Future wide-speed-range scramjet operation will involve complex flame structures across different combustion states, including ignition and blowout. Accurately monitoring the evolution of the flame structure is essential for efficient combustion control and the prevention of unexpected blowout. This study proposes a squeeze-and-excitation reconstruction convolutional neural network (SERCNN) model. The model has been demonstrated to be capable of reconstructing precise flame structures from sparse photoelectric signals. The optical dataset from the scramjet combustor equipped with a strut/cavity flameholder was utilized in the training of this model. The dataset contains eight typical combustion stages ranging from ignition to self-sustained combustion to lean blowout, and encompasses multiple combustion modes including intensive combustion, weak combustion, flame flashback, and flame lift-off. This variety of modes and complex flame structures allows the accuracy of the model to be evaluated under conditions that closely resemble real engine transients. Compared with the existing reconstruction models, the SERCNN model achieves superior performance over the entire test set. Moreover, simply expanding the field of view of each photoelectric sensor further improves reconstruction accuracy. With this optimized input form, the average reconstruction metrics reach the linear correlation coefficient of 0.9302, structural similarity index of 0.7792, and peak signal-to-noise ratio of 24.1641. The reconstructed fields are also used to extract the evolution of one-dimensional flame intensity and flame centroid during unsteady combustion, demonstrating the model's utility for analyzing dynamic flame behavior. These results indicate that the proposed SERCNN-based approach offers an effective, lightweight strategy for monitoring and diagnosing complex combustion states in scramjet engines.
Gao et al. (Sun,) studied this question.