Fault Detection And Effluent Quality Prediction Of Papermaking Wastewater Treatment Process Based On Improved PLS

The emission of wastewater from papermaking industry takes a high proportion in the total industrial effluent,which directly impacts the water pollution control.In papermaking wastewater,the water quality indexes such as chemical oxygen demand concentration,biochemical oxygen demand concentration,microorganism concentration,and toxic substance concentration are critical to reflecting the real-time water quality and optimizing the strategy of online control and management in wastewater treatment process.However,the wastewater treatment process(WWTP)is a typical biochemical process,which is generally characterized by time delay,non-stationarity,strong interference,hostile working environment and complex mechanism.Therefore,the traditional measurement methods are not applicable due to their high cost or low reliability.In this case,the data-driven prediction model provides a better choice for online measurement of water quality indexes.In this paper,for the high dimensionality,nonlinearity,stochasticity,and dynamic features of papermaking WWTP data,the improved partial least squares is used as the main modeling method.The main content of research includes:1.Considering that not all the collected data of WWTP are beneficial to prediction model,the dynamic concurrent kernel partial least squares(DCKPLS)is used for fault detection to ensure the validity of modeling data.First,the augmented matrices and kernel technique are applied to partial least squares(PLS)to achieve the desirable interpretation of nonlinear and dynamic characteristics in process data.Then,the concurrent partial least squares(CPLS)is constructed by dividing the input and output data spaces into five subspaces,which effectively overcomes the inherent limitation of PLS decomposition and realizes the comprehensive process monitoring.Finally,model performance is evaluated by the simulated sensor faults in WWTP.The results show the higher fault detection rates of DCKPLS model than the other counterparts.2.Taking into account the complex characteristics of WWTP data,this paper presents a composite approach integrating the augmented matrices with improved PLS algorithm for the prediction of water quality.First,two probabilistic models,Gaussian process regression and relevance vector machine,are used to rebuild the inner function between each pair of PLS latent variables,which are denoted as GPR-PLS model and RVM-PLS model,respectively.Thus,the improved PLS models can be used to handle the high dimensionality,nonlinearity,and random features of process data.Second,the augmented matrices constructed by time-lagged values are embedded into GPR-PLS and RVM-PLS to construct their corresponding dynamic models.Based on the lagged variables,dynamic features,derived from the correlation between historical data and current data,can be interpreted efficiently.In fact,there are two ways for matrix expansion including finite impulse response(FIR)and autoregressive exogenous(ARX),which provide significantly different improvements for prediction model.Therefore,based on the correlation analysis of lagged variables,an adaptive algorithm of matrix expansion is proposed at last.Experimental results show the superiority of the proposed model in prediction accuracy and generalization capacity.