基于近红外光谱的沼液挥发性脂肪酸含量快速检测

挥发性脂肪酸（Volatile Fatty Acids，VFA）作为厌氧发酵过程的重要中间产物，其在厌氧反应器中的累积能够反映出产甲烷菌的不活跃状态或厌氧发酵条件的恶化。为了实现对农牧废弃物厌氧发酵进行过程分析和状态监控，将近红外光谱（Near Infrared Spectroscopy，NIRS）与偏最小二乘（Partial Least Squares，PLS）相结合构建玉米秸秆和畜禽粪便厌氧发酵液乙酸、丙酸和总酸含量快速检测模型。将竞争自适应重加权采样法（Competitive Adaptive Reweighted Sampling，CARS）与遗传模拟退火算法（Genetic Simulated Annealing algorithm，GSA）相结合构建CARS-GSA算法对沼液中的乙酸、丙酸和总酸进行特征波长优选，原始光谱数据1 557个波长点经预处理和波长优选后，得到乙酸、丙酸和总酸特征波长变量分别为135、101和245个，建立的回归模型验证决定系数分别为0.988、0.923和0.886，预测均方根误差（Root Mean Squared Error of Prediction，RMSEP）分别为0.111、0.120和0.727，相对分析误差分别为9.685、3.685和3.484，与全谱建模相比RMSEP分别减少了17.78%、15.49%和1.22%，能够满足农牧废弃物厌氧发酵过程发酵液中乙酸和丙酸含量的快速检测需求，基本满足总酸的检测需求。结果表明，通过构建CARS-GSA算法优选乙酸、丙酸和总酸的敏感波长变量，参与建模的波长点数量显著减少，有效降低了变量维度和模型复杂度，提升了回归模型检测精度和预测能力，为快速准确检测沼液VFA提供了新途径。

Rapid determination of volatile fatty acids in biogas slurry based on near infrared spectroscopy

Abstract: Volatile fatty acids (VFA), serving as important intermediate products in anaerobic digestion (AD), have been considered as the key variables in most AD monitoring strategies, as they respond to incoming imbalances, indicating the buffer capacity of digesters to process disturbance and imminent digester failure that caused by sudden operational changes. In order to ensure efficient operation of AD while improve the utilization rate of raw materials, it is necessary to accurately monitor and evaluate the operation state of biogas engineering, via detecting the concentrations of VFA in the process of biogas production with corn stover and animal manure as feedstocks. Previously, the rapid detection models of acetic acid (AA), propionic acid (PA) and total acid (TA) in biogas slurry have been constructed, using the near infrared spectroscopy (NIRS) technique combined with the partial least squares (PLS), aiming to overcome the time consuming and high-cost in the traditional chemical analysis method. However, a prediction model can trigger the high complexity and low accuracy, due to the spectroscopic data generally includes quantities of invalid redundant information. In this study, an integrated algorithm was presented, based on the competitive adaptive reweighted sampling (CARS) and genetic simulated annealing algorithm (GSA), to optimize the characteristic wavelength variables of AA, PA, and TA, and thereby to improve the efficiency and precision of NIRS detection models. An AD experiment was carried out with corn stover, pig manure and cow manure as feedstocks, where 155 samples of biogas slurry were collected. The NIRS data of biogas slurry was acquired in a transmittance mode using the AntarisTM II FT-NIR spectrophotometer equipped with a quartz cuvette. A gas chromatography (GC) system was used to measure the VFA of biogas slurry, where 81 valid data of AA, 78 valid data of PA, and 87 valid data of TA were obtained to establish the regression model. One segment of the spectrum with 95 wavelength points was removed from 4 933.02 to 5 295.57 cm-1, and 1462 wavelength variables remained, mainly due to the saturation of spectrum can be caused by the strong combination band of -OH from water. The spectral preprocessing methods were selected, according to the mean relative error of calibration set. Correspondingly, the samples were divided into the calibration set and validation set, using Sample Set Portioning based on Joint X-Y Distances (SPXY) algorithm. The number of characteristic wavelength variables for AA, PA, and TA were 135, 101, and 245, respectively. The PLS regression models were established with the characteristic wavelengths of AA, PA, and TA, where the results were the coefficients of multiple determination for prediction () of 0.988 2, root mean squared error of prediction (RMSEP) of 0.111 4, and the residual predictive deviation (RPD) of 9.685 0 for AA, of 0.922 9, RMSEP of 0.120 0, and RPD of 3.684 8 for PA, of 0.885 6, RMSEP of 0.726 8, and RPD of 3.484 0 for TA. Meanwhile, compared with the whole spectrum model, the RMSEP in the CARS-GSA model decreased by 17.78%, 15.49%, and 1.22%, respectively, showing that the number of wavelengths significantly decreased after the optimization, whereas, the performance of regressive model was obviously higher than that of the whole wavelengths. The results demonstrate that the CARS-GSA model can fulfil the requirement of rapid detection for AA and PA concentrations in biogas slurry during anaerobic fermentation with agricultural waste as feedstocks, while basically meet the detection requirement of TA concentration. The CARS-GSA model also can be used to enhance the forecasting capability of the model, while reduce its complexity. The findings can provide a new way to improve the accuracy and robustness of prediction model, base on optimizing sensitive wavelengths for AA, PA, and TA, further for rapid and accurate measurement of VFA concentrations in biogas slurry.