Yang Dalian, Zhang Yang, Li Renjie, Long Hui, Huang Changzheng
Hunan Provincial Key Laboratory of Health Maintenance for Mechanical Equipment, Hunan University of Science and Technology, Xiangtan, 411201, China.
School of Intelligent Engineering, Shaoguan University, Shaoguan, 512000, China.
Sci Rep. 2025 Jul 2;15(1):22501. doi: 10.1038/s41598-025-06623-4.
Aiming at the problem of inaccurate fault diagnosis caused by limited fault samples and challenging feature extraction in helicopter planetary gear trains, this study proposes a fault diagnosis method based on Bispectrum and Attention Mechanism Deep Convolutional Generative Adversarial Networks (BAMDCGAN). First, to enhance the sample quality generated by the Attention Mechanism Deep Convolutional Generative Adversarial Network (AMDCGAN), bispectral features are adopted as input samples, forming the proposed BAMDCGAN framework. Secondly, by utilizing the experimental data of the planetary gear train, bispectral feature samples under three load conditions and five fault states were constructed. These samples were used as the training data for BAMDCGAN, enabling the adversarial generation of enhanced fault samples. Finally, the Convolutional Neural Network (CNN) and Vision Transformer (VIT) are trained on the augmented dataset respectively for planetary gear train fault diagnosis. Comparative experiments with Envelope Spectrum + AMDCGAN + CNN, Time Domain + AMDCGAN + CNN, Short-Time Fourier Transform (STFT) + AMDCGAN + CNN, Envelope Spectrum + AMDCGAN + VIT and Hilbert-Huang Transform (HHT) + AMDCGAN + CNN methods demonstrate that the proposed BAMDCGAN-based fault diagnosis method achieves the highest diagnostic accuracy, exceeding 97.8% across varying load conditions. Compared to non-augmented samples, diagnostic accuracy is improved by 2.1%.
针对直升机行星齿轮系故障样本有限导致故障诊断不准确以及特征提取具有挑战性的问题，本研究提出了一种基于双谱和注意力机制深度卷积生成对抗网络（BAMDCGAN）的故障诊断方法。首先，为了提高注意力机制深度卷积生成对抗网络（AMDCGAN）生成的样本质量，采用双谱特征作为输入样本，形成了所提出的BAMDCGAN框架。其次，利用行星齿轮系的实验数据，构建了三种负载条件和五种故障状态下的双谱特征样本。这些样本被用作BAMDCGAN的训练数据，实现了增强故障样本的对抗生成。最后，分别在增强数据集上训练卷积神经网络（CNN）和视觉Transformer（VIT）用于行星齿轮系故障诊断。与包络谱+AMDCGAN+CNN、时域+AMDCGAN+CNN、短时傅里叶变换（STFT）+AMDCGAN+CNN、包络谱+AMDCGAN+VIT和希尔伯特-黄变换（HHT）+AMDCGAN+CNN方法的对比实验表明，所提出的基于BAMDCGAN的故障诊断方法实现了最高的诊断准确率，在不同负载条件下超过97.8%。与未增强的样本相比，诊断准确率提高了2.1%。
