A Reliable Virtual Sensing Architecture With Zero Additional Deployment Costs for SHM Systems

Structural health monitoring (SHM) serves to safeguard the operational safety of building structures; however, the high cost of SHM nodes limits its large-scale applications. In this article, we propose a novel computational model that integrates the physical model of SHM sensing to generate "virtual" sensor nodes with reliable data output at zero additional deployment cost, thereby enabling cost-efficient sensing for SHM systems. To achieve this, we build a generative adversarial network (GAN) combined with the physical model and design a discriminator to ensure that the generated virtual sensor node data aligns with the authentic physical characteristics. The generator employs a 1-D convolutional layer in a convolutional neural network (CNN) and a bi-long short-term memory network (LSTM) model to capture spatial-temporal correlations, along with a weighted smoothing algorithm to reduce noise while preserving data integrity. To support the model, we design a spatial-channel attention mechanism to enhance robustness. We conduct tests on the real-world dataset of the Belgian railway bridge KW51, and the results indicate that our system can generate virtual sensor nodes with 98.2% accuracy toward the ground truth without the need to deploy new devices (with no additional deployment cost). Hence, with its reliable sensing and cost-efficient features, we believe that our system could be helpful in facilitating the large-scale application of SHM systems, thereby providing effective safety monitoring for a wider range of buildings.