact faults of permanent magnet synchronous motors. Based on the analysis of simulation and test data, this paper draws the following conclusions: (1) This paper uses data-driven method to realize the non-invasive diagnosis of HRC fault of permanent magnet synchronous motor, avoiding the problem that the mathematical and electromagnetic relationship is too complicated in the model diagnosis method. (2) The two-dimensional recursive graph signal conversion method is innovatively introduced into the field of high resistance fault diagnosis of permanent magnet motors, and one-dimensional three-phase current signals are transformed into two-dimensional recursive images respectively and channel fusion is carried out to enhance fault characteristics, eliminate the influence of artificial feature extraction, and lay a rich fault information foundation for the classification and recognition of HRC faults of permanent magnet synchronous motors. (3) An ECA-ConvNeXt classification model with added attention mechanism was proposed to solve the gradient dispersion problem, improve the adaptability of the model in the channel dimension, and increase the classification accuracy of the model on the experimental data set from 96.53% to 97.35%. The comparison between the proposed method and other classification models proves the superiority of this classification model. © 2024 China Machine Press. All rights reserved.
