| description abstract | Abstract. Structural health monitoring in wetland environments poses significant challenges due to highly variable operational conditions and the difficulty of acquiring accurate, representative load spectra. To address these issues, this study proposes a novel multicondition load spectrum fusion framework that integrates orthogonal experimental design, multiphysics finite element simulation, and a hybrid convolutional neural network (CNN)–bidirectional long short-term memory (BiLSTM)–attention deep learning model. A total of 25 representative working conditions, incorporating variations in slope, friction, and wind resistance, were systematically designed using orthogonal principles. Coupled dynamic-static finite element simulations were then used to generate corresponding stress–strain responses. The multisource load spectra were processed by the proposed deep learning model, which hierarchically extracts and fuses key features. CNN module captures localized spatiotemporal patterns, BiLSTM module models bidirectional long-range dependencies, and attention mechanism dynamically highlights condition-specific features, enhancing predictive performance. Validation against field data from wetland monitoring equipment yielded a prediction accuracy of 95.53% for the composite load spectra, significantly outperforming traditional single-condition models. The fused spectra also improved fatigue life estimation, demonstrating strong robustness and generalizability under complex terrain conditions. Prediction errors remained consistently low, confirming the model's reliability. This study presents a cost-effective, data-driven approach for structural reliability assessment, reducing field measurement requirements by approximately 80% without compromising accuracy. Future work will explore the integration of physics-informed constraints and expand the framework to multiscale and heterogeneous structural systems. | |
