Machine learning-based low-cost sensor calibration
One of the most recently developed research fields is the in-situ calibration of low-cost sensors using machine learning techniques. In the group, we have analyzed our own data sets using linear and non-linear machine learning techniques: multiple linear regression, K-nearest neighbors, random forest and support vector regression. In addition, we have also analyzed the joint use of different sensors measuring the same phenomenon, thus performing sensor fusion using the same machine learning techniques. The different investigations carried out have resulted in the following publications:
Analysis of sensor networks using Graph Signal Processing
In the section on sensor network analysis we have opened the field to the use of graph signal processing techniques to describe and maintain the data quality of a sensor network. Sensor networks, which may contain high precision instrumentation and low cost sensors, are irregular networks with highly complex relationships, so the use of graph signal processing techniques can help in the analysis of signals on such networks. Thus, the main research explains how to use the data of a network to learn the graph and its implicit relationships between nodes, in order to take advantage of the information of neighboring nodes. The following articles are the results of these studies:
Applications to some fields
We have applied our results to other fields such us precision agriculture or the exposure of black carbon to people. In the first paper we have analyzed the calibration of humidity sensors to crop fields, while in the second, we have created a proxy sensor of black carbon, a non-regulated pollutant.
Data sets obtained during the project
We have produced data sets that have been published during the project.