Subsurface Defect Detection Using Deep Convolutional Neural Network And Non-contact Laser Ultrasonics

As an important bearing component,alloy parts are widely used in aerospace,rail transit,bridge and ship,nuclear industry and other fields.In the process of fabrication and long-term service,metal components are easy to cause various internal and external defects and lead to serious safety accidents due to errors in manufacturing process,impact,high temperature and high pressure and other loads.Therefore,there is an urgent need for a detection technology to quantify defects.As an advanced non-contact ultrasonic nondestructive testing technology,laser ultrasound is widely used in the integrity and internal damage evaluation of metal components.However,the existing laser ultrasonic testing methods mainly rely on artificially feature extraction of ultrasonic signal,and then establish the mapping relationship between signal features and defect information to realize defect detection and evaluation.However,this procedure is very subjective and empirical,and in most cases it is very difficult to extract the reliable features that are most effective or relevant with the defects to be detected.Therefore,it is urgent to develop an intelligent detection method which can automatically extract the characteristics of laser ultrasonic signals and realize the quantitative evaluation of defects.With the support of the key project of the National Natural Science Foundation of China "Research on key technologies of intelligent detection of large-scale composite aircraft structures with complex shapes using adaptive phased array three-dimensional focus imaging" and Shenzhen Key Laboratory of Smart Sensing and Intelligent Systems,a new method of subsurface defect detection based on deep convolution neural network is proposed in this study.Firstly,the laser ultrasonic signals with different defect widths are simulated by establishing the finite element model,and then the time-frequency analysis of the signals is carried out by wavelet transform to obtain the scalograms with time-domain and frequency-domain characteristics.Finally,the scalograms are input into the deep convolution neural network model to predict the defect width.The innovation of this research work is to obtain the scalograms of the signals with time-domain and frequency-domain characteristics through wavelet transform.Using the advantages of convolutional neural network,it can automatically extract the signal features and quantitatively analyze the defect width with high precision,avoiding the problem of human factors in traditional methods.The main research work is as follows:(1)The finite element model is established to obtain a large number of laser ultrasonic signals of defects with different widths.According to the principle of laser ultrasonic thermoelastic excitation,the finite element model of laser ultrasonic noncontact detection of aluminum alloy is established by using COMSOL software,and the laser ultrasonic simulation signals with different defect widths are obtained.By comparing the experimental signals with the simulation signals,the effectiveness of the finite element simulation model is verified,and 31 laser ultrasonic signals with different defect widths are simulated by the model.(2)Based on wavelet transform,the laser ultrasonic time-domain signals is transformed into scalograms with time-domain and frequency-domain characteristics.Since the subsequent convolutional neural network needs to take the images as the input,the wavelet transform is used to convert the laser ultrasonic time domain signals into scalograms.According to the waveform of the signals to be measured and the characteristics of wavelet basis function,dB4 wavelet transform is selected to process the simulation signals generated from the finite element simulation model and obtain its scalograms.By analyzing the color,position and brightness of pixels in the scalograms,it is verified that the scalograms of laser ultrasonic signals contain information related to the width of subsurface defects.(3)The deep convolution neural network model is used to realize the automatic recognition and classification of defect width.Based on the characteristics of VGG16,ResNet50 and DenseNet161 deep convolution neural network models,a VRD model is established to evaluate the width of subsurface defects.The scalograms obtained through dB4 wavelet transformation are input into the VRD model for training,and the characteristics of the scalograms are automatically extracted through the VRD model to classify the defect width.By comparing the training effects of VRD model and the above three models separately,the advantages of VRD model are revealed.According to the analysis of four model evaluation indexes and confusion matrix,the accuracy of VRD model in defect width classification is verified,which proves the effectiveness and feasibility of this defect detection method based on deep learning proposed in this study.The method proposed in this study combines laser ultrasound,finite element simulation,wavelet transform and deep learning,which can realize the automatic extraction of defect characteristic signals and high-precision identification of subsurface defects.This method can be applied not only to the detection of subsurface defects,but also to the detection of other types of defects,so as to provide important technical guarantee for the service safety of engineering structures.