Synergistic mechanical parameter regulation and signal processing optimization: a robust framework for defect detection in carbon fiber wet winding

Carbon fiber-reinforced epoxy composites are widely used in high-end manufacturing, while wet winding fabrication is prone to fiber misalignment, porosity and delamination that severely degrade component mechanical performance. Existing defect detection methods developed for this process have critical limitations in real-time in-process monitoring, adaptability to multi-source industrial noise, and classification robustness under inevitable process fluctuations. To address these issues, this work proposes an integrated framework combining adaptive signal processing, multi-domain feature engineering and a robust artificial neural network. A three-step preprocessing sequence improves the signal-to-noise ratio of raw signals by 60.00%. Multi-domain features are extracted and reduced to eight discriminative descriptors via principal component analysis coupled with recursive feature elimination. A neural network integrated with residual blocks and squeeze-and-excitation layers is constructed for accurate defect type and severity classification. Validation on NOL ring samples achieves 96.50% overall accuracy, with a strong negative correlation of −0.9084 between predicted defect severity and tensile strength. The model outperforms conventional benchmark methods, and maintains 94.00% accuracy under industrial process fluctuations. This framework achieves an end-to-end detection latency of less than 25 ms on industrial control hardware, which fully meets the real-time requirements of wet winding production, enabling reliable in-process quality control for carbon fiber wet winding processes.