由于机电伺服系统（Electro-Mechanical Actuator，EMA）在实际应用中故障数据样本少，会影响故障诊断方法的分类效果。针对故障数据缺失下机电伺服系统的故障诊断问题，设计了一种基于门控循环单元（Gated Recurrent Unit，GRU）和卷积注意力的改进辅助分类生成对抗网络（Auxiliary Classifier Generative Adversarial Network，ACGAN）故障诊断方法，能够稳定地生成各故障类别高质量数据。首先，在ACGAN中引入Wasserstein距离与梯度惩罚，优化损失函数，提升对抗训练稳定性。其次，在生成器和判别器中加入GRU和卷积注意力模块，增强网络对关键特征和时序特征的提取能力，克服了卷积网络在处理时序数据时的局限性，提高了生成样本的质量。最后，通过共享分类器与判别器网络参数，利用平衡数据集微调分类器，进一步提高模型的诊断性能。基于搭建的EMA实验台，得到由大量正常数据与少量故障数据组成的不平衡实验数据集，通过对比和消融实验，验证了所提方法的有效性和优越性。

Due to the limited number of fault data samples in practical applications of Electro-Mechanical Actuator (EMA), the classification performance of fault diagnosis methods can be affected. To address the fault diagnosis problem in EMA with missing fault data, an improved Auxiliary Classifier Generative Adversarial Network (ACGAN) fault diagnosis method based on GRU and convolutional block attention module (CBAM) is designed, which can stably generate high-quality data for each fault category. First, Wasserstein distance and gradient penalty are introduced into the ACGAN to optimize the loss function, improving the stability of adversarial training. Second, GRU and CBAM are add-ed to both the generator and discriminator, enhancing the ability of network to extract key region features and tem-poral features, overcoming the limitations of convolutional networks in handling sequential data, and improving the quality of generated samples. Finally, parameter sharing between the classifier and discriminator is implemented, and the balanced dataset is used to fine-tune the classifier, further improving the diagnostic performance of model. Based on the EMA test rig, an unbalanced dataset consisting of a large amount of normal data and a small amount of fault data is obtained, and the effectiveness and superiority of the proposed method is verified through comparison and ablation experiments.