Neural Network Assisted Simultaneous Diagnosis For Soot Multi-parameters Fields In Laminar Flames

Because of the complexity of the optical setup and the inefficiency of data processing in the traditional soot optical diagnosis method,this dissertation attempts to develop machine learning assisted optical diagnosis methods to achieve the purpose of real-time measurement of the actual power equipment in the future.On the one hand,three Compact-Modulated Absorption/Emission(CMAE)diagnosis methods are developed based on Modulated Absorption/Emission(MAE).On the other hand,the neural network model is used to assist optical diagnosis and reshape the traditional optical diagnosis.Contrasted with the original MAE technology,the white LED point light source or surface light source significantly simplifies the realization step of beam homogenization,and the prism camera can record the radiation of both bands of flame at the same time,which reduces the complexity of the detection system.The results show that in order to avoid the abnormal extinction at the edge of the flame,the backlight beam intensity should be more than 2.5 times of the flame radiation intensity.In addition,the robustness and accuracy of the three CMAE diagnosis methods are verified by standard Santoro flame numerically.Furthermore,by the error transfer theory,the measurement errors of soot volume fraction and temperature are calculated to be ±0.05 ppm and ±60 K respectively.It is worth pointing out that CMAE-2 and CMAE-3 can carry out experiments in a relatively harsh experimental environment(narrow experimental space,vibrating and dusty environment),and are expected to be candidates for the measurement of high-fidelity flame soot parameters under the limited space,weight and power supply conditions of Chinese Space Stations.In the aspect of machine learning-assisted optical diagnosis,two kinds of neural network models are mainly developed to assist the flame radiation field to simultaneously predict the soot multi-parameter field.The first model is Bayesian optimized back propagation neural network model(BPNN).Through the flame radiation field,the two-dimensional distribution of local soot volume fraction,temperature and primary particle diameter in laminar diffusion flame can be predicted simultaneously.The second one is the U-net convolution neural network model,which can also predict the local soot volume fraction and temperature field in the laminar diffusion flame through the flame radiation field.Theoretically,the training and prediction processes of the two models are simulated by adding Gaussian random noise to the numerical data,and the robustness of the two models is verified.Furthermore,based on the laminar flame soot experimental database,two models are trained,and the influence of the amount of training sample data on the prediction accuracy of the model is discussed.The feasibility and prediction accuracy of the two models are verified by comparison with the results of optical measurements.These works provide cornerstones for the application of diagnostic methods to the real-time measurement of practical power plants.At present,the machine learning-assisted optical diagnosis technology is thriving,which is promising to provide a more efficient,lower cost and high-fidelity method for the simultaneous diagnosis of multi-parameters of combustion and reaction flow.It also offers essential elements for the practical application of laboratory measurement technology towards industrialization.