Research On Soft Sensor Methods For Industrial Processes Based On Latent Variable Models

The stability and safety of industrial processes have a decisive impact on the high-quality development of the economy and industries.Data-driven soft sensor technology provides an ef-ficient and low-cost solution for the quality variable prediction and process monitoring,which has been widely applied in the process industry.However,traditional data-driven soft sensor models can only capture the direct and deterministic relationships in data,making it difficult to adapt to the increasingly complex production processes of today.Based on latent variable models,this study investigates methods for building soft sensors to better understand the deep information embedded within data.By examining information from these latent variables,it ac-curately identifies the dynamic and uncertain aspects of industrial processes.This study details the development and algorithmic intricacies of the proposed model.The accuracy and applica-bility of the proposed techniques are verified through both numerical simulations and real-world industrial applications.The primary research contributions and innovative advancements of this study are outlined as follows.(1)To address the issues of data noise and process time delays,a soft sensor method based on Finite Impulse Response(FIR)filtering is proposed for predicting quality variables in indus-trial processes.This method utilizes FIR filters as latent variable models to extract refined and valuable latent variables from sequences of easily measurable variables,thereby ensuring the accuracy of soft sensor modeling and online inference.To move away from the traditional re-liance on process mechanisms in FIR filter design,conditional information entropy is employed to quantify the static association between latent variables and quality variables.The parameters of the latent variable model are optimized with the aim of maximizing this static association.The proposed latent variable model effectively eliminates redundant information in the data that is irrelevant to quality variables,thus enhancing the predictive accuracy and noise resistance of the soft sensor model.(2)In response to the lack of universality in the Wiener framework soft sensor models,a dynamic soft sensor method based on latent variable models is proposed.This method uses FIR filters as the latent variable model.Subsequently,a time series model is applied to describe the dynamic relationship between latent variables and quality variables.During offline train-ing,this dynamic relationship is quantified based on the recursive time series model,which has strong dynamic descriptive capabilities.After completing offline training,to reduce potential prediction divergence risks,a nonlinear moving average soft sensor model based on latent vari-ables is constructed.This model,utilizing information from latent variables within a specific time window,achieves real-time online prediction of quality variables,effectively enhancing the accuracy of the soft sensor.(3)From a theoretical perspective,the issues of prediction bias present in traditional re-cursive soft sensor models are analyzed.To address this issue,a recursive soft sensor modeling method based on Sequential Monte Carlo sampling has been designed.This method applies a generative latent variable model to capture the uncertainty in dynamic process changes.Subse-quently,it generates multiple trajectories for predicting the primary variable from a probabilistic inference perspective.By considering each potential scenario,the proposed algorithm can out-put an accurate prediction that reduces the risk of significant prediction biases.Moreover,an online updating method for soft sensors,based on resampling techniques,has been developed.This allows for real-time updates of current predictions,based on the online detection results of quality variables.(4)An invertible neural network is used to model the prediction residual distribution of the soft sensor model,applied to residual-based process monitoring problems.This method maps non-Gaussian residual signals into latent variables in a reversible manner,ensuring the complete retention of residual information after mapping.The reversible neural network is trained in an unsupervised manner,making the distribution of the latent variables approach a Gaussian distribution.Since traditional testing algorithms(like the T~2test and Q test)rely on the Gaussian distribution assumption,the proposed method can serve as a bridge between soft sensor residuals and testing algorithms,thereby improving the accuracy of process monitoring.