The Analysis of Features Importance in Electrical Infrared Images Faults Diagnosis

This paper proposes an effort to study on input features and classifier about fault detection and diagnosis of infrared images within electrical installations and employ the results in fault detection robots with the purpose to improve the accuracy and be more convenient saving the waste of money and workforce. The features are extracted from infrared images of electrical equipments and classified by using random forest algorithm. In the experiments, the classification performances of various input features are evaluated. The commonly used indicators to describe the classification performance, including sensitivity(TPR), specificity(TNR), accuracy(ACC) and area under curve (AUC) are employed to identify the most suitable input feature as well as the best configuration of classifiers. The results of the experimental demonstrate that the combination of features including Skewness, Max, Kurtosis, 0.95 percentile, Gradient Direction Histogram, Max-Min, 0.75 percentile and other features can result in the best effect for infrared images fault diagnosis. In addition, the Random Forest performs better than the support vector machine(SVM) using radial basis kernel function (RBF) or gaussian kernel function. At the end of the paper, we further discuss how those features effect the fault diagnosis with infrared images. And as for employing the results in fault detection robots, it shows a considerably good effect in terms of accuracy and convenience compared with some traditional methods. It endows those robots with the abilities to make a real time monitoring and detect the faults.