Research On Water Quality Anomaly Detection Of Urban River Using Multi-Indicators Time Series Data

In recent years,with the rapid economic and social development,water pollutions occur from time to time,and water shortages have become serious increasingly.Therefore,the establishment of a warning system for water quality anomaly detection is the key of ensuring the safety of the water environment in urban rivers,so as to reduce the degree of manual labor and improve water environment emergency control capacity.However,the existing methods based on conventional water quality indicators are often analyzed from a single perspective,such as physical and chemical properties(turbidity,conductivity,dissolved oxygen),organic content(chemical oxygen demand,ammonia nitrogen),acidity(pH),in which the implied information of water quality time series data cannot be fully exploited;as a result,the accuracy and timeliness could be impaired.Based on the mining of online monitoring water indicators and the relating multiple time series data,this paper presented research on water quality anomaly detection of the urban river,which combines water quality indicators prediction,water quality incident early detection and water pollution sources identification,giving early warning information and decision support for possible pollution events.The main contents and innovations are summarized as follows:(1)The water quality indicators prediction method based on multiple time series data was studied.Considering the water quality time series data show dynamic characteristics covering non-linear and non-stationary,etc.affected by various external factors,a dual attention mechanism with long-short-term memory network(LSTM)method was introduced to predict water quality indicators.The LSTM was used to capture long-time temporal information.The dual attention mechanism was introduced based on LSTM to capture the time-series fluctuations of water quality,and adaptively adjusted attention weight between historical time points and dependencies of multiple series on the current prediction LSTM cell.Thus,key points from the historical water quality time series were extracted to reveal the implicit water quality fluctuations.The performances of different prediction models were compared using the real monitoring data of urban river water quality.Therefore,the method provides support for water quality anomaly detection.(2)The method of detecting water quality anomalies based on water quality indicators prediction models was studied.As the fact that the traditional fixed threshold methods tend to be unreliable caused by unreported and misreported,this paper,combined with the water quality prediction model,presented a water quality anomaly detection method including fluctuation features and Isolation Forest to fully utilize the fluctuation characteristics of multiple water quality time series.To improve the robustness against noise,the fluctuation entropy was applied to obtain the self-similarity of indicator fluctuations,then the fluctuation correlation features were introduced to analyze the correlation similarity of the multiple time series caused by water quality incidents.Finally,the characteristics of multiple fine-grained fluctuations are combined to determine the probability of water anomaly and to excavate the anomalies implied in complex water quality fluctuations so as to improve the detection rate of abnormal events.(3)Based on the results of water anomaly detection,a sample database of common pollution sources was constructed.The non-linear correlation between multiple indicators was excavated by stacking sparse auto-encoder(SAE)according to the joint response characteristic between multiple water quality indicators and different pollution sources.Combined with the K-means++method,a feature database with pollution sources differentiation ability and independence was constructed,which was utilized in the pollution sources identification method based on cosine similarity.Experimental analysis has been implemented to test the performance of the proposed method;after preliminary comparison and verification,the proposed method could identify the pollution sources such as industrial wastewater,domestic sewage,mud water,livestock water and poultry breeding wastewater with different degrees of pollution.It can also detect unknown pollution that would be put into feature database in time.The algorithm has certain adaptability in the actual river water quality monitoring scene.In general,this paper aims to study water quality anomaly detection of the urban river using multiple time series data.Experimental results and analysis demonstrate that the proposed methods in this thesis show advantages on the water quality anomaly detection and pollution sources identification,and this research provides a practical online anomaly detection approach for intelligent and automated real-time early detection systems.