Research On Damage Monitoring Of Plate Structures Based On Convolutional Neural Networ

In the long-term service of critical parts of aircraft such as outer wings,drag panels,and nose cowlings,these plate structure parts are susceptible to threatening behaviors such as hail impacts and bird flock collisions,resulting in damage to the components and seriously threatening the safe operation of the aircraft,so structural health monitoring of plate structures is a research topic of great significance.In this thesis,damage on aluminum plates is monitored using Lamb waves using piezoelectric wafer sensors as actuators and transducers.The traditional processing of the signal,such as Fourier analysis,wavelet analysis,and damage index method,requires manual extraction of features,and although the features are artificially highlighted,many deep-level features are also lost in the process,which does not make full use of the signal data and affects the accuracy while requiring a high level of expertise from the researcher.This thesis investigates the application of convolutional neural networks(CNN)and long and short-term memory(LSTM)networks in the monitoring of structural damage in slabs.In this thesis,we propose two models based on one-dimensional convolution neural network(1-D CNN)and two-dimensional convolutional neural network(2-D CNN-LSTM)for damage monitoring on slab structures and experimentally validate their effectiveness,and explore the influence of the number of sensors and their arrangement on the accuracy of damage localization.The main contents are as follows:(1)This thesis describes the basic principles of CNN model and LSTM model,designs a1-D CNN that can directly process Lamb wave time domain signals,uses the original signal and the difference signal as the input to compare the experimental results respectively,determines the choice of the input signal in the experiment,and then explores the influence of the number and arrangement of sensors on the damage localization accuracy of the plate structure,and finally trains the resulting net model with a compact structure,and the experiment shows that the method can locate the damage area on the aluminum plate relatively quickly.(2)Considering that CNN lacks the attention to signal time dimensional features,this thesis introduces LSTM into CNN,and given that 2-D CNN has excellent feature extraction ability for images,which can effectively improve the accuracy of image data classification or regression,2-D CNN-LSTM is designed for damage localization of plate structures.In this method,the difference signals formed between Lamb wave signals and reference signals at different damage locations are converted into two-dimensional images as training samples.The results show that the method can effectively distinguish the signals collected at different damage locations to achieve the determination and localization of damage,and the improvement of the localization accuracy of damage made by the method in this thesis is verified by comparison experiments.