Deep Learning Methods for Impact Load Identification and Displacement Estimation in Structures

Abstract

Structural Health Monitoring (SHM) is essential for ensuring the safety and longevity of critical infrastructure. Estimating displacement and impact load is paramount for SHM because these metrics often provide direct indicators of structural integrity and performance. Displacement measurements can reveal excessive deflection, lateral movement, or longitudinal movement, which are key signs of potential structural failure. Accurate estimation of impact loads helps in understanding the applied forces during exceptional events, such as vehicle collisions, which is critical for preventing progressive collapse and ensuring the structure's resilience. By understanding and accurately estimating these parameters, timely interventions can be implemented to prevent catastrophic events, optimize maintenance schedules, and make informed decisions regarding repairs or replacements under budget constraints. Estimating both displacement and impact load offers a comprehensive assessment of structural behavior and enhances the robustness of SHM systems. This dissertation explores the use of deep learning methods, specifically Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (BLSTM) networks, to address these challenges by providing indirect measurement solutions using acceleration data. The research is divided into two main objectives. The first objective is the comprehensive identification of impact loads using multi-axis accelerometers and CNN models. This approach addresses the number of sensor requirements, measurement points, and the identification of load location, magnitude, and direction without needing localization measurements. Additionally, a physics-based simulator is developed to generate the necessary training data for the CNN models, incorporating uncertainties in structural parameters. The method was validated in both numerical and laboratory case studies. The physics-based simulator demonstrated its ability to effectively train the CNN model, addressing uncertainty issues in model parameters. Moreover, the applicability of this method was tested by utilizing a single measurement point in different locations, showcasing its versatility and potential for real-world engineering applications. The second objective focuses on the estimation of displacements using BLSTM models with acceleration measurements. This method aims to overcome the challenges associated with low-frequency and zero-mean displacement estimation, such as those caused by moving loads on bridges and earthquake-induced offsets. Laboratory case studies, including a lab-scale bridge and a shaking table for earthquake simulations, validate the effectiveness of the proposed BLSTM models. The dissertation makes significant contributions to the field of SHM by enhancing the practicality and accuracy of impact load identification and displacement estimation. It presents robust frameworks for indirect measurements by utilizing acceleration data. Acceleration measurements are easy to acquire, do not require a fixed reference point, and are relatively low-cost, making them ideal for field applications. These acceleration measurements can be applied to various structures, improving the overall efficiency and reliability of SHM systems. The integration of deep learning techniques with physical modeling and simulation opens new avenues for advanced structural diagnostics and maintenance planning.