레이저 분광 영상 기반 CNN을 이용한 고온로 온도 예측 연구

Abstract

In this manuscript, purpose to predict the internal temperature of a high-temperature furnace using a convolutional neural network (CNN). Experiments data was based on hyperspectral image, rather than the CT-TDLAS method using laser absorption spectroscopy and was constructed to the image of spectral bands by laser spectroscopy that passed through a high-temperature furnace. Through repetitive experiments, a total of 20,000 data were composed of the measurement range of temperature 25 ℃ to 800 ℃. Based on these data, the study was conducted by predict the temperature of spectroscopy image using CNN. Learning was conducted with data obtained by dividing the number of the output layer by 10 instead of 775. When learning the output layer divided into 10, the verification data showed 89.79% accuracy and the test data showed 88.73%. When the Gaussian sub-pixel interpolation was applied to make up for accuracy, the accuracy was 90.49%, it was improve by about 1.75%. When the number of output layers was set to 4, accuracy of the test data was the best, and it was confirmed that the optimal model could be configured by adjusting the number of output layers according to the data. Through these research results, the possibility of industrial application development of a measurement system using laser spectroscopic image was confirmed.