The gearbox is a key component of the transmission system. Its condition directly affects the overall operating efficiency. However, the complex structure of the gearbox results in a coupled and variable transmission path for the vibration signals. This limits the representability of the fault information provided by a single sensor, potentially leading to misinterpretation of the gearbox condition. In particular, the fault information distributed at different measurement points has different sensitivities. Considering the limitations of single sensors in fault detection and the different richness and sensitivity of fault feature information of multiple sensors, a multi-sensor feature fusion network is proposed for gearbox fault diagnosis. Initially, with the implementation of Convolutional Neural Network (CNN), a series of CNN sub-models are respectively built to extract the high-dimensional features hidden in the multi-sensor signals. Subsequently, by combining ensemble learning and the Squeeze-and-Excitation (SE) attention mechanism, the fault features extracted from multiple sensor signals are fused. In this way, the critical features can be well integrated and enhanced. A gearbox transmission system is used to validate the effectiveness of the proposed method. Experimental results demonstrate that the proposed method effectively fuses multi-sensor signals and significantly improves diagnostic accuracy compared to single-sensor approaches.