%0 Journal Article %T 基于GAN和少样本学习的电机故障诊断方法研究
Research on Motor Fault Diagnosis Method Based on GAN and Few-Shot Learning %A 邹琪红 %J Software Engineering and Applications %P 176-188 %@ 2325-2278 %D 2025 %I Hans Publishing %R 10.12677/sea.2025.142017 %X 随着工业自动化的进步，电机故障诊断已成为保证工业设备正常运行的关键技术。传统的故障诊断方法往往依赖于大量标注数据进行训练，但实际应用中，电机的故障种类繁多且故障数据较为稀少，数据不足成了故障诊断的一大瓶颈，为了克服这一挑战，本文提出了基于对抗生成网络的故障诊断方法，旨在解决数据问题，提高模型性能。本文引入了对抗生成网络，通过增强数据来缓解数据不足的问题。具体来说，我们先将振动数据进行频谱化，然后作为对抗生成网络的输入，通过生成器生成更多的数据，用于扩充数据集，生成的数据不仅具有较好的时频特征，还保留故障的多样性，从而更好地提升小样本学习模型的泛化能力。经过增强后的数据再经过LKA (Large Kernel Attention)模块进行特征提取，最后经过一个全局分支一个局部分支处理后进行分类。此外为了进一步提升诊断精度和训练效率，我们结合了KL散度和Wasserstein距离，提出并采用了动态权重调整策略和学习率调整策略，使得训练过程更加稳定，并加速了优化过程。本文在公开数据集CWRU上进行了大量实验，结果表明了我们所提模型的有效性。
With the advancement of industrial automation, motor fault diagnosis has become a key technology to ensure the normal operation of industrial equipment. Traditional fault diagnosis methods often rely on large amounts of labeled data for training. However, in practical applications, motor faults are varied, and fault data is relatively scarce, making data insufficiency a major bottleneck for fault diagnosis. To address this challenge, this paper proposes a fault diagnosis method based on Generative Adversarial Networks (GANs) to solve the data issue and improve model performance. In this study, we introduce GANs to enhance data and alleviate the problem of data insufficiency. Specifically, we first convert vibration data into spectrograms, which are then used as input to the GAN. The generator generates additional data to expand the dataset. The generated data not only possesses good time-frequency features but also retains the diversity of faults, thereby improving the generalization ability of the few-shot learning model. The enhanced data is then processed through the LKA (Large Kernel Attention) module for feature extraction. Finally, after processing by a global branch and a local branch, the data is classified. Furthermore, to further improve diagnostic accuracy and training efficiency, we combine KL divergence and Wasserstein distance, and propose dynamic weight adjustment and learning rate adjustment strategies. These strategies make the training process more stable and accelerate optimization. Extensive experiments on the public CWRU dataset demonstrate the effectiveness of the proposed model. %K 电机故障， %K 诊断模型， %K 深度学习， %K GAN
Motor Fault %K Diagnosis Model %K Deep Learning %K GAN %U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=111075