【目的】针对航空发电机电枢绕组匝间短路故障诊断问题，本文提出一种基于卷积神经网络(CNN)和压缩-激励(SE)注意力机制以及双向长短期记忆(BiLSTM)网络的诊断方法，以提高电枢绕组故障诊断的有效性和鲁棒性。【方法】首先，建立了电励磁双凸极发电机(DSEG)电枢绕组匝间短路的等效解析模型，采用发电机机壳表面振动信号作为故障特征信号，对影响振动特性的气隙磁密和电磁力波的表达式进行了理论推导。然后，通过Workbench有限元软件仿真分析了各种短路条件下电磁力波对振动特性的影响。最后，对振动信号进行采集，将其作为实际试验数据并提取相关故障特征，通过将这些提取的特征输入CNN-SE-BiLSTM，系统有效地实现了故障诊断。其中，CNN和BiLSTM网络可以有效提取振动信号中的局部特征和时序特征;SE注意力机制可以进一步优化特征权重分配，通过选择性放大相关诊断特征，显著增强了模型的故障诊断能力。【结果】试验结果表明，所提基于CNN-SE-BiLSTM的故障诊断方法在多种工况下对电枢绕组匝间短路故障的诊断准确率均超过99%，且在噪声环境中仍能表现出良好的诊断性能。与传统诊断方法相比，所提方法在各种转速和负载条件下具有更强的抗干扰能力、更高的故障识别精度以及更快的诊断速度，进一步验证了该方法在航空航天发电机复杂工况下的适用性。【结论】本文所提CNN-SE-BiLSTM诊断方法有效提高了DSEG系统的故障诊断效率和准确性。通过将用于空间特征提取的CNN、用于选择性特征加权的SE注意力机制以及用于捕捉综合时间信息的BiLSTM相结合，该方法为诊断匝间短路故障提供了一种精简而高效的解决方案。

[Objective] For armature winding inter-turn short circuit fault diagnosis in aerospace generators, this paper proposes a diagnostic method based on convolutional neural network (CNN) with squeeze-excitation (SE) attention mechanism and bidirectional long short-term memory (BiLSTM). It aims to enhance the effectiveness and robustness of armature winding fault diagnosis. [Methods] Firstly, an equivalent analytical model for armature winding inter-turn short circuit fault diagnosis in a doubly salient electromagnetic generator (DSEG) was established, and the vibration signal on the surface of the generator casing was adopted as a characteristic signal of the fault. Theoretical derivations of the expressions for the air gap magnetic flux density and electromagnetic force, which influenced vibration characteristics, were carried out. Secondly, the impact of electromagnetic force on vibration characteristics under various short circuits was analyzed using finite element simulation in Workbench software. Finally, vibration signals were collected and used as actual experimental data to extract relevant fault features. These extracted features were put into the CNN-SE-BiLSTM for efficient fault diagnosis. Among this, the CNN and BiLSTM networks effectively extracted local features and features in time series from vibration signals, while the SE attention mechanism optimized the feature weight distribution by selectively amplifying relevant diagnostic features, significantly enhancing the model’s fault diagnosis capability. [Results] The experimental results showed that the proposed CNN-SE-BiLSTM-based method achieved a diagnostic accuracy of over 99% for inter-turn short circuit faults in armature windings under various working conditions. Moreover, it still exhibited strong diagnostic performance in noisy environments. Compared to traditional diagnostic methods, the proposed method not only demonstrated strong anti-interference ability, high fault recognition accuracy, and fast diagnostic speed under various rotational speeds and load conditions, validating its applicability in aerospace generators with complex operating conditions. [Conclusion] The CNN-SE-BiLSTM diagnostic approach effectively improves the fault diagnosis efficiency and accuracy of DSEG systems. By combining CNN for spatial feature extraction, SE attention mechanism for selective feature weighting, and BiLSTM for capturing comprehensive temporal information, this method provides a streamlined and highly effective solution for inter-turn short-circuit fault diagnosis.

中央高校基本科研业务费专项资金资助(NS2021021);航空科学基金项目(201933052001)