Deep learning-based prediction and classification of corrosion in RC structures via PZT-based EMI signatures and equivalent modeling

The present study introduces a comprehensive experimental–numerical–computational framework for the automated assessment of corrosion-induced damage in reinforced concrete (RC) structures using high-dimensional electro-mechanical impedance (EMI) signatures. Unlike prior studies limited to specific corrosion stages, this work utilizes large-scale EMI datasets generated from accelerated experiments and validated COMSOL simulations to enable data-driven prediction of reinforcement mass loss across a continuous degradation spectrum. A calibrated equivalent spring–mass–damper model is developed to establish a physical link between EMI signatures and structural degradation, accurately capturing impedance behavior in the 150–200 kHz range, and quantifying stiffness loss due to corrosion. To address the challenge of high dimensionality and redundant information inherent in EMI data, the Orthogonal Matching Pursuit (OMP) algorithm is employed for signal compression and reconstruction. This technique preserves critical frequency-domain features while significantly reducing input dimensionality, thereby enhancing both training efficiency and model generalization. A series of supervised machine learning (ML) models, including linear regression, support vector regression (SVR), and ensemble techniques such as random forest and XGBoost, are evaluated for the regression of reinforcement mass loss. Among them, random forest demonstrated the highest performance (R² = 0.949), capturing non-linear relationships between conductance spectra and damage evolution. Furthermore, advanced deep learning (DL) architectures—namely a baseline 1D Convolutional Neural Network (CNN), CNN integrated with Long Short-Term Memory (CNN–LSTM), and CNN combined with Bidirectional LSTM (CNN–BiLSTM)—are developed to exploit both spatially localized patterns and temporal correlations in EMI signals. The CNN–BiLSTM model achieved superior accuracy (R² = 0.9966), particularly when trained on OMP-compressed data, highlighting the effectiveness of compressed EMI representations. A novel 1D CNN-based classification model was developed to autonomously categorize corrosion severity into predefined classes—uncorroded, mild, moderate, and severe—without relying on handcrafted features or intermediate signal processing. The model achieved flawless classification performance (Precision, Recall, F1-score = 1.00) across all severity levels, clearly demonstrating its capability to learn physically meaningful representations of structural degradation without manual feature extraction. This research lays a strong foundation for real-time corrosion monitoring in RC structures by integrating smart sensing, high-fidelity simulations, and intelligent modeling. Future extensions include deployment on edge devices, integration with wireless sensor networks, adaptation to diverse RC systems, and incorporation of uncertainty quantification for more reliable structural health management.