采用极限学习机的流场积分吸光度快速测量方法

发动机是飞行器动力系统的核心组件，发动机流场的动态监测可以掌握发动机内部流场的燃烧情况，对于飞行器状态监测和性能评估具有重要意义。拥有先进的诊断技术是发展发动机技术的基础，也是研制新型航空航天飞行器的必要条件之一。激光吸收光谱技术可以实现燃烧场气体参数的测量，在发动机严苛的流场环境中，吸收光谱波长调制技术（WMS）可以提高信噪比。但基于WMS解算积分吸光度和温度、浓度二维分布的方法都是以模拟退火算法（SA）为核心，因此存在执行时间较长的问题。根据随时间演化的流场光谱参数、光线分布为固定信息这一内在关联性，以及已有的WMS方法可以计算积分吸光度值，采用机器学习方法建立谐波信号（S_2f/1f）与积分吸光度（A）的模型，选择极限学习机算法（ELM），其训练时间短，预测结果快。利用神经网络可以逼近真值的特性，仿真确定光线布局下不同流场模型的S_2f/1f和A，构造数据集对神经网络开展模型训练。在数值仿真验证中，共仿真2 000组数据集，随机选取1 800组作为训练集训练模型，其余200组作为预测集，统计测试集的预测积分吸光度平均相对误差为1.058%，决定系数平均值为0.999，验证了训练模型的可靠性。为进一步探究模型的抗噪声性，采用的方法是在测试集S_2f/1f数据集中分别加入3%，5%和10%的随机噪声，统计预测积分吸光度平均相对误差分别为3.1%，4.6%和8.1%，这一结果可以表明ELM具有较好的抗噪声性。基于该方法，在直连式超燃冲压发动机上开展验证实验，实验有效时长为5 s，采集数据约10 GB，分别采用ELM和WMS两种方法解算积分吸光度，对比发现：结果基本一致，且相比执行时间数小时的WMS方法，ELM预测积分吸光度耗时仅为15 s左右，实现了发动机流场积分吸光度的快速测量。

Rapid Measurement of Integrated Absorbance of Flow Field Using Extreme Learning Machine

The engine is the core component of the vehicle power system. The dynamic monitoring of the engine flow field can grasp the combustion situation of the internal flow field of the engine, which is of great significance for the vehicle condition monitoring and performance evaluation. Therefore, advanced diagnostic technology is the basis for the development of engine technology and one of the necessary conditions for the development of new aerospace vehicles. The laser absorption spectroscopy technique can realize the measurement of gas parameters in the combustion field, and the absorption spectroscopy wavelength modulation technique (WMS) can improve the signal-to-noise ratio in the harsh flow field environment of the engine. However, the WMS-based methods for solving the integrated absorbance, temperature, and concentration are centered on simulated annealing algorithms (SA) and suffer from long execution times. Based on the intrinsic correlation of the spectral parameters of the flow field evolving and the light distribution as fixed information, a machine learning method is used to model the harmonic signal (S_2f/1f) and the integrated absorbance (A), and the extreme learning machine algorithm (ELM) is selected, which has a short training time and fast prediction results. Using the neural network’s property can approximate the true value, the simulation determines S_2f/1f and A for different flow field models under light layout and constructs data sets to carry out model training for the neural network. In the method validation, 2 000 data sets were simulated, 1 800 sets were selected as the training set to train the model, and the remaining 200 sets were used as the prediction set. The average relative error of the predicted integrated absorbance of the test set was 1.058%, and the coefficient of determination was 0.999, which verified the reliability of the training model. Random noise of 3% and 5% was added to the input S_2f/1f data set, and the average relative errors of predicted integrated absorbance were 1.89% and 3.2%, respectively, which showed that ELM has better noise resistance. Based on this method, experimental validation was carried out on a direct-connected scramjet with a practical test duration of 5 s and about 10GB of collected data, and the integrated absorbance was solved by both ELM and WMS methods respectively, and the results were consistent. Compared with the WMS method, which takes several hours to perform, the ELM predicts the integrated absorbance in about 15 seconds, enabling the rapid measurement of the integrated absorbance of the engine flow field.