(305a) Transfer Learning for Soft Sensor Modeling of Industrial Chemical Processes Using Bayesian Inference | AIChE

Other Sites & Tools

Technical Groups

(305a) Transfer Learning for Soft Sensor Modeling of Industrial Chemical Processes Using Bayesian Inference

Authors 

Xie, J. - Presenter, University of Alberta
Huang, B., University of Alberta
Dubljevic, S., University of Alberta
Soft sensors, as alternatives for hardware sensors, have been extensively used for the online estimation of crucial quality variables of industrial chemical processes. The main idea of soft sensors is to use easy-to-measure process variables (e.g., flow rate, pressure, temperature, etc.) and regression algorithms to estimate hard-to-measure quality variables (chemical concentration, emulsion flow water content, etc.) [1]. Compared to first principle-model-based soft sensors, data-driven soft sensors have gained popularity due to their simplicity, flexibility, and fewer requirements on process knowledge. In practice, industrial chemical processes often suffer from data sparsity issues, especially when at the early or new operating stage. On the other hand, the complex time-varying dynamics of industrial processes potentially lead to the distribution variation of collected monitoring data. As a result, the trained soft sensor models based on the original dataset via traditional statistical methods or conventional machine learning methods could not be directly implemented in a new environment. To address this issue, a transfer learning-based soft sensor technique is proposed by using Bayesian inference.

Typically, transfer learning methods can be classified into four categories, including: instance-based, feature-based, parameter-based, and relation-based [2-3]. In terms of how to transfer, namely, transfer mechanism, the methods can be mainly categorized into similarity-based and adversarial-learning-based approaches. Compared to those transfer learning methods that are based on similarity measures and transferring features, the proposed method is based on Bayesian inference and transferring model parameters. More specifically, the proposed transfer learning method is based on a dynamic latent variable model in the form of a linear discrete-time state-space model corrupted by Gaussian noises. Instead of using expectation maximation (EM) for point estimation, Bayesian inference is used, which ensures that the distribution uncertainties of the model parameters can be learned and updated in a dynamic manner, as in [4-6]. Different from the typical transfer learning methods, only source-domain data are needed and input and output measurement data from the target domain of interest are not necessary, during the training stage of the proposed algorithm. In the implementation or validation stage, limited output samples from the target domain can be incorporated without many extra technical difficulties or computation burdens. Moreover, we show that potential constraints in the model parameters or features (such as slowness in slow feature analysis [7-8]) can be incorporated into the proposed algorithm. Finally, we verify the effectiveness and applicability of the proposed method via a numerical example and an industrial steam-assisted gravity drainage dataset.

References

[1]. P. Kadlec, B. Gabrys, S. Strandt, Data-driven soft sensors in the process industry, Comput. Chem. Eng. 33 (4) (2009) 795-814.

[2]. S.J. Pan, and Q. Yang, 2009. A survey on transfer learning. IEEE Trans. Know.l Data Eng., 22(10), pp.1345-1359.

[3]. K. Weiss, T.M. Khoshgoftaar, and D. Wang, 2016. A survey of transfer learning. Journal of Big data, 3(1), p.9.

[4]. Y. Ma, S. Zhao, and B. Huang, 2018. Multiple-model state estimation based on variational Bayesian inference. IEEE Trans. Autom. Control, 64(4), pp.1679-1685.

[5]. Y. Ma, and B. Huang, 2017. Bayesian learning for dynamic feature extraction with application in soft sensing. IEEE Trans. Ind. Electron., 64(9), pp.7171-7180.

[6]. J. Xie, B. Huang, and S. Dubljevic, 2021. Transfer Learning for Dynamic Feature Extraction Using Variational Bayesian Inference. IEEE Trans. Knowl. Data Eng., 34(11), pp.5524-5535.

[7]. L. Wiskott and T. J. Sejnowski, Slow feature analysis: Unsupervised learning of invariances, Neural Comput., vol. 14, no. 4, pp. 715-770, 2002.

[8] R. Turner and M. Sahani, A maximum-likelihood interpretation for slow feature analysis, Neural Comput., vol. 19, no. 4, pp. 1022-1038, 2007.

Topics 

Process Automation & Control
Energy
Learn From Experience

More Conference Links