Session: Structural Health Monitoring I & II, Material Characterization by Ultrasonic wave

Paper Number: 168717

168717 - Stsr-Net: Spatio-Temporal Graph Neural Network for Full-Field Structural Response Prediction 

Abstract: 

We propose a Spatio-Temporal Graph Neural Network for Structural Response Prediction, STSR-Net.
This GNN framework reconstructs time-dependent acceleration responses at unobserved locations or
nodes. It uses partial sensor observations. STSR-Net learns spatial and temporal dependencies effec-
tively across diverse scenarios. The proposed graph representation includes spatial, response, modal,
and structural features to build node embeddings. It also uses physical and dynamic relationships to
define the edge weights. The network extracts localized temporal features using parallel 1D convolutions,
statistical descriptors, Fourier transforms, and wavelet decompositions. It further captures the long-range
temporal features using Gated Recurrent Units with attention mechanisms. These extracted temporal
features at the known nodes are linearly transformed and forwarded to all nodes via an edge-aware graph
attention convolution layer, which utilizes the physical and dynamic relationships between nodes. The
new hidden embeddings are refined via positional encoding and self-attention, following which they are
used in a hierarchical decoding process to reconstruct the acceleration response by separately modeling
the low, and high frequencies via fully connected and transposed convolution layers respectively. The
proposed multi-objective loss function used for training focuses on the graph Laplacian regularization,
temporal smoothness, and reconstruction error. Experimental validations on the American Society
of Civil Engineers (ASCE) SHM benchmark demonstrate that our framework can accurately predict
structural responses across diverse scenarios.

Presenting Author: Anowarul Habib UiT The Arctic University of Norway

Presenting Author Biography: Anowarul Habib is the Group Leader of the Acoustic Sensing and Imaging Lab at UIT The Arctic University of Norway. He earned his BSc and MSc degrees from the University of Leipzig, Germany, and completed his PhD at the University of Siegen, Germany. Following his PhD, Dr. Habib held a postdoctoral position at Goethe University Frankfurt. In 2011, he received the “Best Young Scientist Award” at the Symposium on Ultrasonic Electronics in Japan. He is currently serving as an Associate Editor for the Journal of Nondestructive Evaluation, Diagnostics, and Prognostics of Engineering Systems. His primary research interests include the application and development of scanning acoustic microscopy, point contact excitation and detection methods for wave propagation in piezoelectric materials, high-frequency polymer transducer fabrication, and structural health monitoring.

Authors:

Adarsh Gupta Indian Institute of Technology Guwahati
Shivam Ojha Indian Institute of Technology Guwahati
Shahrul Kadri Ayop Sultan Idris Education University
Anowarul Habib UiT The Arctic University of Norway