Process Monitoring And Quality Control For Batch Processes Based On Data-driven Methods

Having the advantages of quick response to changing market,deep customer-oriented manufacturing mode,and producing high-value-added products,batch processes have experienced rapid development and beome increasingly popular in modern industries,such as fine chemical industry,biomedical,food manufacturing,and semiconductors.With the development of the society and improvement of living standards,the requirements of high-quality products are greatly increased.To solve it,various measures are taken to ensure the product qualities and maintain the profits by enterprises.Especially,the scale of modern industrial systems becomes more and more large and the degree of automation is increased.However,the possibility of process failture is greatly increased with the complexities,which may cause huge economic loss and catastrophic consequence.Therefore,process safety has become an important and indispensable part of modem industrial systems to reduce the risk and improve the product quality.In general,the current process monitoring methods fall into one of the two categories:model-based methods and data-driven methods.For model-based methods,it is difficult and time-consuming to develop accurate mathematical models.Besides,model mismatch is a great challenge for this kind of methods.Therefore,the model-based process monitoring methods may not work well.With the development of data acquisition and storage technologies,data-driven process monitoring methods paly a more and more important role in batch processes and lots of studies have been reported.In fact,batch processes face the complex characteristics of multi-stage,multimode,cumulative quality effect,strong dynamics,and etc.However,the current methods can not deal well with these challenges.For example,phase partition results are influenced by tunable parameters,the data collection is assumed to be from a signle mode,and product qualities are controlled until the end of the batch.To solve the above-mentioned problems,a series of stage-based process modeling,monitoring,quality prediction and control methods have been developed as follows:1.To deal with the problem of phase parition,an iterative two-step sequential phase partition algorithm is proposed in the present work to overcome the disadvantages,including a lack of quantitative index to indicate transition patterns and tunable parameters that cannot be quantitatively determined.In the first step,with a given tunable parameter,initial phase partition results are obtained by checking changes of control limit of squared prediction error.Sequentially,fast search and find of density peaks clustering algorithm is employed to adjust the tunable parameter,i.e.,degradation degree,and update the phase partition results.These two steps are iteratively executed until a proper degradation degree is found for the first phase.Then the remaining phases are processed one by one using the same procedure.Moreover,a statistical index is quantitatively defined based on density and distance analysis to judge whether a process has transitions and when the transition regions begin and end.In this way,the phases and transition patterns are quantitatively determined without ambiguity from the perspective of monitoring performance.2.To solve the process monitoring problem of the multimode and multiphase batch processes,a novel process monitoring method based on between-mode similarity evaluation and discriminative information analysis is proposed without the limitation of single mode.First,by treating the entire batch data as an analysis unit,all batches are automatically classfied into different classes using dynamic time warping algorithm according to the data distribution.In this way,the differences along the batch direction are well distinguished through classification.Considering each mode may be different from phase to phase,a multimode-Fisher discriminant analysis algorithm is developed to investigate the causes between different modes within each phase.Specifically,a sparse manner is conducted to extract discriminative information,in which discriminative variables contributing to the discriminative information are directly identified.In this way,the essential causes between different modes in each phase are revealed.The proposed method provides a detailed insight into the inherent nature of multimode and multiphase batch processes.3.For the quality prediction issues of batch processes,a quality prediction method based on subspace decomposition is proposed with in-depth analysis of process characteristics.In this paper,a quantitative index is defined which can check ahead of time whether the product quality result from accumulation or the addition of successive process variations.For the product indices with cumulative quality effect,a phase-wise cumulative quality analysis method is proposed based on subspace decomposition which can explore the non-repetitive quality-relevant information(NRQRI)from the process variation at each time within each phase.NRQRI refers to the quality-relevant process variations at each time that are orthogonal to those of previous time and thus represents complementary quality information which is the key index to cumulatively explain quality variations time-wise.Treating the phase as basic unit,process-wise cumulative quality analysis is conducted where a critical phase selection strategy is developed to identify critical-to-cumulative-quality phases and quality predictions from critical phases are integrated to exclude influences of uncritical phases.It is feasible to judge whether the quality has the cumulative effect in advance and thus proper quality prediction model can be developed by identifying critical-to-cumulative-quality phases.4.For non-optimality statuses and quality control issue of batch processes,this study proposes an intelligent non-optimality self-recovery method with reinforcement learning.First,the causal variables that lead to the non-optimality are identified by developing a status-degraded Fisher discriminant analysis with consideration of sparsity.In this way,the non-optimal variables are quantitatively and automatically identified.Second,using the self-learning mechanism,a self-recovery method is proposed based on the reinforcement learning to automatically adjust the set-points of the causal controlled variables without accurate process models.The self-recovery action is taken iteratively through the Actor-Critic structure with two interactive neural networks until the expected status is achieved.Therefore,without any process knowledge,non-optimal causes are diagnosed and effective actions are taken to remedy the process to its optimal status.The efficacy of the proposed approaches is validated through two typical batch processes,i.e.,injection molding process and Penicillin fermentation process.In comparsion with other methods,the proposed methods show better performances.On the basis of this,some future research directions are discussed.