(299d) Machine Learning-Based Model Predictive Control of Nonlinear Processes Using Noisy Data

Modeling large-scale, complex nonlinear processes has been a long-lasting challenge in process systems engineering. In recent years, machine learning techniques such as neural networks (NN) have been successfully applied to solve regression/modeling problems based on datasets from industrial process operation or numerical simulations in process engineering, when a first-principles model is difficult to obtain. As machine learning algorithms learn process models from massive training data, the performance of machine learning models depends heavily on the quality of datasets. Training on defective datasets, such as missing values, noisy data, and outliers, may causes performance degradation because NN can result in inaccurate predictions for missing data points and may easily overfit to the label noise or outliers. However, real-world data often involve noisy patterns due to, for example, sensor and actuators noise in chemical plants, and therefore, pose challenges for machine learning modeling and its applications in chemical industry such as chemical process control. Therefore, how to efficiently address the issue of noisy dataset remains an important research topic within the context of NN modeling.

Researchers have been using segmented methods to tackle this problem [1,2,3]. For instance, some methodologies assume that the noise has a specific structure (e.g., Gaussian noise), or they are based on first-principles models and depend on carefully parameter tuning. With regards to neural network training process, noise correction layers have also been introduced to mitigate the impact of noise [4]. In this work, we will start with the implementation of data smoothing techniques to clean the datasets from industrial noisy measurements. Subsequently, we will develop noise-tolerant NN training algorithms by, for example, designing cost functions and NN layers, based on statistics information (e.g., distribution) of noisy data such that the NN models do not overfit to the label noise. The resulting NN models will be used in model predictive control (MPC) to control the operation of industrial processes. Finally, the performances of noise-tolerant NN models and of the novel NN-MPC will be demonstrated through a simulated ammonia plant in Aspen Plus Dynamics [5], for which, we will use industrial time-series data under normal operation to determine noise levels of the process measured variables and add that noise level to the outputs of the Aspen Dynamics simulators for these processes in order to evaluate the performance of the proposed control methods under realistic operation conditions.

[1] Leibman, M.J., Edgar, T.F. and Lasdon, L.S., 1992. Efficient data reconciliation and estimation for dynamic processes using nonlinear programming techniques. Computers & Chemical Engineering, 16, 963-986.

[2] Câmara, M.M., Soares, R.M., Feital, T., Anzai, T.K., Diehl, F.C., Thompson, P.H. and Pinto, J.C., 2017. Numerical aspects of data reconciliation in industrial applications. Processes,5,38 pages.

[3] Patwardhan, S.C., Narasimhan, S., Jagadeesan, P., Gopaluni, B. and Shah, S.L., 2012. Nonlinear Bayesian state estimation: A review of recent developments. Control Engineering Practice, 20(10), pp.933-953.

[4] Li, J., Wong, Y., Zhao, Q., and Kankanhalli, M. S., 2019. Learning to learn from noisy labeled data. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 5051-5059).

[5] Zhang, Z., Wu, Z., Rincon, D. and Christofides, P.D., 2019. Operational safety of an ammonia process network via model predictive control. Chemical Engineering Research and Design, 146, 277-289.