基于LLE-BPNN的小麦岛海水硝酸盐含量分析

Analysis of Nitrate in Seawater of Wheat Island Based on LLE-BPNN

水中过量的硝酸盐会造成部分水生生物难以存活、引发人类尤其是婴儿患病等危害,因此硝酸盐浓度成为水质检测中的一项重要指标。传统的硝酸盐浓度测量方法操作复杂、反应缓慢,近年许多研究人员开始通过紫外可见(UV-Vis)光谱技术结合人工神经网络(ANN)的方法对水中硝酸盐的含量进行测量。提出了一种将流形学习(manifold learning)方法中的局部线性嵌入(LLE)与反向传播神经网络(BPNN)相结合的建模方法,用以得到硝酸盐光谱曲线与浓度间的关系,实现对青岛市崂山区小麦岛海水中硝酸盐浓度快速准确的定量分析。实验选取了过滤后的小麦岛海水配置59组不同浓度的加标溶液,采用实验室自主研制的光谱分析仪采集这些样本的光谱测量值,通过标准正态变换(SNV)方法对测得硝酸盐溶液的光谱数据进行校正处理,有效降低了由仪器本身或环境带来的噪声影响;选取预处理后的光谱数据的前1 500维处理后进行对比实验,以解决使用BPNN对全部2 048维数据建模时内存不足的问题,再通过网格搜索结合十折交叉验证的方法优化LLE中的邻近点数k和嵌入维数d,得到最优参数值k=15,d=3,实现对实验数据的降维处理;通过BPNN将降维后的训练集光谱信息与其对应的浓度信息进行建模,实现对预测集硝酸盐浓度定量分析,引入决定系数(R^2)和预测均方根误差(RMSEP)评价建模效果,与直接使用BPNN建模预测的结果比较,改进方法的R^2由0.926 3提升至0.992 8, RMSEP由0.442 5下降到0.280 4,建模预测程序的运行时间由327 s缩短至0.5 s。采用这59组数据的全部2 048维进行LLE-BPNN建模时,得到R^2=0.995 7, RMSEP=0.136 5,在用时相近的前提下,相比仅使用前1 500维时的建模精度更好。分析结果表明, LLE-BPNN的方法可实现对海水中硝酸盐浓度的快速预测,使预测精度得到显著提升,同时能大幅降低预测时间。

Excessive nitrate in water may influence some aquatic organisms’ survival and cause harm to humans,especially infants.Therefore,nitrate concentration becomes an important indicator in water quality monitoring.Due to the complexity of operation and slow response of conventional methods for measuring nitrate concentration,many researchers have begun to use ultraviolet/visible (UV-Vis) spectroscopy combined with artificial neural network (ANN) methods to measure nitrate content in water.This paper proposes a modeling method combining locally linear embedding (LLE) in manifold learning with back propagation neural network (BPNN).The relationship between the spectral curve of nitrate and the concentration was obtained,so that a rapid and accurate quantitative analysis of the nitrate concentration in the wheat island of Laoshan District,Qingdao was achieved.In the experiment,we selected 59 groups of spiked solutions with different concentrations of filtered wheat island seawater,and collected spectral measurements of these samples using a laboratory-developed spectrum analyzer,with standard normal variate (SNV) method calibrating spectral data of measured nitrate solution to reduce the noise caused by the instrument itself or the environment.First1 500-dimensional of the pre-processed spectral data was used to avoid insufficient memory when using the entire2 048-dimensional data to build BPNN model,and a control experiment was performed.Then the number of neighboring points k and the embedding dimension d in the LLE were optimized by the grid search combined with the ten-fold cross validation method,obtaining the optimal k=15,d=3 .Then the dimension of the experimental data was reduced.The spectral information of the reduced-dimensional training set and its corresponding concentration information were modeled by the BPNN to achieve a quantitative analysis of the nitrate concentration in the prediction set.Coefficient of determination ( R2) and root mean square error of prediction (RMSEP) were introduced to evaluate modeling effects.And compared with the predicted results obtained by only using BPNN modeling,R2 of our improved method increased from 0.926 3 to 0.992 8,and RMSEP decreased from 0.442 5 to 0.280 4,and prediction modeling program run time decreased from 327 s to about 0.5 s.In addition,we used all2 048 dimensions of the 59 data sets for LLE-BPNN modeling,with R2=0.995 7 and RMSEP=0.136 5,which was improved compared to the modeling accuracy when only using the first1 500 dimensions,while elapsed time was similar.The analysis results above showed that using the LLE-BPNN method can achieve a rapid prediction of nitrate concentration in seawater,while significantly improving prediction accuracy and reducing prediction time.