基于长短期记忆神经网络的板裂纹损伤检测方法

Method for Plate Crack Damage Detection Based on Long Short-Term Memory Neural Network

针对板不同位置裂纹损伤的智能分类问题,提出了一种基于长短期记忆(LSTM)神经网络的板裂纹损伤检测方法.采用Abaqus二次开发建立板裂纹损伤模型,计算高斯白噪声激励下板的加速度响应,并通过数据扩充方法生成数据集,同时考虑了噪声对损伤检测的影响.建立基于LSTM的板裂纹智能检测模型,直接将板的加速度响应作为输入,不需要额外的损伤特征提取,并以最小预测误差为目标,选择模型的超参数,优化模型配置.与多层感知机模型和基于小波包变换的多层感知机模型进行对比表明,本文提出的LSTM模型在板裂纹损伤检测中具有更高的损伤定位精度和更好的适用性.

Aimed at the problem of intelligent classification of crack damage in different positions of the plate,a method for plate crack damage detection based on long short-term memory(LSTM)neural network is proposed.The Abaqus secondary development is used to build the plate crack damage model and calculate the acceleration response of the plate under Gaussian white noise excitation.The data set is generated by data augmentation,and the influence of noise on damage detection is considered.An intelligent crack detection model based on LSTM is established,which directly takes the acceleration response of the plate as the input and does not require additional damage feature extraction.With the goal of minimizing prediction error,the hyperparameter of the model is selected and the model configuration is optimized.The comparison of the multi-layer perceptron model and the multi-layer perceptron model based on wavelet packet transform shows that the LSTM model proposed in this paper has a higher damage location accuracy and a better applicability in plate crack detection.