整体改进的基于支持向量机的故障诊断方法

Overall-improved fault diagnosis approach based on support vector machine

为了消除噪声或野值样本对支持向量机分类器推广性能的不利影响,从数据预处理、特征提取和分类器设计等几个方面对现有的基于支持向量机的故障诊断方法进行了整体改进。一方面,在独立分量分析的基础上提出一种残余总体相关分析时域特征提取方法,利用独立分量分析的冗余取消特性以及残余总体相关分析的整体约简能力,抽取描述不同故障模式类的典型低维特征,削减原始数据中的噪声干扰;另一方面,对各模式类特征样本进行模糊C-均值聚类,然后以类内平均距离和类间平均距离共同构建一个有效性判别准则,用于区分特征空间中的有效样本与野值点,去除野值对支持向量机目标函数的影响。在此基础上引入具有可控稀化解的前向最小平方近似支持向量机算法,并采用基于复杂多故障模式分级识别的二分类策略,共同形成一种整体改进的基于支持向量机的故障诊断方法。对齿轮箱故障的诊断结果验证了该方法的有效性,对于受强噪声干扰的小样本数据,所构建的故障分类器也具有良好的推广能力。

In order to eliminate the negative influence of noise or outlier samples on the generalization performance of the classifier based on support vector machine （SVM） , we make an overall improvement on current SVM based fault diagnosis approaches from several as- pects, such as data preprocessing, feature extraction, classifier design, and etc. On one hand, a time domain feature extraction method is proposed using residual total correlation analysis （RTC-A） based on independent component analysis （ICA）, in which the redundan- cy reduction characteristic of ICA and the overall reduction capability of RTC-A are used to extract the typical low-dimensional features describing different fault patterns, and reduce the noise interference in raw data. On the other hand, the fuzzy C-means clustering （FCM-C） of the feature samples of various fault pattern classes is performed. Then, an effectiveness judgment criteria is constructed using both the averaged inner-class distance and inter-class distance, which is used to distinguish the effective samples and outlier points in the feature space, and remove the influence of the outlier points on the objective function of SVM. On the basis above, the forward least squares approximation support vector machine （FLSA-SVM） with controllable sparse solutions is introduced, and the binary-classi- fication strategy of stepwise recognition based on complex fauh patterns is applied, with both of them an overall improved fault diagnosis approach based on SVM is developed. The diagnostic results for gearbox faults verify the effectiveness of the proposed approach. The de- rived fault classifier shows good generalization performance, even if the data set is a small sample with strong noise interference.