Search papers, labs, and topics across Lattice.
This paper introduces a novel intelligent fault diagnosis method for PMSM stator faults by combining Variational Mode Decomposition (VMD) for feature extraction with a parallel BiTCN-Transformer network for fault classification. The parallel architecture leverages BiTCN for local feature extraction and Transformer for global dependency modeling, enhancing the detection of weak early-stage fault features. Experimental results on the KAIST PMSM stator fault dataset demonstrate a high average diagnostic accuracy of 99.42% across 15 fault categories, outperforming traditional methods.
Achieve 99.42% accuracy in diagnosing 15 PMSM stator fault categories by fusing BiTCN's local feature extraction with Transformer's global dependency modeling in a parallel architecture.
To address the challenges of weak early features, challenges in feature extraction, reliance on prior knowledge in traditional diagnostic methods, and the long training time, slow training convergence, and low efficiency of mainstream serial models for inter-turn short circuits (ITSC)and inter-coil short circuits(ICSC) in permanent magnet synchronous motor (PMSM) stators, this paper proposes an intelligent fault diagnosis method combining variational mode decomposition (VMD) with a BiTCN-Transformer parallel model. First, VMD decomposes three-phase current signals into multiple scales to extract intrinsic mode functions (intrinsic mode functions ,IMFs) containing fault-sensitive information. Subsequently, leveraging the complementary advantages of BiTCN in local feature extraction and Transformer in long-range dependency modeling, a parallel dual-branch architecture is designed: the BiTCN branch captures local temporal dynamic features through bidirectional dilated convolutions, while the Transformer branch captures global dependencies via a self-attention mechanism. To reduce computational complexity, a convolutional layer with stride 8 is embedded at the front end for sequence downsampling, and implicit convolutional encoding replaces fixed positional encoding. The core encoder retains only a single-layer multi-head self-attention structure, reducing parameters for model lightweighting. Finally, features from both branches are collaboratively learned in a fusion layer for multi-scale, multi-perspective fault pattern discrimination. Experimental comparisons on the Korea Advanced Institute of Science and Technology (KAIST) PMSM stator fault dataset show that the proposed parallel architecture achieves an average diagnostic accuracy of 99.42% across 15 stator fault categories, outperforming traditional serial models and redundant multi-branch models. The method requires no prior knowledge or massive data, exhibits fast training and convergence, and demonstrates significant potential for practical engineering applications, providing an effective solution for early weak fault diagnosis in motors.
