关键词: 直升机, 尾传动, 滚动轴承, 样本不平衡, 故障诊断, 数据增强

Abstract: To address the problem of dataset imbalance caused by the scarcity of fault samples in the fault diagnosis of bearings in helicopter tail drive systems, we propose the Integrable Physical Information Generative Adversarial Network (IE-PIGAN) model. On the basis of retaining the original adversarial loss and network structure of Generative Adversarial Network (GAN), this model embeds the physical prior knowledge of tail drive fault bearings into the generator loss in the form of integral equations, and combines the data loss of physics-informed neural networks to impose physical constraints on the generation process, thus generating fault samples that take random noise as input and conform to the physical characteristics of bearings. Experimental results show that the signals generated by IE-PIGAN are highly similar to real signals in both time domain and frequency domain, and its percent root mean square error, root mean square error, and frequency-domain correlation coefficient of maximum mean discrepancy are all lower than those of GAN. When the generated data are added to the training set, the diagnostic accuracy of the convolutional neural network reaches 91.7%, which is 8.4% higher than that without generated samples, and the accuracy outperforms five mainstream models such as denoising autoencoders and variational autoencoders. The volume of generated data exhibits a trend of first increasing and then decreasing with respect to the enhancement effect. Compared with other generative models, IE-PIGAN has the slowest accuracy decline rate after the sample volume exceeds the inflection point. This model effectively alleviates the imbalance problem of fault samples for helicopter tail drive bearings, and improves the accuracy and robustness of tail drive bearing fault diagnosis.

Key words: helicopter, tail drive, rolling bearing, sample imbalance, fault diagnosis, data augmentation