整体改进的基于支持向量机的故障诊断方法

In order to eliminate the adverse effects of noise or outlier samples on the generalization performance of the support vector machine classifier, the existing fault diagnosis methods based on support vector machines have been comprehensively improved from several aspects such as data preprocessing, feature extraction and classifier design. On the one hand, a time-domain feature extraction method of residual overall correlation analysis is proposed based on independent component analysis. By utilizing the redundancy cancellation characteristics of independent component analysis and the overall reduction ability of residual overall correlation analysis, typical low-dimensional features describing different fault mode classes are extracted to reduce the noise interference in the original data. On the other hand, the feature samples of each mode class are subjected to fuzzy C - means clustering, and then an effectiveness discrimination criterion is constructed jointly with the average intra-class distance and the average inter-class distance, which is used to distinguish the valid samples from the outlier points in the feature space and remove the influence of outliers on the objective function of the support vector machine. On this basis, a forward least squares approximate support vector machine algorithm with controllable sparse solutions is introduced, and a binary classification strategy based on hierarchical recognition of complex multiple fault patterns is adopted, jointly forming a comprehensively improved fault diagnosis method based on support vector machines. The diagnosis results of gearbox faults verify the effectiveness of this method. For small sample data affected by strong noise interference, the constructed fault classifier also has good generalization ability.

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一均值聚类，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。在此基础上引人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。