Abstract: To address the issues of insufficient information from a single signal and the difficulty of effectively fusing multi-source heterogeneous features in mechanical fault diagnosis of high-voltage circuit breakers, a diagnostic method combining multi-source vibration-current feature fusion and an improved residual network is proposed. Variational mode decomposition and multi-scale dispersion entropy are employed to extract mechanical features from vibration signals, while temporal and amplitude features are extracted from the coil current waveform. One-dimensional feature vectors are mapped into two-dimensional images using a padding repetition method, and the current, vibration, and their pixel-wise average are encoded into the RGB three channels. An improved residual network embedded with a Squeeze-and-Excitation channel attention mechanism (SE-ResNet18) is constructed. Through five-fold cross-validation, the proposed method achieves an average recognition accuracy of 98.3% ± 0.5% on six types of faults, significantly outperforming single-signal and conventional fusion methods.