(576d) Topology-Aware Soft Sensor Modeling Leveraging Graph Neural Networks

Soft sensors play an important role in modern (bio-) chemical process operations. The reason for this is that soft sensor modeling allows to estimate critical process parameter that are hard to measure inline, e.g. product yields. Approaches to soft sensor modelling are divided into two categories, data-driven and first-principle based. A major advantage of data-driven approaches is that their performance scales with training data. As data becomes abundant in modern plants, data-driven approaches also become increasingly attractive. First principle based models allow to incorporate prior physical understanding of the underlying process. This is a current shortcoming of data-driven models, because they are black box in nature.

We propose a graph neural network based modelling framework to incorporate process topology into the soft sensor model. With this hybrid approach, we utilize the process topology to build a directed graph. This allows the application of neural networks to the time series data.

We develop a message passing-based graph neural network to model the graph topology underlying the process. We model unit operations as nodes and streams connecting unit operations as edges. Sensor measurements can then be directly embedded as attribute vectors in the corresponding part of the graph. This reflects their physical location in the plant.

We demonstrate our method on a computational case study and an industrial case. First, we train our method on the Tennessee Eastman process. Then, we consider an industrial spry dryer. Overall, we contribute towards closing the gap between data-driven and first-principle driven soft sensors. To this purpose, we introduce a platform technology based approach using a modelling framework that leverages the process topology as an inductive bias. In the future, we will apply our method online in an industrial settings for different types of processing plants, showing the power of this platform approach.