一种多源光谱融合的水样COD实验检测方法

COD Detection Method of Water Quality Based on Multi-Source Spectral Fusion

提出了一种多源光谱融合的水质化学需氧量(COD)预测算法。该算法利用深度学习方法训练COD预测模型,并通过感知卷积网络确定紫外-可见吸收光谱和三维荧光光谱中各个位置的关注程度,不断移除关注程度较高的特征并重新训练网络,以发现可能被忽视的有效特征。进一步筛选并利用关注程度最高的融合特征位置,通过建立偏最小二乘(PLS)模型来预测COD,以更好地利用光谱数据中的所有有效特征。与支持向量回归(SVR)、PLS、间隔偏最小二乘(IPLS)模型相比,本文的留一法均方根误差(RMSE)分别减小了70.0%、75.1%和56.7%,十折交叉检验RMSE分别减小了64.3%、78.3%和64.6%。

Objective Chemical oxygen demand(COD)refers to the quantity of reducing substances in water requiring oxidation.As the COD concentration becomes higher in water,the organic pollution is more severe.The decomposition of a large amount of organic pollutants excessively consumes dissolved oxygen in water,fostering anaerobic bacterium proliferation and resulting in water discoloration and malodor.Consequently,COD has become an important indicator for water pollution assessment.Spectral analysis for water quality COD assessment is one of the contemporary research focuses.Compared to conventional single-source spectral data prediction,using multi-source spectral data enables the extraction of richer feature information,thereby enhancing prediction accuracy.However,the key issue in detecting COD concentration using spectral methods is how to select appropriate feature wavelengths and establish regression models.Traditional feature extraction techniques(such as particle swarm optimization,ant colony optimization,and other swarm intelligence algorithms)exhibit screening efficacy.However,due to spectral data redundancy,more intelligent individuals are required for feature search,which greatly increases the computational load.If the number of intelligent individuals is reduced,the feature search range of spectral data needs to be narrowed,such as truncating the ultraviolet-visible spectrum to 200 to 400 nm and increasing the excitation and emission intervals of three-dimensional fluorescence spectroscopy.These methods will reduce the utilization range of spectral features.Therefore,we propose a multi-source spectral fusion algorithm for predicting COD concentration in water.The algorithm utilizes deep learning methods to train COD prediction models and determines the attention level of each position in the ultraviolet-visible absorption spectrum and three-dimensional fluorescence spectrum through a perceptual convolutional network.It continuously removes features with high attention levels and retrains the network to discover potentially overlooked effective features.Then,it further screens and utilizes the fused feature positions with the highest attention levels to establish a PLS model to predict COD concentration,aiming to better utilize all effective features in spectral data.Methods We introduce a multi-source spectral fusion method for water quality COD detection.The method establishes a convolutional network that integrates three-dimensional fluorescence and ultraviolet-visible spectra.The structure is depicted in Fig.1.The model initially extracts diverse features from stacked convolutional modules of three-dimensional fluorescence and ultraviolet-visible spectra and then integrates the feature information of three-dimensional fluorescence and ultraviolet-visible spectra through two fully connected layers.Subsequently,a 2×1 fully connected output is used to predict the COD result and then used to calculated the preference of the multi-spectral convolutional network for different features.The network is continuously removed from the training process to remove the features that are highly concerned,and the removed features are used to retrain the network to explore the effective features that have been neglected as much as possible.Ultimately,the PLS model is employed to further screen the key combination features and realize the prediction of COD concentration.Results and Discussions The experimental results of the PLS prediction model established by combining features are presented in Fig.7.The left panel of Fig.7 shows the experimental results using ten-fold cross-validation,revealing a correlation coefficient of 0.99989 and an RMSE of 1.4398.The right panel of Fig.7 illustrates the experimental results using leave-one-out cross-validation,demonstrating a correlation coefficient of 0.99993 and an RMSE of 0.9875.Table 4 summarizes the experimental results,including correlation coefficients and root mean square errors for four modeling methods.From Table 4,we find that the proposed prediction model outperforms the other three prediction models in terms of correlation coefficients and root mean square errors using both leave-one-out cross-validation and ten-fold crossvalidation approaches.The RMSE of leave-one-out cross-validation is 0.9875,which is much lower than that of the other three prediction models.Comparisons show that the prediction model proposed in this paper is superior to the other three prediction models.Conclusions The experimental findings show that the multi-spectral feature-level fusion model achieves better detection performance compared to SVR,PLS,and IPLS,with a reduction of 56.7%in the RMSE of the best IPLS leave-one-out method,reaching 0.9875.The modeling method proposed in this paper demonstrates good feasibility.Using deep learning methods,it can extract effective feature advantages amidst a plethora of redundant attributes while avoiding the challenges of limited generalization capabilities of deep learning models arising from sparse spectral data and water quality labels,which can more accurately detect water quality COD and provide a new means of predicting COD concentration for online water quality detection.At the same time,our multi-spectral fusion-based modeling method holds promise for application in data analysis and model establishment in other detection and recognition fields.