(59u) Learning Dynamical Process Models Using Plant Data: A Real-World Case Study in the Sustainable Manufacturing of Insulation Products

Chemical plants constantly collect vast amounts of sensor data for process monitoring, analysis, and control. Such data can be noisy and incomplete (e.g., due to sensor failure), however, they contain intrinsic information about the process dynamics, especially during transients (startup, shutdown, product transitions, etc.). Therefore, learning models from such data, while challenging, potentially provides an opportunity to develop better controllers as well as optimize the process (e.g., to minimize energy demand). This talk will present the results of an industry-university collaboration to build such data-driven dynamical models.

In particular, in this talk, we will present the development of a neural ordinary differential equation-based model trained on real plant data of insulation products in an oven. The model predicts temperature readings of the product using oven-specific process variables (e.g., fan speed) and other context variables (e.g., operating parameters of upstream processes) as inputs. The development of this model allows for relating temperature-dependent properties with various process parameters.