This paper proposes a new intelligent fault diagnosis approach based on multimodal deep learning to fuse vibration and current signals to diagnose wind turbine gearbox faults. The proposed method typically consists of modality-specific feature learning network and feature fusion network, specifically based on a popular deep learning model named deep belief networks (DBNs). First, two individual DBNs are designed to learn fault-related features directly from raw vibration signals and current signals, respectively. Then, the learned vibration-based features and current-based features are further fused by a third DBN to output the final diagnosis results. The proposed approach is verified on a wind turbine drivetrain gearbox test rig. The experimental results demonstrate that the proposed approach outperformed the compared methods based on single sensor and data-level fusion in terms of diagnostic accuracy, which attributes to the complementary diagnosis information from vibration signals and current signals.