IAS Fellows' Seminar - The challenge for sensors in the â€œInternet of Thingsâ€: how do you know that the data are reliable?
Amongst other benefits, “Internet of Things” is touted as a ubiquitous measurement system, delivering data with high spatial and temporal resolution, from low-cost devices. However, any measurement device needs calibrating, and calibrations drift. The cost of calibration and maintenance can very quickly become very much larger than the cost of installation in the first place. “Machine learning” in its many different manifestations is often presented as a solution to this problem, but then the issue is that there is a hidden model that is being assumed. These approaches also can require very large amounts of training data.
We have been studying this question in application to the development, installation and management of dense networks of air quality measurement instruments. The work described here aims to provide reliable data from high-density, low-cost networks, taking into consideration all aspects of network construction and cost. This endeavour starts with a clear understanding of the network purpose, leading to a specification for the accuracy of the results. This step in fact involves significant social science questions (who are really the users, what are their expectations of the data…) and informs the overall network design. The effort moves through the sensors, with a robust understanding of operating mechanisms leading to understanding of failure modes, interferences in relation to the measurement problem, and causes of drift; then to consideration of the air sampling system and instrument construction; then to methods using network data and understanding of the environmental constraints and correlations to validate the data, manage the maintenance for cost-minimisation, and to perform “blind” calibration using simple, transparent models. It is a total effort, in which each element impinges upon and influences the design decisions at each stage. The overall objective is reliable, high spatial-density, high time-resolution data at minimal cost.
Places are limited and so any academic colleagues or students interested in attending a seminar should register online in advance to reserve a place. Places will be confirmed within 48 hours of receipt (subject to availability). Alternatively please contact the Institute in advance to reserve a place.
Contact email@example.com for more information about this event.