Dissanayake, Susantha ORCID: https://orcid.org/0000-0002-0953-379X, Manivel, Mukundan, Khan, Muhammad Yousaf, Sharma, Atul ORCID: https://orcid.org/0000-0002-6756-4148, Premarathna, Chanaka Prasad, Salehi, Hamid ORCID: https://orcid.org/0000-0002-2516-6619 and Bradley, Michael (2026) Physics-regularised synthetic data generation for predictive maintenance of filtration systems: a hybrid framework for small experimental datasets. Chemical Engineering Science, 339 (Part D):125434. pp. 1-19. ISSN 1873-4405 (Print), 0009-2509 (Online) (doi:10.1016/j.ces.2026.125434)

Preview

PDF (Open Access Article) 54851 DISSANAYAKE_Physics-Regularised_Synthetic_Data_Generation_(OA)_2026.pdf - Published Version Available under License Creative Commons Attribution. Download (6MB) | Preview

Abstract

Predictive maintenance in chemical engineering is often constrained by the limited availability of run-to-failure data, as industrial equipment is typically serviced before catastrophic failure occurs. This challenge is particularly significant for filtration systems, where severe degradation and terminal blockage are rarely captured experimentally. To address this limitation, this study proposes a physics regularised hybrid framework that integrates first principles filtration modelling with machine learning. A small experimental dataset (N=300) of filter pressure drop measurements was used to calibrate a mechanistic filtration model, which was then used to generate synthetic degradation trajectories covering the complete fouling process, including operational pressure-drop threshold. Stochastic noise was incorporated to represent realistic process variability and produce an expanded training dataset. Four physics regularised generative models, PR-VAE, PR-GAN, PR-TimeGAN, and PR-DDPM, were evaluated for their ability to generate physically realistic filtration behaviour. The synthetic datasets were then used to train Long Short-Term Memory (LSTM) and Temporal Convolutional Network (TCN) classifiers using a Train on Synthetic, Test on Real (TSTR) framework. Among the models evaluated, the PR-DDPM with TCN achieved the highest blockage forecasting accuracy of 98.8% on unseen experimental data. PR-TimeGAN effectively preserved local temporal consis ency, whereas PR-DDPM more accurately captured complete degradation trajectories and operational pressure-drop threshold. These results demonstrate that physics regularised generative modelling can overcome limited experimental data and enable reliable predictive maintenance for industrial filtration systems.

Item Type:	Article
Additional Information:	The processed experimental datasets, Python implementation scripts, neural network architectures, hyperparameter configurations, random seed tracking records, and trained model checkpoints generated and analysed during this study are available from the corresponding author upon reasonable request, subject to institutional data repository restrictions and proprietary agreements.
Uncontrolled Keywords:	physics-regularised synthetic data, generative data augmentation, cake filtration, predictive maintenance, blockage forecasting
Subjects:	Q Science > Q Science (General) T Technology > T Technology (General) T Technology > TP Chemical technology
Faculty / School / Research Centre / Research Group:	Faculty of Engineering & Science Faculty of Engineering & Science > School of Engineering (ENG)
Last Modified:	30 Oct 2026 16:16
URI:	https://gala.gre.ac.uk/id/eprint/54851

Actions (login required)

View Item

Downloads