Infrared image change detection of substation equipment in power system using random forest

Early detection of equipment faults plays a crucial role in power system, and automatic change detection of working status of an equipment is an efficient tool for this purpose. In this study, we proposed a novel method to automatically detect temperature change in local region of a substation equipment in power system using bi-temporal infrared images. We considered the change detection as two-class classification problem, and a supervised machine learning algorithm — Random Forest (RF) — was used for forecasting change trend. Various features were extracted from two temporal images for change detection. The features we extracted include gray-level, weighted intensity mean, RGB, LBP, gray-level histogram, and texture originating from the grayscale images and color images of the bi-temporal infrared images of the substation equipment. Cross validation was used to evaluate the robustness of these extracted features. Due to the existence of environmental noise, there are isolated detection points in the change detection results. In order to remove these isolated noise points and improve detection accuracy, we performed a morphological filtering on the detection results. Evaluation indexes such as Dice Similarity Index (DSI), kappa coefficient were used to evaluate the detection performance. Four classical change detection methods i.e. Image Differencing, Image Ratioing, Change vector analysis (CVA) and Principal Component Analysis (PCA) were tested for comparison purpose. Experimental results demonstrated that the proposed algorithm outperformed significantly these classical methods.