Smart water meter systems are large scale wireless sensor networks: water meters installed in thousands of house-holds, collect hourly measurements that are reported over a wireless network to a central database. This paper introduces a new method for activity discovery in real-world, hourly water meter readings. The method addresses the following constraints: 1) observations are unlabelled and so unsupervised learning of activity types is required, 2) only automatically collected readings are used, and 3) coarse-grained hourly readings mask sub-hourly concurrent and sequential activities. Automatic rule-based labelling is combined with hierarchical clustering. New criteria are introduced for evaluating the quality of discovered activity clusters. We demonstrate the utility of our activity discovery and evaluation methods using a real-world case study of over 35,000 example days from a smart water meter trial in the inland Western Australian town of Kalgoorlie Boulder. The results show that the new method is able to discover meaningful and significant activity patterns from coarse-grained hourly readings.
Note: Journal articles and conference papers (and links where available) are available under open access arrangements where possible. Otherwise please contact your institution’s library, the authors, or publishers to organise full access.