A hybrid attention-enhanced denoising autoencoder-SVR method for quantitative assessment of fatigue crack propagation in composite laminates

Accurate quantification of naturally initiated fatigue cracks in carbon fiber reinforced polymer (CFRP) laminates is critical for structural integrity assessment. This paper proposes a multi-stage framework integrating an attention-enhanced denoising autoencoder (Attn-DAE) and a moth-flame optimization (MFO) tuned support vector regressor (SVR). The Attn-DAE is developed to suppress modal interference and extract crack-sensitive latent features from highly dispersive guided wave signals. These features are subsequently mapped to crack lengths using an SVR model, where an MFO scheme is employed to globally optimize hyperparameters and avoid the local-minimum convergence typical of conventional algorithms. To ensure realistic damage evolution, the framework is validated using tension–tension fatigue experiments conducted on open-hole CFRP coupons without artificial pre-cracks. Under a rigorous leave-one-state-out cross-validation, the proposed approach achieves an R 2 of 0.9797 and an RMSE of 1.0246 mm, outperforming multiple baselines including PCA, BPNN, and traditional diagnostic indicators. Results demonstrate that the proposed approach delivers robust feature learning, globally optimized regression, and effective crack length quantification for offline guided-wave data within the tested experimental framework.