【目的】针对电机故障诊断在复杂工况下特征提取不充分、时序依赖关系捕捉不足及噪声干扰下诊断精度不高的问题,提出一种基于多尺度卷积-长短期记忆网络-注意力机制(MCNN-LSTM-Attention)的故障诊断方法。【方法】首先,通过MCNN模块使并行卷积层提取振动信号中的多尺度局部特征,解决难以全面表征故障特征的问题。然后,利用LSTM模块对多尺度特征序列进行时序建模,深入挖掘故障特征的时序依赖关系。在此基础上,引入Attention模块对时序特征进行自适应加权,以聚焦关键故障信息,抑制无关噪声干扰。最后,多个模块协同解决特征表示不全面与诊断模型鲁棒性不强的问题。【结果】为了全面评估所提电机故障诊断方法在实际应用中的鲁棒性与适应性,通过模拟现场复杂工况,引入负载波动、转速扰动等典型非理想因素,并在信号采集过程中叠加不同强度的噪声干扰,以构建贴近真实工业环境的测试场景。在此基础上,利用实际采集的电机运行数据对诊断模型进行验证。试验结果表明,该方法在多种复杂工况和不同噪声水平下均能保持较高的诊断准确率,同时其在跨工况、跨故障类型的测试中表现出良好的泛化能力。【结论】所提方法实现了电机在复杂工况下的准确故障诊断,这一成果为实际工程应用提供了可靠的技术支撑,同时也为后续设备运行中的预防性维护措施提供了有益的参考和新的思路。

[Objective] To address the challenges of insufficient feature extraction, inadequate capture of temporal dependencies, and low diagnostic accuracy under noise interference in motor fault diagnosis under complex working conditions, a fault diagnosis method based on a multi-scale convolutional neural network, long short-term memory, and attention mechanism (MCNN-LSTM-Attention) is proposed. [Methods] First, through the MCNN module, the parallel convolutional layers were used to extract multi-scale local features from the vibration signals, addressing the problem of difficulty in comprehensively characterizing fault features. Then, the LSTM module was utilized to perform temporal modeling on the multi-scale feature sequences, deeply mining the temporal dependencies of the fault features. On this basis, the Attention mechanism was introduced to adaptively weight the temporal features, so as to focus on the key fault information and suppress irrelevant noise interference. Finally, the multiple modules worked collaboratively to solve the problems of incomplete feature representation and weak robustness of the diagnostic model. [Results] To comprehensively evaluate the robustness and adaptability of the proposed motor fault diagnosis method in practical applications, typical non-ideal factors such as load fluctuations and speed disturbances were simulated and introduced by emulating complex on-site operating conditions, and noise interferences of different intensities were superimposed during the signal acquisition process, so as to construct a test scenario close to the real industrial environment. On this basis, the diagnostic model was verified using actual collected motor operating data. The experimental results showed that the method maintained high diagnostic accuracy under various complex operating conditions and different noise levels, and at the same time, it exhibited good generalization capability in tests across different operating conditions and fault types. [Conclusion] The proposed method achieves accurate fault diagnosis of motors under complex operating conditions. This result provides reliable technical support for practical engineering applications, and also offers useful reference and new insights for preventive maintenance measures in subsequent equipment operation.

辽宁省教育厅科研项目(JYTMS20231483)；辽宁省科技厅科研项目(2024011836-JH3/4700）