Data-driven intelligent monitoring system for key variables in wastewater treatm

Honggui Han *,Shuguang Zhu 3,4,Junfei Qiao Min Guo

1 College of Electronic Information&Control Engineering,Beijing University of Technology,Beijing 100124,China

2 Beijing Key Laboratory of Computational Intelligence and Intelligent System,Beijing 100124,China

3 Engineering Research Center of Digital Community,Ministry of Education,Beijing 100124,China

4 Beijing Laboratory for Urban Mass Transit,Beijing 100124,China

Keywords:Data-driven Soft sensor Intelligent monitoring system Data distribution service Wastewater treatment process

ABSTRACT In wastewater treatment process(WWTP),the accurate and real-time monitoring values of key variables are crucial for the operational strategies.However,most of the existing methods have difficulty in obtaining the real-time values of some key variables in the process.In order to handle this issue,a data-driven intelligent monitoring system,using the soft sensor technique and data distribution service,is developed to monitor the concentrations of ef fluent total phosphorous(TP)and ammonia nitrogen(NH4-N).In this intelligent monitoring system,a fuzzy neural network(FNN)is applied for designing the soft sensor model,and a principal component analysis(PCA)method is used to select the input variables of the soft sensor model.Moreover,data transfer software is exploited to insert the soft sensor technique to the supervisory control and data acquisition(SCADA)system.Finally,this proposed intelligent monitoring system is tested in several real plants to demonstrate the reliability and effectiveness of the monitoring performance.

1.Introduction

Rapid growth in population,urbanization,and industrialization has drawn the attention of many researchers towards the scarcity of clean water.Limited waterresourcesdemand engineersto design wastewater treatment systems.Nowadays,the increasing ef fluent quality requirements and tightening discharge regulations encourage the operators to exploit the real-time monitoring measurements of process variables in wastewater treatment process(WWTP)[1,2].However,most of the key variables are hard to be measured,such as the ef fluent total phosphorus(TP),ammonia nitrogen(NH4-N),chemical oxygen demand(COD),biological oxygen demand(BOD)and total nitrogen(TN)[3].Moreover,the concentrations of TP and NH4-N play an indispensable role in the control of ef fluent quality and prevent eutrophication of water body[4,5].Thus,the real-time monitoring values of these process variables are becoming more and more important for better understanding of the operational strategies of WWTP[6-8].In recent years,many techniques have been developed to analyze the key variables[9].For example,Gentili and Fick proposed an ion selective electrode method for determining ammonium concentration in[10].The results show that this selective electrode method is suitable for continuous determination.However,the main problems of this method are low sensitivity,long equilibration time and matrix interference[11].Xue et al.developed a fabricated micro fluidic method for real-time monitoring of ammonium concentration in[12].This method consists of a fluorescence detection system and a LED-induced fluorescence spectrometry technique.This method is more sensitive than the ion selective electrode and colorimetric method.However,this method is also time-consuming[13,14].Moreover,Chen et al.proposed a flow technique system for monitoring the ammonium concentration[15].This system combines the solid-phase extraction with sequential injection and a colorimetric method.This method can determine the ammonium in a simple,low cost and rapid way,as well as reduce the consumption of reagent and generation of toxic wastes[16].Although these above methods have made significant progress for measuring the key variables,the accuracy is still unsuitable for the operation of WWTP[17-19].

Recently,the data-driven soft sensors have been widely used in WWTP for real-time monitoring measurements of key variables[20,21].For example,Aguado et al.introduced a two multi-layer feedforward neural network(MLPNN)to monitor TP concentration in a pilot plant[22].This MLPNN-based soft sensor utilizes the available data ofon-line sensors to overcome the time-delay.Hong etal.designed a softsensor,based on a feed-forward back-propagation neuralnetwork(FBNN),to estimate the nutrient concentrations in the biological nutrient removal process[23].In order to improve the dynamic prediction performance,a splitnetwork structure is used to obtain satisfactory modeling performance.Moreover,Bagheri etal.developed a softsensor,based on the radialbasis function neuralnetwork(RBFNN)and MLPNN,to monitor both TP and NH4-N[24].The results show that the monitoring accuracy of MLPNN is better than that of RBFNN.However,these above soft sensors are failing to achieve available accuracy within the noise conditions[25].Meanwhile,most of these above soft sensors cannot be used in real WWTP[26,27].In fact,the monitoring systems,based on data-driven soft sensors,have been widely accepted and adopted for the industrial processes[28,29].And some researchers are trying to design a cost-effective platform for developing a monitoring system for WWTP in recentyears[30,31].Kim etal.proposed a monitoring system,based on a data-driven model,to measure the ef fluent TP in a full-scale lab plant[32].Moreover,Huang et al.integrated the datadriven soft sensor and the data acquisition and control system(DAC)for the real-time prediction of nutrient concentration[33].Meanwhile,the collected electrode signal was displayed on a computer monitor,and stored in the data file.However,these two systems are tested in a lab-scale reactor.Thus,the development of monitoring systems for the real WWTP is still with some limitations[34].

Motivated by the above discussions,in this paper,a practical intelligent monitoring system combining a data-driven soft sensor and data distribution service is developed for measuring the real-time values of key variables in WWTP.Meanwhile,the hardware is well equipped to ensure enough datasets and the proposed methods are compiled to monitoring software for practical application and to reduce computing complex.The main contributions of this paper are two-fold.

1)Using the data-driven strategy,a novel data-driven soft sensor,based on the fuzzy neural network(FNN)and the principal component analysis(PCA)method,is proposed for monitoring both of the concentrations of TP and NH4-N.

2)Utilizing the data distribution service technique,a hybrid data transfer software is further developed to insert the soft sensor technique to the supervisory control and data acquisition(SCADA)system.Moreover,the proposed intelligent monitoring system was verified in several real plants.

The remainder of the paper is organized as follows.The next section introduces the development of the proposed intelligent monitoring system in detail.In Section 3,several simulation and real application experiments are applied to investigate the effectiveness of the datadriven soft sensor model and the intelligent monitoring system.Finally,Section 4 summarizes the main conclusions.

2.Data-driven Intelligent Monitoring System

2.1.Hardware structure

The main focus of WWTP is to reduce the wastewater to the natural waters,as well as to meet the discharge regulations.Same as the other industrial processes,WWTP should be run in safe operation.In this study,an intelligentmonitoring systemhas been designed for obtaining the real-time values of key variables(TP and NH4-N)in WWTP.Fig.1 shows the schematic diagram of the proposed intelligent monitoring system along with the primary systems.This intelligentmonitoring system is equipped with some on-line sensors,such as the dissolve oxygen(DO),potentialofhydrogen(PH),temperature(T),oxidation-reduction potential(ORP),and total suspended solids(TSS).The output signals from the sensors are integrated and connected to a programmable logic controller(PLC)for transmitting primary indicators.The PLC system is interfaced with equipped sensors and collected reliable data with a fastresponse time.Moreover,the PLCsystemhas been connected through a serial port(RS 232)of the host computer,which uses the real-time data to calculate the values ofkey variables and also stores the data in the form of local file.

Fig.1.Schematic diagram of intelligent monitoring system.

In fact,the on-line sensors,with reliable accuracy,are used for the intelligent monitoring system.These sensors are also with short response time.Moreover,the chosen PLC,which is able to avoid missing data and keep the accuracy of the instrument,is the Siemens S7-200.

2.2.Data distribution service

The on-line sensors(i.e.,DO,PH,T,ORP and TSS),located in the intelligent monitoring system,are relevant to the key variables.The signals of these sensors are transferred to the SCADA system.Then,in order to integrate the soft sensor technique and the data transmission technique,the suitable data transfer software must be taken into account.In this study,a data distribution service,between the monitoring software and the SCADA system,is designed to realize data transmission,while reducing the response time.

The data distribution service is used to distribute data to the intelligentmonitoring systemfromthe SCADAsystem,which integrates several types of data from the hardware sensors and PLC in WWTP.In addition,the software,connecting with the data distribution service,is developed by the ole for process control(OPC)server.In the data distribution service,since the SCADA system and the monitoring software must be concordant,the names of parameters are confirmed through the OPC client in this study.The internet protocol(IP)of host computer and the ports are also configured.Meanwhile,the OPC client is provided to request data for the OPC server based on the OPC general agreement.After the application of the data distribution service,the on-line data can be collected.The configured IP is then used to identify the network address of host computer.

2.3.Soft sensor technique

2.3.1.Variable selection

In fact,the data obtained from the data distribution service are often high dimensional.Therefore,the first and most important step of the soft sensor technique is the data pre-processing.Meanwhile,the choice of secondary variables is a crucial stage for data pre-processing.The secondary variables of the soft sensor should be relevant to the key variables.In the soft sensor technique,the principal component analysis(PCA),which is the most popular method to map a high dimensional space to a low dimensional space,is used for selecting secondary variables.The procedure of PCA is illustrated in the following procedure.

where U is the score matrix,which contains β dominanteigenvectors,μ is the mean of allinput variables.Based on Eq.(1),the standard PCA can be rewritten as

Table 2 The selected input variables

where FAis the covariance matrix,Λis the corresponding diagonalentry in the associated eigenvalue.Then,it will be

where ρiis the contribution rate, λiis the eigenvalue value of the i th variable.In general,the contribution over 85%is accepted to explain the output variables.Therefore,the last few eigenvectors are discarded since their contributions to the key variables are negligible.

Based on the results of PCA and variance cumulative percent in Table 1,the contributions of five variables are over 85%and the key variables can be explained by these variables.Therefore,PH,TSS,ORP,DO and T are selected as the secondary variables of the data-driven soft sensor model for monitoring TP and NH4-N.Meanwhile,Table 2 gives a detailed description of these variables.

2.3.2.Model design

In this paper,a multi-input multi-output(MIMO)FNN is used to design the data-driven softsensor model.The outputs ofFNNare defined by

where y=[y1,y2,…,yM]Tis the output of FNN,xiis the input of the i th neuron in the output layer,M is the number of neurons of the RBF layer,k is the number of neurons of the input layer,W is the weight matrix and wm=[,,…,]is the connection weightbetween the m th outputlayerand the normalized layer,v is the outputofthe normalized layer,cij(j=1,2,3,…,Q)is the center of the neuron and σijis the width.

2.3.3.Adaptive second-order algorithm(ASOA)

To improve the performance of the intelligent monitoring system,the MIMO FNN is trained by an adaptive second-order algorithm(ASOA).The training process will stop when the training errors reach the desired value.In this study,the desired value is 0.01.According tothe computation procedure ofthe Levenberg-Marquardt algorithm,the updated rule of ASOA is

Table 1 The latent and percentage of the collected input variables

Fig.2.Schematic diagram of monitoring process.

where H(t)is the quasi Hessian matrix,G(t)is the gradient vector,I is the identity matrix which is used to avoid the ill condition in solving inverse matrix,and η(t)is the adaptive learning rate[35]

In this ASOA,the output parameter matrix W,the center vector c,and the width vector σ can be optimized simultaneously.The quasi Hessian matrix H(t)and the gradient vector G(t)are accumulated as the sum of related sub-matrices and vectors.

Fig.3.The predicting values of TP in the first plant.

Fig.4.The predicting values of NH4-N in the first plant.

where the sub-matrices Hm(t)and the sub-vectors ωm(t)are

where em(t)is the error of the m th neuron in the output layer at time t,and the Jacobian vector jm(t)is calculated as

With the computation procedure,the elements of the Jacobian vector jq(t)are calculated in[35].

Fig.5.The predicting error of TP in the first plant.

Fig.6.The predicting error of NH4-N in the first plant.

The proposed ASOA-FNN can be used for monitoring TP and NH4-N with better accuracy and computing time since the ASOA-FNN has the advantages of fast convergence and powerful searching ability.Moreover,the ASOA-FNN can reduce the computational complexity of the learning process and reach smaller testing error with much faster speed.

2.4.Intelligent monitoring system

The framework of the intelligent monitoring process is shown in Fig.2.The intelligent monitoring system consists of a SCADA system to collect data from WWTP,a data distribution service to transmit real-time data for monitoring software,a variable selection to select secondary variables,a data-driven soft sensor model based on FNN to calculating the key variables and monitoring software to achieve the monitoring results.In summary,the implementation procedures of the intelligent monitoring system are given as follows

When she recovered her senses she was more than ever convinced that he was dead, since even Melinette was no longer near her, and no one was left to defend her from the odious60 old Enchanter

Step 1 Take the SCADA system as data source.The SCADA system collects the data from the equipped PLC and the OPC server is employed to communicate with data distribution service.

Step 2 Design and configure the data distribution service.When the data distribution service is used to transmit real-time data for monitoring software,it should be configured by data deliver setting,OPC client and UDP.Meanwhile,it saves the real-time data in form of local file in the host computer.All of these parts have been complied into the data transfer software.

Fig.7.The predicting values of TP in the second plant.

Table 3 The performance comparison of different approaches

Step 3 Select the secondary variable.With the saved historical data,input variables associated with TP and NH4-N are selected depending on the PCA algorithm.Then,re-arrange the dataset based on the input variables and divide the dataset into training data and testing data.

Step 4 Adjust the parameters of the soft sensor model.Utilize the historical data to train FNN after initializing the parameters and adaptive learning rate.The parameters and adaptive learning rate are used for initializing the realmonitoring model.

Step 5 Monitor key variables.Acquire the new data samples from SCADA and use the softsensor technique to calculate the values ofTP and NH4-N.Moreover,a graphical user interface(GUI)of monitoring software is developed to display the monitoring results.

The procedure of the proposed data-driven soft sensor is programmed in MATLAB,and then is compiled to the monitoring software by C sharp(C#).Meanwhile,the designed data distribution service is also compiled by C#for implementation.The intelligent monitoring system is mainly combined by the two types of software.

In order to evaluate the performance of the data-driven soft sensor and intelligent monitoring system,the proposed soft sensor is off-line tested on the pilot WWTP and the full scale WWTP datasets to provide comprehensive performance analysis.Then,the designed monitoring system is applied to two real WWTPs and the results were compared with those of the other methods.The monitoring performance is evaluated by the root mean squared error(RMSE)and percentage prediction accuracy(P):

Fig.8.The predicting values of NH4-N in the second plant.

Fig.9.The predicting error of TP in the second plant.

where e(t)=yd(t)-y(t)is the mean error,y(t)and yd(t)are the output of FNN and the desired output at time t(t=1,2,…,N).

3.Results and Discussions

Fig.10.The predicting error of NH4-N in the second plant.

Fig.11.The performance of monitoring TP in WWTP1.

Fig.12.The performance of monitoring NH4-N in WWTP1.

3.1.Lab-scale WWTP

In this part,the proposed intelligent monitoring system is validated in two lab-scale plants.For the lab-scale plant,the biological reaction tank consists of an anaerobic reactor,an anoxic reactor and an aerobic rector.The blowers are installed to supply air to the aerobic reactor and the mixers are installed in the biological reaction tank to raise the efficiency of wastewater treatment.The capacities of these two labscale plantsare 800 m3·d-1and 1000 m3·d-1,respectively.In addition,there are different in fluents which will lead to the different processes and ef fluent parameters.To achieve the monitoring values of TP and NH4-N,the PH,TSS,ORP,DOand T sensors are installed in the two plants to collect the real-time data.

In order to show the effectiveness of the proposed intelligent monitoring system,four different methods are selected for a thorough comparison:RBFNN,back propagation neuralnetwork(BPNN)adaptive network based fuzzy inference system(ANFIS)[36]and MLPNN[24].For a fair comparison,the inputs and outputs of the neural networks are the same(the inputs are the values of PH,TSS,ORP,DO and T,the outputs are the monitoring values of TP and NH4-N)and the experiments have been repeated 20 times using the same data.

For the first plant,the prediction results of TP and NH4-N are shown in Figs.3-6.The prediction values of TP and NH4-N are depicted in Figs.3 and 4.The prediction errors are shown in Figs.5 and 6.The results in Figs.4-7 show that the proposed ASOA-FNN is able to estimate the key variables.To evaluate the performance of the proposed intelligent monitoring system,the testing RMSE and the mean accuracy are listed in Table 3.The details in Table 3 show that ASOA-FNN owns less testing RMSE and higher mean accuracy compared with other methods.

For the second plant,the prediction results of TP and NH4-N are obtained as shown in Figs.7-10.The prediction values of TP and NH4-N are depicted in Figs.7 and 8.The prediction errors are plotted in Figs.9 and 10.It is obvious that the prediction errors of ASOA-FNN are much smaller than those of RBFNN,ANFIS[36]and MLPNN[24].To evaluate the performance of the proposed intelligent monitoring system,the testing RMSE and the mean accuracy are given in Table 3.It can be concluded that ASOA-FNN is more accurate than other methods.

In the experiments,the proposed method predicts the key variables in the two lab-scale WWTP.According to the analysis of the results,the ASOA-FNN is able to monitor the key variables in two lab-scale plants.However,ASOA-FNN based monitoring system must be validated by the full-scale WWTP before its application.

Table 4 Summary of monitoring performance in the full-scale WWTP

Fig.13.The performance of monitoring TP in WWTP2.

Fig.14.The performance of monitoring NH4-N in WWTP2.

3.2.Full-scale WWTP

In this section,the designed intelligent monitoring system is applied to two full-scale WWTPs located in Beijing.The capacities of these two full-scale plants are 400000 m3·d-1in WWTP1 and 1000000 m3·d-1in WWTP2,respectively.The datasets of key variables and secondary variables are achieved by the hardware sensors,PLC and SCADA system.

The intelligent monitoring system achieves the acquisition and processing of the datasets and realizes the monitoring of the key variables,including ef fluent TP and NH4-N.The performance of the monitoring system can be seen from the GUI of the designed monitoring software.The performance of the system is tested from 1st Jan.to 31st Dec.2016.This is regarded as a sufficient length for testing the system since one year covers the seasonal and weather variations.The monitoring results of WWTP1 are shown in Figs.11 and 12.The monitoring values of TP are displayed in Fig.11 and the values of NH4-N are depicted in Fig.12.The red color and green color curve of the GUI represent trend values provide by the datadriven soft sensor technique and the current values of the TP and NH4-N analyzer.And the length of the curve is decided by system setting and running time.The errors of the monitoring results are calculated through the current values and trend values.The values of mean error(e(t)=yd(t)-y(t))and mean accuracy resulting from the monitoring system are listed in Table 4.

Meanwhile,in order to further validate the capability of the developed intelligent monitoring system,this system is also used in WWTP2.The monitoring values of TP are given in Fig.13.The monitoring values of NH4-N are shown in Fig.14.It is obvious that the trend values can track the current values,which represents the monitoring results with a small deviation.The mean error and mean accuracy of the monitoring system are shown in Table 4.It can be seen that the proposed intelligent monitoring system can obtain better accuracy than the other methods in these two full-scale WWTPs.

In addition,the response time of the intelligent monitoring system and analyzers is shown in Fig.15.The shortest response time of analyzers is 600 s since the detection is a complex chemical reaction process.The response time of the system is less than 1 min and the time delay is generated by data updates.Compared with the other methods,the response time of the system is shorter.

Fig.15.Comparison of the response time of the system and analyzers.

4.Conclusions

In this paper,an intelligent monitoring system,containing the soft sensor technique and the data distribution service,has been designed to monitor the key variables in WWTP.The proposed soft sensor model is based on ASOA-FNN and the inputs are selected by PCA.The data distribution service is developed to provide real-time data for the soft sensor model from the SCADA system.The appropriate TP and NH4-N datasets,as well as PH,TSS,ORP,DO and T datasets,are used for the proposed intelligent monitoring system.According to the results in lab-scale WWTPs,the data-driven soft sensor can predict the TP and NH4-N with suitable performance.Moreover,the intelligentmonitoring system is tested in full-scale WWTPs.The results demonstrate that the monitoring values fit the detection values well with small error and high accuracy.Especially,the response time of the proposed system is shorter than the other methods.It increases the ability to deal with the problems of real-time monitoring and control in full-scale WWTPs.