基于信息增益的SFT中故障影响因素降维方法研究

On the dimensional reduction method for the fault-affecting factors based on the information gain

为了研究空间故障树降维方法,参考因素空间中的信息增益法,构建了空间故障树中影响因素的降维方法。通过影响因素对目标因素的信息增益情况判断该因素被删除降维的可能性,分为可被忽略和可等同2种情况并给出了判断此2种情况的条件及降维方法的过程描述和涉及的定义。应用该方法分析了故障状态数据,设使用时间、使用温度和使用湿度为影响因素,元件故障概率为目标因素。分析结果表明,使用时间和使用湿度对故障概率的影响接近;使用温度对故障概率的影响最小,但3个因素都不能作为可降维的影响因素而被删除。

The present paper intends to construct the dimension reduction method for the influential factors by building up a space fault tree in the space fault tree system through the dimension reducing method by using the information gain method of the factor space. For the research purpose, we have built up the possibility to reduce the dimension of the information by taking into account the method of information gain. The situation of the information gain status-in-situ can be judged by using the two states, that is to say, those that can be neglected and those that can be equivalent. It is just on the basis of the aforementioned two states that it is possible to delete directly or to be substituted by. Therefore, the dimension reduction conditions can be summarized as known below： The paper gives the conditions to determine the above two states. Only when the information gain of the two influential factors on the one target factor is similar, can then one of the factors be replaced or directly deleted. The reduced dimension conditions are summarized as follows ： （ 1 ） It is possible to delete the factors that are potentially of little significance to the information gain for the target factor system; （2） It is possible to delete those factors that just have little influence on the information gain. And, next, we have provided the dimension reduction methods and the definitions involved, in which the steps of the dimension subtraction can be illustrated as follows： setting up the research domain; building up the background relationship; the corresponding table to be worked out to indicate the constitutive factor background relationship; the marginal distribution accounting factors; the analysis of the influential factors on the target; the table for the constructive condition distributions; the information gain situation of the influential factors, and the ways and execution steps for the dimension reduction. And, then, it is necessary to analyze the fault status data by using the above mentioned method. Briefly speaking, the influential factors may include the time, tempera- ture and the humidity to be taken, whereas the target factor should be the component fault probability, all of which can be joined into a synthetic analysis system through just one example. As a result, it would be possible to gain the degree and the distribution of the influence, too. Furthermore, it is also needed to make evaluation of the impact of the time and humidity to be consumed on the fault probability, though it shouldn＇ t be taken as the equivalent to the dimension reduction. Besides, the temperature taken has the lest influence on the fault probability evaluation, though it is improper to be neglected.