水质污染源的及时精确定位和精细化的污染防治措施是打赢水污染防治攻坚战的迫切需求, 为解决地表水实际水样高锰酸盐指数准确分类的实际问题, 以光谱降噪和光谱有效信息提取为切入点, 根据紫外-可见光谱数据的特点, 提出使用一维卷积神经网络处理紫外-可见光谱数据。 为验证检测一维卷积神经网络对地表水光谱信号分类的可行性, 选取长江的某段流域作为取样点。 采集当天的长江上游水、 某河水、 嘉陵江水, 生活污水、 500 mg·L-1邻苯二甲酸氢钾溶液来模拟污染水源。 将几种水样按不同的配比来模拟当天该流域的水污染变化情况。 采集现有的单一水样及混合配比水样的光谱数据, 根据各类水样的特征光谱信息进行区分, 实现地表水高锰酸盐指数的预测分类, 快速确定异常水样的污染来源, 通过仿真实验, 优化模型参数并完成优化训练。 与K最邻近法、 支持向量机等传统分类方法相比, 该算法在光谱预处理复杂度和定性分析准确度方面有较大优势, 在没有复杂的数据预处理前提下, 将获取的350条光谱数据建立水质分类模型, 随机选择其中245条数据作为训练集, 另105条数据作为测试集, 模型的混淆矩阵分类精度达99.0%。 不仅简化了整个光谱分析流程, 而且能保留更多的有效光谱信息, 减小人为预处理对紫外-可见光谱数据的影响, 实现地表水高锰酸盐指数的准确分类。 实验结果表明该方法可对不同水体水样进行准确分类, 快速定位污染源, 为无法激发荧光的污染物溯源提供了科学依据, 为与三维荧光技术辅助配合快速精确定位地表水污染源提供了可能, 同时表明了深度学习在紫外-可见光谱法测量实际水样领域有着巨大的应用潜力和研究价值。

The timely and accurate location of water pollution sources and fine pollution prevention and control measures are the urgent need to win the battle of water pollution prevention and control, in order to solve the practical problem of accurate classification of permanganate index of surface water samples, in this paper, based on spectral noise reduction and spectral effective information extraction, according to the characteristics of UV-visible spectral data, one-dimensional convolution neural network is proposed to process UV-visible spectral data. In order to verify the feasibility of detecting a one-dimensional convolution neural network to classify the spectral signals of surface water, a section of the Yangtze River was selected as the sampling point. The water from the upper reaches of the Yangtze River, a river and the Jialing River were collected on the same day, and domestic sewage and 500 mg·L-1 potassium hydrogen phthalate solution were used to simulate the polluted water source. Several kinds of water samples were used to simulate the basin’s changes in water pollution on the same day according to different proportions. Collect the spectral data of existing single and mixed water samples, and distinguish them according to the characteristic spectral information of all kinds of water samples. Realize the prediction and classification of surface water permanganate index, quickly determine the pollution source of abnormal water samples through simulation experiments, optimize the model parameters and complete the optimization training. Compared with traditional classification methods such as the K nearest neighbor method and support vector machine, this algorithm has great advantages in spectral preprocessing complexity and qualitative analysis accuracy. 350 spectral data obtained are used to establish a water quality classification model, of which 245 data are randomly selected as the training set and 105 data as the test set. The confusion matrix classification accuracy of the model is up to 99.0%. It not only simplifies the whole spectral analysis process but also retains more effective spectral information, reduces the influence of artificial pretreatment on UV-Vis spectral data, and realizes the accurate classification of the permanganate index of surface water. The experimental results show that this method can accurately classify water samples from different water bodies, locate pollution sources quickly, and provide a research basis for tracing the sources of pollutants that can not stimulate fluorescence. It provides the possibility for rapid and accurate location of surface water pollution sources with the aid of three-dimensional fluorescence technology. It shows that depth learning has great application potential and research value in the UV-vis spectroscopy measurement of actual water samples.