电工技术学报2025，Vol.40Issue(19)：6223-6241,6256,20.DOI:10.19595/j.cnki.1000-6753.tces.241537
基于小波散射变换和Shuffle-PMDA的永磁同步电机故障识别
Permanent Magnet Synchronous Motor Fault Identification Based on Wavelet Scattering Transform and Shuffle-PMDA
摘要
Abstract
Permanent magnet synchronous motor(PMSM)is prone to various faults in the complex and changing working environment,and serious faults may lead to equipment damage or scrap.If the turn-to-turn and winding-to-winding short-circuit faults are not detected early,they may evolve into more serious faults,such as phase-to-phase short-circuits.However,detecting the same-phase turn-to-turn and winding-to-winding short-circuit faults at the early stage is challenging.The existing deep-learning-based fault diagnosis models have high computational costs and difficulty mining deep hidden features.Therefore,this paper proposes a new fault identification method combining the Wavelet scattering transform(WST)and a lightweight convolutional network Shuffle module with parallel multi-dimensional attention(PMDA). The U-phase current signals of different fault types of PMSM are sampled through a sliding window,and the length of each sample is set to be 3 000.80%of the samples are used for model training and 20%for model testing.WST processes each sample sequence to generate a feature matrix for fault identification.WST enhances the weak features in the turn-to-turn and winding-to-winding short circuit faults,reduces the data dimensionality,and maintains the distinguishability between different fault levels.Based on the lightweight convolutional network,the Shuffle module and PMDA combination model can deeply mine and learn the fault feature information.Effective differentiation of feature importance is achieved by introducing a multi-scale attention mechanism. Experiments on real data show that the proposed method performs well in several aspects.Firstly,the method achieves significant recognition results regarding inter-winding short-circuit faults of different degrees.Its recognition accuracy is as high as 99.75%,better than the traditional fault diagnosis methods like CWT time-frequency diagram(96.70%),VMD component matrix(95.16%),VMD component pseudo-color diagram(91.42%),and Gram's angle and field map(80.42%).Secondly,the recognition accuracy is 99.67%under low SNR(20 dB)and maintains a high recognition accuracy of 94.83%under high SNR(5 dB),exhibiting good robustness in complex noise environments.In addition,compared with other lightweight pre-trained deep learning models,the proposed model shows obvious advantages in multiple metrics,such as training time,sample average test time,recall rate,specificity,and F1 score.Ablation experiments of different attention mechanism combination models show that the deep learning combination model significantly improves PMSM fault recognition.Finally,the experimental results on dataset Ⅰ(the number of fault types is 30)and dataset Ⅱ(the number of fault types is 45)under the increased rated power of 1.5 kW and 3.0 kW show that the recognition accuracy of dataset Ⅰ is 97.00%,and dataset Ⅱ is 94.19%.The proposed model has a strong generalization and adaptation ability for power and multiple-fault identification.关键词
永磁同步电机/故障识别/小波散射变换/深度学习/Shuffle-PMDA模型Key words
Permanent magnet synchronous motor/fault identification/wavelet scattering transform/deep learning/shuffle-PMDA model分类
信息技术与安全科学引用本文复制引用
毕贵红,李玉洪,赵四洪,陈仕龙,张靖超..基于小波散射变换和Shuffle-PMDA的永磁同步电机故障识别[J].电工技术学报,2025,40(19):6223-6241,6256,20.基金项目
国家自然科学基金资助项目(51767012). （51767012）
