AAE-CycleWGAN fusion framework for generating fused strain data from sparse to dense domains in bridge monitoring systems

Effective monitoring of pedestrian bridges is challenged by noisy measurements, missing data, and time-varying crowd excitation–critical issues for infrastructure and crowd management where timely safety monitoring is paramount. In this work, we quantify human–structure interaction to enable a monitoring system that detects abnormal loads and supports bridge safety management. We also model structural responses under dense-occupancy conditions to predict behavior in busy scenarios, providing decision-ready insights that reduce the risk of serviceability loss. We propose a fusion-centred framework that (i) derives informative, noise-robust features from raw noisy structural strain responses via an optimized adversarial autoencoder (AAE) for denoising and dimensionality reduction, (ii) mitigates data scarcity by synthesizing missing fused modalities using a Cycle-Consistent Wasserstein GAN with Gradient Penalty (CycleWGAN-GP), and (iii) performs downstream condition assessment and crowd–structure interaction analysis with an optimized 2D-CNN. Using only structural sensors, the system infers aspects of crowd movement (e.g., speed and weight proxies), supports real-time safety decisions, and generalizes from unpaired training conditions to predict responses under future or unseen regimes. Validation is conducted on a laboratory-scale pedestrian timber bridge instrumented at midspan with three Fiber Bragg Grating (FBG) strain sensors under multiple scenarios that mimic moving human-induced loads. We evaluate generation quality with standard criteria and assess classification performance on both multiclass and binary tasks. Comparative studies include standard ML baselines and an ablation without CycleWGAN-GP. Results show improved missing-data generation with CycleWGAN-GP and robust condition monitoring on held-out, unseen real data (avoiding leakage): binary classification accuracy improved from 85.21 % to 95.40 %, while multiclass accuracy increased from 85.50 % to 96.65 %. The proposed framework enhances the predictive capability and reliability of bridge monitoring systems by jointly addressing noise, missingness, and crowd-induced variability within a unified SHM pipeline.