基于ResNet和领域自适应的轴承故障诊断研究

Bearing Fault Diagnosis Based on ResNet and Domain Adaptation

轴承故障诊断在实际工业场景中意义重大。基于信号处理方法和机器学习方法,往往非常依赖先验知识,难以保证特征提取的有效性,深度学习方法要求训练集和测试集满足同一分布,这在工业现场难以满足,使得模型性能大幅下降。提出一种基于多层领域自适应的故障诊断方法,能够实现多种类、多尺寸的轴承故障诊断。首先,采用预训练好的ResNet18(Residual Network)作为特征提取器,并对每个残差块提取的特征计算MK-MMD(Multiple Kernel-Maximum Mean Discrepancy)距离,通过同时匹配高层和低层特征以有效匹配边缘分布差异。其次,每个残差块提取的特征都进入与之匹配的分类器中,通过Softmax层计算的预测概率分布,并转化为伪标签,缩小条件分布差异。最后,引入Adam优化器,对整体模型参数进行优化,加快模型训练,提高模型收敛速度。实验结果表明,所提出的方法能够有效提取可迁移特征,在负载变化的场景下达到了较高的诊断精度,并具有一定的泛化能力。

Bearing fault diagnosis is of great significance in practical industrial scenarios.Methods based on signal processing and machine learning often rely heavily on prior knowledge, and it is difficult to guarantee the effectiveness of feature extraction.Deep learning methods require the training set and test set to meet the same distribution, which is difficult to meet in industrial sites, greatly reduces model performance.A fault diagnosis method based on multi-layer domain adaptation is proposed, which can realize various types and degrees of bearing fault diagnosis.Firstly, pre-trained ResNet18(Residual Network) is used as a feature extractor, and the MK-MMD(Multiple Kernel-Maximum Mean Discrepancy) distance is calculated for the features extracted from each residual block.In this way, high-level and low-level features are simultaneously matched to effectively match the marginal distribution differences.Secondly, the features extracted by each residual block are entered into a matching classifier, and the predicted probability distribution calculated by the Softmax layer is converted into pseudo labels to reduce the conditional distribution differences.Finally, the Adam optimizer is introduced to optimize the overall model parameters, speed up model training, and improve model convergence speed.The experimental results show that the proposed method can effectively extract transferable features, achieve high diagnostic accuracy under load changing scenarios, and have a certain generalization ability.