Research On Dynamic Early Warning Efficient Analysis Method Based On Water Quality Anomaly Detection

The quality of water is related to the safety of people’s lives and properties.The state has long attached importance to the water environment and has gradually established a complete water quality control system to deal with emergencies such as water pollution and the proliferation of water pollutants.It has accumulated a large number of effective research results.With the development of Internet of Things technology,various high-frequency water quality acquisition sensors realize automatic monitoring and upload data in real time,accumulating a large amount of water quality related data.Through deep learning technology,the water quality big data is mined,and its characteristics are extracted to better control water quality.This thesis explores the application of deep learning in water quality warning,and applies dynamic and efficient water quality warning models to the system.The use of deep learning methods to predict and analyze water quality time series data has good results.Traditional methods require a large number of precise parameters for water quality simulation,such as river width,depth and other parameters,while deep learning methods only use water qualityrelated time series data.Feature extraction from data has the advantages of simplicity and efficiency.Through the research and exploration of water quality early warning methods,this article splits water quality early warning into two models to improve the accuracy.Firstly,water quality prediction model.In the water quality time series prediction model,this thesis compares the accuracy of the three smoothing index,support vector regression,and Recurrent Neural Networks variants in water quality prediction;to explore the influence of the seasonality of water quality time series data on the water quality prediction results,the introduction empirical mode decomposition and loessbased Seasonal and Trend decomposition using Loess（STL）are adopted;and the methods are selected from the perspective of data leakage for model training.To ensure the timeliness of early warning exploring the prediction step size and accuracy,this work is determined that the combination of STL decomposition and long and shortterm memory network based on codec has a good effect in all aspects of water quality prediction.Secondly,water quality anomaly detection model.In the water quality anomaly detection model,this thesis compares the effects of PauTa criterion,isolated forests,and multivariate Gaussian distribution anomaly detection on water quality anomaly analysis;by studying the seasonality of water quality time series data,we use seasonal segmentation to find the best dynamic early warning method of abnormal threshold avoids the problem of unreasonable early warning frequency caused by the traditional method using fixed threshold.Therefore,this work is determined that the isolated forest model is better than multivariate Gaussian distribution anomaly detection on multidimensional data,and the PauTa criterion has the ability to analyze single-dimensional data anomaly.Finally,this thesis combines the above two models to realize a dynamic and efficient water quality early warning model,and applies the model to a water quality early warning system.The system design is divided into data module,deep learning module,and task scheduling module.The task scheduling module implements continuous iterative training of the deep learning model.The model can learn the characteristics of new data in time to ensure prediction accuracy.The system takes the water quality early warning function as the core,realizes water quality real-time monitoring and water quality prediction functions,and effectively provides auxiliary decision support for water quality controllers.