Transformers are critical components in metro traction power supply systems. Due to their complex structure and variable states, any malfunction can have wide-reaching impacts. Timely evaluation, fault diagnosis, and health prediction of transformers are crucial for developing effective maintenance strategies, reducing fault incidence, and enhancing the reliability of power systems. Traditional transformer fault diagnosis typically relies on gas data analysis. In the era of artificial intelligence, most fault predictions have employed classical machine learning algorithms. However, transformer faults are not solely indicated by gas data; relying on a single modality provides limited information, thereby restricting accuracy. Therefore, in the field of deep learning, this paper proposes the construction of a multimodal fault dataset and the application of multimodal fusion for transformer fault diagnosis. Experimental results demonstrate that the accuracy of multimodal information fusion surpasses that of traditional dissolved gas analysis and the SVM method in machine learning.