Fault Diagnosis of Planetary Gearbox Based on Random Forest and Singular Value Difference Spectrum

Different fault modes of gears have unique data characteristics in the time domain waveforms of their vibration signals. However, the vibration signals of planetary gearbox with complex components always suffer from strong background noise, thus directly utilizing the original signals to identify the faults may produce large errors. In this study, a fault diagnosis method based on random forest and singular value difference spectrum is proposed. First, the singular value difference spectrum method is used to denoise the original signals of different fault modes, and then the reconstructed signal sequences are used as fault features for fault diagnosis. Second, increasing the data overlap degree can strengthen the model’s learning ability for the data characteristics of different fault modes. Finally, the bootstrap resampling method is adopted to randomly extract samples from the training set to train the parameter-optimized random forest model, and the test is performed on the gear transmission rig to identify the five different fault modes (healthy state, root crack, broken teeth, missing teeth and wear) of the planetary gear. The results indicate that random forest has a significantly improved accuracy and efficiency compared to the BP neural network, and the average recognition accuracy for the five fault modes can be reached up to 96.30%.