Detecting abnormal sensors via machine learning: An IoT farming WSN-based architecture case study

• Published in: Measurement : journal of the International Measurement Confederation Vol. 164; p. 108042
• Main Authors: de Souza, Paulo Silas Severo, Rubin, Felipe Pfeifer, Hohemberger, Rumenigue, Ferreto, Tiago Coelho, Lorenzon, Arthur Francisco, Luizelli, Marcelo Caggiani, Rossi, Fábio Diniz
• Format: Journal Article
• Language: English
• Published: London Elsevier Ltd 01.11.2020; Elsevier Science Ltd
• Subjects: Actuation; Agricultural management; Algorithms; Anomalies; Anomaly detection algorithms; Architecture; Computer architecture; Crop yield; Data analysis; Decision making; Heuristic methods; Information management; Machine learning; Outlier detection; Power consumption; Safety factors; Sensors; Studies; Target detection; Wireless communications; Wireless networks; Wireless sensor network; Wireless sensor networks; Wireless sensor network; Anomaly detection algorithms; Outlier detection
• ISSN: 0263-2241; 1873-412X
• DOI: 10.1016/j.measurement.2020.108042

•Analysis of Algorithms able of detecting sensors with abnormal behavior in IoT WSN.•Platform to choose the best anomaly detection algorithm according the application.•A case study to validate the proposal on a WSN agricultural environment. Precision Agriculture is a broad, systemic, and multidisciplinary subject, dealing with an integrated information and technology management system, based on the concepts that the variability of space and time influence crop yields. Precision farming aims at more comprehensive management of the agricultural production system as a whole. It uses a set of tools, instruments, and sensors to measure or detect parameters or targets of interest in the agroecosystem. Sensors are distributed in the environment and are usually communicated through a Wireless Sensor Network (WSN). Due to this dispersion of the sensors, errors could occur in Byzantine form or could be caused by safety factors, which can lead to a misinterpretation by the system of data analysis and actuation over the environment. Anomaly detection algorithms can detect such problem sensors by allowing them to be replaced, or the wrong data is ignored. Therefore, this work presents a reference architecture and a heuristic algorithm that aid the decision of which anomaly detection to use based on the demands of agricultural environments. We performed a preliminary evaluation, analyzing different anomaly detection algorithms regarding execution time, accuracy, and scalability metrics. Results show that the decision-making supported by the proposed architecture reduces edge devices’ power consumption by 18.59% while minimizing the device’s temperature in up to 15.94% depending on the application workload and edge device characteristics.