Abstract
This thesis develops and validates an integrated, low-cost platform combining Internet of Things (IoT) hardware, multimodal sensing, and machine learning (ML) to provide continuous, non-invasive monitoring and predictive endpoint management for traditional African fermentations: amasi (fermented milk beverage) and mahewu (fermented maize beverage). The research addresses the variability and safety risks inherent in artisanal processing by replacing intermittent, subjective assessments with a real-time digital twin architecture.
The study implemented a temperature-compensated electrical conductivity (EC) calibration framework that successfully maps ionic shifts to total titratable acidity (TTA). In the amasi case study, convolutional neural networks (CNN) achieved a global prediction accuracy of R2 = 0.9475 for real-time acidity estimation. Furthermore, integrating non-invasive computer vision with sensor streams demonstrated that multimodal fusion consistently improves state estimation; in mahewu fermentations, the multi-output models achieved high predictive fidelity, with cross-validated R2 = 0.985 and test R2 = 0.988 for joint physicochemical targets. Process scheduling was further enhanced through a time-to-target forecasting model for amasi, achieving R2 = 0.98 and a mean absolute error of less than 144 minutes.
To address post-fermentation quality, the research established a proof-of-concept transfer learning framework. By leveraging labelled yoghurt data as the source domain, the adapted Random Forest models achieved classification accuracies of 0.85–0.88 and a consistent F_1 score of approximately 0.85 for predicting spoilage onset in amasi. While the study is limited by laboratory-scale vessel geometries and a reliance on proxy spoilage labels, the results establish a scalable pathway for digital quality assurance in African fermented foods.
The uniqueness of this research lies in delivering the first open-source prototype that uses EC as a high-fidelity proxy for TTA in artisanal systems, outperforming existing commercial platforms that lack TTA-specific modelling or autonomous control actuation. By achieving predictive accuracies exceeding 94% across both dairy and cereal matrices, this work demonstrates that affordable IoT-ML frameworks can materially reduce batch variability while preserving the cultural and sensory integrity of traditional food systems.