基于深度学习的贫燃熄火预测方法和实验研究

Aimed at the flame beating, local extinction and reignition before the blow-off of lean burn swirling flames, a method was proposed to predict the lean burn blow-off of premixed flames based on deep learning. Deep network VGG19 was em-ployed to extract the key features of flames and the flame states were classified. The RGB flame images containing random noise generated by vid2vid model were used to verify the robustness and generalization performance of beating flame's state classification(BFC) model. Through the quantification of flame stagnation degree, more details are added to the classification results of flame states, thus reducing the misjudgment of the flame state. In addition, the future frame state of the oscillating flame before the blow-off is predicted from the angle of video prediction to enhance the understanding of near flameout. The results show that the recognition accuracy of the BFC model is 95.44%, the F_1-score is 94.54% and the average prediction accuracy of the future 2ms frame state of the flame is also over 84%. Therefore, the method can achieve a reliable prediction of flame state classification and future state.