采用规范化局部保持投影的轴承故障诊断方法

A Novel Diagnostic Approach for Bearing Faults Using Regularized Locally Maintaining Projections

针对轴承故障检测中特征融合导致的维度高、相关性强等问题,提出一种采用规范化局部保持投影算法(LPP)的轴承故障诊断(En-LPP)方法。首先,采用熵规范化的方法将相似度矩阵结合到传统LPP算法的优化函数中,与投影向量一并求解,得到一种规范化LPP降维算法;然后对原始轴承振动信号进行小波变换和经验模式分解得到10条信号分量,每个分量通过计算平均值、均方根等,提取12维统计特征,经归一化后生成特征向量;然后将特征向量输入到规范化LPP降维算法中进行迭代共同求解,得到满足终止条件的相似度矩阵和投影向量;最后利用降维后的特征集训练极限学习机模型确定轴承最终工作状态以实现故障检测。实验结果表明:与传统LPP方法以及其他降维方法相比,所提出的En-LPP方法对于轴承故障诊断的性能更好;在小波变换72维特征集合以及经验模式分解48维特征集合下的分类精度平均提升了7%以上;在4种不同分类器组合下的分类精度平均提升了17%以上;较好的降维特征区分能力使得En-LPP方法的故障诊断性能在不同条件组合下均具有很好的鲁棒性。

A novel diagnostic approach for bearing faults using regularized locally maintaining projections (En-LPP) is proposed to solve the problems of high dimension and strong relevance caused by feature fusion in bearing fault detection domains. Firstly, the entropy normalization method is used to combine a similarity matrix into the optimization function of the traditional locality preserving projections (LPP) algorithm, and to solve it along with the projection vector to obtain a regularized LPP algorithm with reduced dimension. Then, the original bearing vibration signal is decomposed into ten signal components through wavelet transform and empirical mode decomposition, and 12-dimensional statistical features for each component are extracted. Final high-dimensional feature vectors are generated by combining and normalizing these statistical features of all components. These feature vectors are then inputted into the proposed algorithm, and iteration is performed to obtain the similarity matrix and the projection vector that satisfy termination conditions. Finally, the reduced dimension feature set is used to train an extreme learning machine model for determining the final working state of a bearing to achieve fault detection. Experimental results and comparisons with the traditional LPP algorithm and other dimension reduction algorithms show that the proposed En-LPP algorithm has better performance for bearing fault diagnosis. The classification accuracy is improved by more than 7% on average under the 72-dimensional feature set of the wavelet transform and 48-dimensional feature set of empirical mode decomposition, and the averaged classification accuracy of the regularized LPP algorithm with combinations of 4 classifiers is about 17% higher than those of other dimension reduction algorithms. The proposed approach has better ability to distinguish dimension-reduced features, which makes its fault diagnosis performance robust under different combinations of conditions.