Soft Sensor Based On Bayesian Network

Nowadays,industrial production has become more and more complicated and large-scale.To monitor the state of production and guarantee product quality,it is crucial to measure quality variables accurately and timely.As a matter of fact,the measurement of quality variables is usually obtained by analyzer or laboratory analysis.Unfortunately,the instruments are expensive while laboratory analysis often introduces large delay,which have brought great difficulties to real-time acquisition of quality-relevant variables.The aim of soft sensor is to establish an input-output mathematical model,in order to predict hard-to-measure variables such as component and content through easy-to-measure variables such as temperature,pressure and flow rate,etc.As computer technology is developing,industrial plants have accumulated a huge amount of history data reflecting the true operating conditions,which provides the essential data basis for soft sensor.Generally industrial processes are intricate,and strong nonlinearities and uncertainties are always presented.How to describe the mathematical relationship between quality-relevant and easy-to-measure variables has become a research hotspot in both industry and academia.Bayesian network is a kind of directed acyclic graph consisting of observable and hidden nodes.It is regarded as a combination of probability,causal reasoning and graph theory.Moreover,it has a gift for dealing with nonlinearities and uncertainties,and possesses a desirable ability of learning causality.Therefore,Bayesian network has been applied for soft sensor to estimate quality-relevant variables accurately and timely in the paper.Research has been made on the basis of Bayesian network,and our novel works can be summarized as follows:(1)To deal with the performance degradation resulting from variations of working condition and process drifts,a Bayesian network for adaptive soft sensor modeling is proposed.Our proposed method introduces adaptive methods to Bayesian network,applying a probabilistic model to accommodate process uncertainties,and updating models to adapt to time-varying changes.Case studies show our proposed methods have the potential to cope with missing data,and satisfactory performances have been achieved in spite of the existence of missing data.(2)To deal with high model complexity and accuracy deterioration caused by redundant variables,K2 algorithm has been utillized to complete structure learning and variable selection,such that the network structure can be simplified to some degree.Unfortunately,as the number of variables is increasing,the performance of K2 algorithm deteriorates significantly.To this end,the mutual information with K2 arlgorithm(MI-K2)is proposed to enhance the interpretation of K2 algorithm.Compared to the primary method,MI-K2 can shrink the scale of network rapidly,and the relationships will be demonstrated more clearly.Furthermore,it may reduce the time cost of structure learning and speed up the training process.(3)Distributed Bayesian network is proposed for quality prediction of large-scale plant-wide processes.The proposed method adopts a two-stage strategy--local modeling and model fusion to build a model.It is beneficial for dealing with the phenomenon that key information may be covered up in a complicated process.Due to fewer variables in a local model,the size of network can be much smaller.Thus the risk of underfitting can be reduced sharply.Moreover,the precision of prediction can be further promoted by means of model fusion.Finally,a series of research work is systematically summarized,and the future work is demonstrated in detail.