基于Fisher Score与最大信息系数的齿轮箱故障特征选择方法

因果网络定向问题实质是一个“多对多”因果关系发现过程,传统的V-结构定向方法只能确定一组马尔可夫等价类而非最终的因果关系.为解决该问题,从柯氏复杂度的因果推断原理视角出发,利用贝叶斯链式法则推导出局部网络因果定向规则,并在此基础上提出高维全局网络因果定向方法.同时,将前者运用于改进基于局部条件独立信息搜索学习马尔可夫毯典型算法,后者运用于改进基于约束的因果网络结构学习典型算法.实验结果表明,改进后算法在保证较高准确率的同时可有效提升执行效率.

Fault feature selection method of gearbox based on Fisher Score and maximum information coefficient

Aiming at the problem that it is difficult to select multiple fault features of gearboxes in industrial environment, a new fault feature optimization selection method combining Fisher Score and maximum information coefficient(MIC) is proposed. First, considering about uneven distribution and overlapping of multi-fault features, the Fisher Score calculation method is used to construct the ranking rules of the importance of the feature indicators. Second, based on the impact of redundant features on the effective feature representation, the maximum information coefficient is used to update and rank redundant features. Then, taking classification accuracy as the judgement basis, using the support vector machine(SVM) theory, a fault feature optimization selection method combining Fisher Score and maximum information coefficient is established. Finally, the UCI standard data set and the gear failure simulation data set are used to verify the effectiveness and engineering practicability of the proposed algorithm. Comparative analysis of simulation experiments shows that compared with the traditional mRMR and reliefF methods, the number of feature subsets proposed is moderate and the accuracy is higher.