Low-power neuromorphic sensor fusion for elderly care

Yu, Zheqi (2022) Low-power neuromorphic sensor fusion for elderly care. PhD thesis, University of Glasgow.

Abstract

Smart wearable systems have become a necessary part of our daily life with applications ranging from entertainment to healthcare. In the wearable healthcare domain, the development of wearable fall recognition bracelets based on embedded systems is getting considerable attention in the market. However, in embedded low-power scenarios, the sensor’s signal processing has propelled more challenges for the machine learning algorithm. Traditional machine learning method has a huge number of calculations on the data classification, and it is difficult to implement real-time signal processing in low-power embedded systems. In an embedded system, ensuring data classification in a low-power and real-time processing to fuse a variety of sensor signals is a huge challenge. This requires the introduction of neuromorphic computing with software and hardware co-design concept of the system. This thesis is aimed to review various neuromorphic computing algorithms, research hardware circuits feasibility, and then integrate captured sensor data to realise data classification applications. In addition, it has explored a human being benchmark dataset, which is following defined different levels to design the activities classification task. In this study, firstly the data classification algorithm is applied to human movement sensors to validate the neuromorphic computing on human activity recognition tasks. Secondly, a data fusion framework has been presented, it implements multiple-sensing signals to help neuromorphic computing achieve sensor fusion results and improve classification accuracy. Thirdly, an analog circuits module design to carry out a neural network algorithm to achieve low power and real-time processing hardware has been proposed. It shows a hardware/software co-design system to combine the above work. By adopting the multi-sensing signals on the embedded system, the designed software-based feature extraction method will help to fuse various sensors data as an input to help neuromorphic computing hardware. Finally, the results show that the classification accuracy of neuromorphic computing data fusion framework is higher than that of traditional machine learning and deep neural network, which can reach 98.9% accuracy. Moreover, this framework can flexibly combine acquisition hardware signals and is not limited to single sensor data, and can use multi-sensing information to help the algorithm obtain better stability.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Additional Information: Supported by a joint industrial scholarship from University of Glasgow and the Transreport Ltd (UK) Company.
• Subjects: T Technology > T Technology (General); T Technology > TA Engineering (General). Civil engineering (General)
• Colleges/Schools: College of Science and Engineering > School of Engineering
• Supervisor's Name: Abbasi, Dr. Qammer, Heidari, Professor Hadi and Imran, Professsor Muhammad
• Date of Award: 2022
• Depositing User: Theses Team
• Unique ID: glathesis:2022-83001
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 22 Jun 2022 13:30
• Last Modified: 22 Jun 2022 13:33
• Thesis DOI: 10.5525/gla.thesis.83001
• URI: https://theses.gla.ac.uk/id/eprint/83001

Related URLs

• Article DOI
• Article DOI
• Article DOI
• Article DOI
• Article DOI