2. An intelligent fault diagnosis method for electrical equipment using infrared images

Collecting pictures by portable thermal imager is the major form of IRT applied in electrical systems for on-site diagnosis or subsequent fault analysis. Most of IRT cameras today have their own analysis software and have the capability to prepare the inspection reports. However, the evaluation processes are very time-consuming as they are still given by manual. Furthermore, the analysis for a large electrical system with hundreds of equipment becomes more complicated, as image collecting need large amount of labor and it is not easy to collect images in rainy, windy or other bad weather. In [7] infrared images are collected by an intelligent robot and then transferred to the base station or control room. The usual way of estimating the condition of an equipment is to set a temperature threshold. When the actual temperature is higher than the threshold, an alarm is given that the device is in disorder. This method is quite simple but sometimes it is inaccurate as a simple threshold is inadequate in many times. Moreover, for a voltage heating equipment, the surface temperature is not very high even when it is under serious defects. Therefore, an intelligent diagnosis system is needed which can make a faster and more accurate decision. With the rapid development of computer vision techniques and artificial intelligence, some intelligent diagnosis systems for electrical equipment by analyzing infrared images automatically are constructed. They usually consist of three major steps as Fig.1 illustrates[8]. The first step is to find the region of interest(ROI), and then the feature information in the ROI is extracted, followed by the classification step at last. The key step of intelligent diagnosis systems is to find the ROI which decides whether the correct information can be extracted.

Currently, automatic inspection of IRT images is widely used in medical imaging, non-destructive testing, structure defect detection and so on. However, automatic diagnosis of IRT images employing intelligent systems is still in early stages in power systems due to the complex analysis and various factors that need to be considered in developing such a system. Many scholars at home and abroad have done a lot of work on infrared diagnosis of electrical equipment[8]. A. S. N. Huda et al. [9] proposed a semi-automatic approach for thermographic inspection of electrical installation within building. Initially the interest regions of the images are manually segmented and then the statistical features are extracted between hot and reference regions. Principle component analysis is then applied for the best features selection. At the final stage, the condition of electrical equipment is classified using a multilayered perceptron neural network. References [10] and [11] study contamination degree of insulator based on the basic steps of an intelligent diagnosis system. As their images are obtained in lab conditions and the backgrounds of these images are not very complex, they can easily find the ROI using a simple image segmentation algorithm. In Reference [12] remote images of equipment are captured and moment invariant of images are calculated as a characteristic vector and input to a support vector machine for equipment type recognition. Subtracting the former image from the current image is used to judge whether the running state is in order. This method is simple and effective, but the investment is huge. In [14] suitable features selection for monitoring thermal conditions of electrical equipment are studied. Three sets of features are extracted, namely, the first order histogram, the grey level co-occurrence matrix and the component based intensity. These features are then input to a neural network for classification. Binary image segmentation techniques are used to find ROI in references [13] and [14]. The problem with this image segmentation method is that the image trends to be over segmented due to its over-centralized intensity distribution and low intensity contrast.

From above we can see that the difficulties of intelligent fault diagnosis for electrical equipment based on infrared images are to find the accurate ROI and to extract features which can characterize the machine conditions. In this paper a new intelligent diagnosis system are proposed in order to overcome these difficulties. In the first stage of our method, an infrared image is clustered to five regions using K-means algorithm. Then temperature information in each region is extracted and combined as image features. Finally, these image features are input to a SVM classifier. To reinforce the classification performance, a parameter optimization approach is used.

The rest of this paper is arranged as follows. Image feature extracting based on K-means algorithm is presented in Section 2. Parameter optimization method of SVM is introduced in Section 3. Experimental results and data analysis are given in Section 4. Section 5 is the conclusions.

K-means is a clustering algorithm which is also named as K-average. The mean of all data in each cluster subset is the representative point of that cluster. The key idea of K-means algorithm is to divide dataset into separate categories by iterative process, aiming that the criterion function for evaluating clustering performance achieves its optimum. The commonly used criterion function is distance (e.g., Euclidean distance, Manhattan distance or Minkowski distance), error sum of squares, or mean of a cluster [16]. In this paper the last one is employed as our criterion function.

In our method, an infrared image without temperature scale is converted into a greyscale image firstly. Then this greyscale image is clustered using K-means algorithm. By comparing various numbers of clustering on many images, we find that 5 clustering can well reflect temperature difference in different parts in images. It is easy to find out the minimum, mean, medium and maximum of grey value in each cluster.

Feature extraction using K-means algorithm is described below:

1. Convert an infrared image into a greyscale image, and initialize cluster centers for 5 groups;

2. Cluster all points in the image according to the minimum distance to 5 groups;

3. Calculate the mean value of each cluster, and use it as the new cluster center;

4. Repeat step 2) and 3) until the cluster centers don't change;

5. Regions in each image are obtained, and the grey value from small to large are marked with black, red, green, blue and white respectively.

In this paper, a new intelligent diagnosis method using infrared images is proposed to classify four electrical equipment conditions, i.e., normal, general failure, serious failure and emergency breakdown. Our method consists of two subsequent procedures: image feature extraction using K-means algorithm and support vector machine for classification with parameter optimization from coarse to fine. From experiments we can see that our method can solve the problems of finding ROI and feature extraction in intelligent fault diagnosis for electrical equipment using infrared images. In addition, we can obtain temperature information in each region in infrared images and the white region represents the fault location when the power equipment is in failure.