Probabilistic reconstruction for spatiotemporal sensor data integrated with Gaussian process regression

Abstract

The effective health management of sensor networks is very important for the reliability assessment of engineering structures. Sensor failure and data missing occur frequently due to the influences of signal noise and adverse environment. This paper proposes a probabilistic reconstruction framework of missing data using spatiotemporal correlation of synchronous sensors. Faulty sensors in multi-sensor network are detected by projecting high-dimension feature into a visualization optimal discriminant vector space. The Gaussian process regression (GPR) machine learning is developed to reconstruct the structural dynamic nonlinear response by integrating with temporal and spatial information. The Bayesian posterior probabilistic output rather than point estimation is used to quantify the inherent uncertainty induced by non-stationary stochastic process. Various types of prior kernel functions are modeled to obtain the optimal function according to the characteristic of sensor data. A subset of sensor networks with different correlation coefficients is proposed to obtain the optimal selection strategy. The proposed framework is demonstrated by accelerator sensors data collected from Canton tower in Guangzhou. The results show that the reconstructed data agree well with the measured values in time and frequency domain. The GPR data-driven method can achieve a higher accuracy than artificial neural network approach. The selection of sensors has a significant impact on missing data reconstruction. Selecting some highly correlated sensors is as accurate as applying the entire network sensors.