化学需氧量(COD)是反映水体受有机物污染程度的重要指标。 紫外吸收光谱法是目前水体COD检测研究中应用最为广泛的方法, 具有样品无需预处理, 成本低, 无污染, 测定速度快等优点。 但是, 原始光谱数据维数高, 光谱信息中包含大量冗余变量, 直接将全光谱数据进行建模存在精度低, 计算复杂等问题。 针对紫外吸收光谱全光谱建模精度低, 光谱数据存在大量共线性的问题, 提出了一种基于粒子群算法(PSO)结合偏最小二乘(PLS)优选特征波长建立预测模型的方法, 以提高紫外吸收光谱预测模型的精度和适用性, 简化模型。 利用搭建的紫外吸收光谱装置, 采集29份不同浓度的COD标准溶液的紫外光谱数据, 每份标准溶液采集5次取平均值并对其进行平滑处理, 减少仪器和环境带来的误差。 考虑到标准溶液在200~310 nm的光谱范围内存在吸收, 故选取该波段范围内246个波长点作为建模数据, 每个波长点下的吸光度数据作为一个粒子并按照顺序编号, 以PLS为建模方法, 相关系数 r和均方根误差(RMSE)为评价指标, 设置粒子群算法适应度函数 f( x)=min(RMSE), 取粒子初始种群数为20个, 惯性权重 w=0.6, 自我学习因子 c1=1.6, 群体学习因子 c2=1.6, 最大迭代次数为200次, 算法终止条件为达到最大迭代次数。 算法输出全局最优变量取值为168, 94, 181, 183, 175, 209, 106和142。 采用粒子群算法优选的8个波长点建立PLS预测模型的相关系数 r和预测均方根误差RMSE分别为0.999 98和0.155 1。 为了验证PSO-PLS建立的预测模型效果, 建立了PLS, iPLS和SVR三种预测模型进行对比。 验证结果表明, PSO-PLS模型的相关系数 r和均方根误差RMSE均优于其他三种预测模型, 说明粒子群算法能有效的提取用于PLS建模的特征波长, 消除子区间变量的共线性, 提高预测模型的精度。 该方法为实现水体COD实时在线监测提供了一种有效途径。

Chemical oxygen demand (COD) is an important indicator of the degree of water pollution by organic matter. Ultraviolet absorption spectroscopy is the most widely used method for COD detection in water. It has the advantages of no pretreatment of samples, low cost, no pollution, and fast measurement speed. However, the original spectral data has high dimensions, and the spectral information contains a large number of redundant variables. Modeling the full spectral data has problems such as low accuracy and complicated calculations. Aiming at the low accuracy of UV absorption full-spectrum modeling and a large amount of collinearity in spectral data, this paper presents a method based on particle swarm optimization (PSO) and partial least squares (PLS) to select characteristic wavelengths to establish a prediction model. Improve the accuracy and applicability of the UV absorption spectrum prediction model and simplify the model. The UV spectrum data of 29 different concentrations of COD standard solutions were collected. Each standard solution was collected 5 times and averaged and smoothed to reduce the errors caused by the instrument and the environment. Taking into account the absorption of the standard solution in the spectral range of 200~310 nm, 246 wavelength points in this wavelength range were selected as modeling data, and the absorbance data at each wavelength point was used as a particle and numbered in order. PLS was used as the model Method, the correlation coefficient r and the root mean square error (RMSE) are used as evaluation indicators. The particle swarm algorithm fitness function f( x)=min (RMSE) is set. The initial population of particles is 20, the inertia weight w=0.6, and the self The learning factor c1=1.6, the group learning factor c2=1.6, the maximum number of iterations is 200, and the algorithm termination condition is to reach the maximum number of iterations. The output value of the optimal global variable of the algorithm is 168, 94, 181, 183, 175, 209, 106, 142. The correlation coefficient r and the predicted root mean square error RMSE of the PLS prediction model established by the eight wavelength points selected by the particle swarm optimization algorithm were 0.999 98 and 0.155 1, respectively. In order to verify the effectiveness of the prediction model established by PSO-PLS, three prediction models of PLS, iPLS and SVR were established for comparison. The verification results show that the correlation coefficient r and the root mean square error RMSE of the PSO-PLS model are better than those of the other three prediction models, which shows that the particle swarm algorithm can effectively extract the characteristic wavelengths used for PLS modeling and eliminate the common of sub-interval variables Linear, improving the accuracy of the prediction model. This method provides an effective way for real-time online monitoring of COD in water bodies.