Abstract:
This thesis aims to present efforts to solve tea fermentation monitoring using Machine Learning
(ML) and the Internet of Things (IoT). This is a thesis by a publication containing five articles.
Each of the articles presents a contribution to achieving the thesis’s objectives. The design
science approach guided the research process carried out in this study.
The first article analyzed the proposed computing models for monitoring tea fermentation.
It was evident that most of the proposed models were in the form of feasibility studies; there was
noexistingdatasetonteafermentation. Furthermore, mostoftheworkappliedmachinelearning
and vision because of the fair cost of using these technologies and their ease of implementation.
In the second article, we collected and released a tea fermentation dataset since there was
none existing as reported in the first article. We describe the article and release it for use by the
community. We expect that it will contribute to the progress in the field as researchers have a
dataset to train, validate and test their models.
In the third article, we performed a feasibility study on the applicability of machine learners
in classifying the tea fermentation dataset and LabelMe dataset, where the performance of a
deep learner dubbed ”TeaNet”, a simplified version of AlexNet, was compared to the standard
machine learning models: Random Forest (RF), K-Nearest Neighbor (KNN), Decision Tree
(DT), Support Vector Machine (SVM), Linear Discriminant Analysis (LDA) and Naive Bayes
(NB). The Alexnet model was chosen for inspiration as it has shown promising performances
in classification tasks across the field and is one of the most adopted deep learners for transfer
learning. Further, AlexNet is one of the earliest developed deep learners; thus, we expected it
to be stable. The deep learner outperformed the other classifiers.
In the fourth article, we adopted the TeaNet model that reported promising classification
results in the third article for real-time detection of tea fermentation in the Sisibo tea factory,
Kenya. The majority voting techniques were adopted to improve the model’s performance as it
did not show good performance compared to results reported in article three. The model was
xxvii
made up of raspberry pi 3 models with a Pi camera to take real-time tea images as fermentation
progresses. The incorporation of the majority voting technique to aid in the decision greatly
improved the model and made the model usable for the task.
During the deployment of the model, high latency intermittent internet and electricity chal
lenges were encountered. Thus we offloaded the solution from the cloud to the edge and fog
and powered the solution using a photo-voltaic energy source. Further, we applied duty-cycling
where idle components were allowed to sleep, which saved 50.6559Wh during the deployment
and reported the experiments in the fifth article.
