Research On Anomaly Detection Methods Of Urban Water Quality

With the development of the economy, accidents of water pollution occur frequently, the safety of drinking water becomes an urgent problem in our society. At present, the online water quality monitoring system based on threshold alarm usually can’t meet the needs of intelligent detection of water pollution events. Detection of water quality anomalies, early or timely warnings of the emergency water pollution events are very important and necessary research topics to guarantee the quality and safety of drinking water and to reduce the damage caused by water pollution events.The water quality prediction, water pollution event detection, and water anomaly information source and category judgment methods for urban water monitoring and management are studied in the thesis. The main work and results are listed as follows:(1) In order to master and further analyze water quality status and trends, and to find the abnormal of water quality and make early warnings, the prediction method of water quality is researched and discussed. The prediction method is based on a Self-optimizing RBF neural network model. An improved method by using a differential evolution algorithm for optimizing two parameters automatically is investigated to overcome the difficulty of determining the input order and expansion speed of the RBF neural network prediction model. Prediction results indicate that the method can reduce the effects of normal cyclical fluctuations of water quality, identify non-stationary fluctuation signals, and extract and provide suspicious abnormal signals for the further analysis. Simultaneously, the experiment results show that the method has high prediction accuracy and can reduce the cumbersome of setting parameters. The intelligent prediction of water quality is achieved.(2) Water quality anomaly detection method has a large rate of false positive results, especially for water quality data mixed with impulse noises. To solve this problem, the wavelet transform modulus maxima de-noising method is studied. The de-noising residual time series are obtained by comparing the real measurement values with the prediction values and using the wavelet transform method. If the new residual at one time is greater than the specific threshold, water quality is deemed as abnormal. Experimental results prove that the proposed algorithm has a lower false alarm rate.(3) Baseline changes of water quality data may cause false positive results. Therefore an anomaly classification method to distinguish baseline changes and water pollution events is presented. The polynomial fitting method is used to extract time series, which will be clustered into different groups by the FCM clustering method. And then the classification information would be saved in the water quality change knowledge base. If the new change matches with the one in the models knowledge base, the water change is claimed normal baseline change. Anomaly information discrimination experimental analysis is carried out by using water quality time series including flow rate, residual chlorine, pH and conductivity. The analysis results show that this method can distinguish baseline change and water quality event anomaly, thus further improve the performance of anomaly detection algorithms.Anomaly detection method of urban water quality researched in the thesis will contribute to the realization of real-time monitoring of water quality anomaly, and lay the foundation for constructing a highly automatized and informationized water quality early warning system.