基于神经网络方法的鸟撞飞机风挡冲击载荷反演

以鸟撞实验中传感器实测信号为基础,结合有限元正问题计算方法与神经网络理论,构造小波动态延时反馈神经网络,并详细分析了该网络的结构参数、对比了网络单点应变输入法、两点应变输入法以及三点(多点)应变输入法的训练效率与反演精度.构造的神经网络可以高精度地反演出鸟撞飞机风挡过程中冲击载荷时间历程,同时具有较高的抗干扰能力,且训练过程平稳、训练效率高.根据已有的研究成果,提出了鸟撞实验应变传感器建议布置,可以在满足实验测量要求的基础上简化实验过程,提高实验效率.

NEURAL-NETWORK BASED BIRD STRIKE LOADINGS INVERSE TO AIRCRAFT WINDSHIELD

It is difficult to accurately capture the transient history of bird impact to aircraft windshield throngh conventional experimental methods. Based on the measured real time signals of bird strike experiment and finite element numerical solutions, this paper constructs a dynamically delayed feed wavelet (DDFW) neural network to inverse the impact loadings. The structural parameters, training efficiency and inverse precision of the network are studied in detail by comparing single point, bi point and triple point strain input methods. As a result, the DDFW neural network is effective for the impact loading inverse of bird strike windshield with high precision and strong anti jamming capability, as well as smoothly training process and high efficiency. Based on the research, this paper also suggests a strain sensors layout scheme for bird strike experiment, which can simplify experimental measures, and improve the experimental efficiency.