Research On Data-driven Process Monitoring And Fault Diagnosis Methods For Wastewater Treatment Process

With the rapid development of society,the pollution of industrial has been received more and more attention.The safe and stable operation of wastewater treatment plants(WWTPs)is very important for the energy conservation and environmental protection.However,the actual WWTPs usually work in the harsh environment.The sensor failure,controller failure and the process failure(sludge bulking,toxic shock)are easy to occur.In particular,the most medium-sized WWTPs of rural areas are more prone to accidents.If the fault can not be checked in time,it will not only cause the huge economic losses,but also cause the secondary pollution to the environment.Moreover,the wastewater treatment system is a complex industrial system,which covers physical,chemical and biological reactions.In addition,the uncertain external disturbance makes the process monitoring more difficult.Therefore,this paper makes an in-depth study on the process monitoring of wastewater treatment from three aspects:data,variables and models.According to the demand of WWTP,the paper based on the data-driven method to carry out the research,the corresponding research results and innovations can be summarized as follows:1.To solve the problem that quality-related faults of WWTPs disturbed by noise are not easy to detect,the unsupervised robust adaptive canonical correlation analysis process monitoring method is proposed,termed as Rab-CCA.Rab-CCA combines a robust decomposition technique,which can decompose the corrupted matrix into a low rank matrix and a sparse kernel matrix.Subsequently,a new criterion function is established to achieve the purpose of multi-objective optimization,in such a way that the information of the low-rank matrix and sparse matrix can be fully extracted.In order to monitor the quality-related faults,the T~2 and SPE control charts are reconstructed,and a new adaptive statistical control limit is embed in Rab-CCA.Rab-CCA can monitor the effluent quality-related faults under the strong noise interference.2.To deal with the non-stationary and over-complex(with nonlinear,multiplicative faults,etc.)characteristics of process industry,the process monitoring method of unsupervised extended nonlinear canonical correlation analysis,termed as M-NAKCCA,is proposed.Firstly,the traditional CCA method is re-boosted as a new method,M-NAKCCA,to better nonlinear fault detection.Moreover,a matter-element model(MEm)is assimilated into M-NAKCCA to make the information more refined.Moreover,to solve the time-consuming problem of high-dimensional matrix decomposition,the technique of Nystro?m is embed in M-NAKCCA to estimate the low rank approximation of Gram matrix effectively.Finally,the T~2 control chart is reconstructed by using the residual of input-output space to monitor the system quality-related faults.The experimental results show that M-NAKCCA is more effective than other three methods in quality related fault detection.3.To solve the problem of time-consuming and maintaining the performance of the model under complex working conditions,a process monitoring method of supervised forecastable component-support vector machine is proposed(OFC-SVM).Firstly,to reduce the time-consumption of the SVM model,the forecastable component analysis is embeded in OFC-SVM for feature extraction.In addition,the Fore CA algorithm uses information entropy to measure the amount of information,which relaxes the condition that variable data must obey Gaussian distribution.Secondly,in order to keep the performance of the monitoring model stable.The quadratic grid search(GS)algorithm is utilized to optimize the parameters of the proposed method.Finally,to properly evaluate the method performance,a new evaluation index is proposed,named Pre-alarm Rate(PAR),aiming to achieve the quantitative trade-off between false alarm rate(FAR)and missed alarm rate(MAR).4.To deal with the problems of missing labeled training data and the difficulty in locating faults,the paper carried out in-depth research from the data level,variable level and model level.Firstly,to solve the problem of missing labeled data,a supervised cross-spatiotemporal relevance vector machine is proposed(named EAdsp B-TLM).EAdsp B-TLM improves and expands the traditional RVM model,and has the following advantages:Firstly,the probabilistic relevance vector machine(Pr RVM)under the Bayesian framework is re-derived so that it can be used to forecast the plant operating conditions.Secondly,the Pr RVM method is extended under the sparse Bayesian learning framework,and the transfer learning is integrated into the sparse Bayesian learning framework to provide its transfer ability.Finally,the source domain(SD)data are re-enabled to alleviate the issue of insufficient training data.In addition,in order to solve the problem that it is difficult to locate the root variable of sludge bulking in wastewater treatment process.The moving average residual reconstruction contribution plot-Granger causality analysis method is developed,termed as Mard RCP-GC.Mard RCP-GC first reconstructs the traditional contribution plot from the perspective of improving the signal-to-noise ratio.Then,Mard RCP-GC uses the new VS-R variable selection method and Granger causality analysis to locate the root cause of the abnormality.The EAdsp B-TLM model and Mard RCP-GC used for sludge bulking detection and root cause location in WWTPs effectively.5.To solve the problem of negative transfer caused by the mismatch between the source domain(SD)data and the target domain(TD)data,and the performance of model is difficult to maintain stability under time-varying conditions.A novel supervised optimization cross-spatiotemporal canonical correlation-vector machine for process monitoring,termed as CS-Ad Boost Tr LM,is developed.Firstly,a cross-spatiotemporal canonical correlation analysis(CCA)method is designed,which is a feature-based transfer learning method constructed on the basis of CCA method.Then,the instance-based transfer learning and feature-based transfer learning are embedded into the CS-Ad Boost Tr LM to make it have the stronger learning ability.Finally,the domain adaptation between SD data and TD data is realized to overcome the problem of negative migration.In addition,the invariant parameters of CS-Ad Boost Tr LM cannot be used to handle the time-varying conditions.Moreover,the sub-models generated by iteration are too sensitive to the parameters.Particle swarm optimization is selected to optimize the local detection models,in such way that the integrated detection model can converge to the optimality globally.Finally,Conlcusions,and the future research has prospected.