The global correlated features of signal data could be extracted in multi-head attention fault diagnosis methods, but the weight of strongly correlated features are usually diluted by weakly correlated features. A residual multi-head attention multilevel integrated network is proposed in this paper, to enhance the fault characterisation ability of strongly correlated features included in multimodal signal. Firstly, the current signal and vibration signal data are effectively fused, using a feature fusion technique based on the gating mechanism. Subsequently, the fused features are input into the residual multi-head attention network and the residual multi-head sparse attention network with different thresholds, respectively. Thirdly, the faults of the motor system are diagnosed using the residual multi-head attention network and the residual multi-head sparse attention network with different thresholds. The diagnostic results are then integrated to obtain the final result. The proposed network strengthens the strong correlation features, retains the influence of the weak correlation features, and thereby enhances the model’s fault characterisation ability. The feasibility of this method is demonstrated through comparative experiments with other models.