Method for detecting internal defects of epoxy glass fiber composite material based on terahertz spectrum

Abstract

The application discloses a kind of based on terahertz spectrum's epoxy glass fiber composite internal defect detection method, comprising the following steps: (1) to epoxy glass fiber, to prepare the epoxy glass fiber sample with different aperture or depth with fine carving;(2) each sample is placed into terahertz system and is measured using transmission mode, and the terahertz spectrum data corresponding to sample is collected;(3) the terahertz spectrum data of each sample is corrected by pretreatment method;(4) band selection method is used to the feature extraction of terahertz spectrum data;(5) the defect detection model based on terahertz transmission spectrum is established using chemometrics method;(6) the visualization analysis of defect area is carried out by different threshold segmentation method.The application uses the method that chemometrics and threshold segmentation are combined to quantify the depth and area of epoxy glass fiber internal defect, realizes the quantitative detection of defect.

Description

Technical Field

[0001] This invention relates to the field of chemical testing technology, and in particular to a method for detecting internal defects in epoxy glass fiber composite materials based on terahertz spectroscopy. Background Technology

[0002] Composite materials are novel materials created by combining multiple materials of different components through a composite process. They retain the key characteristics of the original constituent materials while acquiring properties not possessed by the original components through the composite effect. Material design can be used to make the properties of each component complement and interact with each other, thereby achieving better performance. With the improvement of the overall performance of rail vehicles and the rapid development of high-speed rail, composite materials have been widely used in the rail transit field. Components made of composite materials are lightweight, high-strength, and rigid, playing a vital role in reducing vehicle weight, noise and vibration, improving safety and comfort, and reducing maintenance. Therefore, composite materials have become ideal structural components for high-speed rail transit.

[0003] In Japan's Shinkansen high-speed trains, GFRP (glass fiber reinforced plastic) composite materials are used for window interiors, washrooms, toilets, and sinks. In my country, high-speed trains like the "Harmony" series use composite materials for roof covers and battery boxes, while the "Fuxing" series uses carbon fiber composite materials for side panels. However, composite materials are one-piece molded components with minimal secondary processing. Their advanced nature is accompanied by inconsistent quality and high cost. In practical applications, even with well-designed and researched processes, defects can still occur during manufacturing, leading to quality problems and, in severe cases, the scrapping of entire structures, resulting in significant economic losses. Furthermore, after composite structures are applied in real-world applications, they are subject to erosion from long-term exposure to light, heat, humidity, and biological agents during service. They may also be damaged by impacts from flying stones. Additionally, existing defects may expand, creating safety hazards. Therefore, routine non-destructive testing of composite structures in the field is essential to ensure that they do not suffer structural damage that could compromise safety.

[0004] Terahertz (THZ) technology is a novel detection method that offers advantages such as being non-destructive, non-electric, and non-contact. Furthermore, terahertz spectroscopy features high penetration, high bandwidth, and short wavelengths, exhibiting particularly good penetration through many dielectric and non-polar materials. Its spectra contain comprehensive information, revealing a significant portion of the internal structure of matter, and possessing unique properties not found in traditional spectroscopy.

[0005] In recent years, some researchers have begun to use terahertz spectroscopy to detect composite materials. However, current research only performs terahertz imaging on material samples with built-in defects and does not quantify the location and area of ​​internal defects in the material. Summary of the Invention

[0006] In view of the above, the purpose of this invention is to provide a method for detecting internal defects in epoxy glass fiber composite materials based on terahertz spectroscopy. This method uses a combination of chemometrics and threshold segmentation to quantify the depth and area of ​​internal defects in epoxy glass fibers, thereby achieving quantitative detection of defects.

[0007] A method for detecting internal defects in epoxy glass fiber composites based on terahertz spectroscopy includes the following steps:

[0008] (1) The epoxy glass fiber is finely sculpted to prepare epoxy glass fiber samples with different pore sizes or depths.

[0009] (2) Place each sample into the terahertz system and measure it using the transmission mode to collect the terahertz spectral data of the corresponding sample.

[0010] (3) The terahertz spectral data of each sample were corrected by preprocessing methods;

[0011] (4) Feature extraction of terahertz spectral data is performed using a band screening method;

[0012] (5) A defect detection model based on terahertz transmission spectroscopy was established using chemometric methods;

[0013] (6) Visual analysis of defect area using different threshold segmentation methods.

[0014] The above-mentioned method for detecting internal defects in epoxy glass fiber composite materials based on terahertz spectroscopy, wherein step (1) specifically includes:

[0015] Epoxy glass fiber was precision-carved using CNC machine tools to simulate pore defects. The first sample was left unprocessed, representing a defect-free sample. The second sample had nine pores machined, each with a diameter of 10 mm and a depth increasing from 0.3 mm to 2.7 mm, representing pore defects of the same diameter but different depths. The third sample had nine pores machined, each with a depth of 2.5 mm and a diameter increasing from 3 mm to 11 mm, representing pore defects of the same depth but different diameters.

[0016] The above-mentioned method for detecting internal defects in epoxy glass fiber composite materials based on terahertz spectroscopy, wherein step (2) specifically includes:

[0017] During terahertz time-domain spectroscopy measurements, the laboratory ambient temperature was controlled at 23±0.5℃ and the humidity was controlled below 10%. After the experimental equipment was preheated for half an hour, it was used to collect terahertz transmission images. The sample was placed on a moving platform, and then the height of the platform was adjusted to obtain the optimal terahertz signal. A metal was used as the reflective layer, and the image was scanned in a point-to-point mode. The sample was translated on the XY two-dimensional translation stage with the platform in steps of 0.3mm. The terahertz imaging camera was controlled by software settings to acquire terahertz images of the sample. The images were scanned point by point to complete the acquisition of sample image information.

[0018] The above-mentioned method for detecting internal defects in epoxy glass fiber composite materials based on terahertz spectroscopy, wherein step (3) includes:

[0019] The terahertz spectral data of each sample are corrected using any one of the following methods: asymmetric least squares correction method, adaptive iterative reweighted penalized least squares correction method, sparsity-based baseline estimation and denoising correction method, and standard normal transformation.

[0020] The above-mentioned method for detecting internal defects in epoxy glass fiber composite materials based on terahertz spectroscopy, wherein step (4) includes:

[0021] Feature extraction of terahertz spectral data is performed using any one of the following methods: competitive adaptive weight sampling algorithm, continuous projection method, or non-information variable elimination method.

[0022] The above-mentioned method for detecting internal defects in epoxy glass fiber composite materials based on terahertz spectrum, wherein the model established in step (5) is a partial least squares model or a least squares support vector machine model.

[0023] The above-mentioned method for detecting internal defects in epoxy glass fiber composite materials based on terahertz spectroscopy, wherein step (6) specifically includes:

[0024] The terahertz RGB image of the sample is converted into a corresponding grayscale image, and then binarized. The binarization process includes manual thresholding, Otsu's method, grayscale histogram bimodal method, and double thresholding. Finally, the connected component labeling method is used to calculate the number of pixels of the hole defects of different apertures. The defect area is calculated by the ratio of the number of holes defect pixels to the number of pixels in the entire sample image, thus completing the detection of the hole defect area.

[0025] The above-mentioned method for detecting internal defects in epoxy glass fiber composite materials based on terahertz spectroscopy, wherein the formula for calculating the defect area in step (6) is as follows:

[0026]

[0027] Where S represents the area of ​​the hole defect, S zThe area of ​​the entire sample is represented by n. i N represents the number of pixels in a single hole defect sample, and N represents the total number of pixels in the entire sample.

[0028] The present invention provides a method for detecting internal defects in epoxy glass fiber composites based on terahertz spectroscopy. First, the terahertz spectral characteristics of pores at different depths in epoxy glass fiber composites are analyzed. Second, different preprocessing methods are applied to the experimentally collected data. A PLS fully interactive model is established and evaluated for the terahertz spectra of pores at different depths with the same pore size, and the optimal preprocessing method is selected. Band selection is performed using the preprocessed data, and the optimal band selection is used to establish PLS and LS-SVM quantitative detection models respectively. Finally, different threshold segmentation methods are used to process the terahertz images of defect samples with different pore sizes at the same depth. The area of ​​defects with different pore sizes is calculated using pixels, and the established quantitative defect depth detection model is combined to complete the detection of pore defects in the samples. This invention can detect the location, depth, and area information of internal pore defects, avoiding accidents caused by internal material defects. It provides a technical basis for defect detection in the composite material production process, offers new methods and means for defect detection and quality supervision of other types of composite materials, and provides a certain guarantee for the safe operation of transportation facilities. Attached Figure Description

[0029] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0030] Figure 1 Images of epoxy glass fiber samples: (a) No internal defects; (b) Internal pore defects of the same diameter but different depths; (c) Internal pore defects of the same depth but different diameters.

[0031] Figure 2 A schematic diagram of selecting the region of interest in a terahertz image;

[0032] Figure 3 (A) Time-domain spectrum and (B) Frequency-domain spectrum of pore defects at different depths in epoxy glass fiber samples;

[0033] Figure 4 These are images showing the effects of different preprocessing steps on time-domain spectral data.

[0034] Figure 5 The image shows the feature extraction results of the THz spectrum of the epoxy glass fiber pore defect sample using the CARS algorithm.

[0035] Figure 6 The image shows the results of feature extraction from the THz spectrum of an epoxy glass fiber porous defect sample using the UVE algorithm.

[0036] Figure 7 This is a schematic diagram of the selected variables in the original spectrum after wavelength variable selection using the SPA algorithm, where (a) shows the trend of RMSECV as the number of wavelength variables selected by the SPA algorithm changes; and (b) shows the feature extraction results of the spectrum using the SPA algorithm.

[0037] Figure 8 A comparison of prediction results for two different quantitative models, PLS and LS-SVM, for pore defects in epoxy glass fiber.

[0038] Figure 9 The images show terahertz imaging results of epoxy glass fiber samples with pore defects. In (a), the pore defects are of the same diameter but different depths; and in (b), the pore defects are of the same depth but different diameters.

[0039] Figure 10 Image comparison of sample two: (a) actual object image, (b) terahertz imaging image, (c) manual thresholding segmentation, (d) Otsu, (e) grayscale histogram bimodal method, (f) double thresholding segmentation;

[0040] Figure 11 This is a comparison chart showing the detected defect area and actual area of ​​the hole defect in Sample 2 using different binarization processing methods. Detailed Implementation

[0041] To make the objectives, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Several embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that the disclosure of the present invention will be thorough and complete.

[0042] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0043] This invention provides a method for detecting internal defects in epoxy glass fiber composite materials based on terahertz spectroscopy, comprising the following steps (1) to (6):

[0044] (1) The epoxy glass fiber is precision sculpted to prepare epoxy glass fiber samples with different pore sizes or depths.

[0045] The sample used in this embodiment is an epoxy glass fiber reinforced composite material, which has a temperature resistance of 200 degrees Celsius, a high voltage resistance of approximately 45 kV, a flexural strength of 480–510 MPa, and an insulation resistance of 10¹⁴ Ω after immersion in water. This material is produced using a lamination molding process and is composed of glass fiber and epoxy resin. The glass fiber content is approximately 30%, and it consists of unidirectional glass fiber cloth arranged in a 0 / 90 degree cross-layout direction. The sample dimensions are 45 mm × 45 mm × 3 mm.

[0046] like Figure 1 To simulate pore defects, CNC machine tools were used to precisely carve epoxy fiberglass samples. The first sample was left unprocessed, representing a defect-free internal sample. Figure 1 As shown in (a), the second sample was machined with nine holes, each 10 mm in diameter, with depths increasing sequentially from 0.3 mm to 2.7 mm (0.3 mm, 0.6 mm, 0.9 mm, 1.2 mm, 1.5 mm, 1.8 mm, 2.1 mm, 2.4 mm, 2.7 mm). These represent internal defects in the sample where the holes have the same diameter but different depths. Figure 1 As shown in (b), the third sample was machined with nine holes of 2.5 mm depth and diameters increasing sequentially from 3 mm to 11 mm (3 mm, 4 mm, 5 mm, 6 mm, 7 mm, 8 mm, 9 mm, 10 mm, 11 mm). These represent internal pore defects of the same depth but different diameters. Figure 1 As shown in (c). There are certain errors in the actual sample processing; subsequent data processing should be based on the actual dimensions. Figure 1 The aim is to reveal internal defects in the sample by performing terahertz imaging on another intact surface of the sample during the actual experiment.

[0047] The epoxy glass fiber samples with internal pore defects were divided into two groups, with nine pore defects in each group. The pores in the first group were numbered: 1-1, 1-2, 1-3, 1-4, 1-5, 1-6, 1-7, 1-8, 1-9, and so on for the second group. Terahertz transmission imaging was used to image the back side of the glass fiber samples. Figure 1 (b) Taking this as an example, when placing the samples, the hole depths are arranged in an orderly manner. The first row, from left to right, is 1-1, 1-2, 1-3; the second row is 1-4, 1-5, 1-6; and the third row is 1-7, 1-8, 1-9. The depth and diameter of each hole are measured 5 times using vernier calipers, and the average value is calculated. The hole defect numbers and specifications for the two different types of defect samples are shown in Table 1. The sample bags are already labeled with serial numbers and hole depths and diameters. The defects of sample one are arranged according to... Figure 1 (b) Arrange samples from lightest to darkest, according to sample two. Figure 1 (c) Arrange the samples in the correct order and position. Sample placement method as follows: Figure 1 As shown.

[0048] Table 1. Number and specifications of pore defects in two different types of samples.

[0049]

[0050] (2) Place each sample into the terahertz system and measure it using the transmission mode to collect the terahertz spectral data of the corresponding sample.

[0051] The terahertz system used in this embodiment is the QT-TO1000 instrument provided by Qingdao Fengda Terahertz Technology Co., Ltd., with a frequency measurement range of 0.1-4.0 THz, a maximum scanning area of ​​100×100 mm, and an imaging speed of 60 pixels / s. The femtosecond pulses generated by the femtosecond laser are split into pump and probe beams by a beam splitter. The pump beam enters the terahertz emitter, generates a terahertz pulse, and after being focused, passes through the sample. The terahertz pulse carrying sample information and the probe beam are collinear and reach the detector crystal, transmitting the signal to the computer.

[0052] During the experiment, the laboratory ambient temperature was controlled at 23±0.5℃, and the humidity was controlled below 10%. The experimental equipment was preheated for half an hour before being used to collect terahertz transmission images. The sample was placed on a moving platform, and the platform height was adjusted to obtain the optimal terahertz signal. A metal was used as the reflective layer, and images were scanned in a point-to-point mode. The sample was translated on an XY two-dimensional translation stage with the platform in 0.3mm increments. The terahertz imaging camera was controlled by software settings to acquire terahertz images of the sample, scanning point-to-point to complete the information acquisition of the sample image.

[0053] Terahertz spectroscopy contains comprehensive spectral information, such as absorption coefficient, refractive index, and absorbance. Therefore, terahertz spectroscopy can reveal a great deal of information about the interior of a substance. The original terahertz spectral data of the sample is based on the time-domain signal. The frequency spectrum distribution of the THz pulse can be obtained by using Fast Fourier Transform (FFT), which can be expressed as formula (1):

[0054]

[0055] Where A(ω) represents the amplitude of the pulse signal, Let E(t) be the phase of the signal, E(ω) be the time-domain signal, and E(ω) be the frequency-domain signal.

[0056] In this embodiment, each pixel of the acquired terahertz image contains a complete terahertz waveform, thus also possessing spectral measurement capabilities, truly achieving image-spectrum integration. The original sample data is a three-dimensional array; each pixel in the XOY plane contains one spectral line, with the Z-axis representing wavelength, a time domain range of 0-90 ps, ​​corresponding to 2250 band points. Within the defect-free portion of the sample, the information contained in each spectrum is essentially consistent; therefore, it is necessary to extract the region of interest, i.e., the region containing defect information. For example... Figure 2 For the selection of regions of interest in terahertz imaging, this paper extracts and studies terahertz time-domain spectral data in the range of 20-36 ps.

[0057] The time-domain and frequency-domain spectra of epoxy glass fiber samples with pores at different depths were analyzed. To avoid the influence of noise and redundant information on subsequent experimental data, terahertz spectra of the region of interest were manually extracted, focusing only on the terahertz time-domain spectra in the 20-36 ps range and the terahertz frequency-domain spectra in the 0.1-1.4 THz range. For pores at different depths, 121 spectral data points were collected for each pore, totaling 1089 spectral data points for nine pores at different depths. 400 terahertz time-domain spectral variables were included in the 20-36 ps range, and 118 terahertz frequency-domain spectral variables were included in the 0.1-1.4 THz range.

[0058] from Figure 3 In (A), a relatively uniform stepped peak distribution can be observed. As the depth of the epoxy glass fiber pore defects increases (i.e., the thicker the glass fiber through which the terahertz wave penetrates), the peak value of the time-domain spectrum of the pore gradually decreases, while the time of the peak position increases. From... Figure 3 As can be seen in (B), when the pores of epoxy glass fiber are shallow, the terahertz waves of epoxy glass fiber are distinct. As the pore depth increases, the number of terahertz waves that can penetrate the sample gradually decreases, resulting in a decrease in the intensity of the terahertz waves.

[0059] (3) The terahertz spectral data of each sample were corrected by preprocessing method.

[0060] Among them, any one of the following methods can be used to correct the terahertz spectral data of each sample: Asymmetric Least Squares Correction (ALS), Adaptive Iterative Reweighted Penalized Least Squares Correction (AirPLS), Sparsity-Based Baseline Estimation and Denoising Correction (BEADS), and Standard Normal Transform (SNV).

[0061] Asymmetric least squares can be applied to baseline estimation. With asymmetric least squares, the least squares penalty function denoises and balances the baseline trend without prior information about the waveform or baseline; parametric time distortion is its only relevant factor. Slowly changing baseline estimates can be obtained using Whittaker smoothing. Effective baseline estimation is achieved based on asymmetric weighted smoothing of trend bias, which estimates the baseline x based on Whittaker smoothing, as shown in Equation (2):

[0062]

[0063] In the formula, x represents spectral data, Δ is the difference operator, and the weight w i The choice is asymmetric: when y i >z i At that time, w i =p, otherwise, w i =1-p; μ is the regularization parameter.

[0064] Adaptive Iterative Reweighted Penalized Least Squares: This method is based on penalized least squares. During the iterative filtering process, it adaptively adjusts the weights of the sum of squares of the baseline residual and the original signal, which can quickly and flexibly detect and remove irregular changes in the baseline.

[0065] Sparsity-based baseline estimation and denoising is based on modeling a series of chromatographic peaks with sparse derivatives and converting the baseline into a low-pass signal. This method constructs a convex optimization problem to generalize these nonparametric models.

[0066] The Standard Normal Transform (SNV) is mainly used to eliminate the influence of solid particle size, surface scattering, and optical path variation on the spectrum. The algorithm processes a spectrum (based on the rows of the spectral array), and formula (3) is the calculation formula for the SNV transformed spectrum.

[0067]

[0068] in, m is the number of wavelength points, and k = 1, 2, ..., m.

[0069] To minimize the impact of instrument vibration, temperature variations, sample inhomogeneity, and noise during the experiment, four different preprocessing methods—AirPLS, ALS, BEADS, and SNV—were used to correct the data before establishing the quantitative model. A fully interactive PLS model was established for each method, and the optimal preprocessing method was selected. The correlation coefficient R0 was verified through full interaction. cv We used root mean square error (RMSECV) to explore the optimal spectral preprocessing method. Figure 4 The images show the effects of different preprocessing steps on time-domain spectral data.

[0070] After correcting the acquired terahertz time-domain spectral information, a partial least squares (PLS) model was established, and the preprocessing effect was evaluated. Table 2 shows the PLS models established for the terahertz spectrum of Sample 1 after different preprocessing steps. The study found that ALS preprocessing yielded the best results, and the PLS model established using ALS performed best, with a high correlation coefficient R0. cv The value is 0.9926, and the root mean square error (RMSECV) is 0.0920.

[0071] Table 2. Correction and PLS modeling results of THz spectra of epoxy glass fiber samples with porosity defects.

[0072]

[0073] (4) Use the band screening method to extract features from terahertz spectral data.

[0074] The band selection methods used in this embodiment include Competitive Adaptive Weighted Sampling (CARS), Continuous Projection Method (SPA), and Uninformative Variable Elimination (UVE). CARS is a wavelength selection method based on regression coefficients. This method mimics the "survival of the fittest" principle in Darwinian evolution, treating each wavelength as an individual and gradually eliminating them to effectively select the optimal wavelength combination. SPA is a forward cyclic selection method that uses vector projection analysis to select effective wavelengths with minimal redundancy and collinearity. UVE is a wavelength selection method based on PLS regression coefficients. This method uses regression coefficients as a measure of wavelength importance, integrating noise and concentration information when selecting wavelengths, making it more intuitive and practical, and capable of selecting effective spectral information.

[0075] Terahertz spectrometers have a wide spectral range and collect a large number of spectral variables, which also contain a significant amount of redundancy, collinearity, and background information. This useless information can affect the accuracy of modeling. Therefore, it is necessary to filter effective signals from the large amount of spectral data for subsequent analysis. The Competitive Adaptive Weighted Algorithm (CARS), Uninformative Variable Elimination (UVE), and Continuous Projection Method (SPA) are employed for feature variable extraction.

[0076] Wavelength variable selection based on CARS algorithm

[0077] like Figure 5 This is the result of screening terahertz spectra of epoxy glass fiber samples with porosity defects using the CARS algorithm. From... Figure 5As shown in (a), the number of wavelengths is negatively correlated with the number of sampling times; the more sampling times, the fewer wavelengths there are. When the number of sampling times is between 0 and 10, the rate of decrease in the number of wavelengths gradually increases, and the number of wavelengths drops sharply. When the number of sampling times is between 10 and 13, the rate of decrease in the number of wavelengths gradually decreases, and the number of wavelengths decreases slowly. When the number of sampling times is greater than 13, the number of wavelengths essentially stops changing. This phenomenon is caused by the CARS algorithm's process of coarsely selecting wavelength variables and then finely selecting them. Figure 5 (b) shows the change in RMSECV values ​​during the screening process. The graph shows that when RMSECV reaches its minimum value of 0.0866, the number of samples is 13. When the number of samples is between 0 and 13, the RMSECV value decreases with increasing sample count, showing a negative correlation. When the number of samples is greater than 13, the RMSECV value increases with increasing sample count. This phenomenon indicates that before 13 samples, the CARS algorithm effectively removes irrelevant variables. However, after 13 samples, the CARS algorithm not only fails to remove irrelevant variables but also removes relevant variables. Therefore, setting the number of samples is crucial when using the CARS algorithm. Figure 5 (c) This graph shows the changes in the regression coefficients of the wavelength variable during the screening process. In the graph, "*" indicates the number of samples corresponding to the minimum value of RMSECV, which is 13 samples. Figure 5 As shown in (a), the number of wavelengths at this time is 109.

[0078] Wavelength variable screening based on UVE algorithm

[0079] like Figure 6 The figure shows the results of terahertz spectrum screening of epoxy glass fiber porous defect samples using the UVE algorithm. The two horizontal lines in the figure represent the threshold lines for UVE wavelength screening, with upper and lower thresholds of 62.3591 and -62.3591, respectively. The green vertical line is the boundary between stable and unstable wavelength distributions; the wavelengths to the left of the boundary are stable, and those to the right are unstable. Based on the screening results, spectral variables exceeding the upper and lower thresholds are used as input variables for the model, while other spectral variables are discarded. After UVE screening, the number of spectral variables was reduced from 400 to 50, eliminating 87.5% of the data, significantly reducing model complexity and improving model performance.

[0080] Wavelength variable selection based on SPA algorithm

[0081] The continuous projection algorithm selects the terahertz spectral intensity variable as input based on the orthogonal projection information of the terahertz wavelength variable. The number of variables selected for SPA is set to 10 and 35 respectively. The SPA band selection results are as follows: Figure 7The process is as shown. After the terahertz spectral intensity was run, 20 wavelength variables were selected as input vectors. After PLS modeling, the root mean square error was 0.1432. Figure 7 This is a schematic diagram of the selected variables in the original spectrum after wavelength variable selection using the SPA algorithm.

[0082] Table 3 shows the results of the terahertz spectral quantitative detection model for epoxy glass fiber porous defect samples based on band selection and PLS. Three band selection methods (CARS, UVE, and SPA) were compared, and the established PLS quantitative model was evaluated based on the results of cross-validation. Firstly, among the UVE, CARS, and SPA methods, CARS screening showed the best performance, with a high correlation coefficient R0 in the cross-validation. cv The correlation coefficient was 0.9936, and the root mean square error (RMECV) was 0.0854. Compared to the original data, CARS performed better, with an increased correlation coefficient and a decreased RME. UVE and SPA performed worse, with lower correlation coefficients and RME. cv The error decreased, and the root mean square error increased. The possible reason is that too few variable data were selected, and the PLS algorithm could not build an effective model based on a small amount of variable information and true values.

[0083] Table 3. Results of the THz spectral quantitative detection model for epoxy glass fiber porous defects based on band selection and fully interactive PLS.

[0084]

[0085] (5) A defect detection model based on terahertz transmission spectroscopy was established using chemometric methods.

[0086] The method used in this embodiment to establish a quantitative model of the depth of pore defects in epoxy glass fiber includes partial least squares (PLS) and least squares support vector regression (LS-SVM).

[0087] Partial Least Squares (PLS) is an effective statistical method for multivariate factor regression. It can effectively analyze spectral information and is widely used to build linear models with a large number of highly collinear variables, allowing variables to be correlated with a small sample size. Its model can be expressed as:

[0088] Y = bX + e (4)

[0089] In the formula: b represents the regression coefficient; e represents the residual matrix of the model.

[0090] Least Squares Support Vector Regression (LS-SVM) is a machine learning method developed based on statistical learning theory. Its key parameters are the input vector, kernel function type, and corresponding parameters. The objective optimization function of the LS-SVM algorithm is shown in Equation (5), and the function constraints are shown in Equation (6).

[0091]

[0092]

[0093] In the formula, w is the weight vector; γ is the regularization parameter; e i For error; x i and y i , i = 1, ..., n, where n is the number of samples in the calibration set.

[0094] In the formula These are called kernel functions, and there are two standard types: linear (Lin) kernel functions and radial basis function (RBF) kernel functions. Equation (7) is the formula for linear kernel functions, and Equation (8) is the formula for nonlinear kernel functions.

[0095] K(x, x) k )=x T x k +1 (7)

[0096] K(x, x) k ) = exp(-||xx k || 2 / 2σ 2 (8)

[0097] Where x represents a sample point, x k Let σ be the center point of the kernel function, γ be the distribution parameter, and σ be the center point of the kernel function. 2 These are kernel parameters, representing the variance of the radial basis functions.

[0098] This embodiment first preprocesses the spectral data of epoxy glass fiber samples with pores at different depths. A PLS model is then established through full interactive validation to evaluate the preprocessing method and select the optimal one. Using the preprocessed spectral data, both the PLS model and the LS-SVM model are established to quantitatively analyze the depth of the pores. The correlation coefficient R between the modeling set and the prediction set is compared. c R pThe root mean square error (RMSEC) and root mean square error (RMSEP) are used to evaluate the model. A higher correlation coefficient and a smaller RMSEC indicate higher model accuracy, and closer RMSEC and RMSEP values ​​indicate better model robustness. Finally, for terahertz images of pore defects with different pore sizes, image processing techniques are used to determine the area of ​​the pore defects. The detected values ​​and actual values ​​of pore defects with different diameters are compared, and a superior quantitative model for pore defect depth is used to assess the ability of terahertz spectroscopy to detect pore defects in epoxy glass fibers.

[0099] Correlation coefficient R of evaluation indicators c R p This can be expressed as formula (11):

[0100]

[0101] The root mean square error (RMSEC) and RMSEP can be expressed as formula (12):

[0102]

[0103] In formulas (11) and (12), y i，actual y represents the actual defect depth of the i-th hole defect sample in the modeling or prediction set. i，predicted This represents the predicted defect depth of the i-th hole defect sample in the modeling or prediction set. is the average actual defect depth of all hole defect samples in the model set or prediction set, where n is the number of samples in the model set or prediction set.

[0104] Establishment of PLS ​​quantitative model for terahertz time-domain spectroscopy of defective samples

[0105] Partial least squares (PLS) is a multivariate correction method based on linear regression. It can simultaneously decompose the true values ​​and the variable matrix, incorporating information from the true values ​​into the variable matrix decomposition process. Furthermore, it maximizes the correlation between principal components and concentration, making full use of the relationship between spectral variables and concentration. The modeling involved 1089 data samples, divided into a 3:1 ratio for the modeling set and the prediction set, with 817 samples in the modeling set and 272 samples in the prediction set.

[0106] Table 4 compares the modeling effects of different wavelength variable selection methods combined with the PLS model. As can be seen from Table 4, modeling using terahertz spectra of epoxy glass fiber porous defect samples with different numbers of variables corresponds to the full-spectrum PLS, CARS-PLS, UVE-PLS, and SPA-PLS models, respectively.

[0107] Table 4. Results of the THz spectral quantitative detection model for epoxy glass fiber porous defect samples based on band selection and PLS.

[0108]

[0109] By comparing the R values ​​of the modeling set and the prediction set c R p RMSEC and RMSEP were used to evaluate the effectiveness of PLS ​​quantitative models established using different wavelength variable selection methods. A comprehensive evaluation of the correlation coefficient and root mean square error of the modeling and prediction sets showed that CARS modeling performed best, selecting 109 variables. The R-value of the modeling set was [not specified]. c The R and RMSEC values ​​are 0.9852 and 0.1257 respectively, and the R of the prediction set is... p The RMSEP values ​​were 0.9873 and 0.1236, respectively, showing improvements in evaluation metrics for both the prediction set and the modeling set compared to full-band data modeling. Although the RMSEP of the UVE algorithm on the modeling set was... c RMSEC outperforms CARS, but its prediction set performance is slightly worse than CARS, and its model performance is less stable. While SPA's modeling performance is worse than the original data, it selects the fewest variables. In summary, for PLS quantitative models, CARS and UVE perform better, SPA performs worse, and CARS band selection improves model accuracy and stability while reducing data volume.

[0110] Establishment of LS-SVM quantitative model for terahertz time-domain spectroscopy of defective samples

[0111] Least Squares Support Vector Machine (LS-SVM) uses a least squares linear system as the loss function, which reduces computational complexity. Its key parameters are the input vector, the type of kernel function, and its corresponding parameters. Two typical kernel functions are the Radial Basis Function (RBF) kernel and the Linear Lin kernel. In this modeling, 1089 samples were divided into a modeling set and a prediction set at a 3:1 ratio, with 817 samples in the modeling set and 271 in the prediction set. Wavelength variables selected using UVE, CARS, and SPA variable selection methods were used as inputs to the LS-SVM model. Table 5 compares the prediction performance of LS-SVM models built using different wavelength variable selection methods.

[0112] Table 5 Results of the THz spectral quantitative detection model for pore defects in epoxy glass fiber based on band selection and LS-SVM

[0113]

[0114] Table 5 shows that the LS-SVM quantitative detection models established using the terahertz spectra of epoxy glass fiber porous defect samples correspond to the original spectral LS-SVM, CARS-LS-SVM, UVE-LS-SVM, and SPA-LS-SVM models, respectively. Each model uses two kernel functions, RBF and Lin, for modeling. The R-values ​​of the prediction set are then analyzed. p The effectiveness of different wavelength variable methods in screening data to establish PLS quantitative models was evaluated using RMSECP and RADP. Among these, SPA showed the best modeling performance when using radial basis function (RBF) as the kernel function. p The R value is 0.9998, and the RMSECP value is 0.0152; when using linear Lin as the kernel function, the modeling effect of the original spectrum is the best, and R... p The correlation coefficient (RCC) is 0.9937, and the RMSECP is 0.0827. Comparing the radial basis function (RBF) kernel and the linear Lin kernel, the RBF kernel performs better than the linear Lin kernel. Comparing the PLS quantitative model and the LS-SVM model, the LS-SVM model performs better than the PLS model. In summary, the LS-SVM models established by several band selection methods all show good modeling results, reducing the amount of data computation and improving the model accuracy. In particular, the SPA-LS-SVM model selects only 20 wavelength variables, and the correlation coefficient reaches 0.9998, with the root mean square error reduced to 0.0152.

[0115] Quantitative model analysis of defective samples using PLS and LS-SVM

[0116] like Figure 8 This figure compares the prediction results of two different quantitative models, PLS and LS-SVM, for epoxy glass fiber porosity defects. Using the wavelength variable selection method reduces the computational cost of model building. Using the SPA algorithm combined with LS-SVM significantly reduces the computational cost and improves model accuracy. As shown in the figure, the LS-SVM model is more accurate and has a greater advantage over the PLS model. The R-value of the SPA-LS-SVM model is... p The value is 0.9998, and the RMSEP is 0.0152.

[0117] (6) Visual analysis of defect area using different threshold segmentation methods.

[0118] Image binarization transforms a digital image into a matrix of binary black and white pixels, separating local features from the background. Typically, an appropriate threshold is selected to obtain an image that reflects both the overall and local features of the image. Gray values ​​above the threshold are set to white, and gray values ​​below the threshold are set to black, resulting in a binarized image. The binarization methods used in this embodiment include manual thresholding, Otsu's method, the bimodal gray-level histogram method, and dual-thresholding segmentation.

[0119] Manual thresholding segmentation involves manually selecting a reasonable threshold through visual observation. The Otsu method, also known as the maximum inter-class variance method, is unaffected by image contrast and brightness under certain conditions. Its basic idea is to use the image's gray-level histogram to find the optimal segmentation threshold when the inter-class variance of the target is maximized. The bimodal gray-level histogram method first obtains the image's average gray-level histogram, which has two distinct peaks. A minimum frequency value exists between these two peaks, and the corresponding gray-level value is the optimal threshold obtained by the bimodal method. Dual-threshold segmentation sets two thresholds, manually selecting the two most suitable thresholds to avoid errors caused by single-threshold segmentation.

[0120] Region labeling is a process that assigns the same label to connected pixels and different labels to different connected components. This paper converts the original RGB image of epoxy glass fiber pore defects into a corresponding grayscale image, performs binarization, and finally calculates the number of pixels in the defect area using region labeling.

[0121] Analysis of Terahertz Imaging Results of Defective Samples

[0122] like Figure 9 The image shows the results of the test using terahertz transmission imaging. Figure 9 (a) is a terahertz transmission image of sample one. From Figure 9 As shown in (a), the color of the 2.66 mm deep hole defect is very similar to that of the 3 mm thick original sample. The color of the 0.3 mm deep hole is most significantly different from the other holes. The terahertz intensity decreases with increasing depth of the hole defect. This is because the terahertz band that can pass through the sample gradually decreases with increasing sample thickness, resulting in poor imaging of deeper hole defects. Although the terahertz imaging effect of deeper hole defects is not obvious, their outlines can still be clearly shown. Figure 9 (b) is a terahertz transmission image of the pore defects in sample two. From Figure 9 As shown in (b), defects with different pore sizes are clearly presented. This indicates that terahertz transmission imaging technology has great potential in the visualization of detected pore defects in epoxy glass fibers.

[0123] Detection of the area of ​​pores in defective samples

[0124] Terahertz imaging can clearly identify the location of pores and defects inside epoxy glass fiber samples. Combined with the established quantitative detection model for pores and defects, the depth of the pores and defects can be determined, but the area of ​​the defects cannot be detected. Therefore, image processing technology is used to detect the area of ​​the defects.

[0125] First, the terahertz RGB image of sample two is converted into a corresponding grayscale image. Then, binarization processing is performed, including manual thresholding (threshold set to 38), Otsu's algorithm, grayscale histogram bimodal method, and dual thresholding (thresholds set to 140 and 250). Finally, the connected component labeling method is used to calculate the number of pixels for each pore defect of different diameters. The defect area is calculated by the ratio of the number of pore defect pixels to the total number of pixels in the entire sample image, thus completing the detection of the pore defect area. Figure 10 The terahertz imaging of sample two is processed using different binarization methods. Since this embodiment uses back-side imaging of the defect, the defect location in the imaging image is reversed compared to the actual object image.

[0126] Because terahertz nondestructive testing methods often exhibit diffraction phenomena, which can even lead to blurred boundaries, the binarization of the terahertz image of sample two has a certain impact and causes some error in the calculation of the defect area. There were minor errors in the sample processing; the actual sample dimensions are 44.81 × 44.84 mm, with an area of ​​2009.2804 mm². 2 When performing region labeling, the total number of pixels in the entire sample image is 145×145, totaling 21025. Table 6 shows the detection area and actual area results of different binarization processing methods for the two-hole defects in the sample.

[0127] Table 6 shows the detected defect area and actual area of ​​the sample's two-hole defect binarization processing method.

[0128]

[0129] Figure 11 This is a comparison chart showing the detected defect area and actual area of ​​the pore defects in Sample 2 using different binarization methods. Figure 11 As can be seen, the optimal binarization method is the gray-level histogram bimodal method. The correlation coefficient between the detected defect area and the actual defect area is 0.9997, and the root mean square error is 2.2849, indicating that the area of ​​internal pore defects in epoxy glass fiber can be detected by combining terahertz imaging with image processing technology.

[0130] In summary, the terahertz-based method for detecting internal defects in epoxy glass fiber composites provided by this invention first analyzes the terahertz spectral characteristics of pores at different depths in epoxy glass fiber composites. Secondly, different preprocessing methods are applied to the experimentally collected data. A PLS fully interactive model is established and evaluated for the terahertz spectra of pores at different depths with the same pore size, and the optimal preprocessing method is selected. Band selection is performed using the preprocessed data, and the optimal band selection is used to establish PLS and LS-SVM quantitative detection models respectively. Finally, different threshold segmentation methods are used to process the terahertz images of defect samples with different pore sizes at the same depth. The area of ​​defects with different pore sizes is calculated using pixels, and the established quantitative defect depth detection model is combined to complete the detection of pore defects in the samples. This invention can detect the location, depth, and area information of internal pore defects, avoiding accidents caused by internal material defects. It provides a technical basis for defect detection in the composite material production process, offers new methods and means for defect detection and quality supervision of other types of composite materials, and provides a certain guarantee for the safe operation of transportation facilities.

[0131] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for detecting internal defects in epoxy glass fiber composite materials based on terahertz spectroscopy, characterized in that, Includes the following steps: (1) The epoxy glass fiber is finely sculpted to prepare epoxy glass fiber samples with different pore sizes or depths. (2) Place each sample into the terahertz system and measure it using the transmission mode to collect the terahertz spectral data of the corresponding sample. (3) The terahertz spectral data of each sample were corrected by preprocessing methods; (4) Feature extraction of terahertz spectral data is performed using a band screening method; (5) A defect detection model based on terahertz transmission spectroscopy was established using chemometric methods; (6) Visual analysis of defect area using different threshold segmentation methods.

2. The method for detecting internal defects in epoxy glass fiber composite materials based on terahertz spectroscopy according to claim 1, characterized in that, Step (1) specifically includes: Epoxy glass fiber was precision-carved using CNC machine tools to simulate pore defects. The first sample was left unprocessed, representing a defect-free sample. The second sample had nine pores machined, each with a diameter of 10mm and a depth increasing from 0.3mm to 2.7mm, representing pore defects of the same diameter but different depths. The third sample had nine pores machined, each with a depth of 2.5mm and a diameter increasing from 3mm to 11mm, representing pore defects of the same depth but different diameters.

3. The method for detecting internal defects in epoxy glass fiber composite materials based on terahertz spectroscopy according to claim 2, characterized in that, Step (2) specifically includes: During terahertz time-domain spectroscopy measurements, the laboratory ambient temperature was controlled at 23±0.5℃ and the humidity was controlled below 10%. After the experimental equipment was preheated for half an hour, it was used to collect terahertz transmission images. The sample was placed on a moving platform, and then the height of the platform was adjusted to obtain the optimal terahertz signal. A metal was used as the reflective layer, and the image was scanned in a point-to-point mode. The sample was translated on the XY two-dimensional translation stage with the platform in steps of 0.3mm. The terahertz imaging camera was controlled by software settings to acquire terahertz images of the sample. The images were scanned point by point to complete the acquisition of sample image information.

4. The method for detecting internal defects in epoxy glass fiber composite materials based on terahertz spectroscopy according to claim 1, characterized in that, Step (3) includes: The terahertz spectral data of each sample are corrected using any one of the following methods: asymmetric least squares correction method, adaptive iterative reweighted penalized least squares correction method, sparsity-based baseline estimation and denoising correction method, and standard normal transformation.

5. The method for detecting internal defects in epoxy glass fiber composite materials based on terahertz spectroscopy according to claim 1, characterized in that, Step (4) includes: Feature extraction of terahertz spectral data is performed using any one of the following methods: competitive adaptive weight sampling algorithm, continuous projection method, or non-information variable elimination method.

6. The method for detecting internal defects in epoxy glass fiber composite materials based on terahertz spectroscopy according to claim 1, characterized in that, The model established in step (5) is a partial least squares model or a least squares support vector machine model.

7. The method for detecting internal defects in epoxy glass fiber composite materials based on terahertz spectroscopy according to claim 1, characterized in that, Step (6) specifically includes: The terahertz RGB image of the sample is converted into a corresponding grayscale image, and then binarized. The binarization process includes manual thresholding, Otsu's method, grayscale histogram bimodal method, and double thresholding. Finally, the connected component labeling method is used to calculate the number of pixels of the hole defects of different apertures. The defect area is calculated by the ratio of the number of holes defect pixels to the number of pixels in the entire sample image, thus completing the detection of the hole defect area.

8. The method for detecting internal defects in epoxy glass fiber composite materials based on terahertz spectroscopy according to claim 1, characterized in that, In step (6), the formula for calculating the defect area is as follows: Where S represents the area of ​​the hole defect, S z The area of ​​the entire sample is represented by n. i N represents the number of pixels in a single hole defect sample, and N represents the total number of pixels in the entire sample.

Images