CN-122015669-A - Composite material terahertz thickness measurement method based on physical interpretable sparse expansion network

CN122015669ACN 122015669 ACN122015669 ACN 122015669ACN-122015669-A

Abstract

A terahertz thickness measurement method for composite materials based on a physical interpretable sparse expansion network relates to the crossing field of terahertz nondestructive detection technology and artificial intelligence, and comprises the steps of acquiring terahertz time-domain echo signals of the composite materials to be measured, establishing a sparse inverse problem model, constructing a physical-guided interpretable sparse expansion network, designing a dynamic sparse weighted focus loss function and calculating thickness of each layer.

Inventors

XU YAFEI
LI PEIHAN
ZHANG HUA
LIU DATONG
LIU CHANG

Assignees

哈工大郑州研究院
哈尔滨工业大学

Dates

Publication Date: 20260512
Application Date: 20260213

Claims (10)

1. The terahertz thickness measurement method of the composite material based on the physical interpretable sparse unfolding network is characterized by comprising the following steps of: s1, acquiring a terahertz time-domain echo signal of a composite material to be tested; s2, establishing a sparse inverse problem model based on a terahertz wave propagation mechanism and a convolution sparse coding theory; s3, constructing a physical-guided interpretable sparse expansion network, and embedding a physical mechanism into the deep neural network through an optimization process of an expansion iteration shrinkage threshold algorithm, wherein the interpretable sparse expansion network comprises a relaxation dictionary learning module of physical constraint, a mixed scale dense connection strategy, a dynamic threshold mechanism and an effective compression-excitation module; s4, designing a dynamic sparse weighted focus loss function, and solving the problem of weak signal missing detection caused by extremely sparse interlayer reflection signals; s5, calculating the thickness of each layer according to the flight time sequence output by the network.
2. The terahertz thickness measurement method of composite materials based on a physically interpretable sparse deployment network according to claim 1, wherein in step S1, the step of obtaining terahertz time-domain echo signals of the composite materials to be measured includes: S11, preparing a composite material sample comprising a single layer, multiple layers and a damaged structure; S12, acquiring terahertz data through an ultrafast femtosecond laser source, an optical delay line, a central control unit, a terahertz transmitter-receiver and an x-y motion platform by using a terahertz time-domain spectroscopy system in a reflection mode.
3. The method for terahertz thickness measurement of composite materials based on a physically interpretable sparse deployment network according to claim 1, wherein in step S2, the step of establishing a sparse inverse problem model includes: s21, converting time domain waveforms of an incident signal and a reflected signal into a frequency domain by utilizing Fourier transformation; s22, analyzing reflection and transmission characteristics of terahertz waves at a medium interface by using a Fresnel formula; S23, modeling the signals as sparse convolution combinations of a plurality of dictionary atoms, and solving an L1 regularization problem to restore sparse vectors.
4. The method for terahertz thickness measurement of composite materials based on a physically interpretable sparse deployment network according to claim 1, wherein in step S3, the step of constructing a physically guided interpretable sparse deployment network comprises: s31, adopting a physical constraint relaxation dictionary learning strategy, constructing a parameterized physical kernel function to simulate an ideal terahertz echo, and introducing a relaxation term to adaptively fine tune waveform details; S32, designing a deep sparse expansion framework, wherein each layer corresponds to one iteration step of an optimization algorithm and comprises physical initialization, cascade iteration updating and feature fusion based on an attention mechanism; S33, introducing a mixed scale dense connection strategy and a dynamic attention threshold mechanism into each layer, and improving the capability of network fitting sparse vectors.
5. The terahertz thickness measurement method of composite material based on a physically interpretable sparse deployment network of claim 4, wherein in step S31, a parameterized physical kernel function is constructed based on a terahertz wave propagation mechanism A linear superposition of two gaussian derivative functions is used to simulate an ideal physical echo: ; Wherein, the Representing a learnable set of physical parameters, controlling the amplitude ratio, the pulse width and the time delay respectively; Introducing an additional slack term over the physical core Dictionary atoms The definition is as follows: 。
6. The terahertz thickness measurement method of composite materials based on a physically interpretable sparse expansion network according to claim 4, wherein in step S32, the sparse codes are initialized by a physically constrained relaxation dictionary and are realized by transpose convolution when physical initialization is performed: 。
7. the method of terahertz thickness measurement of composite materials based on a physically interpretable sparse deployment network according to claim 4, wherein in step S32, the physically guided interpretable sparse deployment network end incorporates a compression excitation module: ; Wherein, the The Sigmoid function is represented as a function, Represents an average pooling of the data in the pool, Representing the weights of the fully connected layers.
8. The terahertz thickness measurement method of composite material based on a physically interpretable sparse deployment network of claim 4, wherein in step S33, a feature map is input Outputting a and inputting graph through dynamic threshold module Threshold values with the same dimension : ; Wherein the method comprises the steps of Is a base threshold parameter; is a lightweight attention module comprising two layers of convolution.
9. The method of terahertz thickness measurement of composite materials based on a physically interpretable sparse unfolding network according to claim 4, characterized in that in step S4, the step of designing the dynamic sparse weighted focal point loss function comprises: S41, introducing a focus loss mechanism, and reducing the dominant position of a large number of simple negative samples in gradient updating; s42, designing a dynamic weight coefficient, and dynamically adjusting the loss weight according to the proportion of positive samples in the real label to prevent weak signal missing detection; The dynamic sparse weighted focus loss function is ; Wherein the method comprises the steps of Representing network output Applying a Sigmoid function to obtain a result; Is a positive sample indication function; is a dynamic weight coefficient; ; Wherein the method comprises the steps of Scaling factors for sensitivity; Definition of focusing terms The method comprises the following steps: ; Wherein the method comprises the steps of In order to predict the probability of a probability, Is a focus parameter.
10. The method of terahertz thickness measurement of composite materials based on physically interpretable sparse unfolding network according to claim 1, characterized in that in step S5, when terahertz waves are perpendicularly incident, the thickness is defined as: ; Wherein, the In order to measure the thickness of the sheet, In order to achieve the light velocity, the light beam is, For the time difference of adjacent pulses in the time-of-flight sequence output by the network, Is the refractive index of the sample.

Description

Composite material terahertz thickness measurement method based on physical interpretable sparse expansion network Technical Field The invention relates to the crossing field of terahertz nondestructive testing technology and artificial intelligence, in particular to a terahertz thickness measuring method for a composite material based on a physical interpretable sparse unfolding network. Background The composite material has high specific strength, corrosion resistance and other excellent performances, and is widely applied to the fields of aerospace, wind power generation and the like. However, defects such as thickness variation, layering and the like are easy to generate in the manufacturing and service processes, and the structural safety is seriously affected. Terahertz time-domain spectroscopy is an important means for nondestructive testing of composite materials due to the advantages of non-ionization, good penetrability to nonpolar materials and the like. By analyzing the flight time of terahertz pulse propagating in the material, the measurement of the thickness of the structural layer, the thickness of the defect and the depth can be realized. However, in practical applications, achieving high-precision composite terahertz thickness measurement faces extremely serious challenges. First, composite materials typically have heterogeneous and anisotropic structural characteristics in which terahertz waves experience strong frequency-dependent attenuation and dispersion effects as they propagate. This results in severe stretching, asymmetric distortion and peak shifting of the reflected echo pulses in the time domain, which renders conventional methods based on pulse peak positioning ineffective. Secondly, when detecting a thin layer structure or a multi-layer glued structure, due to the fact that the thickness of each layer is small, the reflected echoes of the front interface and the rear interface are subjected to serious time domain overlapping (namely signal aliasing) due to pulse broadening, and complex composite waveforms are formed, so that independent reflection interfaces cannot be directly distinguished. In addition, the actual industrial detection environment is complex, scattered noise and system noise often submerge weak interlayer reflection signals, and the low signal-to-noise ratio environment further increases the difficulty of high-precision flight time extraction. In order to solve the above signal processing problem, researchers have proposed various signal restoration and parameter estimation methods, which can be mainly classified into deconvolution methods based on physical models, optimization algorithms based on sparse representation, and data-driven deep learning methods. The first is deconvolution methods based on physical models, such as frequency domain deconvolution, wiener filtering (WIENER FILTERING), and so on. The method tries to recover an ideal pulse sequence by eliminating the influence of a system response function, so that the pulse width can be compressed to a certain extent, and the resolution is improved. However, these methods are limited in that they are extremely sensitive to high frequency noise and are prone to noise amplification and Gibbs ringing effects, and at the same time they generally assume that the system is linear and ignore the complex dispersion characteristics of the material, making it difficult to accurately handle waveform distortions caused by frequency dependent attenuation. The second category is optimization algorithms based on sparse representation, which is also the dominant direction of current enhancement of terahertz resolution. The method uses the sparse characteristic of the terahertz interlayer reflection signal in the time domain to model the terahertz interlayer reflection signal as convolution of dictionary atoms and sparse coefficients, and representative algorithms comprise basic Pursuit (basic Pursuit), matching Pursuit (Matching Pursuit), iterative contraction threshold algorithm (ITERATIVE SHRINKAGE-Thresholding Algorithm, ISTA) and the like. These methods have the advantage of having a solid mathematical theory basis that can recover sparse reflective interfaces from undersampled or aliased signals. However, the method has obvious defects that firstly, the calculation efficiency is low, hundreds of iterations are needed for convergence by traditional iterative algorithms such as ISTA and the like, the real-time requirement of industrial online detection is difficult to meet, secondly, the parameter self-adaption capability is poor, regularization parameters (such as sparseness weight) and step length are usually adjusted manually and well, the variable detection environment is difficult to adapt, thirdly, dictionary mismatch is caused, and a preset analysis dictionary (such as Gaussian wavelet) cannot perfectly represent complex dispersion echo in an actual composite material, so that reconstruction accuracy is limi