Characteristic mapping using HTTGCN in structural materials

Existing multi-source heterogeneous data fusion methods either flatten the high-order structure into vectors, losing critical cross-mode interactions，or adopt a two-stage 'decomposition-then-prediction, paradigm that prevents end-to-end optimisation.To address these gaps, this paper proposes hierarchical tensor Tucker-graph convolutional network, a novel framework that integrates hierarchical Tucker decomposition with graph convolutional networks in an end-to-end learnable manner.Multi-source monitoring data is organised as a fourth-order tensor; hierarchical Tucker decomposition explicitly extracts high-order correlation features, while a sensor topology graph enables graph convolutional networks to aggregate spatial dependencies.On the Z24 bridge benchmark, the framework reduces the root mean square error of first-order frequency prediction by 15.8% compared to the state-of-the-art spatial-temporal transformer, with p < 0.001 and a large effect size (Cohen's d = 1.92).The framework provides a highly accurate and interpretable path from damage-sensitive data to quantitative performance assessment in building structural health monitoring.