基于一维卷积迁移学习的跨工况机床轴承故障诊断

滚动轴承作为机床的重要核心零件，对保证机床的正常运转至关重要。然而在实际工作中，机床的工况经常根据不同的工作要求产生相应的变化，对机床轴承的转速以及负载产生一定的影响，从而导致轴承的机械振动信号呈现出非平稳性、非线性和非周期性等特点。目前基于深度学习的轴承故障诊断方法对数据具有一定的依赖性，要求训练（源域）和测试（目标域）数据集具有相同的数据特征且存在足够多的带有故障信息的标签数据。然而，由于机床常在非平稳工况下运行，因此在某一工况上建立的训练模型无法直接用于其他工况。为了解决这一问题，基于迁移学习（TL）技术，设计一维卷积神经网络（1-DCNN）与迁移学习相结合的模型。该模型利用一维卷积网络直接从原始振动信号中提取故障特征信息，并利用对抗策略迁移技术提取两域的公共特征。利用域分布差异度量拉近两域的特征分布，实现轴承跨工况迁移故障诊断。最后通过构建的12组迁移任务对比实验，验证所设计模型的优越性。结果表明：设计的基于一维卷积的迁移学习神经网络模型可直接实现对机床轴承故障的实时监测；设计的模型通过结合对抗策略迁移与度量域分布差异两种迁移策略，大大提高了迁移故障诊断性能，可更好地提取源域与目标域的公共特征；在实验构建的12组迁移任务中优于其余两种迁移策略，能完美完成迁移故障诊断任务。

One-Dimensional Convolutional Migration Learning-Based Bearing Fault Diagnosis for Cross-Condition CNC Machine Tools

As an important core part of machine tools,rolling bearings plays an important role in ensuring the normal operation of machine tools.However,in the actual working environment,the working condition of machine tools often changes according to different working requirements,which will have a certain impact on the speed and load of machine tool bearings,resulting in the mechanical vibration signal of the bearing showing non-stationarity,nonlinear and non-periodic characteristics.At present,the bearing fault diagnosis methods based on deep learning are data-dependent,requiring the training (source domain) and test (target domain) data sets to have the same data characteristics and the presence of sufficient labeled data with fault information.However,since machine tools operate under non-stationary conditions,the training model built on one condition cannot be directly applied to another condition.In order to solve this problem,based on transfer learning (TL) technology,a model combining one-dimensional convolutional neural network (1-DCNN) and transfer learning was designed.In this model,one-dimensional convolutional network was used to extract the fault feature information directly from the original vibration signal,and the common features of the two domains were extracted by adversarial strategy migration.The difference measurement of domain distribution was used to narrow the feature distribution of the two domains and realize the fault diagnosis of bearing migration across working conditions.Finally,12 groups of migration tasks were constructed to verify the superiority of the designed model.The results show that the designed transfer learning neural network model based on one-dimensional convolution can directly achieve real-time monitoring of machine tool bearing faults.The designed model greatly improves the performance of migration fault diagnosis by combining adversarial strategy transfer and measurement domain distribution difference transfer strategies,which can better extract common features of the source and target domains.Among the 12 migration tasks constructed in the experiment,it is superior to the other two migration strategies and can perfectly complete the migration fault diagnosis task.