Dynamic Process Monitoring And Fault Diagnosis For Wastewater Treatment Process With Mixed Data Characteristics

As a country suffering from the shortage of water resources,China takes the prevention of water pollution very seriously.This issue would directly affect the health of residents,interfere with the industrial and agricultural production,and even threaten the national security.Therefore,it is particularly important to ensure the smooth operation of wastewater treatment plants.However,wastewater treatment processes(WWTPs)frequently encounter with abnormal events such as sensor failures,sludge bulking,toxicity shocks and ammonia inhibitions.These exceptions can seriously hinder the healthy and stable running of WWTPs.So,it is necessary to detect faults and eliminate their adverse effects in time by the means of process monitoring.Nevertheless,WWTPs are complex industrial processes,and often exhibit mixed data characteristics such as dynamic correlations,strong non-Gaussianity,strong nonlinearity and high dimension.These issues make traditional methods unsuitable to monitor WWTPs effectively.Therefore,this thesis proposes methodologies based on independent component analysis(ICA),slow feature analysis(SFA),long short-term memory(LSTM),variational auto-encoder(VAE),Granger causality analysis(GCA)and other multivariate statistics,artificial intelligence and signal processing algorithms for monitoring WWTPs with mixed data characteristics,which are described detailly as follows.1.A complex-valued slow independent component analysis(CSICA)is proposed to detect incipient faults raising from the dynamic processes with non-Gaussian and dynamic characteristics.This algorithm can extract potential features from the complex-valued matrix containing the raw data and their changing rates by resorting to a complex-valued ICA operation and a batch of phase shifts.These latent features are not only statistically independent of each other,but also capture slowly-changing patterns.Therefore,this algorithm not only considers the auto-correlations and con-correlations among monitored variables,but also takes the highorder statistics into account,thereby deeply mining hidden information.Finally,this algorithm together with the novel statistics proposed in this paper,namely ,, and their corresponding control limits,can effectively detect incipient faults.2.To deal with the issue of the strong nonlinearity,non-gauss and dynamics facing by process monitoring,a non-Gaussian temporal sequence variational auto-encoder(NG-TSVAE)is proposed in this paper.First,the overcomplete ICA algorithm is utilized to extract sufficient latent source signals containing high order statistic information from the raw data.Then,hidden states are further refined from temporal sequences of the latent source signals through a VAE network which is composed of multilayer LSTMs,thereby constructing the latent-state subspace and residual subspace.Finally,statistics LF and R are constructed for the latent-state subspace and residual subspace,respectively,to realize fault detection.In this paper,a numerical simulation and a real WWTP study case are used to verify the superiority of the proposed method in monitoring nonlinear non-Gaussian dynamic processes comparing with other traditional methods.3.This paper presented a novel distributed process monitoring method,which is termed as spatiotemporal-causality-based Multiblock WDCVA(SCMB-WDCVA),to tackle the issue of complicated coupling relationships and characteristics of observations extracted from largescale dynamic processes.First,this method can decompose a complex system into multiple subsystems within two steps,with helps of GCA and a directed topology containing the spatial information of the monitored variables.Then,a local canonical variate analysis mode and corresponding Wasserstein-distance based fault detection statistics can be constructed for each subsystem,thereby revealing local fault detection results of subsystems.Finally,local fault detection results are fused into a global conclusion by Bayesian inference.The superiority of this scheme in phase of fault detection is verified by case study of benchmark simulation model1(BSM1)and Tennessee Eastman process(TEP).4.To realize fault diagnosis and isolation for non-Gaussian dynamic and large-scale dynamic processes,the differential mapping reconstructed contribution plot(DM-RCP)and a novel hierarchical fault diagnosis and isolation method are proposed,respectively.One the one hand,the DM-RCP can eliminate the negative influence of smearing effect and inherent contributions derived from normal working conditions on fault diagnosis by reconstruction and difference operations.Therefore,the DM-RCP together with the CSICA algorithm can provide reasonable diagnosis conclusions for minor faults in non-Gaussian dynamic processes.One the other hand,the proposed hierarchical method can construct the variate faulty indices and block faulty indices for diagnosing the faults detected by SCMB-WDCVA at the systematic and variable levels,respectively.Further over,with the assistance of GCA,the two abovementioned fault diagnosis method can further locate the root cause of faults and visualize their propagation paths.The effectiveness of these two fault-diagnosis-and-location methods are verified by numerical model,BSM1 and TEP.In the conclusion,this paper also discusses other unsolved issues of process monitoring and prospects future research directions,while summarizing all the research contributions.