Spectroscopy and Spectral Analysis, Volume. 45, Issue 2, 532(2025)
A Generalized Combustion Characteristic Prediction Model Based on Flame Hyperspectral Technology
Accurately measuring combustion characteristics can provide important references for optimizing combustion technology. This study predicts multiple combustion properties simultaneously using high spectral images of methane flames, addressing the challenge of monitoring combustion characteristics in engineering, which typically requires various collection devices and on-site sampling. A flame characteristic prediction model based on the 3D-CNN method is proposed. Initially, this method is validated on laminar flame high spectral images and further applied to turbulent flame high spectral images. Subsequently, a universal model and training method are established. Finally, the Grad-CAM method is used to conduct interpretability studies of the model. In the laminar flow model, we investigated the influence of different parameters on the accuracy of flame characteristic prediction models. In the turbulent flow model, we explored the feasibility of using transfer learning and the impact of adding wavelet denoising techniques and altering the number of convolutional layers on the model results. Within the general model, we addressed the issue of inconsistent labels in different datasets, enabling the blending of laminar and turbulent flame data for mixed training of the model. This facilitated the establishment of a combustion characteristic prediction model, ranging from laboratory-scale flames to real-world engineering flames. In the study of model interpretability, the Grad-CAM method was employed to investigate the operational principles and decision-making processes of neural network models. The results indicate that the average accuracy of the laminar flow prediction model for different predictive parameters exceeds 96%, with the Mish+Mae parameter combination achieving the highest average accuracy of 98% and demonstrating enhanced model stability. The average accuracy in the turbulent flow prediction model is consistently above 95%. Transfer learning, wavelet denoising, and altering the number of convolutional layers all contribute to improving model performance to some extent, contingent upon appropriate adjustments. The general model achieves an average of 97% and 94% accuracy for laminar and turbulent parameter predictions, respectively, while exhibiting higher robustness. In the interpretability study, information from Grad-CAM heatmaps suggests that this method can effectively identify flame-sensitive regions, extract crucial component information from spectra, and accomplish flame characteristic predictions. This study supports the intelligent development of experimental testing techniques in combustion science, holding significant implications for investigating combustion characteristics.
Get Citation
Copy Citation Text
WANG Ming, HE Hong-juan, WANG Bao-rui, AI Yu-hua, WANG Yue. A Generalized Combustion Characteristic Prediction Model Based on Flame Hyperspectral Technology[J]. Spectroscopy and Spectral Analysis, 2025, 45(2): 532
Received: Jan. 22, 2024
Accepted: May. 21, 2025
Published Online: May. 21, 2025
The Author Email: WANG Bao-rui (brwang@iet.cn)