Research on temperature field prediction method in an aero-engine combustor with high generalization ability

Using the inlet flow parameters to get the temperature field of the aero-engine combustor can help researchers quickly learn about the combustion state of the combustor, which is essential to aero-engine combustor design and optimization. This study puts forward a fast-predicting scheme on the temperature distribution of aeroengine combustor by deep learning method. Different networks are trained to gain multiple predicting models, and the prediction performance of temperature field models under different dataset processing methods and different network structures are compared. The results show that both temperature field prediction models constructed by fully-connected networks and fusion convolutional networks have good predictive capabilities. However, when the equivalent ratio conditions deviate significantly from the training dataset, the model performance deteriorates seriously. By introducing reference data to process the dataset, the models' prediction ability for equivalent ratio conditions far from the training dataset is significantly improved. Further research shows that this data processing method has ability to extrapolate to some extent.

Automated summary

This study proposes a fast-predicting scheme for the temperature distribution of aeroengine combustor using deep learning. Different network structures are trained and compared to predict the temperature field under different dataset processing methods. The results show that fully-connected networks and fusion convolutional networks have good predictive capabilities. Introducing reference data to process the dataset significantly improves the models' prediction ability for equivalent ratio conditions far from the training dataset.