基于声发射的铝蜂窝板超高速撞击损伤模式识别方法

摘要

A damage pattern recognition method based on neural network is proposed to recognize the damage state of aluminum honeycomb core sandwich under hypervelocity impact (HVI) through acoustic emission.A variety of experimental signals are obtained,10 characteristic parameters related to damage are presented by test of nonparametric analysis the relationship with damage pattern,combining with precise source localization,time-frequency analysis,wavelet transformation and modal acoustic emission technology.The BP neural net mode based on Bayesian regularization is established by analyzing the relationship with damage pattern using nonparametric analysis.After establishing the optimal network model,two optimal combinations are selected by analyzing the recognition ability of different parameter combinations,the damage pattern recognition ability is verified with non-same source sample.The result shows that propagation distance is a significant parameter but irrelevant to damage pattern.Automatic window wavelet energy ratio within 125-250 kHz frequency range decrease the ability of damage pattern recognition.Using a Bayesian regularization neural network with combination of 9 parameters,including propagation distance,rise time,hold time,cut-off frequency,4 kinds of automatic window wavelet energy ratio and wavelet energy entropy,presents 9.38％ wrong point rate to a group of random non-same source sample.%为通过声发射技术识别铝合金蜂窝板超高速撞击(HVI)的损伤状态,提出一种基于神经网络的损伤模式识别方法.通过超高速撞击实验获取声发射信号,结合精确源定位技术、时频分析技术、小波分析技术及模态声发射技术,提出了10个与损伤相关的特征参数,通过非参数检验分析其与损伤的关系,设计了一种基于贝叶斯正则化BP神经网络的超高速撞击损伤模式识别方法.建立最优网络模型,通过不同参数组合识别能力分析,优选出2种特征参数组合,通过非同源样本对其损伤模式识别能力进行验证.结果表明:传播距离与损伤模式无关,却是识别损伤模式的重要参数;125～250 kHz频域的自动加窗小波能量比会降低损伤模式的识别能力;采用贝叶斯正则化的BP神经网络可以较好地识别蜂窝板超高速撞击损伤模式,参数组合为传播距离、上升时间、持续时间、截止频率、4个自动加窗小波能量比及小波能量熵,共9个参数,对任意选取非同源样本识别错分率仅为9.38％.