Activity Recognition in Monitored Environments using utility Meter disaggregation

Wonders, Martin (2016) Activity Recognition in Monitored Environments using utility Meter disaggregation. Doctoral thesis, Northumbria University.

Text (PhD thesis)
Wonders.Martin_phd.pdf - Submitted Version

Download (5MB) | Preview


Activity recognition in monitored environments where the occupants are elderly or disabled is currently a popular research topic and is being proposed as a possible solution that may help maintain the independence of an aging population within their homes, where these homes are adapted as monitored environments. Current activity recognition systems implement ubiquitous sensing or video surveillance techniques which inherently, to varying degrees, impinge on the privacy of the occupants of these environments. The research presented in this thesis investigates the use of Ubiquitous sensors within a smart home setting with a view to establishing whether activity recognition is possible with a reduced, less intrusive subset of sensors that can be realised using utility meter disaggregation techniques. The thesis considers the selection of sensors as a feature selection problem and concludes that data produced from water, electricity and PIR sensors contribute significantly to the recognition of selected activities. With an established method of
activity recognition that implements a reduced number of sensors it can be argued that occupants of the monitored environment maintain a greater level of privacy. This level of privacy, however, is dependent on such systems being practically implementable into homes that are designed to assist and monitor the residents, and as such configuration and maintenance of these systems are also considered here. The utility meter disaggregation technique presented proves to perform exceptionally well when trained with large quantities of data, but gathering and labelling this data is, in itself, an intrusive process that requires significant effort and could compromise the practicality of such promising systems.
This thesis considers methods for implementing synthesised, labelled training data for both disaggregation and activity recognition systems and shows that such techniques can significantly reduce the quantity of labelled training data required. The work presented shows a significant contribution, in the areas of sensor selection and the use of utility meter disaggregation for activity recognition, and also the use of synthesised labelled training data to reduce significant system training times. The work is carried out using a combination of publicly available datasets and data collected from a purpose built smart home which includes water and electricity meter disaggregation. It is shown that a system for non-intrusive monitoring within an ambient environment, occupied by a single resident, is achievable using repurposed versions of the standard domestic infrastructure. More specifically it is demonstrated that a minimum baseline accuracy of 93.45% and F1-measure of 91.22 can be achieved using disaggregation at the water and electricity meters combined with locality context provided by home security PIR sensors. Methods of speeding up the deployment and commissioning process are proven to be viable, further demonstrating the potential practical application of the proposed system.

Item Type: Thesis (Doctoral)
Uncontrolled Keywords: machine learning, synthesis of training data for machine learning, sensor selection in ambient environments, synthetic minority oversampling
Subjects: G900 Others in Mathematical and Computing Sciences
Department: Faculties > Engineering and Environment > Computer and Information Sciences
University Services > Graduate School > Doctor of Philosophy
Depositing User: Ay Okpokam
Date Deposited: 16 Aug 2017 10:55
Last Modified: 31 Jul 2021 23:05

Actions (login required)

View Item View Item


Downloads per month over past year

View more statistics