Adaptive Health Monitoring Using Aggregated Energy Readings from Smart Meters

Worldwide, the number of people living with self-limiting conditions, such as Dementia, Parkinson’s disease and depression, is increasing. The resulting strain on healthcare resources means that providing 24-hour monitoring for patients is a challenge. As this problem escalates, caring for an ageing population will become more demanding over the next decade, and the need for new, innovative and cost effective home monitoring technologies are now urgently required. The research presented in this thesis directly proposes an alternative and cost effective method for supporting independent living that offers enhancements for Early Intervention Practices (EIP). In the UK, a national roll out of smart meters is underway. Energy suppliers will install and configure over 50 million smart meters by 2020. The UK is not alone in this effort. In other countries such as Italy and the USA, large scale deployment of smart meters is in progress. These devices enable detailed around-the-clock monitoring of energy usage. Specifically, each smart meter records accurately the electrical load for a given property at 10 second intervals, 24 hours a day. This granular data captures detailed habits and routines through user interactions with electrical devices. The research presented in this thesis exploits this infrastructure by using a novel approach that addresses the limitations associated with current Ambient Assistive Living technologies. By applying a novel load disaggregation technique and leveraging both machine learning and cloud computing infrastructure, a comprehensive, nonintrusive and personalised solution is achieved. This is accomplished by correlating the detection of individual electrical appliances and correlating them with an individual’s Activities of Daily Living. By utilising a random decision forest, the system is able to detect the use of 5 appliance types from an aggregated load environment with an accuracy of 96%. By presenting the results as vectors to a second classifier both normal and abnormal patient behaviour is detected with an accuracy of 92.64% and a mean squared error rate of 0.0736 using a random decision forest. The approach presented in this thesis is validated through a comprehensive patient trial, which demonstrates that the detection of both normal and abnormal patient behaviour is possible.

Name dc:type.qualificationname

phd

Level dc:type.qualificationlevel

doctoral

Grantor dc:publisher.institution

Liverpool John Moores University

Year dc:date.issued

2017

Author dc:creator

• Chalmers, C

Contributors dc:contributor

• Hurst, W
• Mackay, M
• Fergus, P

• QA75 Electronic computers. Computer science
• QA76 Computer software
• RA0421 Public health. Hygiene. Preventive Medicine

DOI dc:identifier.doi

https://doi.org/10.24377/LJMU.t.00007543

OAI identifier oai:identifier

oai:researchonline.ljmu.ac.uk:7543