A High-Accuracy Cross-Device Data Integrity Framework for Trustworthy SHM Sensing

Abstract

The dense networked deployment of structural health monitoring (SHM) sensors is driving engineering structures toward comprehensive digital monitoring, enabling real-time health assessment through multipoint data acquisition. However, their extensive distribution exposes sensor networks to complex security risks. Factors, such as protocol discrepancies among heterogeneous nodes, resource constraints, and open deployment environments, pose significant challenges to data integrity. To address this, this article proposes TVCAX, a novel anomaly detection model composed of four essential components: transformer, variational autoencoder (VAE), cross-attention, and XGBoost. The model’s core philosophy lies in predicting the time series of sensor data and comparing it with actual values by exploiting the correlation of direct sensing information from each node, thereby enabling cross-device data integrity verification. Additionally, to improve detection accuracy in complex environments, an adaptive threshold algorithm is designed to dynamically adjust thresholds based on the statistical characteristics of real-time data. Validation using the real-world KW51 railway bridge dataset shows that the system achieves an error rate of only 0.09% in generating sensor data replicas and a precision rate of up to 99.8% in anomaly detection. Experimental results demonstrate that the TVCAX model efficiently realizes anomaly detection and provides a robust foundation for SHM security.