Development of an end-to-end deep learning model for acceleration data restoration of a pedestrian bridge

The management of anomalous data presents significant challenges due to its high- volume, randomness, and the requirements for data interpretation precision in structural health monitoring (SHM) systems. This paper presents a novel end-to-end deep learning architecture for the restoration of single-channel acceleration data in SHM. The proposed model directly inputs abnormal data segments into the network for efficient restoration, whereas traditional methods require additional localization of anomalies within smaller sub-segments, a process that is both time-consuming and prone to error. The framework demonstrates autonomous processing capabilities for multiple anomaly types including outliers, gain distortions, linear offsets, precision degradation artifacts, missing data instances, and composite anomaly combinations, without requiring manual intervention. Comprehensive validation using the publicly accessible Dowling Hall pedestrian bridge dataset reveals superior restoration performance across diverse anomaly scenarios, demonstrating its robustness and effectiveness in data restoration. Furthermore, the study emphasizes the critical role of network parameters, positional encoding, and a newly proposed loss function in enhancing model efficacy. This research establishes a paradigm shift toward automated, high-fidelity data restoration techniques in SHM systems, with significant implications for enhancing the reliability and operational efficiency of structural integrity assessments.