基于监督学习的紫外-可见光光谱水质在线异常检测方法研究

水资源关系到国计民生，近年来时有发生的水污染事件使污染物入侵预警得到了广泛的社会关注。针对现有基于紫外-可见光光谱的水质异常检测方法存在的检出下限偏低的问题，提出一种基于监督学习的紫外-可见光光谱水质异常检测方法。该方法首先获取不同数据集中的正常样本差异性空间，再使用正交投影方法去除差异性空间中的光谱数据分量，以达到基线校正的目的；然后采用偏最小二乘判别分析从校正后的光谱中提取特征，利用训练集得到的最优阈值确定离群点；最后采用序贯贝叶斯滚动更新每个时刻上的异常概率，确定水质报警序列。实验选用苯酚作为模拟污染入侵事件的注入试剂，采样2周内的紫外-可见光光谱数据，在实验平台上对提出的方法进行了验证。实验结果表明，采用的正交投影基线校正方法可以消除不同批次水质光谱的背景差异，更为充分的利用了光谱信息，降低了对特征污染物的检出下限。

In-Situ Detection of Water Quality Anomaly with UV/Vis Spectrum Based on Supervised Learning

Water resources are related to national economy and people’s livelihood, detection of water quality anomaly has attracted more attention because of the water pollution events happened in recent years. In this paper, the detection method of water quality anomaly with UV-Vis Spectrum based on supervised learning was proposed to solve the problems of existing methods, which behaved as a high detection limit and poor adaptability method. The pretreatment of orthogonal projection was used to correct the gap between different batches of spectral data. Afterwards the Partial Least Squares Discriminatory Analysis was adopted to extract the features from the data set. Outliers were found by comparing the alarm signal with the best threshold from the training set. Finally, Sequential Bayesian Method was used to update the probability of Contaminate Intrusion Events and to get the alarm sequence. The results showed that the proposed method had the lower detection limit than unsupervised method and the pretreatment of orthogonal projection improved the adaptability of detection method based on supervised learning for baseline changing.