(522h) Integrating Machine Learning-Based Sensing with Model Predictive Control

With the great strides in applications of artificial intelligence in recent years, an important question in chemical engineering has been the utility of machine learning for chemical processes [1]. Interesting applications of artificial intelligence have been in areas such as image recognition [2] and natural language processing [3], which perform a type of sensing for which controllers in the process industries have not traditionally been made responsive. For smart manufacturing [4], it is interesting to consider whether these modes of sensing may enhance efficiency and safety by, for example, identifying hazards which would typically only be noticed by humans and not by process sensors, or by enabling controllers to adapt their behavior to the sentiments of engineers with respect to the process response under the control actions. Developing controllers with appropriate responses to sensing techniques related to vision and speech, however, introduces fundamental challenges, in particular for handling the uncertainty in the results of image or language processing techniques derived from training models with large data sets.

We develop a framework for the design of speech and image-responsive control for the chemical process industries by elucidating various control-theoretic questions for model predictive control which arise in this context and providing potential answers for processes with dynamics described by classes of nonlinear first-order ordinary differential equations. We focus on model predictive control designs with Lyapunov-based stability constraints [5] to examine stability and feasibility considerations in this context. As an initial example, we consider the case that an image recognition algorithm is trained to recognize leaks in piping. An MPC that is responsive to this method of leak detection may be responsible for selecting whether or not it should modify the process dynamics after a leak is detected by preventing flow from entering the leaking pipe. If the leak has not significantly changed the process dynamics, then the MPC’s actions must maintain closed-loop stability and recursive feasibility before the switch in the dynamics and after the switch in the dynamics; however, if the leak has changed the process dynamics but the image recognition procedure does not detect the leak with certainty, the MPC must seek to maintain closed-loop stability in a case where one of several models might describe the current process dynamics and the future process dynamics if the MPC adjusts the flow into the pipe. Through this and other examples, we lay the groundwork for analyzing the use of speech and image recognition for enhancing process safety and efficiency through control in the process industries.

[1] V. Venkatasubramanian. The promise of artificial intelligence in chemical engineering: Is it here, finally? AIChE Journal, 65:466-478, 2019.

[2] K. He, X. Zhang, S. Ren and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 770-778, Las Vegas, Nevada, 2016.

[3] A. Kumar, O. Irsoy, P. Ondruska, M. Iyyer, J. Bradbury, I. Gulrajani, V. Zhong, R. Paulus and R. Socher. Ask me anything: Dynamic memory networks for natural language processing. In Proceedings of the 33^rd International Conference on Machine Learning, New York, New York, 2016.

[4] J. Davis, T. Edgar, R. Graybill, P. Korambath, B. Schott, D. Swink, J. Wang and J. Wetzel. Smart manufacturing. Annual Review of Chemical and Biomolecular Engineering, 6:141-160, 2015.

[5] M. Heidarinejad, J. Liu and P. D. Christofides. Economic model predictive control of nonlinear process systems using Lyapunov techniques. AIChE Journal, 58:855-870, 2012.