数模联合驱动的动态对抗自适应轴承故障诊断方法

A new simulation-data driven dynamic adversarial adaptive fault diagnosis method for bearings

针对现实工业场景下,故障数据样本稀缺,服役工况复杂导致的滚动轴承诊断准确率低下的问题,提出了一种数模联合驱动的动态对抗自适应轴承故障诊断方法。首先,提出用于快速产生具有明确时频域轴承故障特征的四自由度动力学仿真模型。随后,探讨了实测数据分布与动力学仿真信号之间的共性和差异性。最后,建立可以提取隐层域不变特征并自动对齐仿真源域数据、目标域待诊断数据分布的动态对抗自适应网络。设计滚动轴承故障诊断试验,探讨了由动力学模型产生带有标签信息的仿真信号、带标签信息的其他数据集实测信号与极少数带标签信息的待诊断实测信号构成的源域数据基础上,设定的3类任务中神经网络的诊断效果,完成了对大量无标签样本的分类识别。结果表明,仿真信号包含轴承故障的特征信息,可以对真实的轴承数据进行表征,并且所提出的动态对抗自适应网络相较于其他诊断方法能更准确实现轴承的故障诊断。同时,源域数据中包含极少数的带标签目标域数据可使得提出方法的识别准确率大幅提升。

A dynamic adversarial adaptive bearing fault diagnosis method driven by a combination of simulation and data was proposed to solve the problem of low diagnostic accuracy of rolling bearings due to the scarcity of fault data samples and complex service conditions in real industrial scenarios.Firstly,a four-degree-of-freedom dynamics simulation model was proposed for the rapid generation of bearing fault characteristics with a well-defined time-frequency domain.Subsequently,the commonalities and differences between the measured data distribution and the dynamics simulation signal were explored.Finally,a dynamic adversarial adaptive network that can extract invariant features in the hidden domain and automatically align the distribution of data in the source domain of the simulation and the data to be diagnosed in the target domain was developed.The fault diagnosis experiment of rolling bearing was designed,and the source domain data composed of the signal with label information generated by the dynamic model,the measured signal of other dataset with label information and a small amount of signal in target domain with label information.The diagnostic effect of the neural network in the three categories of tasks was discussed,and the classification and recognition of a large number of unlabeled samples were completed.Results show that the simulated signal contains the characteristic information of bearing faults,which can characterize the real bearing data,and the proposed dynamic confrontation adaptive network can realize the bearing fault diagnosis more accurately than other diagnosis methods.At the same time,the source domain data contains very few labeled target domain data,which can greatly improve the recognition accuracy of the proposed method.