Global-local feature cross-fusion network for ultrasonic guided wave-based damage localization in composite structures

The precise and stable detection of damage in carbon fiber composite material structures holds immense significance in ensuring the continuous, secure functioning of equipment used for long-term services and averted calamities. This paper proposes a novel damage localization method for composite materials structures named global-local feature cross-fusion network (GLFCFN), which utilizes advanced deep learning methodologies, namely convolutional neural network (CNN), gated multi-layer perceptron (gMLP), and multi-head attention mechanism, coupled with ultrasonic guided wave (UGW) based structural health monitoring (SHM) technique, in order to improve the damage localization performance. Firstly, a novel feature extraction method is proposed, which constructs local feature extractor and global perception feature extractor based on CNNs and MLPs, respectively, to extract more comprehensive damage indices. Secondly, a feature fusion method based on the multi-head attention mechanism is developed to allow interactive fusion of local features and global perception features to automatically summarize the important feature representations of damage information, thereby effectively improving the accuracy and stability of damage localization. Finally, the performance of the method is evaluated using simulated damage experiments on a composite material laminate with only four transducers. The results indicate that the proposed method achieves a mean relative error of 3.54 % on the 100 unseen samples, with a standard deviation of localization errors at 5.83 mm, compared with other deep learning-based damage localization methods, the proposed GLFCFN method has superior performance.