基于常规水质参数的供水管网特征污染物分类方法研究

发布时间：2018-05-01 04:39

本文选题：常规水质参数 + 特征污染物　；参考：《浙江大学》2017年硕士论文

【摘要】：随着城市供水安全受到越来越严峻的挑战,构建能够对城市供水管网水质进行持续在线监控的预警系统意义重大。在检测出水污染事件之后,为了更好地提供污染物特性等应急信息,需要进一步识别污染物的具体类别。可能引起水体污染的物质种类繁多,且很多没有针对性的检测仪器。面对这一状况,本文研究了污染物与常规水质参数响应之间的关系,并基于此开展了污染物分类识别研究。论文主要工作和创新点如下:(1)研究了常规水质参数与某些重金属盐、有机盐和无机盐污染物之间的相关响应规律,分析了不同监测数据时间序列幅值变化特性,提出了利用这些因不同污染物而不同的变化特性及其组合信息,进行不同污染物的分类与识别的技术架构。(2)研究了通过度量常规水质参数组合信息之间的相似性判别污染物类型的技术方法。该方法首先采用自回归模型进行水质背景信号估计,再利用K均值聚类算法融合多个指标的预测残差获取污染物引起的水质参数响应类别中心,最后采用相似性度量方法进行污染物识别。其中重点针对污染物识别过程中,常规水质参数响应幅值受污染物浓度影响的问题,从理论上分析了余弦距离的特性,其主要度量的是水质参数向量之间的夹角,因此受幅值改变的影响较小,在污染物分类识别中具有较好的效果。通过污染物注入实验比较了欧式距离,马氏距离,余弦距离等不同相似性度量方法在五种特征污染物上的识别效果,验证了理论分析的正确性。(3)针对常规水质参数与污染物浓度变化之间的非线性、各参数之间变化趋势不一致以及训练样本不足等问题,提出基于SVM多分类模型进行污染物分类的方法。考虑到污染物注入初始阶段错分率高,论文引入分类概率,通过研究最大分类概率以及分类概率标准差,对样本进行区分,避免在水质参数波动信息不显著情况下做出错误的单一决策。最后对相似性度量方法和SVM多分类模型在不同情况下的性能进行了详细对比分析,明确了各自的性能优势和适用场合。(4)利用所研究的基于相似性度量的分类方法和基于SVM多分类模型的分类方法结合C#与MATLAB混合编程技术,在实验室模拟水质监测系统基础上设计开发了管网水质污染物分类软件。该软件具有特征污染物分类判别,特征库动态更新,分类算法管理,分类结果展示等功能。
[Abstract]:As the security of urban water supply is facing more and more serious challenges, it is of great significance to construct an early warning system which can continuously monitor the water quality of urban water supply network. After the detection of water pollution events, in order to provide better emergency information such as pollutant characteristics, it is necessary to further identify the specific types of pollutants. There are many kinds of substances which may cause water pollution, and many untargeted detection instruments. In this paper, the relationship between pollutants and the response of conventional water quality parameters is studied, and the classification and identification of pollutants are carried out. The main work and innovation of this paper are as follows: (1) the correlation response between conventional water quality parameters and some heavy metal, organic and inorganic salt pollutants is studied, and the variation characteristics of time series amplitudes of different monitoring data are analyzed. It is proposed to use these information, which vary from pollutant to pollutant, and their combinations, The technical framework for classification and identification of different pollutants. Firstly, the autoregressive model is used to estimate the background signal of water quality, and then K-means clustering algorithm is used to fuse the prediction residuals of multiple indexes to obtain the response class center of water quality parameters caused by pollutants. Finally, the similarity measurement method is used to identify pollutants. Aiming at the problem that the response amplitude of conventional water quality parameters is affected by pollutant concentration in the process of pollutant identification, the characteristics of cosine distance are analyzed theoretically. The main measure is the angle between water quality parameter vectors. Therefore, the effect of amplitude change is relatively small, and it has better effect in pollutant classification and identification. The effects of different similarity measures, such as Euclidean distance, Markov distance and cosine distance, on the recognition of five characteristic pollutants were compared by pollutant injection experiments. The correctness of the theoretical analysis is verified. (3) aiming at the nonlinearity between the conventional water quality parameters and the change of pollutant concentration, the variation trend between the parameters and the shortage of training samples, etc. A method of pollutant classification based on SVM multi-classification model is proposed. Considering the high misclassification rate in the initial stage of pollutant injection, the classification probability is introduced, and the sample is distinguished by studying the maximum classification probability and the classification probability standard deviation. Avoid making a single wrong decision when the fluctuation information of water quality parameters is not significant. Finally, the performance of similarity measurement method and SVM multi-classification model in different cases are compared and analyzed in detail. It is clear that their respective performance advantages and applicable situation. (4) the classification method based on similarity measure and the classification method based on SVM multi-classification model are used to combine C # and MATLAB hybrid programming technology. Based on the laboratory simulation water quality monitoring system, the classification software of water pollution in pipe network is designed and developed. The software has the functions of distinguishing the characteristic pollutants, updating the feature database dynamically, managing the classification algorithm and displaying the classification results.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TU991.2

【参考文献】