AbstractMachine Learning (ML) on low powered devices is still in its infancy, and the use of ML for continuous linear movement analysis is limited. There is a need to find an effective way to reduce the amount of computational load on low power devices. One of the ways to do so is to reduce the number of features within the ML model. This research took on that challenge and proposed a novel feature reduction approach by automatically selecting suitable and relevant feature subsets that enable ML models to achieve acceptable performance when deployed to low powered devices.
At the core of this thesis, it focuses on the methodologies that can help to reduce the complexities of the machine learning model to improve its efficiencies and enable it to operate successfully on low powered devices with limited computational resources. Internet of things (IoT), wearable devices and data-driven techniques offer the ability for practitioners to collect vast volumes of data and process them promptly drawing useful insights. At the same time, while the information is still helpful, practitioners can propose suitable interventions based on the information extracted by the physical activity monitoring device.
The need for Lightweight ML is because the objective of this research is to deploy a suitable solution on low powered devices that can effectively operate within the limited computational resources available. This thesis has successfully proposed, developed and tested a novel lightweight ML approach for linear machine learning problems. The proposed method has also achieved its aims of automatically selecting appropriate features for machine learning applications. The thesis shows how the novel technique can significantly improve computational efficiency by exploiting the correlation co-efficient and variance between elements to eliminate irrelevant features. This method has been piloted and tested in one publicly available dataset and two case studies - a study of children's physical activities and energy expenditure, and a swim classification study. These case studies demonstrate that the proposed novel approach is useful in selecting appropriate features and reducing model complexity while performing physical activity recognition and monitoring. The thesis also demonstrated the effectiveness of deploying the lightweight ML method on low powered devices by significantly increasing the computational efficiency. The approach was evaluated using several supervised machine learning algorithms including; decisions trees, linear & logistic regression, random forests, multilayer perceptron, support vector machines.
The result of the "children's physical activity" case study has shown that this newly proposed approach can effectively predict children's energy expenditure and perform well within 5% of the baseline model. The best performing model using boosted trees achieved a 90% predictive accuracy inline with other research. The research has also shown chat it can significantly increase the model's computational efficiency by up to 75% enabling it to operate on low powered devices with limited computational resources successfully. The "swim classification" case study explored the novel lightweight model's performance on physical activity classification for swimming. The best performing model showed that the novel approach effectively reduced the feature complexity and model dimensionally to achieve 78% classification accuracy while being within 5% of the baseline and gaining above 70% computational efficiency running on the low powered device (a device limited in computational resources). This novel lightweight approach can significantly reduce the models feature dimensionality enabling suitable ML algorithms to operate on low powered devices, saving time and computational resources. It has been incredibly effective at dealing for linear triaxial time-series based Physical Activity monitoring and recognition problems.
Although this proposed approach was specifically proposed for linear machine learning problems, the thesis has further explored this novel approach in a non-linear, time-series, activity-based situation in a "smart care home" case study. As anticipated, the experiment has found that this novel approach was unsuitable and ineffective, and an alternative technique using Conditional Random Fields was recommended for future expansion of the approach to tackle datasets nonlinearin nature. Finally, the contributions to the body of knowledge from this thesis and future directions of work have been provided.
|Date of Award||Mar 2021|
|Supervisor||Yanguo Jing (Supervisor) & Xin Lu (Supervisor)|