基于属性约简的绝缘节破损预测

On the damage forecast for the insulation joints based on the study of the attribute reduction

准确的绝缘节破损预测能够保证铁路运输安全和经济效益。支持向量机算法能够处理轨道电路测试数据,对其进行分类,预测可能存在隐患的绝缘节,但支持向量机预测模型的原始样本多有冗余,基于此,提出了一种基于粗糙集和支持向量机的绝缘节破损预测模型。通过改进主分量启发式属性约简算法,降低样本维数,同时选用模拟退火算法完成SVM自动参数选优。实例分析和仿真结果表明,与单一支持向量机算法相比,属性约简后的粗糙集-支持向量机算法提高了分类器的分类性能,与采用网格搜索技术的SVM预测方法相比,模拟退火算法有效提高了SVM的预测精度。

In accordance with the track circuit testing data,this paper attempts to combine the rough sets theory with the support vector machine algorithm to make a forecast study for the possible existence of the so-called hidden insulation joints in the railway transportation.Since more detailed insulation joints should be able to better reflect the damaged degree or degrees of the insulation joints,it would be more convenient for on-site staff members to initiate more rational and timely co-operative response measures against the likely-resulted damage degrees according to the actual situation.The reason for this is that the rough sets theory is fit for reducing the dimensionality and selecting available features through the so-called attribute reduction.And,for reducing such damage degrees,we have to work out the so-called nine condition attributes,such as the track circuit,the limitation resistors,the power-delivery voltage,the rail joints,the broken trough,the rail voltage,insulation,etc.as the electrical features and the related influential factors.On the other hand,the fluctuations in the due functions are related to the states of insulation joints,too.Therefore,we have to choose the necessary working states,such as the decision attributes,for instance,the normal states,the slight damaged,medium damaged and seriously damaged states.It is just for such kinds of needs,information entropy is needed as the first step for discretizing the attributes for the discrete attribute data for the rough sets to process.At the same time,it is also necessary to get the core sets in terms of the concept of consistency degrees,on whose basis we can realize the optimal reduction set by using the improved principal component heuristic attribute reduction algorithm.And,next,forecast and classification can be done to support the vector machine algorithm after the reduction.And during the above said process,it would be possible to perform the SVM automatic parameter optimal selection in terms of simulated annealing algorithm.And,finally,by means of confusion matrix,calculation can be done to work out the accuracy,precision,as well as recall,whereas the prediction models can be compared and assessed through F-measure process.The theoretical analysis and simulation results we have gained demonstrate that the proposed SA-SVM algorithm process to be superior to the SVM in the prediction accuracy.The Accuracy,Precision,Recall and F-measure indicators of SVM classifiers have all been increased by 1%,1.41%,1% and 0.98%,respectively,through the above said reduction.