E-Thesis 13 views 1 download
Steel Surface Parameterisation for On-line Quantification with Machine Learning Models / ALEXANDER MILNE
Swansea University Author: ALEXANDER MILNE
-
PDF | E-Thesis – open access
Copyright: the author, Alex Milne, 2026, Distributed under the terms of a Creative Commons Attribution NonCommercial 4.0 License (CC BY-NC 4.0)
Download (19.97MB)
DOI (Published version): 10.23889/SUThesis.71779
Abstract
Accurately estimating steel surface roughness parameters in real-time during industrial manufacturing is a critical challenge. Traditional post-production stylus-based measurements of surface roughness, specifically the arithmetic mean roughness (Ra), are slow and sample only small sections of a stee...
| Published: |
Swansea
2026
|
|---|---|
| Institution: | Swansea University |
| Degree level: | Doctoral |
| Degree name: | Ph.D |
| Supervisor: | Xie, X., and Tam, G. |
| URI: | https://cronfa.swan.ac.uk/Record/cronfa71779 |
| Abstract: |
Accurately estimating steel surface roughness parameters in real-time during industrial manufacturing is a critical challenge. Traditional post-production stylus-based measurements of surface roughness, specifically the arithmetic mean roughness (Ra), are slow and sample only small sections of a steel coil, leading to production inefficiencies, costly rework or downgrading of coils, and potential downstream delays in demanding just-in-time manufacturing environments due to re-production when quality is not met. While the on-line, non-contact laser reflection measurement device used by our partners offers a continuous monitoring solution, its accuracy is often insufficient compared to stylus-based ground truth measurements, negating its utility for process control.This research explores the application of machine learning (ML) to enhance the accuracy of surface roughness estimation from raw laser reflection data. By leveraging both limited la-belled and abundant unlabelled industrial datasets, this work investigates various ML-driven methods. These include: data imputation techniques aimed at improving existing calculation methods by addressing data quality issues; end-to-end models designed for direct mapping from sensor input to the stylus-equivalent Ra value; efficiency enhancements for promising model architectures; and the incorporation of representation learning via self-supervised learning (SSL) to effectively utilise unlabelled data and overcome labelled data scarcity.The findings demonstrate the substantial advantages of ML approaches over conventional methods. End-to-end deep learning models consistently achieve significantly higher accuracy in predicting Ra compared to the baseline sensor calculations. Our comprehensive comparative studies reveal that architectures such as Temporal Convolutional Networks (TCNs) and certain 2D Convolutional Neural Networks provide a robust combination of high predictive ac-curacy and computational feasibility for this specific industrial data type. Furthermore, lever-aging large unlabelled datasets via pretraining proved highly effective; representation learning strategies, particularly a novel task-specific coil classification pretext task and a TCN-based autoencoder, yielded models with enhanced generalisation and accuracy upon fine-tuning with limited labelled data. Methodological insights also include the necessity of a statistics-aware loss function for imputation tasks requiring the preservation of surface texture characteristics, and the successful GPU acceleration of the ROCKET model for enhanced throughput.The key contributions of this thesis therefore encompass not only the successful application and comparison of various ML techniques (imputation; end-to-end regression; and pretraining strategies like contrastive learning, autoencoding, and coil classification) to this specific industrial problem, but also methodological advancements such as the statistics-aware imputation loss and efficiency improvements to the ROCKET model.These findings show that machine learning models, particularly those enhanced with representation learning techniques, can significantly improve the accuracy of on-line surface rough-ness estimation. This paves the way for reliable real-time quality control, enabling faster process feedback, reducing waste, facilitating process optimisation, and ultimately opening possibilities for automated closed-loop control systems in steel manufacturing. |
|---|---|
| Keywords: |
Machine learning, Surface roughness, Ra, On-line measurement, Steel manufacturing, Temper rolling, Time Series Extrinsic Regression (TSER), Deep Learning, Self-Supervised Learning, Data Imputation |
| College: |
Faculty of Science and Engineering |
| Funders: |
EPSRC Industrial Case award |