A deep learning reconstruction method based on 3D CNN for combustion flow field in a supersonic combustor

Accurately obtaining the 3D combustion flow field in a supersonic combustor is crucial for health monitoring and intelligent control of combined cycle engines. A deep learning reconstruction method based on 3D CNN for combustion flow field in a supersonic combustor is proposed. To establish the dataset for model training, numerical simulations are conducted on a typical combined cycle engine. Then the results are preprocessed to simulate the image characteristics of high-speed cameras in ground tests. Based on 3D convolutional neural networks (3D CNN), the single-path 3D CNN (3D SG-CNN) and the multi-path coupled with gradient information 3D CNN (3D MGI-CNN) are developed. The trained model successfully established the mapping relationship between the 3D combustion flow field and the 2D flame image. Meanwhile, due to the integration of gradient information, 3D MGI-CNN achieves higher reconstruction accuracy and superior flame boundary identification compared to other models. Additionally, to demonstrate the effectiveness of the proposed models in practical ground tests, reconstruction accuracy under noise disturbance is validated. The results confirm that the proposed models can maintain high accuracy even when the signal-to-noise ratio (SNR) is reduced to 10 dB. The results demonstrate robustness and significant practical application potential of the proposed models.