In response to the problem that single sensor data is easily affected and interfered with by the environment, making it difficult to collect accurate gearbox fault signals, resulting in low fault recognition accuracy, this paper proposes a multi-sensor self-learning weighted fusion method. This method not only effectively improves the accuracy of fault diagnosis but also has good interpretability. Firstly, the signals of different sensors are transformed into time–frequency images through sliding window and continuous wavelet transform to improve the signal to noise ratio; In order to fully extract fault features from the time–frequency images, a feature extraction module is constructed using model transfer method for feature extraction; Then, the self-learning weighted fusion module amplifies the fault features in the feature matrix while suppressing interference features to improve the accuracy of gearbox fault diagnosis; Finally, the classification is carried out through the classification module. The average accuracy of this method under different operating conditions on the gearbox dataset of Southeast University and the WT planetary gearbox dataset of Beijing University of Technology is higher than 99.9%.