Vibration-based Monitoring of the Zwartewaterbrug: a Machine Learning Approach

Abstract

In the Netherlands there are many bridges that crosses waterways, which are key nodes in the transportation system. The safety and integrity of these bridges is nowadays assessed by (visual) inspections at regular intervals. This approach cannot provide information about damage development in between inspections, leading to potential failure or high intervention costs. A maintenance strategy for bridges based on continuous vibration-based monitoring provides a possible cost-effective solution that may replace the periodic (visual) inspections. Vibration-based structural health monitoring of civil engineering structures is receiving increasing attention in recent years. This is due to new developments in related areas such as sensing technology, system identification, data mining and condition assessment. A particular damage scenario that frequently occurs in steel orthotropic deck structures is the development of fatigue cracks in the welds between the longitudinal U-shaped stiffeners and the transverse beams or the steel deck plate. This damage scenario was also observed in the steel deck structure of the Zwartewaterbrug, located in Hasselt in the Netherlands. Vibration tests were performed on the bridge to develop a damage detection algorithm for early detection of the fatigue cracks in the bridge deck. In order to simulate the effect of damage, small masses were attached to the bottom of the bridge deck. The measurements were performed over a relatively short time period to exclude – in as far as possible – the effects of environmental variability. However, the bridge was kept open to traffic during the testing and structural response was primarily triggered by traffic load. The vibration data (accelerations) obtained during these tests was used for validation of the proposed methodology in this thesis. The aim of the thesis is to detect the added masses from the vibration data. To achieve this, a data-based approach for damage detection was proposed, which, besides the data acquisition, consists of four main stages: data preprocessing, feature extraction, pattern recognition and decision making. The analysis is focused on the high-frequency acceleration response, because the associated higher-frequency local modes are more sensitive to the small changes induces by the added mass. The similarity filtering preprocessing procedure was applied to filter out the variability related to the operational loading, and the Frequency Domain Decomposition system identification technique was used to extract damage indicators. Subsequently, a support vector machine algorithm was applied to “learn” patterns in the extracted damage indicators, which enables the algorithm to assign a damage state to a given measurement vector. Alternatively, a novelty detection technique was used to distinguish data from the “healthy” structure from data corresponding to a “damage” state. It was not possible to reliable detect the added masses from the high-frequency acceleration data due to the large variations in the damage indicators of a mass class. These variations are attributed to the changing environmental conditions over the measurement period. The principal component analysis was used in an attempt to remove these variations from the damage indicators, but the results were still not satisfactory. In conclusion, it was not possible to properly test the proposed methodology for damage detection using the data of the Zwartewarterbrug, mainly because the data was not suitable for the machine learning algorithms employed. Recommendations are given related to the methodology for the similarity filtering, determination of the high-frequency range, and machine learning algorithm to be used. Additionally, recommendations are given for future measurement campaigns to obtain data that is more suitable for successful application of the methodologies employed in the present project.