Noise-Suppressed Temperature Measurement Based on Machine Learning in a Scramjet Combustor

Abstract

An algorithm of noncontact portable temperature measurement sensors that can be used for scramjet combustion chamber measurement is developed. The two-wire colorimetric temperature measurement is analyzed and found unable to be applied in the scramjet combustor due to its weak noise resistance. Then the inversion capabilities of the OH (hydroxyl) radical emission spectrum database of the multilayer perceptron and the convolutional neural network are compared. Considering the influence of noise, the influence of adding different proportions of random Gaussian noise on the network prediction results is compared. After adding 5% random Gaussian noise to the convolutional neural network, the regression error of the temperature prediction in the range of 500–4000 K is less than 50 K. As a method verification, the experiment of the ground combustion chamber of the scramjet under $M=6$ working condition is processed. Compared with the tunable diode laser absorption spectroscopy measurement result, the measurement temperature of the convolutional neural network is about 400 K higher, and the root mean square error is close to the measurement result of tunable diode laser absorption spectroscopy.