【目的】异步电机在现代工业变工况及跨设备应用场景下,因数据分布偏移导致传统故障诊断模型性能严重退化。针对现有单一域对齐策略泛化能力不足以及诊断模型计算复杂度高的问题,本文提出了一种融合领域对抗网络(DANN)与关联对齐(CORAL)的轻量化故障诊断方法。【方法】基于Maxwell与Simulink搭建联合仿真平台,模拟矢量控制下电机的动态行为,获取多工况定子电流信号。采用时域差分与连续小波变换协同的预处理策略,抑制基波分量并生成高辨识度的时频图像特征。采用轻量化网络GhostNetV2作为特征提取骨干,以降低计算成本。在标准域对抗训练的基础上引入CORAL损失,通过显式对齐源域与目标域特征的二阶统计量,构建隐式对抗与显式统计量匹配的双重域对齐机制。【结果】仿真和试验结果表明,本文所提方法在保证模型轻量化的同时显著提升了诊断性能。在同一电机变工况的迁移任务中,目标域平均分类正确率达到99.1%;在更具挑战的不同电机跨型号迁移任务中,目标域平均分类正确率保持在81.9%,显著优于对比算法。【结论】本文所提双重域对齐策略能够在一定程度上缓解复杂工况下的特征迁移难题,在降噪效果、模型轻量化与泛化性能之间取得了良好平衡,为异步电机的非侵入式在线监测与跨域故障诊断提供了可靠的解决方案。

[Objective] In modern industrial applications with variable operating conditions and cross-equipment scenarios, the performance of traditional fault diagnosis models for induction motors severely degrades due to data distribution shifts. To address the limitations of existing single-domain alignment strategies (insufficient generalization) and high computational complexity in diagnostic models, this paper proposes a lightweight fault diagnosis method integrating domain adversarial neural network (DANN) with correlation alignment (CORAL). [Methods] A co-simulation platform was established based on Maxwell and Simulink to simulate the dynamic behavior of motors under vector control, and stator current signals under multiple operating conditions were obtained. A preprocessing strategy combining time-domain differencing and continuous wavelet transform was adopted to suppress the fundamental component and generate highly distinguishable time-frequency image features. The lightweight network GhostNetV2 was employed as the feature extraction backbone to reduce computational costs. On the basis of standard domain adversarial training, a CORAL loss was introduced, where the second-order statistics of source and target domain features were explicitly aligned, constructing a dual-domain alignment mechanism that implicit adversarial learning with explicit statistical matching. [Results] The simulation and experimental results demonstrated that the proposed method significantly improved diagnostic performance while maintaining model lightweightness. In the transfer task under varying operating conditions of the same motor, an average classification accuracy of 99.1% was achieved in the target domain. For the more challenging cross-model transfer task between different motors, the average classification accuracy remained at 81.9% in the target domain, which was significantly superior to comparative algorithms. [Conclusion] The proposed dual-domain alignment strategy can alleviate the feature transfer challenges under complex operating conditions to a certain extent, achieving a balanced performance in noise reduction, model lightweighting, and generalization capability. It provides a reliable solution for non-intrusive online monitoring and cross-domain fault diagnosis of induction motors.

福建省自然科学基金项目(2025J01524)