Abstract A combustion regime identification based on convolutional neural networks (CNNs) is developed using the recently proposed gradient-free regime identification (GFRI) approach applied to two turbulent CH4/air jet flames featuring multi-regime characteristics. The training and the subsequent application of the CNN rely on the processing of one-dimensional Raman/Rayleigh line measurements of species mass fractions and temperature (CNN input). The combustion regime index is then readily predicted at every point along the measured line (CNN output). For training the neural network, the combustion regime index is first determined using the GFRI method (Hartl et al., 2018) based on the chemical explosive mode analysis (CEMA). Six classes of combustion regimes, including premixed (P), dominantly premixed (DP), multi-regime (MR), dominantly non-premixed (DNP), non-premixed (NP), and lean back-supported (LBS), are well detected by the trained CNN, with a pixel-wise accuracy of more than 85% for burner operating conditions unseen during training (different free-stream equivalence ratios). The quasi instantaneous neural network response provides a perspective towards real-time global combustion regime identification for pollutants emission control. From the results, it is also concluded that introducing physical insight, by combining advanced experimental (Raman/Rayleigh line measurements) and numerical analysis (GFRI), allows for reducing the amount of data needed to train neural networks.

中文翻译：

---

通过拉曼/瑞利线测量训练的机器学习识别燃烧状态

摘要 基于卷积神经网络 (CNN) 的燃烧状态识别是使用最近提出的无梯度状态识别 (GFRI) 方法开发的，该方法应用于具有多状态特征的两个湍流 CH4/空气喷射火焰。CNN 的训练和后续应用依赖于对物种质量分数和温度（CNN 输入）的一维拉曼/瑞利线测量的处理。然后可以很容易地预测沿测量线（CNN 输出）的每个点的燃烧状态指数。为了训练神经网络，首先使用基于化学爆炸模式分析 (CEMA) 的 GFRI 方法 (Hartl et al., 2018) 确定燃烧状态指数。六类燃烧状态，包括预混（P）、显性预混（DP）、多状态（MR）、显性非预混 (DNP)、非预混 (NP) 和瘦背支撑 (LBS) 被训练的 CNN 很好地检测到，对于训练期间未见过的燃烧器操作条件，像素精度超过 85% （不同的自由流当量比）。准瞬时神经网络响应为污染物排放控制的实时全球燃烧状态识别提供了一个视角。从结果中还得出结论，通过结合高级实验（拉曼/瑞利线测量）和数值分析 (GFRI)，引入物理洞察力可以减少训练神经网络所需的数据量。对于训练期间未见的燃烧器操作条件（不同的自由流当量比），像素精度超过 85%。准瞬时神经网络响应为污染物排放控制的实时全球燃烧状态识别提供了一个视角。从结果中还得出结论，通过结合高级实验（拉曼/瑞利线测量）和数值分析 (GFRI)，引入物理洞察力可以减少训练神经网络所需的数据量。对于训练期间未见的燃烧器操作条件（不同的自由流当量比），像素精度超过 85%。准瞬时神经网络响应为污染物排放控制的实时全球燃烧状态识别提供了一个视角。从结果中还得出结论，通过结合高级实验（拉曼/瑞利线测量）和数值分析 (GFRI)，引入物理洞察力可以减少训练神经网络所需的数据量。