水体中的硝酸盐浓度过高不仅会造成水环境污染而且会对人类身体健康造成很大威胁, 传统的检测硝酸盐的方法检测时间长且操作复杂。 针对水体中硝酸盐氮难以快速在线检测的问题, 基于紫外吸收光谱, 提出了一种混合预测模型结合光谱积分快速定量检测水体中硝酸盐浓度的方法。 混合预测模型为低浓度样本建立的双波长法预测模型与高浓度样本建立的偏最小二乘支持向量机(LS-SVM)预测模型数据融合之后的模型。 按照合适的浓度梯度配备了19组硝酸盐氮标准溶液, 通过实验测得不同浓度硝酸盐氮样本的光谱数据。 首先基于双波长法对所有样本进行回归分析, 按照 A= A₂₂₀-2 A₂₇₅计算不同实验样本的吸光度 A, 其中 A₂₂₀和 A₂₇₅是220和275 nm处样本的吸光度, 将吸光度 A与样本浓度值进行线性回归, 拟合出样本浓度的预测值。 结果显示当样本浓度较小时, 相关性很好, r为0.997 4, 随着实验样本浓度的上升, 曲线发生严重的非线性漂移, 因此双波长法只适合低浓度样本预测模型的建立。 对于高浓度样本, 光谱重叠严重, 适合建立非线性的预测模型, 支持向量机(SVM)与LS-SVM都适合小样本的非线性数据建模, LS-SVM预测精度稍高, 运算速度稍快。 通过对所有的实验样本进行全波长光谱积分, 比较相邻样本光谱积分的变化率可以筛选出样本的临界浓度值, 4 mg·L^-1的硝酸盐样本积分值前后变化率最大, 因此选择4 mg·L^-1作为临界浓度值较为合适。 浓度高于4 mg·L^-1的实验样本建立LS-SVM预测模型, 通过交叉验证的方法选择出合适的参数, 正则化参数 γ=50, 核函数选择高斯核, 核函数宽度 σ²=0.36, 训练样本之后进行回归; 其余样本建立双波长法预测模型, 最后进行两种模型的数据融合, 形成从低浓度到高浓度的水体中硝酸盐浓度的检测。 为了验证混合预测模型的预测精度, 另外建立了SVM, LS-SVM, 偏最小二乘(PLS)等模型, 并求出 r, 预测值与真实浓度值平均绝对误差(MAE)和均方根误差(RMSE)对模型进行评价。 验证结果表明, 相比于SVM, LS-SVM和PLS等模型, 提出的混合模型回归的相关系数为0.999 86, 分别提高了1.8%, 1.6%和0.45%, 预测值与真实浓度的平均绝对误差为2.55%, 分别降低了6.27%, 4.49%和1.01%, 均方根误差为0.303, 为四种预测模型中最小, SVM与LS-SVM的相对误差相对较高, PLS模型相对误差上下波动比较大, 混合预测模型相对误差最为稳定, 并保持在较低水平, 由此可见混合预测模型的预测效果明显优于其他几种模型。 并与文献[5—7]中的测量方法进行对比, 该混合预测方法可以简单快速的测量水体中硝酸盐氮的浓度, 且不需要试剂, 无二次污染, 与文献[9]中的预测模型相比, 预测精度明显提高。 因此提出的混合模型可正确快速地预测水体中硝酸盐氮的浓度, 可为在线监测水体中硝酸盐浓度提供有效的技术参考。

High concentration of nitrates in water will not only cause water environment pollution but also pose a great threat to human health. The traditional methods for detecting nitrates have a long detection time and are complex to operate. In view of the difficulty in rapid on-line detection of nitrate nitrogen in water, a method combined a mixed prediction model with spectral integration was proposed to rapidly detect nitrate concentration in water based on ultraviolet absorption spectroscopy. The mixed prediction model is a model after data fusion of the dual wavelength prediction model established by low-concentration samples and the partial least-squares support vector machine (LS-SVM) prediction model based on the high concentration samples. According to the appropriate concentration gradient, 19 sets of nitrate nitrogen standard solution were equipped, and the spectral data of nitrate nitrogen samples of different concentrations were measured by experiment. First, Regression analysis was performed on all samples based on the dual wavelength method. Absorbance A was calculated for different experimental samples according to A= A₂₂₀-2 A₂₇₅, where A₂₂₀ and A₂₇₅ were the absorbance of the samples at 220 and 275 nm. The values were linearly regressed to fit the predicted values of the sample concentrations. The results showed that when the sample concentration is small, the correlation is very good, and r is 0.997 4. The two-wavelength method is only suitable for the establishment of low-concentration samples prediction model with a serious nonlinear drift in the rising curve of the experimental samples concentration. For high-concentration samples, spectral overlap is severe and it is suitable for establishing nonlinear prediction models. Both support vector machine (SVM) and partial LS-SVM are suitable for nonlinear data modeling of small samples. The LS-SVM has a slightly higher prediction accuracy and a slightly faster operating speed. By performing full-wavelength spectral integration on all experimental samples and comparing the rate of change of the spectral integrals of adjacent samples, the critical concentration of the sample can be selected. The 4 mg·L^-1 nitrate sample has the largest change rate before and after the integrated value, so it is appropriate to select 4 mg·L^-1 as the the critical concentration value. The LS-SVM prediction model was established for experimental samples with concentrations higher than 4 mg·L^-1. Cross-validation methods were used to select the appropriate parameters. The regularization parameter was γ=50, and the Gaussian kernel function width was σ²=0.36. The other samples were used to establish the dual-wavelength prediction model, and finally performed the data fusion of the two models, which formed the detection of nitrate from low concentration to high concentration. In order to verify the prediction accuracy of the mixed prediction model, the model of SVM, LS-SVM and PLS was established, and evaluated the model with mean absolute error (MAE), correlation coefficient ( r), and root mean squared error (RMSE). The verification results showed that compared with other models, the correlation coefficient of the proposed mixed model regression is 0.999 86, which is increased by 1.8%, 1.6%, and 0.45% respectively, and the average absolute error between the predicted value and the true concentration is 2.55%, which decreased by 6.27%, 4.49%, and 1.01% respectively, and the root-mean-square error is 0.303, which is the smallest of the four prediction models. The relative error of SVM and LS-SVM is relatively high, and PLS model fluctuates up and down relatively. The relative error of mixed forecasting model is the most stable and remains at a low level, and the forecasting effect of mixed forecasting model is obviously better than that of other models. Compared with the measurement method in [5-7], this hybrid prediction method can simply and quickly measure the nitrate nitrogen concentration in water without reagents and no secondary pollution, andthe prediction accuracy is significantly improved compared with the model in [9]. Therefore, the proposed mixed model can correctly and quickly predict the concentration of nitrate in water, and provide an effective technical reference for on-line monitoring of nitrate concentration in water.