Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
On-line Final Quality Prediction for Multiphase Batch Processes with Uneven Durations
Thesis (MEng)--Stellenbosch University, 2026.
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
Stellenbosch : Stellenbosch University
2026
|
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1867613763027337216 |
|---|---|
| access_status_str | Open Access |
| author | Orsmond, Jedd Liam |
| author2 | Louw, Tobias Muller |
| author_browse | Louw, Tobias Muller Orsmond, Jedd Liam |
| author_facet | Louw, Tobias Muller Orsmond, Jedd Liam |
| author_sort | Orsmond, Jedd Liam |
| collection | Thesis |
| description | Thesis (MEng)--Stellenbosch University, 2026. |
| format | Thesis |
| id | oai:scholar.sun.ac.za:10019.1/135969 |
| institution | Stellenbosch University (South Africa) |
| language | English |
| last_indexed | 2026-06-10T12:41:18.607Z |
| license_str | Not specified — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from SUNScholar — Stellenbosch University Repository |
| publishDate | 2026 |
| publishDateRange | 2026 |
| publishDateSort | 2026 |
| publisher | Stellenbosch : Stellenbosch University |
| publisherStr | Stellenbosch : Stellenbosch University |
| record_format | dspace |
| source_str | SUNScholar — Stellenbosch University Repository |
| spelling | oai:scholar.sun.ac.za:10019.1/135969 On-line Final Quality Prediction for Multiphase Batch Processes with Uneven Durations Orsmond, Jedd Liam Louw, Tobias Muller Bradshaw, Steven Martin Stellenbosch University. Faculty of Engineering. Dept. of Chemical Engineering. Thesis (MEng)--Stellenbosch University, 2026. Orsmond, J. L. 2026. On-line Final Quality Prediction for Multiphase Batch Processes with Uneven Durations. Unpublished masters thesis. Stellenbosch: Stellenbosch University [online]. Available: https://scholar.sun.ac.za/items/1438aeb2-fd51-47c3-8b7e-c44886d4d480 Industrial batch processes are essential for producing high-value, low-volume products – especially in pharmaceutical, biotechnology, and specialty chemical industries. However, their inherent characteristics present significant challenges for on-line quality monitoring – transient, uneven batch durations, multicollinearity of variables, multiphase behaviour, and lack of availability for real-time quality sensors. Off-line quality control strategies are insufficient since no timely corrective interventions can be implemented. Data-driven techniques based on latent methods such as Principal Component Analysis (PCA) and Partial Least Squares (PLS) have been successfully extended for use on-line in batch processes as Multiway-PCA/PLS by (Nomikos & MacGregor, 1994, 1995a). However, these techniques assume a dataset of batches with equal length (rarely the case for industrial processes) and the use of single monolithic models do not adequately capture the shifting variable relationships in multiphase processes. Although multi-model approaches have been proposed, these typically rely on extensive prior process knowledge or physical processing units to define the phase boundaries. These can be both expensive and time-consuming to develop and do not consider phase-like behaviour within stages or the transition dynamics in regions between adjacent phases. The aim of this thesis was to implement an accurate inferential soft-sensor for on-line final quality prediction in industrial multiphase batch processes – demonstrated on a penicillin fermentation case study. To address unequal durations and common asynchronisms, both a modified Dynamic Time Warping (DTW) and Indicator Variable (IV) technique was successfully implemented to align trajectories (376 – 420 hours) with comparable results. DTW introduced minor distortions (especially in the flat regions) and performed best when using the geometric mean to calculate variable weightings (González-Martínez et al., 2011). IV required less computational intensity but challenges in terms of variable selection were noted that may reduce practical implementation. Multicollinearity of input variables was addressed with PCA by projecting the 11-dimensional input space to a latent space with only 8 uncorrelated principal components while still retaining > 95% of cumulative variance. To investigate the effect of different phase division techniques, three approaches based on traditional Global (single monolithic model), Operational-Stage (based on physical processing units), and a data-driven Gaussian Mixture Model (GMM) based on variable correlation clusters were built. The unsupervised GMM technique enabled identification of distinct phase boundaries based on evolving variable correlation structures without the need for prior expert process knowledge. Bayesian- and Akaike Information Criteria (BIC/AIC) were incorporated to inform on model complexity and fit to overcome the major limitation of specifying the number of components a priori. Different covariance configurations were investigated with a full-unshared structure yielding the lowest BIC scores (< 3) compared to other options (> 7.5) without any significant computational challenges. Six phases were identified and Bayesian Model Averaging (BMA) strategy based on dynamic weighting of adjacent models (through their posterior probability distributions) was implemented to achieve smooth transitions in ‘fuzzy’ regions between phases. In each identified phase, local predictive PLS models were built using k-fold cross validation (k=10) to determine the appropriate number of components retained. Only 3-5 latent components were needed to capture 66-88% of the variance in final quality. The overall performance of all three approaches was assessed on-line using 30 independent test batches. The GMM approach achieved Root Mean Squared Error (RMSE) of 0.0165 g/L representing a 21% and 50% improvement over the Operational/Global models respectively (𝑝 < 0.05). Further, the GMM predictions were more consistent and rapidly approached the true value with smooth transitions between phases. In contrast, the Global model showed poor accuracy and high variation for most of the batch duration. The Operational model yielded somewhat more accurate predictions but tended to oscillate with large error spikes at (and near) the stage boundaries due to the strict boundary definitions. The study demonstrated that data-driven techniques can avoid the need for expensive, timeconsuming modelling in developing inferential soft-sensors. Overall, the probabilistic GMM approach accurately captures shifting variable relationships (even within stages) and demonstrated overall superior predictive performance with smooth transitions in ‘fuzzy’ regions. Further, the manner in which phase boundaries are defined, and the transitions between them, are critical for accurate predictions in industrial batch processes. Masters 2026-04-16T12:55:29Z 2026-04-16T12:55:29Z 2026-03 Thesis https://scholar.sun.ac.za/handle/10019.1/135969 en 155 pages application/pdf Stellenbosch : Stellenbosch University |
| spellingShingle | Orsmond, Jedd Liam On-line Final Quality Prediction for Multiphase Batch Processes with Uneven Durations |
| title | On-line Final Quality Prediction for Multiphase Batch Processes with Uneven Durations |
| title_full | On-line Final Quality Prediction for Multiphase Batch Processes with Uneven Durations |
| title_fullStr | On-line Final Quality Prediction for Multiphase Batch Processes with Uneven Durations |
| title_full_unstemmed | On-line Final Quality Prediction for Multiphase Batch Processes with Uneven Durations |
| title_short | On-line Final Quality Prediction for Multiphase Batch Processes with Uneven Durations |
| title_sort | on line final quality prediction for multiphase batch processes with uneven durations |
| url | https://scholar.sun.ac.za/handle/10019.1/135969 |
| work_keys_str_mv | AT orsmondjeddliam onlinefinalqualitypredictionformultiphasebatchprocesseswithunevendurations |