Outlier Detection And Prediction Of Water Quality Data Streams Using Wireless Sensor Networks

Water quality directly determines the quality and production in aquaculture.Aquaculture water quality monitoring is essential in intensive aquaculture.With the development of automation and intelligence of aquaculture,higher requirements of water quality monitoring in aquaculture are proposed.Based on the wireless sensor network water quality monitoring platform,this paper studies outlier detection,data fusion and prediction of the water quality data streams and other key factors in monitoring process.The machine learning,information theory and statistical method are utilized in the real-world water quality monitoring.The main contributions of this paper are as follows:(1)Outlier detection method in water quality monitoring.To detect the outlier of data stream caused by sensor faults or occasional events,we propose a relative density compensated improved SVDD(ID-SVDD)outlier detection based on observing and analyzing the feature of water quality data stream.Firstly,for compensating for the weakness of SVDD in analyzing data distribution,we import the relative density to the traditional SVDD.Then the improved Parzen-windows function is utilized to compute the relative density,and construct the ID-SVDD outlier detection model.Finally,the real-world water quality data streams are used to verify the performance of ID-SVDD,and different existing methods are used for comparison to test the detection performance.The experimental results indicate that ID-SVDD method has better detection performance than other methods.(2)Weighted support function data fusion method of dissolved oxygen.Takeing the dissolved oxygen as an example,the problem of inaccurate water quality monitoring with single sensor was analysized and studied.Based on analyzing the potential relation between homologous sensors,we propose a novel weighted support function fusion method(IDTW-ISD).Using the grey correlation method for reference,we got the ISD support function by adjusting the parameter K and ?.Different from the similarity measurement of two elements at a moment,IDTW-ISD measures the similarity between two time series in a time slot,and dynamic time warping(DTW)is used to compute the similarity.Meanwhile,the time series segment strategy is applied in IDTW-ISD.The experimental results show that the proposed IDTW-ISD weighted support function fusion method can achieve higher accuracy and efficiency.(3)Prediction of dissolved oxygen content using improved online sequential Extreme Learning Machine(OSELM).To solve the shortage of stochastic parameters,low stability and poor flexibility of traditional prediction algorithms,we propose a EFIG-OSELM hybrid prediction model.After analyzing the characteristics of dissolved oxygen time series,we get multi-modal information by decomposing the time series with EMD algorithm.To reduce the computation of prediction algorithm,fuzzy entropy is utilized to realize data reconstruction.And chaos sequence is applied to optimize the genetic algorithm.Finally,the EFIG-OSELM prediction model is established with reconstructed components.The hybrid models based on least square support vector machine(LSSVM),BP neural network are used as comparison,the prediction result,correlation coefficient and precision index of EFIG-OSELM are superior to the counterpart prediction models,which can satisfied the need of dissolved oxygen prediction in real world.(4)Prediction of dissolved oxygen content in aquaculture using clustering-based softplus extreme learning machine.To solve the shortage of low accuracy,more data interference and low efficiency of traditional prediction algorithms in precision and intellectualized aquaculture,we propose another dissolved oxygen prediction model based on clustering-based softplus function extreme learning machine(CSELM).This prediction model analyzed the relation between water parameters and meteorological parameters.And factor analysis method is used to evaluate the meteorology index.Meanwhile,DTW distance is utilized to calculate a custom variable time series similarity.Then cluster the data sets and construct similar datasets.Finally,based on the clustering,we obtain several sub-clusters.In these sub-clusters,sub-CSELM models are used to realize dissolved oxygen prediction respectively.Compared with PLS-ELM and ELM prediction models,the proposed CSELM prediction model can achieve higher prediction precision with fast convergence peed.(5)The development of water quality monitoring pre-warning system.To realize the intelligent aquaculture,we build a water quality monitoring platform.Based on the above outlier detection method and prediction method,the water quality monitoring system is developed which includes the personal computer client and mobile client monitoring softwares.It includes 6 modules,i.e.,water quality and meteorology data acquisition,information search,data analysis,pre-warning processing,equipment management and system management.The early warning system operates very well,which indicates that proposed algorithms can realize detection and prediction efficiently and reasonably.