To address the difficulty in distinguishing between inter-turn short circuit faults and local demagnetization faults in permanent magnet synchronous motors (PMSMs), this paper proposes a method for fault diagnosis based on multi-signal symmetrical dot pattern (MSDP) and hybrid attention improved deformable convolutional residual network (HADRN). First, considering the different current fluctuation features of the two fault types, fault features are extracted from three-phase current signals using the improved two-dimensional symmetrical dot pattern analysis method. Second, an improved residual network model incorporating deformable convolution and hybrid attention modules is constructed to extract weak features and perform fault category mapping. Finally, the proposed algorithm is validated by collecting current signal data through simulation experiments. Comparative studies with various neural network algorithms demonstrate that the proposed method exhibits stronger feature extraction capability and higher diagnostic accuracy.