【摘要】 针对主轴轴承监测数据中正常样本数量与故障样本数量失衡导致的智能模型精度低、鲁棒性差的问题，提出互信息（mutual information，简称MI）改进的奇异谱分解（singular spectrum decomposition，简称SSD）联合生成对抗网络（generative adversarial networks，简称GAN）的数据生成方法（记作MISSD）。首先，对轴承振动信号进行SSD分解，利用MI理论对分解分量进行筛选和归一化处理；其次，使用一维卷积神经网络（convolutional neural network，简称CNN）融合筛选后的信号，作为模型的噪声输入，以保证信息最大化，生成器采用长短时记忆（long short-term memory，简称LSTM）网络代替传统CNN，使生成器结构更适合一维振动信号；然后，引入具有梯度的Wasserstein距离改进损失，并结合谱归一化以提高模型稳定性，对抗生成具有真实样本特征的虚拟样本；最后，构建具有注意力机制（attention mechanism，简称ATT）的CNN模型（记作CNN-ATT）完成故障诊断。结果表明：所提方法生成的故障样本在各项指标上均优于其他模型，对比一维CNN（记作1D-CNN）模型，3个数据集准确率分别提高了7.6%、6.8%和5.7%。所提方法有效提升了故障样本不足情况下的故障诊断精度。更多还原

【Abstract】 Aiming at the problem of low accuracy and poor robustness of intelligent models caused by the imbalance between the number of normal samples and the number of faulty samples in the spindle bearing monitoring data, a data generation method of mutual information( MI) improved singular spectrum decomposition( SSD) combined with generative adversarial networks( GAN) is proposed and denoted as MISSD. Firstly, the bearing vibration signals are decomposed by SSD, filtered and normalized using MI theory, and the filtered signals are fused using a 1D convolutional neural network(CNN) as the noise input to the model to ensure information maximization. Secondly, the generator adopts long short term networks instead of CNN to make it more suitable for 1D vibration signals. Then, Wasserstein distance improvement loss with gradient is introduced and combined with spectral normalization to improve model stability against generating virtual samples with real sample characteristics. Finally, a CNN network model with attention mechanism( ATT) is constructed and denoted as CNN-ATT to complete fault diagnosis. The spindle bearing experiments show that the fault samples generated by the proposed method are better than other models in various indexes. Compared with the 1D-CNN model, the accuracy of the 3 datasets is improved by 7.6%, 6.8%, and 5.7%, respectively. The proposed method effectively solves the problem of fault diagnosis under the condition of imbalanced fault samples.更多还原