Abstract: Considering the complexity and variability of the vibration signals from gearboxes under actual operating conditions, which contribute to low diagnostic accuracy and suboptimal training convergence in traditional convolutional neural network-based gearbox fault diagnosis methods, we propose an enhanced attention mechanism along with a soft threshold activation function derived from this mechanism. Building upon these innovations, we develop a convolutional neural network fault diagnosis model that incorporates the improved attention mechanism. Subsequently, through experiments utilizing data from an open-source gearbox dataset, we demonstrate that both the enhanced attention mechanism and the soft threshold activation function significantly enhance diagnostic accuracy and training convergence performance of the model. Finally, we apply this constructed model to real-world gearbox fault diagnosis scenarios. Results indicate that it fulfills enterprise factory inspection requirements while exhibiting superior diagnostic accuracy and training convergence.