Flame Temperature Detection and Estimation Model Based on Deep Learning and Ordinary RGB Images

To solve the problems of traditional flame temperature accurate measurement that requires complex equipment, high cost, and difficulty in popularization, it proposes a method of rapid flame temperature monitoring and estimation based on deep learning. To establish a data training set based on RGB estimation of flame temperature, firstly, ordinary RGB and high-speed infrared cameras were used to capture the combustion process at the same time to obtain RGB images and corresponding temperature fields. Secondly, an adjustable modular network structure is established, which includes a video interpolation network, a flame detection network, and a flame temperature estimation network, and the switch training method is used to train the network. To prevent overfitting, a network training algorithm based on genetic algorithm is proposed, so that the training of the flame temperature estimation network is completed efficiently. Finally, the reliability of the calculation model is verified by a typical high-temperature combustion test of the solid propellant. The results show that the error between the calculated value and the measured temperature is only ±5.73%.