采用集约化开放集差异分布对齐策略的轴承故障诊断方法

Bearing Fault Diagnosis Method Based on Intensive Distribution Alignment with Open Set Difference

针对当前迁移学习算法应用于轴承故障诊断领域时,目标域中可能存在未知故障类别进而导致识别准确率过低的问题,引入了开集域自适应方法。针对现有的开放集域自适应算法在进行分布对齐时,多注重跨域对齐,很少关注领域内部的分布情况,从而导致对未知类别识别率较低的问题,提出了一种采用集约化开放集差异分布对齐(IDAOD)策略的轴承故障诊断方法,集成了跨域散度对齐和域内分布异化方法。使用开放集最近邻类验证方法获取目标域伪标签,然后构造源域和目标域的总体散度矩阵,进行跨域的散度对齐。基于分布适配加权条件分布,进一步异化同领域内不同类别的空间分布。在结构风险最小化框架下,基于开放集差异理论构造的损失函数,引入正则化项,得到最优解和新的目标域伪标签。通过在Office-31数据集上进行试验,验证了IDAOD算法的可行性。分别在CWRU和JNU轴承数据集上进行故障诊断试验,所提方法对未知故障类别的识别率均高于其他对比开放集算法,验证了所提方法在轴承数据集上能有效提高目标样本含有未知样本时的准确率。

In view of the facts that the current transfer learning algorithm was applied to the field of bearing fault diagnosis,there might be unknown fault categories in the target domain,which led to the problems of low recognition accuracy,and an open set domain adaptive method was introduced.Aiming at the problems that the existing open set domain adaptive algorithm paid more attentions to cross-domain alignment and paid little attention to the distribution within the domain when performing distribution alignment,which led to the low recognition rate of unknown categories.A method was proposed based on intensive distribution alignment with open set difference(IDAOD),which integrated cross-domain divergence alignment and intra-domain distribution alienation method.Firstly,the open set nearest neighbor class verification method was used to obtain the pseudo-label of the target domain.Then,the overall divergence matrix of the source domain and the target domain was constructed,and the cross-domain divergence alignment was performed.Based on the distribution adaptation weighted conditional distribution,the spatial distribution of different categories was further alienated in the same domain.Finally,under the framework of structural risk minimization,the loss function constructed was introduced based on the open set difference theory,and the regularization term was introduced to obtain the optimal solution and the new target domain pseudo-label.The feasibility of the IDAOD algorithm was verified by experiments on Office-31 dataset.The fault diagnosis experiments were carried out on CWRU and JNU bearing data sets respectively,and the recognition rate of unknown fault category is higher than that of other comparative open set algorithms.It is verified that the method proposed herein may effectively improve the accuracy of the bearing data set when target sample contains unknown samples.