Aiming to address the limitations of traditional power transformer fault diagnosis methods, which often rely on manual experience and lack generalization capability, this paper proposes an intelligent fault diagnosis model that integrates a residual network (ResNet) with an attention mechanism. Infrared thermal imaging technology is used to collect thermal distribution images from nine types of transformer conditions, including normal operation and eight levels of short-circuit faults. The ResNet-50 backbone network is employed to automatically extract multi-scale spatial features, while a dual attention mechanism—operating across both channel and spatial dimensions—is introduced to enhance the salience of fault regions. On a dataset of 180 samples, the proposed method achieves an average accuracy of 100%, outperforming both InceptionV3 and VGG16 models. Experimental results demonstrate that the model can effectively identify fault types, significantly reduce misjudgments caused by complex background interference, and offer a high-precision, low-latency intelligent solution for transformer condition monitoring.