基于半监督LLTSA维数约简的故障诊断

FAULT DIAGNOSIS BASED ON SEMI-SUPERVISED LLTSA FOR DIMENSION REDUCTION

线性局部切空间排列(LLTSA)为无监督的维数约简方法,在对高维故障特征集进行维数约简时,不能利用部分样本的类别标签信息,使得获得的低维特征仍出现混叠的情况。针对这个问题,提出了半监督线性局部切空间排列(SS-LLTSA)的维数约简方法,即利用部分标签信息来调整样本点与点之间的距离以形成新的距离矩阵,通过新的距离矩阵进行邻域构建,实现了数据本质流行结构和类别标签信息的结合,能够提取区分度更好的低维特征。此外,还通过支持向量机(SVM)来建立低维特征与故障类别的对应关系,实现故障诊断。SS-LLTSA维数约简增强了故障特征的辨识能力,而SVM优异的模式识别能力能够进一步提高故障诊断精度。滚动轴承的故障诊断实例验证了所提故障诊断方法的有效性。

Linear local tangent space alignment （LLTSA） is an unsupervised dimension reduction method, which will lends to remaining overlaps between faults when it is used to high-dimension fault feature for dimension reduction due to its incapacity of using part sample class label information. Aiming at this problem, semi-supervised linear local tangent space alignment （ SS- LLTSA） dimension reduction method is proposed in this paper. In SS-LLTSA, the distance between different points is adjusted by utilizing part class label information, thereby a new distance matrix is formed and the neighborhood is construct through this new distance matrix. The improved method realized the combination of data intrinsic manifold structure and class label information, and more discriminative low-dimension features can been obtained. And then, the corresponding relationship between low-dimension feature and fault classes are established by using support vector machine （SVM）. Dimension reduction with SS-LLTSA can effectively increase the discrimination of fault feature, and furthermore, SVM can further improve fault diagnosis accuracy with its excellent pattern recognition capacity. Finally, the effectiveness of the proposed method was verified through the fault diagnosis experiment of rolling bearing.