Interpretable hybrid deep learning model for defect detection and tensile strength classification in ultrasonic vibration-assisted friction stir welding

Ensuring consistent quality in friction stir welding (FSW) remains challenging because defect formation and mechanical performance arise from complex thermo-mechanical interactions. This study presents an interpretable convolutional neural network-bidirectional long short-term memory (CNN–BiLSTM) framework for patch-level surface-defect detection and weld-level tensile-strength classification in ultrasonic vibration-assisted FSW AA2060-T8E30 joints. Surface images from 54 welds were divided into six ordered patches to preserve contextual information along the welding direction. DenseNet121 and VGG16 were evaluated as standalone CNNs and hybrid CNN–BiLSTM models. DenseNet121–BiLSTM achieved the best defect-detection performance, with 98.15% fixed-test accuracy and 95.54% mean leave-one-weld-out accuracy, compared with 92.59% and 90.23% for VGG16–BiLSTM. The DenseNet121–BiLSTM model was then applied to three-class tensile-strength classification relative to base-material ultimate tensile strength, achieving 88.89% fixed-test accuracy and 77.78% ± 7.41% under repeated stratified grouped cross-validation. Gradient-weighted Class Activation Mapping (Grad-CAM) visualisations identified surface regions associated with the predictions, while a graphical user interface integrated image input, classification, confidence reporting, and interpretation. The results demonstrate the feasibility of using standard weld-surface images for both visible defect detection and preliminary tensile-strength categorisation to support post-weld decision-making.