The purpose of this paper is to propose a simple, fast, and effective method for detecting measurement errors in data collected with low-cost environmental sensors typically used in building monitoring, evaluation, and automation applications.
The method combines two unsupervised learning techniques: a distance-based anomaly detection algorithm analyzing temporal patterns in data, and a density-based algorithm comparing data across different spatially related sensors.
Results of tests using 60,000 observations of temperature and humidity collected from 20 sensors during three weeks show that the method effectively identified measurement errors and was not affected by valid unusual events. Precision, recall, and accuracy were 0.999 or higher for all cases tested.
The method is simple to implement, computationally inexpensive, and fast enough to be used in real-time with modest open-source microprocessors and a wide variety of environmental sensors. It is a robust and convenient approach for overcoming the hardware constraints of low-cost sensors, allowing users to improve the quality of collected data at almost no additional cost and effort.
The author sincerely thank to all members of the C.H.A.O.S. research group led by Professor Forrest Meggers, and in particular, Hongshan Guo and James Coleman for providing the raw data sets, and to Maria Ferrara for labeling the test data set.
Loyola, M. (2019), "A method for real-time error detection in low-cost environmental sensors data", Smart and Sustainable Built Environment, Vol. 8 No. 4, pp. 338-350. https://doi.org/10.1108/SASBE-10-2018-0051Download as .RIS
Emerald Publishing Limited
Copyright © 2019, Emerald Publishing Limited