Transfer Learning-Based Fault Diagnosis Method for Marine Turbochargers

Abstract

To address the issues of the high cost of marine turbocharger fault simulation testing and the difficulties in obtaining fault sample data, a multi-body dynamics model of a marine turbocharger was developed. The simulation approach was used to acquire the turbocharger vibration signals. The result shows that the amplitude of the 1× vibration signal power spectrum drops as the bearing surface roughness increases. However, the amplitude of the 2× and 9× vibration signal power spectra increases as the roughness increases. The TrAdaBoost transfer learning method is used to develop a marine turbocharger diagnosis model. The validation results of 2040 simulated fault samples reveal that when the desired sample number is 20, the diagnostic model has an accuracy of 87%. When the desired number of samples is 40, the diagnostic model’s accuracy is 96%. The diagnosis model may perform diagnosis information transfer between the actual turbocharger and the simulation model.