Convolutional neural network for highway bridge indirect structural health monitoring

Abstract

Ensuring the integrity of special structures such as bridges and overpasses is a challenge in the design and management of infrastructure, as the eventual failure of these components entails severe economic losses and humanitarian tragedies. Given the impact on the safety of these components, there is a need to establish efficient strategies for monitoring structural integrity, with the intention of minimizing interruptions in traffic flow and maximizing the safety of the community dependent on the operation of these systems. There are two predominant approaches to data acquisition for structural integrity assessment: direct monitoring and indirect monitoring. In indirect monitoring, sensors are installed in the vehicle to capture responses from the dynamic interaction of the vehicle-structure system, while in direct monitoring instrumentation, sensors are installed directly on the structure. The advantages of the indirect approach over the direct involve the ability to obtain spatial information from the entire continuity of the bridge without the need for traffic interruption, as well as a significant reduction in the monitoring cost of all bridges and overpasses along a roadway. Technological development and the refinement of computational capacity have enabled innovative solutions for engineering problems, through artificial neural network algorithms with the potential to adapt and extract information from a given data set. This work addresses the use of artificial intelligence as a solution for the structural health of road bridges and overpasses, exploring the effectiveness of convolutional neural networks (CNNs) for damage identification in highway bridges through indirect monitoring data of acceleration measures. The dataset is generated from a numerical finite element code that simulates the behavior of the dynamic interaction between the vehicle and the structure system. Damage is modeled as stiffness reduction at the midspan. The CNN's ability to identify damage is evaluated by its proficiency in detecting patterns and anomalies in the data, which correlate with different levels of structural impairment.