A quantitative damage diagnosis method for CFRTP/Ti6Al4V hybrid welded/bolted joints based on an attention-enhanced capsule network

Abstract To achieve high-precision quantitative damage identification for CFRTP/Ti6Al4V hybrid welded/bolted joints (HWBJ) after low-velocity impact (LVI), this work developed a damage diagnosis method based on acoustic emission (AE) and a self-attention enhanced capsule network (SA-CapsNet). First, AE waveform signals from HWBJ were collected through LVI tests and converted into time–frequency images using short-time Fourier transform as input features. On this basis, data augmentation techniques were employed to construct the dataset, effectively mitigating the learning challenges under few-shot conditions. The proposed SA-CapsNet was further applied for quantitative damage diagnosis and compared with traditional capsule network (TR-CapsNet) as well as classical models such as residual network (ResNet), convolutional neural networks (CNN), visual geometry group (VGG), and multilayer perceptron (MLP). The results indicate that the SA-CapsNet, which integrates self-attention into the dynamic routing process, achieves a higher recognition accuracy of 99.47%, compared to TR-CapsNet (98.44%), ResNet (96.35%), CNN (95.31%), VGG (94.79%), and MLP (93.23%). Moreover, comprehensive evaluation metrics show that the average precision, recall, and F1-score of SA-CapsNet reached 0.994846, 0.994792, and 0.994792, respectively—all outperforming the corresponding values of other comparative models. Notably, SA-CapsNet not only demonstrates superior statistical performance but also exhibits significantly enhanced misclassification control, highlighting its remarkable potential for high-reliability structural health monitoring applications.