关键词: 生成对抗网络, 电机异常检测, 数据增强, 时序模型, 不平衡数据

Abstract: In practical industrial scenarios,motor anomaly data is scarce and costly to acquire,leading to severe data imbalance issues for anomaly detection models based on deep learning, which in turn results in insufficient generalization capabilities. To address this challenge, this paper proposes an enhanced generative adversarial network architecture incorporating a global-local fusion module. This innovation significantly strengthens the model ’ s ability to capture both global temporal dependencies and local similarity features within time series data. The method aims to learn the highdimensional distribution and temporal dynamics of vibration signals under both normal and rare abnormal motor states, thereby generating high-quality, diverse synthetic abnormal time series data. Experimental results demonstrate that the synthetic data generated by this method exhibits high consistency with real data in both time-domain and frequency-domain features,effectively expanding the training dataset. Ultimately,the anomaly detection model trained on the augmented data achieved an accuracy of 98. 7%, a recall rate of 97. 9%, and an F1 score of 98. 3%—all representing significant improvements. This approach provides an effective technical solution for addressing the challenges of small-sample and
imbalanced data in industrial equipment monitoring.

Key words: generative adversarial networks, motor anomaly detection, data augmentation, time series models, imbalanced data