样本不充分条件下基于支持向量机的故障诊断方法研究

Support vector machines (SVM) have been used frequently as tools in the fault diagnosis field, where they can avoid the need for precise mathematical modeling. Instead, SVM-based approaches make accurate diagnoses by using classifiers for all possible faults, which are constructed by learning based on high quality fault samples. However, the acquisition of ideal sample sets is extremely difficult. Most of the samples acquired in practical conditions have a normal status and only a small proportion of these samples have a faulty status. Moreover, these faulty samples are usually related to common faults, whereas rare faults, concurrent faults, and multi-faults are scarce. To address this problem, the current project will attempt to improve the diagnostic accuracy in conditions where insufficient training samples are available. The major work includes:.1) If the quantities of normal and faulty samples are imbalanced, there may be a drift in the classification hyper-plane, which could lead to misclassification during diagnosis. This project will aim to improve the diagnostic accuracy by specifying efficient penalty factors for the two types of samples to limit any drift..2) An advanced framework for diagnosis will be constructed to cope with situations where some types of faults are missing. The framework will be based on a single-class SVM classifier..3) The intrinsic relationship between single faults and concurrent faults will be determined by mapping them into a higher dimension using a SVM. This will provide the basis for the diagnosis of multi-faults when only single fault samples are available.

支持向量机是故障诊断经典工具之一，基于支持向量机的故障诊断无需精确机理建模，只需对高质量的故障数据集进行学习，构建出全部故障类别的分类器即可实现准确诊断。然而，理想数据集的获取异常困难，在工程实际中采集到的海量数据绝大多数均为正常数据，故障数据偏少；同时，采集到的故障数据中，往往只有常见故障的数据，罕见故障及多部件并发故障情况的样本缺乏。针对上述问题，项目拟对样本不充分条件下基于支持向量机的故障诊断展开提效探讨。包括：1）正常样本与故障样本数量相差较大时，分类超平面会发生漂移，进而引起误判，项目从惩罚因子加权角度出发对该问题进行攻关，抑制分类超平面漂移，进而提高诊断准确率；2）部分故障类别样本缺失情况下，构建基于SVM单分类器的新型决策流程，实现有效诊断；3）探讨单故障与并发故障样本在高维空间的位置关系，为仅有单故障样本条件下开展并发故障的诊断提供依据。

如何在工业大数据的背景下开展有效的故障诊断是当前工业领域的一个前沿热点，而支持向量机则是该类方法中最常用的技术手段。基于支持向量机的故障诊断无需精确机理建模，只需对高质量的故障数据集进行学习，构建出全部故障类别的分类器即可实现准确诊断。然而，理想数据集的获取异常困难，在工程实际中采集到的海量数据绝大多数均为正常数据，故障数据偏少；同时，采集到的故障数据中，往往只有常见故障的数据，罕见故障及多部件并发故障情况的样本缺乏。针对上述问题，项目针对样本不充分条件下基于支持向量机的故障诊断展开提效探讨。包括：1）正常样本与故障样本数量相差较大时，分类超平面会发生漂移，进而引起误判，如何在样本量不对称时，有效抑制分类超平面漂移，进而提高诊断准确率；2）部分故障类别样本缺失情况下，构建基于SVM单分类器的新型决策流程，实现有效诊断；3）探讨单故障与并发故障样本在高维空间的位置关系，探讨有效的故障特征提取方法，便于提高诊断准确率。相关研究目前已发表或收录论文12篇，出版专著1本，申请专利7项。相关技术被用于光伏发电关键部件的故障诊断，效果良好，并通过第三方软件测试。