Data-Driven Diagnosis of PMSM Drive with Self-Sensing Signal Visualization and Deep Transfer Learning

Harmonics in the three-phase currents of PMSM systems are caused by inverter non-linearity, mismatched controller parameters, and sampling errors, which increase torque ripples. To address this issue, this article proposes a data-driven diagnosis method based on self-sensing signal visualization and deep learning. A comprehensive database was established by amalgamating data from simulations and experiments, encompassing a wide range of fault categories. Using self-sensing current signals as the original data, without requiring external test instruments. The short-time Fourier transform is used to obtain the spectrum image of current. And the spectrum images of the three-phase currents are combined into a single image to reduce the data dimension using image fusion. After classifying and labeling the data, the samples were trained using SqueezeNet's transfer learning. The test results show that the fault diagnosis accuracy of 99.03 percent. Compared with traditional methods, the method proposed in this article realizes system-level fault diagnosis and is more practical.