Neuromorphic signal processing for wearable devices

Ding, Yuqi (2025) Neuromorphic signal processing for wearable devices. PhD thesis, University of Glasgow.

Abstract

Wearable health devices have a strong demand in real-time biomedical signal processing. Over the past few decades, advances in Artificial Intelligence (AI), and particularly the development of neural networks, have significantly impacted the field of signal processing, enabling more efficient and accurate analysis of complex biomedical signals. However, continuous monitoring could generate massive amounts of data, resulting in an information bottleneck that challenges data transfer and subsequent post-processing. Neuromorphic computing, an emerging brain-inspired computational architecture, has garnered significant research attention in recent years. In contrast to the conventional von Neumann architecture, which relies on a separation between memory and processing units, neuromorphic systems integrate computation and data storage within a unified physical framework. This design emulates the synaptic dynamics observed in biological neural networks. Such methods can process data proximate to the sensor with reduced power consumption, latency and bandwidth, providing new solutions to signal processing for wearable devices. Among neuromorphic systems, Physical Reservoir Computing (PRC) has emerged as a compelling solution. PRC harnesses the intrinsic dynamics of physical systems to accelerate and reduce the energy consumption of machine learning computations.

This thesis focuses on modelling and implementing PRC frameworks in signal processing applications. In the initial stage, the potential of PRC as a predictor for biomedical applications was explored. The model successfully maps the Magnetomyography (MMG) signal to Electromyography (EMG) with an acceptable normalized root mean square error (NRMSE) of 0.3894. In addition, an average NRMSE of 0.3690 was obtained for predicting Electrocardiography (ECG) to Phonocardiography (PCG).

However, practical applications of signal processing present more complex challenges. While the neuromorphic signal processing underperforms compared to state-of-the-art Deep Learning (DL) algorithms, and relatively few studies have investigated its application in classification tasks, the research is extended to examine the use of PRC in a complex biometric identification task. An overall classification accuracy of 89.03% in identifying twelve testing subjects was achieved during the intermediate stage.

Nevertheless, a significant concern emerged with respect to maintaining precision when implementing high-resolution signals via electronic circuits in PRC-based systems, particularly due to limitations in electronic components. This challenge motivated the research to the next stage that explores a hybrid approach of combining PRC and Spiking Neural Network (SNN), which allows for the conversion of signals from the high-precision analogue domain into a low-precision discrete spiking domain, thereby reducing hardware and storage costs. Consequently, an event-driven PRC framework was proposed and validated in the final stages to address this concern. An average classification accuracy of 80.3% was obtained in classifying 50 gestures which outperforms current SNN-based methods. The results pipeline a new insight into processing real-time signals at the edge for wearable devices, promising compact and ultra-low power electronic systems for temporal signal processing in wearable devices.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Subjects: T Technology > T Technology (General)
• Colleges/Schools: College of Science and Engineering > School of Engineering
• Supervisor's Name: Heidari, Professor Hadi
• Date of Award: 2025
• Depositing User: Theses Team
• Unique ID: glathesis:2025-85340
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 15 Jul 2025 08:44
• Last Modified: 15 Jul 2025 08:44
• URI: https://theses.gla.ac.uk/id/eprint/85340