This paper proposes a fault trace method based on thermal images to solve the difficulty of obtaining fault information, and environments that require diverse and large amounts of fault information, such as offshore wind farms. This paper applies this method to high-voltage die-casting transformers, and proposes a monitoring and diagnosis technology for winding short-circuit faults. The method of classifying thermal image faults is based on the convolutional neural network (CNN) deep learning architecture, and no longer distinguishes between feature extraction algorithms. Unlike classifiers, traditional machine learning methods are used to identify thermal images. The test results show that the fault trace method successfully simulates the fault image on the original thermal image, and the thermal image is presented in color and grayscale methods to train the model and identify performance, with an accuracy rate of over 99.8%, which can be used as a condition maintenance (CBM).