CTFT-based Approach for Parameter Identification in Reinforced Concrete

Abstract

Reinforced concrete structures are critical to numerous infrastructures, making quality inspection essential. Electromagnetic induction techniques, such as pulsed eddy current (PEC), are widely applied in nondestructive testing (NDT) due to their high metal sensitivity and strong anti-interference capability. However, electromagnetic induction signals exhibit strong local perturbations, complex temporal nonlinearities, and significant time–frequency coupling. Existing methods often depend on neural networks that struggle to capture such nonlinear temporal patterns, or require additional sensors, which increase complexity and cost. To this end, this paper constructs a hybrid dataset of simulated and experimental data and proposes CTFT, an end-to-end regression model integrating CNN, Multi-stage Sparse Transformer, Task-Specific Attention (TSA), and Frequency Encoding. By synergizing CNN’s local feature extraction with Transformer’s global modeling, the model effectively adapts to the characteristics of electromagnetic induction data. A task-specific attention mechanism is introduced to adaptively fuse branch features, while a mean-based sparse attention and multi-stage aggregation strategy mitigate the quadratic computational cost of Transformers. Furthermore, a frequency-encoding mechanism embeds excitation frequency priors, enhancing modeling under multi-frequency conditions. Optimized via a ‘simulation pre-training and mixed-data fine-tuning’ strategy, the model achieved ±1 mm accuracies of 92.65% for rebar diameter and 97.07% for cover thickness. With ±2 mm accuracies reaching 98.93 and 99.05%, the model demonstrates superior performance and significant engineering potential.