Non-ferrous metals are very important strategic resources, electrolysis is an essential step in refining non-ferrous metals. In the electrolysis process, plate short circuit is the most common fault, which seriously affects output and energy consumption. The rapid and accurate detection of faulty plates is of great significance to the metal refining process. Given the weak generalization ability and complex feature rule design of traditional object detection algorithms, and the poor detection effect of existing deep learning models in infrared images with many interference factors, an improved Mask R-CNN based fault detection algorithm is--proposed. Improve the generation strategy and the NMS algorithm of proposals to reduce the missed detection; propose the Globally Generalized IoU loss function to characterize better the position and scale relationship between the predicted box and the target box, which is beneficial to the bounding box regression. The experimental results show that the improved model has an accuracy rate of 10.4% higher than the original model, reaching 86.8%. Compared with the common one-stage and two-stage object detection models, the improved model has a stronger detection ability. This algorithm has some reference value for the accurate detection and location of electrolytic cell faults.