With advances in sensing and communication technologies, engineering systems are now commonly instrumented with sensors for system monitoring and management. Occasionally, when sensors become malfunction, it is advantageous to automatically determine faulty sensors in the system and, if possible, recover missing or faulty data. This paper investigates the use of machine learning techniques for sensor data reconstruction and anomaly detection. Specifically, bidirectional recurrent neural network (BRNN) is employed to build a data-driven model for sensor data reconstruction based on the spatiotemporal correlation among the sensor data. The reconstructed sensor data can be used not only for recovering the data of the faulty sensors, but also for detecting anomalies based on an analytical redundancy approach. The proposed method is tested with vibration data based on a numerical simulation of a sensor network for bridge monitoring application. In terms of prediction accuracy, the results show that the BRNN-based sensor data reconstruction method performs better than other existing sensor data reconstruction methods. Furthermore, the sensor data reconstructed can be used to detect and isolate the anomalies caused by faulty sensors.

© (2019) COPYRIGHT Society of Photo-Optical Instrumentation Engineers (SPIE). Downloading of the abstract is permitted for personal use only.

Seongwoon Jeong, Max Ferguson, and Kincho H. Law, "Sensor data reconstruction and anomaly detection using bidirectional recurrent neural network," Proc. SPIE 10970, Sensors and Smart Structures Technologies for Civil, Mechanical, and Aerospace Systems 2019, 109700N (Presented at SPIE Smart Structures + Nondestructive Evaluation: March 05, 2019; Published: 27 March 2019); https://doi.org/10.1117/12.2514436.