A Digital Twin Simulator of a Pastillation Process with Applications to Automatic Control based on Computer Vision

TL;DR

This work develops a digital-twin simulator for pastillation that generates realistic thermal images to train a CNN-based soft sensor (PaNet) predicting instantaneous temperature and flow, enabling automatic control of conveyor speed via a PID controller. PaNet-1D outperforms PaNet-2D in both temperature and flow predictions, with RMSEs around $1.25$–$1.38^{\circ}F$ for temperature and $0.33/\Delta t$ for flow, respectively, and saliency analyses show focused attention on the leading row. The control framework, tuned by VP-BO, achieves rapid convergence to temperature setpoints under different conditions, though some transient instability and belt-speed fluctuations indicate room for controller enhancements such as model-predictive or adaptive schemes. Overall, the digital twin demonstrates a data-driven pathway to automate pastillation processes, with potential extensions to multi-sensor sensing, energy efficiency, and multi-objective optimization in manufacturing.

Abstract

We present a digital-twin simulator for a pastillation process. The simulation framework produces realistic thermal image data of the process that is used to train computer vision-based soft sensors based on convolutional neural networks (CNNs); the soft sensors produce output signals for temperature and product flow rate that enable real-time monitoring and feedback control. Pastillation technologies are high-throughput devices that are used in a broad range of industries; these processes face operational challenges such as real-time identification of clog locations (faults) in the rotating shell and the automatic, real-time adjustment of conveyor belt speed and operating conditions to stabilize output. The proposed simulator is able to capture this behavior and generates realistic data that can be used to benchmark different algorithms for image processing and different control architectures. We present a case study to illustrate the capabilities; the study explores behavior over a range of equipment sizes, clog locations, and clog duration. A feedback controller (tuned using Bayesian optimization) is used to adjust the conveyor belt speed based on the CNN output signal to achieve the desired process outputs.

Figures (7)

• Figure 1: Overview of digital twin simulator for pastillation process. (a) Schematic of a typical pastillation process: a rotating shell deposits pastilles onto a conveyor belt; random clogging events introduce product flow variations. (b) A thermal camera sensor captures spatial temperature distributions of the conveyor belt in real-time; the simulator used models to generate realistic image data. (c) A trained computer vision model (the convolutional neural network PaNet) processes thermal images to predict the average instantaneous flow and temperature of the pastilles. (d) A feedback controller adjusts the conveyor speed to maintain the desired flow and temperature.
• Figure 2: Schematic representation of pastillation system. A rotating shell with speed $\omega$ and K evenly-spaced rows of $h$ nozzles extrudes pastilles at a temperature $u_0$ along the width, $L_x$ of a belt. As they move down the length, $L_y$, of the belt, the pastilles are cooled via cooling water sprayed onto the underside of the belt (at temperature $u_W$) and the ambient air (at temperature $u_{\infty}$).
• Figure 3: Simulation of nozzle clogs. (a) Empirical probability distribution used for determining the occurrence of a clog at each nozzle. (b) Representative distribution of pastille placement onto the belt.
• Figure 4: Performance of PaNets in predicting average temperature and pastille flow rate. (a, c) Regression parity plots for PaNet-1D and PaNet-2D, with the diagonal line indicating ideal predictions. The colorbar represents the logarithm of density. (b, d) Root mean square error (RMSE) for each fold in five-fold cross-validation.
• Figure 5: Saliency maps of PaNets for predicting average temperature. Panels (a) and (b) correspond to PaNet-1D, whereas (c) and (d) correspond to PaNet-2D. In (a) and (c), the earliest row of pastilles has reached the end of the conveyor belt; in (b) and (d), the earliest row is still in transit. For each panel, the top image shows the saliency map (darker regions indicate areas of greater relevance to the model), and the bottom image displays the corresponding thermal camera snapshot.