Real-time outlier detection in integrated GNSS and accelerometer structural health monitoring systems based on a robust multi-rate Kalman filter

Structural health monitoring (SHM) is important in ensuring safety of large civil engineering structures. Global Navigation Satellite Systems (GNSS)-based technology has been commonly used in SHM systems due to its unique ability to obtain real-time 3D displacement information. GNSS-based SHM systems are often integrated with other sensors (e.g., accelerometers) to take advantage of the complementary natures of different technologies and to enhance the overall performance of the integrated systems (e.g., more accurate displacement information). The conventional multi-rate Kalman filter (conventional MRKF) has been frequently used to integrate data from different technologies as the data sampling rate of a GNSS system is usually much lower than that of most other sensors. It is, however, well known that GNSS observations often contain outliers that can significantly reduce the accuracy and reliability of either the GNSS-only or integrated SHM systems. We propose a new robust multi-rate Kalman filter-based approach to integrate observations from GNSS and other sensors more robustly. The approach allows for the first time to mitigate the influence of outliers in the observations of an integrated GNSS and accelerometer SHM system, especially those in the GNSS observations in a real-time mode, which is critically important for an SHM system. Extensive experiments with data from a shaking table and from Stonecutters Bridge, a large cable-stayed bridge in Hong Kong, captured during a strong typhoon have demonstrated that the proposed method can reduce the effects of outliers effectively. The shaking table test has shown that when the method was used with an integrated GNSS and accelerometer system and outliers were present in the observations, the overall accuracy of the system was increased by up to about 70% compared with a GNSS-only system and up to about 44% higher compared to the same integrated system, but the data were processed with the conventional MRKF. It should be noted that the improvements depend on the number and magnitudes of the outliers in the observations. From the frequency perspective, the new method can capture a much broader band and more accurate vibration frequency.