基于主元分析与支持向量机的制冷系统故障诊断方法  

Fault Diagnosis for Refrigeration Systems Based on Principal Component Analysis and Support Vector Machine

作　　者：韩华[1] 谷波[1] 任能 

机构地区：[1]上海交通大学制冷与低温工程研究所,上海200240 [2]江森自控楼宇设备科技(无锡)有限公司,江苏无锡214028

出　　处：《上海交通大学学报》2011年第9期1355-1361,1373,共8页Journal of Shanghai Jiaotong University

基　　金：国家自然科学基金资助项目(50876059)

摘　　要：利用主元分析（PCA）法提取制冷系统特征向量,对典型人工智能方法所建故障诊断模型的性能进行理论研究与应用分析,确定了以支持向量机（SVM）算法为基础的故障诊断模型;针对SVM直接解决多种分类问题的困难,分析了3种多类SVM算法,确定了基于＂一对其他（Onevsothers）＂多类SVM算法的故障诊断模型,并提出基于PCA与SVM组合的PCA-SVM故障诊断模型,同时,利用实验数据加以验证.结果表明：PCA-SVM模型可将16个原始变量转化为相互独立的主元,并可提取前4个主元用于故障诊断而将正常与故障的模式分离,对故障的诊断率不低于98.57%,优于单纯SVM模型,且PCA-SVM模型的训练速度比SVM模型快约130~350倍;PCA-SVM模型对小样本的处理能力优于BP神经网络模型,其诊断正确率较高,训练耗时较少（约1/240）.Principal component analysis（PCA） was employed to make feature extraction from the vast data pool,and the PCA＋SVM（support vector machine） model was established for the fault detection and diagnosis（FDD） of refrigeration systems.Considering that SVM can not be used to solve multi-class classification problems directly,several classical multi-SVMs algorithms were analyzed and compared,and the ＂One vs others＂ algorithm was adopted.The hybrid PCA-SVM FDD model was presented and validated by the historical data from specially designed experiments.The results show that it can isolate normal from faulty modes（detection） and has a diagnostic rate of no less than 98.57%,which is better than the SVM model without PCA;model training is about 130—350 times faster than the latter;it also has better performance on dealing with small sample problem than BP neural network with higher diagnostic rate and much less training time（about 1/240）.

关 键 词：制冷系统 故障诊断 主元分析 支持向量机 

分 类 号：TB66[一般工业技术—制冷工程]