Efficient Multimodal Motor Demagnetization Diagnosis Framework With Half-Wave Fundamental Wave Division Transformation

The gradual adoption of hybrid permanent magnet (HPM) motors has driven increased demand for effective demagnetization fault diagnosis methods. Current techniques for diagnosing demagnetization are either vulnerable to interference or incapable of extracting discernible features. Accordingly, this paper aims to identify clearer features for demagnetization diagnosis and develop a lightweight intelligent model to reduce sample size dependence. In this paper, the halfwave fundamental division-relative position matrix (HFD-RPM) is proposed. This transformation involves initially capturing the positive halfwave of back electromotive force and extracting the fundamental wave for reconstruction. Then, the positive halfwave is divided by the reconstructed halfwave fundamental to substantially alter the sequence data and yield effective time-frequency redundancy features. The resulting sequence data is then normalized to a 1D series, which is subsequently input into a 1D convolutional neural network (CNN) for classification. Through the training and interpretability analysis of the classifier, the proposed framework demonstrates 100% accuracy, a loss of 0.0126, and the proposed transformation enables exceptional lightweight performance. Compared to other feature extraction methods, the proposed approach generates redundant features that contribute more distinctly to classification, enabling the diagnosis of subtle demagnetization signals. Finally, cross-validation using Gaussian signals and real measured signals from the prototypes confirms the proposed framework’s robustness to interference noise.