Transferring Graph Neural Networks for Soft Sensor Modeling using Process Topologies

TL;DR

This paper tackles the challenge of transferring data-driven soft sensors across chemical plants with different process topologies. It introduces a topology-aware spatio-temporal graph neural network that represents a plant as a flowsheet graph (unit operations as nodes, streams as edges) and uses a message-passing GNN per time step followed by a transformer to capture dynamics, enabling transfer learning. In a case study with two differently arranged ammonia synthesis loops, a model pretrained on one loop demonstrates zero-shot transfer to the other and superior data efficiency when fine-tuned with limited target data, achieving a $RMSE$ improvement of up to $24.15\%$ over training from scratch. The work suggests a practical path to reusable, plant-agnostic soft sensors and motivates future validation on industrial data and extension to multi-process deployments.

Abstract

Data-driven soft sensors help in process operations by providing real-time estimates of otherwise hard- to-measure process quantities, e.g., viscosities or product concentrations. Currently, soft sensors need to be developed individually per plant. Using transfer learning, machine learning-based soft sensors could be reused and fine-tuned across plants and applications. However, transferring data-driven soft sensor models is in practice often not possible, because the fixed input structure of standard soft sensor models prohibits transfer if, e.g., the sensor information is not identical in all plants. We propose a topology-aware graph neural network approach for transfer learning of soft sensor models across multiple plants. In our method, plants are modeled as graphs: Unit operations are nodes, streams are edges, and sensors are embedded as attributes. Our approach brings two advantages for transfer learning: First, we not only include sensor data but also crucial information on the plant topology. Second, the graph neural network algorithm is flexible with respect to its sensor inputs. This allows us to model data from different plants with different sensor networks. We test the transfer learning capabilities of our modeling approach on ammonia synthesis loops with different process topologies. We build a soft sensor predicting the ammonia concentration in the product. After training on data from one process, we successfully transfer our soft sensor model to a previously unseen process with a different topology. Our approach promises to extend the data-driven soft sensors to cases to leverage data from multiple plants.

Figures (5)

• Figure 1: Visualization of flowsheet graph representation of an illustrative flowsheet.
• Figure 2: Illustration of the model used for spatio-temporal modeling, here with a look back window of 5 time steps.
• Figure 3: Two ammonia synthesis loops are used in this work. In Process A (left) the feed is first sent to the reactor and then flashed. In Process B, the feed is first flashed and then sent to the reactor.
• Figure 4: Pretrained vs training from scratch for varying number of datapoints. The shaded area shows the standard deviation due to different weight initializations
• Figure 5: Visualization of test set (blue) vs fine-tuned model (orange, green). We visualize both the zero-shot performance (orange) and the model fine-tuned on 51 datapoints.