Data-driven Fault Diagnosis And Multiple-step Prediction Of Wastewater Treatment Processes

With the acceleration of industrialization and urbanization as well as the continuously bloom of population,water pollution is becoming more and more serious,which can be alleviated is to purify wastewater properly.Wastewater treatment process(WWTP)is a complex biochemical reaction process,involving a large number of automated instruments,machinery and electrical equipments,but also vulnerable to interference and influence of seasons,weather,flow and many other factors,resulting in failure frequency with wastewater treatment processes,excessive wastewater discharge,even drinking water pollution.To strengthen the process monitoring and fault prediction in advance,there is a need to solve these problems for wastewater treatment process.In this paper,data-driven fault diagnosis and multi-step prediction are studied,including data quality variables and input process variables of wastewater treatment processes.1.A four-layer radial basis function(RBF)neural network with prior knowledge based soft sensor is proposed.In order to make use of lost information in hybrid modeling process,a priori layer is introduced between the input and the first implicit layer of the traditional three-layer RBF neural network,thereby facilating effectively description of the input-variable obtained by mechanism,statistics,or artificial intelligence algorithm.The weights between input layer and priori layer,which directly reflect the input-variable importance,can change clustering shape of RBF neural network,thus giving more reasonable cluster.Then,an adaptive four-layer RBF neural network for soft sensor using recursive partial least aquares(PLS)is proposed.The weights between input layer and priori layer of the adaptive model is updated online by recursive PLS,which is different from constant weights of the normal four-layer RBF neural network.2.An adaptive soft sensor using multiple heterogeneous model ensemble learning is proposed.In order to describe the mixed-linear-nonlinear characteristics of wastewater treatment processes,a set of linear and nonlinear heterogeneity models is constructed.One of issues in combining a set of learned models using averaging is multi-collinearity problem.To deal with time-varying behavior of a process on one hand and to cope with multi-collinearity problem when implementing ensemble learning on the other,a recursive PLS method is used as the weighted combination method.3.A multi-variable fault diagnosis method based on variational Bayesian mixture factor analysis(VBMFA)is proposed.In order to describe the mixed-linear-nonlinear characteristics of wastewater treatment process,multiple factor analysis models are established for global data.The variational Bayesian mixture factor analyzers is constructed by weighted average of different local factor analyzers.The proposed method is not only able to automatically determine the number of local models,but also can avoid the over-itting problem.The traditional monitoring indices T~2 and SPE are improved using weighted average strategy based on the VBMFA.The improved weighted factors can be changed adaptively according to the real-time dynamic process behavior.The implementation of adaptive mechanism for proposed method is not by their mean and variance,but rather depending on their weight update,which reduces on-line computational complexity on one hand and avoids off-line poor dynamic performence on the other.What's more,the delay data is introduced into the original data matrix,which can enhance the dynamic performence of the model.4.A Gaussian process multi-step fault prediction with direct-recursive strategy is proposed.In order to describe uncertainty arising from wastewater treatment processes,The Gaussian process multi-step prediction model for effluent index is established.The standard deviation of Gaussian process can describe the uncertainty of the process.Therefore,the interval of effluent variable can be predicted by multiple standard variance,which used to estimate the upper and lower time limit of fault occurrence.Then,the nonlinear fitting ability of Gaussian processes with different kernel functions is studied.At last,the performance of multi-step prediction based on Gaussian processe,RBF and ARMA models is compared using direct,recursive and direct-recursive methods respectively.5.A multi-step fault prediction based on auto-associative neural networks is proposed.Firstly,the monitoring model and monitoring index are analyzed,and a nonlinear mapping model is constructed using the shallow and deep auto-associative neural networks.The high-dimensional vectors are projected onto the low-dimensional principal component vectors to extract non-linear features.A multi-variable fault diagnosis model is established by means of monitoring statistical indexes.Secondly,in order to alleviate specific assumptions about the form of the underlying distribution,kernel density estimation is used to estimate the control limit,which is a non-parametric estimation.Thirdlly,considering the case of incomplete data,the minimization strategy was used to estimate the missing values,and the impact of the missing data composition and ratio on the diagnostic ability was discussed.Finally,the BSM1 model and the wastewater treatment dataset from the University of California database were used to validate the multi-step fault prediction model.At last,the summaries are obtained and pay the way for further research of this field.