Integrity-aware multi-sensor health monitoring of bridges via denoising variational autoencoder

Abstract

Accurate bridge health monitoring is essential for safety, yet existing methods still suffer from false alarms and limited multi-sensor integration. This study proposes DAVE-PI (Denoising Autoencoder–Variational Embedding Probabilistic Index), a unified and reliable alarming framework for bridge health monitoring. The proposed method constructs a probabilistic health index by fusing multi-source sensor data using a DAVE and integrating latent-space reconstruction distances and divergence measures to quantify deviations from normal behavior. An adaptive integrity-aware alert mechanism suppresses false positives while reliably detecting structural anomalies. Comparative analyses against convolutional neural networks, one-class classification, variable cumulative error anomaly detection, and empirical machine learning show that DAVE-PI identifies the key anomalies, reduces false alerts, and achieves a 25.3% increase in total anomalies and a 22% improvement in average anomaly density over the next-best models. Compared with existing approaches, DAVE-PI provides a more reliable and scalable alarming strategy, offering a practical solution for intelligent bridge health monitoring.