基于联合矩条件分布匹配的动车组轴承故障诊断方法

Fault Diagnosis Method for Electric Multiple Units Bearings Based on Joint Moment Conditional Distribution Matching

动车组轴承运行跨度大、工况复杂多变、安全要求高、数据分布存在差异的问题造成难以学习到高泛化性能的故障诊断模型。为了克服数据分布差异导致的模型泛化性能降低的问题,提高模型跨工况诊断能力,提出了一种基于联合矩条件分布匹配的领域自适应方法,用于动车组轴承的跨工况故障诊断。该方法的基本思想是通过提出的特征空间线性映射方法,将输入数据的类别信息耦合到提取的特征表示中,提高特征对数据特性的表达能力。进一步基于融合了条件分布信息的特征构造联合一阶和二阶统计量的指标(联合矩),增强对数据分布的描述,同时基于联合矩进行特征匹配以对齐领域分布,促使模型学习共享特征表示和诊断知识迁移。建立在两个公开滚动轴承数据集上的跨工况故障诊断案例的实验结果表明,所提方法相比其他方法具有更有效的故障特征学习能力和知识迁移性能,能够在多个任务上取得最佳的诊断结果。

It is difficult to learn a fault diagnosis model with high generalization performance due to the large operating span of electric multiple units bearings,complex and variable working conditions,high safety requirements and different data distribution.In order to overcome the problem of decreasing model generalization performance caused by the difference of data distribution and improve the ability of model cross-working condition diagnosis,a domain adaptive method based on joint moment condition distribution matching is proposed for cross-working condition fault diagnosis of electric multiple units bearings.The basic idea of this method is to couple the category information of input data into extracted feature representation through the linear mapping method of feature space,so as to improve the ability of feature to express data characteristics.Based on features that incorporate conditional distribution information,the index(joint moment) of the combined first and second order statistics is constructed to enhance the description of the data distribution.Meanwhile,the feature matching based on the joint moment is carried out to align the domain distribution,and promote the model to learn the shared feature representation and the transfer of diagnostic knowledge.The experimental results of cross-working fault diagnosis cases based on two public rolling bearing data sets show that the proposed method has more effective fault feature learning ability and knowledge transfer performance than other methods,and can obtain the best diagnosis results on multiple tasks.