Deep Learning-Driven Non-Destructive Testing: Optimizing Defect Detection with Ensemble Strategies

Non-destructive testing (NDT) forms an important protection mechanism in ensuring the structural integrity of composite materials used in aerospace, automotive, energy and environmental engineering fields; defects such as delamination or fibre fracture if left undetected can seriously affect structural performance. This investigation introduces a sound deep learning based approach for defect detection in composite materials, by transfer learning and ensemble techniques applied to the publicly available data set USimgAIST with 7004 ultrasonic inspection images. Sixteen pre-trained convolutional neural network (CNN) models, such as VGG16, MobileNetV2 and DenseNet169, were tested, and the six top performing models were used to perform a combinatorial ensemble analysis. The triple ensemble (VGG16+DenseNet121+MobileNetV2) and six-way ensemble (VGG16+VGG19+DenseNet121+DenseNet169+InceptionResNetV2+MobileNetV2) achieved an outstanding test accuracy of 0.9936, recall, precision and F1-scores higher than 0.99, surpassing the state-of-the-art performances. Comprehensive performance metrics, which include ROC - AUC scores, training dynamics, and computational efficiency, highlight the reliability and scalability of the proposed models. The current results highlight the transformative power of ensemble-based deep learning for the improvement of NDT with the possibility to enhance the quality control of composite material production on a scalable basis.