AEFORMER: A LIGHTWEIGHT CONV1D-TRANSFORMER MODEL FOR ACOUSTIC EMISSION BASED REAL-TIME STRUCTURAL HEALTH MONITORING

Abstract

Structural Health Monitoring (SHM) plays a critical role in ensuring the long-term safety and sustainability of civil infrastructure, including bridges, tunnels, dams, and reinforced concrete systems. Among various non-destructive evaluation (NDE) techniques, Acoustic Emission (AE) sensing has emerged as a reliable approach for capturing micro-crack propagation and energy release during stress-induced fracture. However, deploying deep learning models for real-time AE classification on embedded SHM platforms remains challenging due to hardware limitations and computational constraints. To address this challenge, we propose AEFormer, a lightweight hybrid deep learning architecture that integrates 1D Convolutional Neural Networks (Conv1D) for local fea ture extraction with Transformer encoders to capture long-range temporal dependen cies. Unlike conventional CNN and Tiny ANN approaches, AEFormer is explicitly op timized for edge-based deployment, achieving high predictive accuracy while maintain ing a compact computational footprint. Experiments were performed on a benchmark dataset comprising 15,000 AE signals sampled at 5 MHz, representing tensile, shear, and mixed-mode cracking. AEFormer achieves 99.82% test accuracy with per-class F1 scores above 0.998, outperforming lightweight CNN and TinyML baselines with fewer than 28,000 trainable parameters. The results demonstrate that AEFormer provides a highly dependable and efficient solu tion for real-time, embedded SHM applications, offering strong potential for deployment in safety-critical monitoring of concrete infrastructure.