Abstract
Structural Health Monitoring (SHM) is critical for ensuring the safety and longevity of civil infrastructure. Sensor-based SHM systems generate vast amounts of data, making traditional anomaly detection methods increasingly challenging due to their complexity and volume (Neu et al., 2016; Buethe et al., 2013). The emergence of deep learning (DL) offers powerful capabilities to process such data effectively (Baduge et al., 2022). This paper aims to provide a comprehensive overview and propose an integrated framework for leveraging deep learning techniques to enhance anomaly detection in sensor-based SHM systems. We explore various deep learning architectures, including autoencoders for unsupervised learning (Toufigh & Ranjbar, 2023), convolutional neural networks for feature extraction (Atha & Jahanshahi, 2017), and recurrent neural networks for temporal data analysis (Bao et al., 2018). The integration of transfer learning (Pan et al., 2023), deep reinforcement learning (Kang et al., 2023), and multimodal data processing (Nong et al., 2023) is discussed, alongside different sensor modalities such as fiber optics (Jayawickrema et al., 2022) and acoustic emissions (Haile et al., 2019). The integration of deep learning significantly improves the accuracy, robustness, and real-time capabilities of anomaly detection, enabling early identification of structural damage (Khani et al., 2019; Arafin et al., 2023), sensor faults (El-Shafeiy et al., 2023), and other critical deviations. Physics-informed deep learning is also highlighted for its potential in scattered wavefield reconstruction (Zargar & Yuan, 2024). Deep learning offers transformative potential for enhancing structural integrity and durability by providing advanced, data-driven anomaly detection solutions. Future research should focus on model interpretability, computational efficiency, and validation on diverse real-world datasets.