Statistical analysis of infrared thermogram for CNN-based electrical equipment identification methods

It is essential to develop infrared (IR) thermogram identification technologies to establish automatic diagnosis systems in power substations. The convolutional neural network (CNN) based methods show the highest accuracy in this field. The IR thermograms of electrical equipment are very different from general digital images, which means the present methods need further improvements. For data-driven CNN methods, it is necessary to study the characteristics of the IR data. This paper collected 11817 thermograms from substations and structured the dataset according to equipment types. The statistical features of mean, variance, skewness, kurtosis and contrast are analyzed and compared with other five image datasets. Several tricks are revealed from the analysis and tested on CNN models. Firstly, greycaling the Iron pseudo-color images extracts the temperature information and makes it possible to design models with fewer channels. The test shows it could reduce over 35% computational costs. Secondly, the sparse information of color and edges of thermograms makes it necessary to keep the original aspect ratio. The image preprocessing method of cropping shows better performance than padding and rescaling. Thirdly, the 0-1 normalization can boost the training process for about 100 epochs, which is related to the particular background of thermograms.