【摘要】 在齿轮箱运行过程中，难以及时获取其充足的故障样本，且生成对抗网络过程难以监控，使得生成样本质量不稳定，导致数据训练不充分以及诊断准确率低等问题。针对上述问题，为了提高模型诊断的准确率，提出一种基于多头注意力机制的条件生成对抗网络与时间卷积网络小样本齿轮箱故障诊断模型。首先，通过在条件生成对抗网络模块前添加多头注意力机制，确保将输入数据的特征向量映射到独立子空间中，对数据特征重新赋予权重；然后，在生成对抗网络对抗过程中引入条件信息提升模型的数据生成和判别精度能力，改善生成样本质量不稳定的问题；最后，将数据合成后统一输入到时间卷积网络中进行模型训练，使模型充分学习数据特征，提高故障诊断的准确率。实验结果表明，本模型针对小样本齿轮箱故障诊断能够有效监控对抗过程生成高质量数据，平均准确率达到98.6%,损失值仅为0.07,验证了该方法的可行性。文中提出的方法为解决小样本齿轮箱故障诊断难题提供了一种新思路，具有良好的工程实践意义。更多还原

【Abstract】 In gearbox operations, obtaining a sufficient number of fault samples in a timely manner is often a significant challenge. Furthermore, the adversarial training process in generative adversarial networks is difficult to supervise, resulting in unstable sample quality, inadequate model training and reduced diagnostic accuracy. To address these issues and improve diagnostic performance, a small-sample gearbox fault diagnosis model is proposed based on a conditional generative adversarial network and a time convolutional network, enhanced with a multi-head attention mechanism. First, the multi-head attention mechanism is introduced before the conditional generative adversarial network to map the feature vectors into independent subspaces, re-weighting the data features for better representation. Next, conditional information is incorporated into the generative adversarial networks framework to improve both the generation and discrimination capabilities, stabilizing the adversarial training process and enhancing sample quality. Finally, both synthetic and real data are fed into the time convolutional network for model training, allowing the model to effectively capture temporal dependencies and thus improve diagnostic accuracy. Experimental results demonstrate that the model successfully monitors the adversarial generation process, producing high-quality data with an average accuracy of 98.6% and a loss rate of only 0.07. The proposed approach offers a novel solution to the small-sample gearbox fault diagnosis problem and has significant practical value for engineering applications.更多还原