Fault diagnosis method of planetary gearbox based on MobileNetV3-AHFF and MS-HNNE models with fused time-frequency transformations

To address the limitations of existing specific time-frequency transformation methods in extracting complex dynamic features from vibration signals, as well as the feature selection bias in the channel attention mechanism and the improper pooling strategy design in the traditional MobileNetV3 network leading to information loss, a planetary gearbox fault diagnosis method based on MobileNetV3-AHFF and MS-HNNE (Mahalanobis distance hierarchical nearest neighbor graph embedding for efficient dimensionality reduction) with fused time-frequency transformations was proposed. By integrating short-time Fourier transform, continuous wavelet transform, and chirplet transform image coding techniques, the vibration signals of the planetary gearbox were transformed into multidimensional time-frequency images, which were then fused to construct a comprehensive feature image representing the signal characteristics. An adaptive hierarchical feature fusion (AHFF) module was designed to enhance the representational ability of the deep learning network. The supervised MS-HNNE algorithm was used to replace the pooling layer before the fully connected layer of MobileNetV3. During the process of dimensionality reduction, the internal structure and key information of the data were retained. The Softmax function was used to complete the classification task for the low-dimensional data. Compared with the existing fault diagnosis models, experimental results from the DDS (drivetrain diagnostics simulator) and Southeast University planetary gearbox fault diagnosis showed that this method not only significantly improved the diagnostic accuracy, but also enhanced the model generalization ability. Its highest diagnostic accuracy reached 99.9% with certain application prospects.