In the past few years, image recognition and computer vision with deep learning have evolved rapidly to become a commercial application that assists traditional human decision making with the engineering of data professionals. However, there are a number of problems with existing deep learning. Deep learning is a black box method that suffers from over-fitting problems more seriously than traditional learning methods and cannot be used to understand the decision-making process in the middle layer. It is used skeptically for applications that require reliability and transparency such as safety. The recent generation model gives a user understanding of the data manifold and is emerging as a new unsupervised learning method. In particular, we focus on the unsupervised learning of these generation models and propose experiments that can be used to detect abnormalities in power distribution facilities. We propose techniques to mitigate dataset shortages by learning and analyzing dataset in power environments with extreme shortages of abnormal dataset, and providing an assessment of the abnormal dataset. The module is still experimental and has no results. But in the future, it is expected to be used as a convergence technology of the power industry and artificial intelligence.