Dynamic Data Reconciliation for Improving the Prediction Performance of the Data-Driven Model on Distributed Product Outputs

Abstract

The product quality of some chemical processes is affected by those output variables exhibiting distributed characteristics. To enable process monitoring and quality analysis, effective measurements of distributed product outputs are particularly important. Because of the limitation and high cost of measurement techniques, soft sensors are gradually used in distributed process monitoring. In this work, for the soft sensor model developed by ensemble just-in-time kernel learning (EJKL), dynamic data reconciliation (DDR) is proposed and combined with EJKL to improve the accuracy of model prediction. First, the EJKL-based model is designed to implement online prediction of distributed outputs. Based on offline measured outputs with Gaussian distributed measurement noise, DDR is then developed for distributed production processes to reduce the prediction errors of the EJKL-based model. Furthermore, the improvement of DDR for prediction results under non-Gaussian measurement noise is also presented. Using the information from the model-based prediction and the offline measurements, DDR can reduce prediction errors through the maximum posterior probability estimation. The case studies of the model predictions on molecular weight distribution and crystal size distribution verify that the accuracy of the predicted outputs can be improved by the proposed method.

Enhancement of DDST-MFAC for tracking performance by using dynamic data reconciliation

“…Results demonstrate the significant noise suppression capability of DDR-DDST-MFAC, especially in scenarios with increased noise variance, surpassing KF-based methods. Notably, the presence of contaminated Gaussian noise in system output underscores the complexity of noise distributions in real chemical processes, necessitating consideration for future control method development [41][42][43]. Additionally, future efforts will focus on integrating DDR technology with advanced control methods to devise simpler and more efficient control strategies.…”

Particle Swarm Optimization-Based Model-Free Adaptive Control for Time-Varying Batch Processes

“…The coefficient may decrease when the bacteria progress from the exponential growth phase to the stationary phase, as shown in the plot at time 60 hours to 80 hours. In addition, considering the control problems of batch processes that may exist in the cases of expectation setting, repetitive operation, and under incomplete process information, further research directions include parameter optimization for data-driven setting tuning MFAC [35], data-driven control in a two-dimensional framework [36], control methods under incomplete information [37], control methods based on active learning [38] dynamic data reconciliation [39], etc.…”

Domain Compensation-Assisted Quality Inference Enhancement of Chemical Processes with Distributed Outputs

“…14 Moreover, dynamic data reconciliation is integrated to improve its prediction performance. 19 However, these methods may struggle to balance the model prediction performance and suitable hyperparameters. GPR is favored for its capacity to capture uncertainty with a few hyper- parameters.…”

“…Liu et al proposed an ensemble just-in-time kernel learning method for the two chemical processes with distributed outputs . Moreover, dynamic data reconciliation is integrated to improve its prediction performance . However, these methods may struggle to balance the model prediction performance and suitable hyperparameters.…”