Engine State Prediction Based on a Dynamic Model and Transfer Learning

In the actual diesel engine engineering scenario, fault data are scarce. Insufficient samples cause the model to over-memorize noise in the training data or details of specific samples rather than learning generalization features, resulting in overfitting. By making full use of the physical information network, a small sample state prediction method based on dynamic modeling and transfer learning is proposed. The virtual fault data generated by the dynamic model are highly similar to the physical fault data collected on the entity, so using the virtual data as the source domain can greatly improve the migration effect. Firstly, a high-fidelity internal combustion engine dynamics simulation model is established and solved, and enough virtual fault samples are obtained as the source domain for deep neural network model training. Secondly, based on the antinoise network, the TrAdaBoost method is applied to carry out sample transfer learning, and the input weights of the source and target samples are dynamically adjusted to obtain the optimal parameters of the model. Finally, the application examples of this method in fault diagnosis of internal combustion engine under two kinds of speed are given. Experimental results show the superiority and feasibility of the proposed method.