Intelligent defect detection in composite materials: the role of deep learning and machine learning fusion

This study presents a novel integration of deep learning (DL) and machine learning (ML) for non-destructive testing (NDT) of composite materials, achieving superior defect detection performance. By combining pre-trained convolutional neural networks (CNNs), such as DenseNet169 and ResNet152, with ML algorithms like Multi-Layer Perceptron (MLP) and K-Nearest Neighbours (KNN), we leverage transfer learning and feature selection (Variance Threshold and XGBoost) to enhance classification accuracy. Experiments on the USimgAIST dataset (7,004 ultrasonic images) demonstrate that the DenseNet169 + MLP model achieved a state-of-the-art accuracy of 99.57%, Cohen’s Kappa of 0.9912, and only 2 false negatives against 4 false positives on the test set, minimising the risk of undetected structural defects. Feature selection significantly reduces computational complexity while maintaining high sensitivity (99.71%) and specificity (99.45%). These results highlight the potential of DL-ML fusion for automated quality control and structural integrity assessment in composite materials. Future work will focus on optimising real-time applications and extending the approach to diverse datasets.