Experimental Data Anomaly Detection at Edge Sensor Nodes Using Physics Laws

Abstract

Reported measurements by sensor nodes in wireless sensor networks (WSNs) are subject to various anomaly due to sensor/node failures or some events in the environment. Resource scarcity in sensor nodes makes the implementation of sophisticated data anomaly detection techniques challenging in WSNs. We present a multivariate physics-based data anomaly detection technique in WSNs, and we implement the technique to measure the accuracy and efficiency of the proposed technique in detail. The proposed technique examines the natural relationship between various types of sensor measurements to determine the existence of any data anomaly in the node. Furthermore, the technique enables the distinction of data errors and events where a data anomaly is flagged. The technique does not require any exchange of information between neighboring sensor nodes or a central node for data anomaly detection. In a case study, we have applied the proposed technique on temperature and humidity sensor readings and have implemented using TI CC3200 Development Kit, TI BOOSTXL-SENSORS, and TI IMETER-BOOST. To evaluate the accuracy and efficiency of the technique, we have manufactured various data errors and events, measured required memory footage, and metered the amount to current consumption by a sensor node. The detailed analyses have confirmed the high accuracy and efficiency in detecting and separating data errors and environmental events.