针对变工况环境下采集到的滚动轴承振动数据特征分布不一致及待诊断样本标签难获取,导致轴承故障难诊断的问题,提出一种基于特征解纠缠和联合域对齐的滚动轴承多源域迁移诊断方法.首先,为更好提取源域和目标域的通用特征,利用卷积自编码器和正交约束实现域共享特征和域私有特征的解纠缠,筛除域私有特征并保留域共享特征进行域间对齐;其次,为缩小源域与目标域间的特征分布差异,采用多核最大均值差异(multiple kernel maximum mean discrepancy,MK-MMD)和相关对齐(correlation alignment,CORAL)方法构建融合度量准则;最后,为避免多源域差异带来的负面影响导致诊断精度下降的问题,采用源对抗模块和迁移对抗模块实现源域间及源域与目标域间的域混淆增强,并采用协同决策方式进行特征加权融合,降低弱相关域特征的干扰,实现最终的故障诊断识别.通过两种跨工况下的滚动轴承故障数据集对所提方法开展试验验证,并与单源域诊断方法及其它多源域诊断方法进行了对比分析,证明了所提方法的有效性和优越性.
Multi-source domain transfer diagnosis method for rolling bearings based on features disentanglement and joint domain alignment
Here,aiming at problems of inconsistent feature distribution of vibration data collected under variable operating conditions and difficulty in obtaining labels for samples to be diagnosed to make bearing fault diagnosis uneasy,a multi-source domain transfer diagnosis method for rolling bearings was proposed based on features disentanglement and joint domain alignment.Firstly,to better extract common features of source and target domains,convolutional autoencoders and orthogonal constraints were used to disentangle domain shared features and domain private features,filter out domain private features,and preserve domain shared features for inter-domain alignment.Secondly,to reduce difference in feature distribution between source domain and target domain,a fusion metric criterion was constructed using multiple kernel maximum mean discrepancy(MK-MMD)and correlation alignment(CORAL)method.Finally,to avoid negative effects of multi-source domain differences on diagnosis accuracy,source adversarial and transfer adversarial modules were used to enhance domain confusion between source domains and between source and target domains.Collaborative decision-making was employed for feature weighted fusion to reduce interference of weakly correlated domain features and realize final fault diagnosis and recognition.Test verification was performed for the proposed method with rolling bearing fault data sets under 2 cross-operating conditions.Contrastive analyses were performed for the proposed method with single-source domain diagnosis method and other multi-source domain diagnosis methods,respectively to demonstrate its effectiveness and superiority.
fault diagnosismulti-source domain transfer learningfeatures disentanglementjoint domain alignment