A novel intelligent fault diagnosis method for electrical equipment using infrared thermography

• Statistical characteristics are extracted using K-means algorithm.
• A coarse-to-fine parameter optimization approach of SVM is proposed.
• Seven sets of features are used as input data for the SVM classifier separately.
• The classification performance of SVM is compared with that of BP neural network.

Infrared thermography (IRT) has taken a very important role in monitoring and inspecting thermal defects of electrical equipment without shutting down, which has important significance for the stability of power systems. It has many advantages such as non-contact detection, freedom from electromagnetic interference, safety, reliability and providing large inspection coverage. Manual analysis of infrared images for detecting defects and classifying the status of equipment may take a lot of time and efforts, and may also lead to incorrect diagnosis results. To avoid the lack of manual analysis of infrared images, many intelligent fault diagnosis methods for electrical equipment are proposed, but there are two difficulties when using these methods: one is to find the region of interest, another is to extract features which can represent the condition of electrical equipment, as it is difficult to segment infrared images due to their over-centralized distributions and low intensity contrasts, which are quite different from those in visual light images. In this paper, a new intelligent diagnosis method for classification different conditions of electrical equipment using data obtained from infrared images is presented. In the first stage of our method, an infrared image of electrical equipment is clustered using K-means algorithm, then statistical characteristics containing temperature and area information are extracted in each region. In the second stage, in order to select the salient features which can better represent the condition of electrical equipment, some or all statistical characteristics from each region are combined as input data for support vector machine (SVM) classifier. To improve the classification performance of SVM, a coarse-to-fine parameter optimization approach is adopted. The performance of SVM is compared with that of back propagation neural network. The comparison results show that our method can achieve a better performance with accuracy 97.8495%.