| Тойм: | <p>This thesis aims to develop systems for intent sensing – the measurement and prediction of what it is that a user wants to happen. Being able to sense intent could be hugely beneficial for control of assistive devices, and could make a great impact on the wider medical device industry.</p>
<br>
<p>Initially, a literature review is performed to determine the current state-of-the-art for intent sensing, and identifies that a holistic intent sensing system that properly captures all aspects of intent has not yet been developed. This is therefore followed by the development of such a novel intent sensing system.</p>
<br>
<p>To achieve this, algorithms are developed to combine multiple sensors together into a modular Probabilistic Sensor Network. The performance of such a network is modelled mathematically, with these models tested and verified on real data. The intent sensing system then developed from these models is tested for sensing modalities such as Electromyography (EMG), motion data from Inertial Measurement Units (IMUs), and audio. The benefits of constructing a modular system in this way are demonstrated, showcasing improvement in accuracy with a fixed amount of training data, and in robustness to sensor unavailability – a common problem in prosthetics, where sensor lift-off from the skin is a frequent issue.</p>
<br>
<p>Initially, the algorithm is developed to classify intent after activity completion, and this is then developed to allow it to run in real-time. Different classification methods are proposed and tested including K-nearest-neighbours (KNN), before deep learning is selected as an effective classifier for this task. In order to apply deep learning without requiring a prohibitively large training data set, a time-segmentation method is developed to limit the complexity of the model and make better use of the available data. Finally, the techniques developed in the thesis are combined into a single continuous, multi-modal intent sensing system that is modular in both sensor composition and in time.</p>
<br>
<p>At every stage of this process, the algorithms are tested against real data, initially from non-disabled volunteer participants and in the later chapters on data from patients with Parkinson’s disease (a group who may benefit greatly from an intent sensing system). The final system is found to achieve an accuracy of 97.4% almost immediately after activity inception, increasing to 99.9918% over the course of the activity. This high accuracy can be seen both in the patient group and the control group, demonstrating that intent sensing is indeed viable with currently available technology, and should be further developed into future control systems for assistive devices to improve quality of life for both disabled and non-disabled users alike.</p>
|
|---|