In the field of industrial automation and intelligent monitoring, real-time detection of substation faults using advanced image classification techniques has important research and application value. A convolutional neural network model called LKTrans-Net is designed to solve this problem by enhancing detection efficiency and accuracy through sophisticated image classification methods. The model adopts a large size 31 × 31 convolutional kernel, which breaks through the limitation of the traditional 3 × 3 convolutional kernel, and optimizes the large-size deep convolutional layer by the re-parameterization technique to achieve high-performance image feature extraction. LKTrans-Net achieves a top-1 accuracy of 83.07% on the substation fault dataset. The experimental results show that a large convolutional kernel helps to capture more spatial information to improve the accuracy, but it also increases the computational load. LKTrans-Net strikes a good balance between accuracy and computational cost, which is suitable for the substation fault detection task that requires high accuracy, and provides an effective solution in the field of intelligent detection.