Semi-supervised consistency models for automated defect detection in carbon fiber composite structures with limited data

Abstract Deep learning has increasingly been adopted in the non-destructive testing industry for detecting defects in carbon fiber composite structures (CFCS), particularly in CFCS-cored aluminum conductor composite core (ACCC) wires. However, the effectiveness of these approaches is often limited by the availability of sizable annotated failure datasets for training purposes. Addressing this challenge, this paper presents a semi-supervised model employing consistency strategies to automate defect detection in CFCS, compensating for the real-world scarcity of samples. It proposes a multi-faceted approach combining synthetic sample generation, transformer-based feature fusion, and a DenseNet architecture-based detection module. Initially, the model generates a large set of synthetic data to mitigate the issue of limited real-world sample availability. These synthetic samples, produced through consistency strategies, are then engineered to complement actual test data. The following step involves a transformer architecture that blends features from synthetic and real samples, refining the dataset for improved damage identification. The final stage features a detection module based on DenseNet architecture, particularly designed for assessing the integrity of CFCS within ACCC wires. Experiments conducted in real-world field scenarios have shown the model’s effectiveness, demonstrating that the hybrid use of synthetic and real samples substantially enhances the training process and damage detection capabilities.