Skip to main content

Modelling Virtual Sensors for Indoor Environments with Machine Learning.

Polanski, D. M. and Angelopoulos, C. M., 2022. Modelling Virtual Sensors for Indoor Environments with Machine Learning. In: 18th International Conference on Distributed Computing in Sensor Systems (DCOSS), 30 May - 1 June 2022, Marina Del Rey, LA, California, 222-228.

Full text available as:

[img]
Preview
PDF
Camera - Modelling Virtual Sensors for Indoor Environments with Machine Learning.pdf - Accepted Version
Available under License Creative Commons Attribution Non-commercial.

704kB

Official URL: https://dcoss.org/dcoss22/

DOI: 10.1109/DCOSS54816.2022.00046

Abstract

Virtual Sensors model the sensing operation of physical sensors deployed in an area of interest by generating sensory data with accuracy and precision close to those collected by physical sensors. Their use in applications such as augmenting the infrastructure of IoT facilities and test beds, monitoring and calibrating the operation of physical sensors, and developing Digital Twins of physical systems have led virtual sensors to attract research attention. Machine learning provides methods for modelling patterns in complex and big data generated by IoT sensing devices, allowing to model the behaviour of these devices. In this work, we investigate ML methods as means of implementation for virtual sensors. In particular, we evaluate the performance of six ML methods in terms of their effectiveness, accuracy and precision in generating sensory data based on data from physical sensors. In our study, we use a multi-modal dataset comprising IoT sensory data for temperature, humidity and illumination collected over a period of two years in an office space at University of Geneva. Our results show that the best performing model at predicting an output of a missing sensor is the Random Forest method, achieving MAPE error below 3%, 5% and 18% respectively for temperature, humidity and illuminance. The worst performing models were the linear radial basis function neural network and linear regression. In future research, we plan to deploy the best performing models natively on IoT devices, making use of tinyML and extreme edge computing methods.

Item Type:Conference or Workshop Item (Paper)
ISSN:2325-2936
Uncontrolled Keywords:Internet of Things; Sensor; Virtual Sensor; Machine Learning
Group:Faculty of Science & Technology
ID Code:37755
Deposited By: Symplectic RT2
Deposited On:11 Nov 2022 16:28
Last Modified:11 Nov 2022 16:28

Downloads

Downloads per month over past year

More statistics for this item...
Repository Staff Only -