On-load tap changer (OLTC) is the only movable component in a DC converter transformer. Due to its frequent operations, mechanical faults are highly likely to occur. To achieve mechanical fault diagnosis of OLTC, a method combining Multimodal Fusion Technique (MFT), Dual-Branch Convolutional Neural Network (DBCNN), and K-Nearest Neighbor algorithm (KNN) is proposed. Firstly, Markov Transfer Field (MFT) and Fast Fourier Transform (FFT) are used to convert OLTC vibration signals from the time domain to Markov transfer relation graphs and frequency domain graphs. On this basis, DBSCNN is used to extract features of the two modalities, namely the temporal relationship and frequency domain, and fuse them. Finally, KNN is used to determine the similarity between features to improve classification accuracy. The results show that the proposed method has high accuracy and robustness.