Asymmetric multi-task deep learning with location attention for structural damage detection

Abstract

Data-driven structural damage detection is a promising research area with significant potential in structural health monitoring. Current studies typically aim to develop a multi-class classifier to detect various types of damage, including identifying damage locations and assessing damage sizes. However, as the number and types of damages increase, training a high-performing multi-class classifier becomes more challenging. Although deep learning and multi-task learning have been utilized to address this issue due to their strong capabilities in information representation and knowledge transfer, a challenge known as negative information transfer can arise between different tasks, resulting in decreased effectiveness. To mitigate the negative effects on different tasks and fully leverage the benefits of multi-task learning, we propose an asymmetric multi-task deep learning model with a location-based attention mechanism to enhance the performance of structural damage detection. Specifically, we categorize the tasks of identifying damage locations and quantifying damage sizes into two groups and design an asymmetric multi-task deep learning network to jointly learn these related yet distinct tasks. Applying asymmetric information sharing, our model can improve the performance of related tasks while minimizing negative transfer of learned information between different tasks. Furthermore, we implement a location-based attention mechanism to emphasize location features, thereby enhancing the model’s damage identification capability. Our model is evaluated on an actual structure. Extensive experimental results demonstrate that our proposed approach can improve the performance of both identifying damage locations and quantifying damage sizes compared to state-of-the-art methods.