An attention-based architecture for predicting thermal stress on turbine blade film hole surface using partial temperature data

The thermal stress of film holes is crucial for the life of turbine blades with double-wall cooling systems. Finite Element Analysis (FEA) is widely used for thermal stress analysis, which requires the entire temperature field as input. However, only partial temperature data can be measured in engineering experiments, which is insufficient for FEA. This study proposed an attention-based deep learning architecture to map partial temperature data to film hole surface thermal stress. A thermal stress dataset with 300 samples of double-wall cooling units was constructed using conjugate heat transfer and FEA simulations and the thermal stress prediction model was trained using this dataset. Compared with the simulation results, the mean relative error (MRE) of the predicted results on the test dataset is 7.43%, and the MRE of peak thermal stress is 2.16%. The impact of the model input composed of different surface groups and sampling schemes on prediction performance was analyzed, and a reference for temperature measurement arrangement that balances accuracy and cost-effectiveness was proposed. Additionally, The impact of input noise on model performance was explored and the results demonstrate that the model trained with noise demonstrated significantly enhanced resistance to noise, and the MRE was about 8%.