【摘要】 滚动轴承在实际运行时，出现故障时间段通常较短，所以采集到的轴承故障振动数据较少，导致轴承正常振动信号的数量远多于故障振动信号，造成数据集的严重不平衡，这显著降低深度学习网络模型的泛化性能和识别准确性。针对该问题，本文提出一种自动生成故障轴承数据的方法，并结合SEMobileNetV2轻量化卷积神经网络，改善由于数据的不平衡性导致的网络性能损失，提高故障轴承诊断的识别率，采用梯度惩罚生成对抗网络，使用改进的SEMobileNetV2轻量化卷积神经网络进行故障分类。多种实验表明，本文采用的WGAN-GP生成对抗网络能够生成与真实数据高度相似的样本。将平衡与不平衡数据集样本分别输入至改进前后的MobileNetV2网络中进行训练。改进网络前，在CRWU和XJTU平衡数据集上得到的平均测试准确率分别为97.84%和98.04%,均高于不平衡数据集上的89.6%与76.70%;改进网络后，在CRWU和XJTU平衡数据集上得到的平均测试准确率分别为98.72%和99.80%,均高于不平衡数据集上的91.60%与87.40%。通过4组实验对比，说明本文所提方法能够有效扩充真实样本集，并提高轴承故障诊断的识别准确率。更多还原

【Abstract】 The time of rolling bearings’ fault operation in actual work is very short, so the collected bearing fault vibration data is usually small, and the collected normal bearing vibration signals are far more than the fault vibration signals, resulting in a serious imbalance in the data set, which will greatly affect the generalization performance and identification accuracy of deep learning network model. To solve this problem, this paper proposes a method to automatically generate fault bearing data, and combines SEMobileNetV2 lightweight convolutional neural network to improve the network performance loss caused by data imbalance, improve the recognition rate of fault bearing diagnosis, and generate a countermeasure network using gradient punishment. An improved SEMobileNetV2 lightweight convolutional neural network is used for fault classification. Experiments show that the WGAN-GP generative adversarial network used in this paper can generate samples that are highly like real data. The balanced and unbalanced data set samples were input into the MobileNetV2 network before and after the improvement for training. Before the improvement, the average test accuracy of CRWU and XJTU balanced data sets were 97.84% and 98.04, respectively, higher than 89.60% and 76.70% on unbalanced data sets. After improving the network, the average test accuracy of CRWU and XJTU balanced data sets are 98.72% and 99.80%, respectively, which are higher than 91.60% and 87.40% on unbalanced data sets. Through the comparison of four groups of experiments, it can be shown that the proposed method effectively expand the real sample set and improve the identification accuracy of bearing fault diagnosis.更多还原