Abstract
As of early 2024, the maintenance of civil infrastructure has shifted from traditional periodic inspections toward continuous, sensor-based Structural Health Monitoring (SHM). This transition has been catalyzed by the proliferation of the Internet of Things (IoT) and high-fidelity sensor networks, leading to a 'sensor mania' that generates vast volumes of heterogeneous data. However, the efficacy of these systems is often hampered by data anomalies, sensor faults, and the challenge of distinguishing between structural degradation and environmental noise. This research proposes a robust framework for anomaly detection in SHM by integrating deep learning (DL) architectures, specifically focusing on multimodal deep neural networks and unsupervised learning paradigms. We evaluate the performance of a hybrid model combining Deep Autoencoders (DAE) with Isolation Forests and compare it against traditional Support Vector Machine (SVM) and standard Convolutional Neural Network (CNN) approaches. Utilizing datasets comprising strain, acceleration, and temperature readings, the study demonstrates that multimodal integration significantly reduces false-positive rates caused by environmental fluctuations. Our results indicate that the DAE-based reconstruction error provides a high-sensitivity metric for identifying subtle structural changes, while the integration of transfer learning enhances model adaptability across different structural types. The findings underscore the potential of physics-informed deep learning to bridge the gap between data-driven patterns and mechanical principles, ensuring the long-term durability and integrity of critical infrastructure. This study provides a comprehensive benchmark for DL-based anomaly detection as of January 2024, offering insights into the transparency and scalability of these models in real-world deployment scenarios.