| Terahertz technology is a cutting-edge nondestructive testing technology.It has been widely used in object imaging,safety monitoring,nondestructive testing and other fields.As terahertz signal is an unstable nonlinear signal,it is difficult to recognize and extract features,and low efficiency of signal analysis.The principal reason for these problems is the high dimension of data,complex features and large amount of calculation.In order to solve the above problems,the sparse representation based terahertz signal feature extraction and recognition algorithm research was carried out.Firstly,the sparse representation algorithm was used to decompose and reconstruct the terahertz signal to achieve the purpose of signal dimension reduction and compression.Secondly,the flight time of the signal was extracted to realize the thickness characterization and defect recognition of the terahertz signal.At last,the multi-adhesive structure sample is taken as the research object,combined the sparse representation model and terahertz imaging technology to identify the disbonding defects of adhesive layers.The main research work of this paper is as follows:(1)In view of the problems of low accuracy of feature recognition and difficulty of feature extraction caused by a large amount of clutter and noise in the actual acquisition of terahertz signals.Based on the generation,detection and propagation mechanism of terahertz,a method suitable for dimensionality reduction of terahertz signal data is studied and discussed theoretically.Through sparse reconstruction of signals,the accuracy of signal feature recognition is improved and effective features of signals are accurately extracted.Thus,the defects of the sample can be detected and identified better.In order to effectively achieve the dimensionality reduction of terahertz signal features,this paper firstly decomposed and reconstructed terahertz signals based on two sparse decomposition algorithms,matching pursuit and orthogonal matching pursuit.Considering the large number of atoms in the fixed overcomplete dictionary.Dictionary learning algorithm was added on the basis of sparse decomposition algorithm,improving the denoising performance and analysis efficiency of the algorithm.Decomposed and reconstructed simulation signals with different signal-to-noise ratios and measured signals.The performance of the algorithm was evaluated by four indexes,including signal-to-noise ratio,signal-to-noise ratio increment,root mean variance and correlation coefficient.The results show that the reconstruction performance of sparse decomposition algorithm is significantly improved after the addition of learning dictionary.(2)In view of the problems such as low signal-to-noise ratio echo and overlapping echo in the actual detected terahertz echo signal,which make it difficult to accurately extract the time of flight and low thickness measurement accuracy.Proposed a sparse representation thickness measurement algorithm based on convolutional model in this paper.Firstly,constructed an over-complete atomic dictionary based on the convolution model of terahertz echo signal.Secondly,reconstructed the impulse response signal,and retrieved the peak value by the threshold method to solve the flight time difference.Finally the signal thickness characterization was realized.The thickness measurement experiments of sparse representation algorithm and traditional time-of-flight extraction method are carried out.The results show that the measurement accuracy of sparse representation algorithm is higher than that of traditional time-of-flight direct extraction method by 15.78%,and that of frequence-wavelet domain deconvolution method by 4.97%.Based on the convolutional sparse representation algorithm,the application technology of thickness detection is studied.Measured the thickness of the low SNR echo signal snd overlapping echoes.Sparse representation algorithm is further verified the effectiveness of the thickness and robustness.(3)In view of the large amount of invalid redundant feature information in the signal obtained from the detection of multi-layer bonded structures with debunking defects by the terahertz time-domain spectral system,the correlation between over-complete dictionary atoms decreases,the number of dictionary atoms increases,and the efficiency of signal analysis decreases in sparse representation algorithm.In this paper,a gradient threshold adaptive sparse algorithm is proposed based on the time-domain characteristics of multibonded terahertz signals,and the sparse representation algorithm is improved.Firstly,effective features of terahertz signals are extracted based on the second-order gradient method.Secondly,sparse thresholds are determined adaptively according to time-domain characteristics of terahertz signals with time-of-flight errors as the restriction condition,so as to reduce redundant invalid information of signals.The signal compression rate of this algorithm can reach 81%,the relative root mean square error is less than 2%,and the data computation time is reduced by 20% compared with the traditional signal compression algorithm,and the data memory space occupation is reduced by 95%,realizing the effective terahertz signal compression.Finally,combining terahertz imaging technology and binary image segmentation method to identify the defect area of terahertz image and calculate the defect area,the results show that the gradient threshold adaptive sparse algorithm can improve the efficiency of terahertz signal analysis and ensure the accuracy of defect recognition. |
