基于机器学习的乙烯火焰温度和炭黑体积分数反演研究

Research on Retrieval of Ethylene Flame Temperature and Soot Volume Fraction Based on Machine Learning

对于碳氢火焰,从辐射信号中反演炭黑温度和体积分数分布,需要求解非线性、病态且高维度的方程,该方程的求解通常使用数值反演方法进行,但由于数据吞吐量大,这是低效耗时的,然而机器学习提供了一种高效且能在线反演的算法。本研究发展了适用于从乙烯层流扩散火焰发射的可见光单色辐射强度中同时反演炭黑温度和体积分数分布的多层感知机(multi-layer perceptron,MLP)模型,该模型从已有数据中获取信息,并建立起炭黑温度和体积分数与火焰辐射强度之间的关系。对氧体积分数为19%~30%下的乙烯层流扩散火焰进行了反演,结果表明所发展的模型不仅有较高的精度,而且计算成本较低,该模型反演一次结果所需时间小于100 ms,另外该模型的泛化能力较好。此外,还使用该模型进行了测量误差下的反演计算,验证了其抗干扰性。

The inversion of the soot temperature and volume fraction profiles in hydrocarbon flames from radiation signals requires the solution of a nonlinear,ill-conditioned and high-dimensional equation,which is usually performed by a numerical inversion method.Due to large data throughput,this is inefficient and time-consuming.However,machine learning provides an efficient and online inversion algorithm.This paper developed a MLP(multi-layer perceptron)model which is applicable to the simultaneous inversion of the soot temperature and volume fraction distribution from the visible monochromatic radiation intensity of ethylene laminar diffusion flame.This model obtains information from existing data and establishes the relationship between soot temperature and volume fraction and flame radiation intensity.The inversions of the ethylene flame with an oxygen volume fraction range from 19%to 30%were performed.The results show that the developed model not only has higher accuracy,but also has lower computational cost,the time required for this model to invert a result is less than 100 ms.Furthermore,the developed MLP model has good generalization ability.In addition,the model was also used to perform inversion calculations under measurement errors to verify its anti-interference ability.