Vehicle-assisted bridge damage assessment using deep learning

Abstract

This thesis introduces innovative methodologies for vehicle-assisted bridge health monitoring, aiming to improve maintenance procedures of ageing infrastructure, a critical concern for transport network owners. By taking advantage of advancements in sensing technology and the increasing interconnectivity between vehicles and infrastructure, these methodologies focus on developing an automated bridge assessment method that efficiently evaluates the current condition of bridge structures. This approach enables more accurate and timely maintenance decisions. The primary objective of this thesis is to create an automated bridge assessment framework for existing bridges by harnessing the synergy between sensors installed on structures and signals transmitted by passing vehicles. By gathering comprehensive information from various sources, including vehicles and the bridge itself, and fusing this data using deep learning techniques, the framework efficiently evaluates the current condition of bridge structures, facilitating more precise and prompt maintenance decisions. The thesis comprises several studies investigating deep learning techniques, such as deep autoencoders (DAE) and probabilistic temporal autoencoders (PTAE), for extracting features and capturing temporal relationships in the data. This enables accurate identification and quantification of potential damage in bridge structures. The first study (Paper IA IB) examines an indirect bridge monitoring system using vertical acceleration responses from a fleet of vehicles passing over a healthy bridge. This study’s findings reveal that the error in signal reconstruction from the trained DAE is sensitive to damage, considering the distribution of results from multiple separate vehicle-crossing events. The proposed method proves effective in detecting damage under operational conditions and demonstrates potential as a new tool for cost-effective bridge health monitoring. The second study introduces a methodology for assessing bridge conditions using a PTAE and multi-sensor data from a fixed sensing framework, collected during train crossings. The study’s results indicate that the proposed method can detect damage with a limited number of sensors, making it a valuable approach to enhance bridge safety. An Exponentially Weighted Moving Average (EWMA) filter and a control chartbased threshold mechanism are applied to further refine the damage assessment process, distinguishing between healthy and progressively deteriorating damage cases. The third study proposes a Probabilistic Deep Neural Network framework for damage assessment, combining vehicle and bridge responses to extract damage-sensitive features for classifying different damage states. The findings of this study demonstrate that incorporating multiple sensor information reduces uncertainties in damage detection and localisation. The results also suggest that the proposed method is robust in handling measurement noise and varying environmental conditions. In conclusion, this thesis advances knowledge in the field of structural assessment through structural health monitoring by providing insights and improvements in techniques and methodologies. By taking advantage of the combined strengths of sensors mounted on structures and signals transmitted by moving vehicles, the developed methodologies provide reliable and precise damage evaluation capabilities. These valuable insights enhance bridge safety, improve resource allocation, and contribute to the overall performance of transport networks. Ultimately, this approach leads to more sustainable and resilient infrastructure, better equipped to handle modern society’s growing demands.