LSA-HELM: A Boosted Configuration to Optimize the Reconstruction Model of Dynamic Force on Aviation Composite Skin

Abstract

The reconstruction of impact forces is critical for structural health monitoring of aero-composite wings. However, due to its complex structure with a limited sensor array, the impact force cannot be accurately determined simply by using an equal-weight transfer function. Meanwhile, the introduction of complex models can improve the accuracy of reconstruction but also increase the computational complexity and running time. To address this issue, a method combining lightweight spatiotemporal attention mechanism and extreme learning machine (ELM) (LSA-HELM) is proposed. By introducing a lightweight spatiotemporal attention mechanism, the input data are weighted to capture key features effectively. Then, the mapping relationship between impact force and strain array is constructed by using Hermite polynomials as the ELM of activation function. The suggested method is verified on an aircraft composite plate. The experimental results show that the peak relative error (PRE) is 4.62% for LSA-HELM, 11.03% for Bayesian, 13.33% for convolutional neural network (CNN), 7.31% for Tiny1DCNN and 9.82% for transformer. It shows under the condition of limited sample number and scarce data features, the proposed method has obvious advantages in terms of reconstruction accuracy and real-time performance and is superior to other methods based on machine learning and traditional analysis methods.