Method for verifying the orientation of carbon fiber composite material plies based on polarized imaging

Abstract

This invention discloses a method, apparatus, device, storage medium, and system for verifying the layup orientation of carbon fiber composite materials based on polarization imaging. The method acquires light intensity images from multiple polarization angles, reconstructs the surface microscopic normal vector field, employs von Mises-Fischer probability distribution statistical modeling, utilizes a dual neural network model to map the roughness sequence, constructs an angular domain power spectral density function after data correction, extracts the actual layup orientation angle, and completes quality grading verification. This invention achieves high-precision, non-destructive, and automated detection of the layup state of carbon fiber composite materials, effectively solving the problems of destructive, inefficient, and insufficient precision in existing detection technologies, and meeting the quality control needs of high-end equipment manufacturing and industrial mass production of carbon fiber composite components.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and optical detection technology, specifically to a method, apparatus, device, storage medium, and system for verifying the layup orientation of carbon fiber composite materials based on polarization imaging. Background Technology

[0002] Carbon fiber reinforced polymer (CFRP) composites possess excellent properties such as high specific strength, high specific modulus, fatigue resistance, and corrosion resistance, and are widely used in the core load-bearing structures of high-end equipment. The mechanical anisotropy, overall structural stiffness, and fatigue service performance of CFRP components are entirely determined by the orientation and regularity of the fiber layup. High-end industrial fields have stringent requirements for CFRP layup angle tolerances, with strict limits on even minute layup offset errors. During the entire process of prepreg manual laying, mold positioning and clamping, and thermosetting curing, process defects such as prepreg slippage, layup angle deviation, local fiber wrinkling disorder, and curing heat deformation offset are prone to occur. These process defects directly cause a decrease in local stiffness of the component, an aggravation of stress concentration, and a reduction in fatigue service life. In severe cases, they can lead to structural failure, posing significant application safety hazards.

[0003] Existing CFRP fiber layup orientation detection technologies have significant inherent limitations, making it difficult to simultaneously meet the practical application requirements of non-destructive testing, high-precision quantification, and batch industrial quality control. Specific technical shortcomings are as follows: 1. Microscopic section inspection is the mainstream method of industry inspection, but it is a typical destructive inspection. It requires cutting, grinding, polishing and microscopic observation of finished components. The inspection cycle is long and the inspection cost is high. It can only achieve small-scale sampling inspection and cannot achieve 100% full inspection of finished products. It is difficult to adapt to industrial mass production quality control scenarios. 2. Manual visual inspection is highly subjective and lacks quantitative standards. It can only identify large ply deviations, but cannot identify minor ply skews and local minor fiber disorder defects, resulting in poor consistency and reliability of inspection. 3. Traditional one-dimensional roughness detection can only characterize the amplitude of surface micro-undulations, without any spatial directional dimensional information. It cannot characterize the inherent anisotropic texture features of CFRP surface, nor can it establish the correlation between micro-morphology and fiber layup orientation. 4. Conventional ultrasonic and infrared non-destructive testing are only effective against internal macroscopic defects such as delamination, voids, and debonding. They have extremely low sensitivity to minor orientation deviations and microscopic arrangement disorders of surface fibers, making it difficult to meet the process requirements of precision layup verification. 5. Existing high-precision surface orientation detection technologies mostly rely on white light interferometers and laser contour scanning equipment, which are expensive, time-consuming point-by-point scanning detection, and have low detection efficiency. Moreover, they can only obtain surface height information and lack fine features in the microscopic normal vector dimension, making it difficult to effectively distinguish subtle anisotropic differences in fiber texture.

[0004] In addition, traditional visible light machine vision and conventional laser optical inspection methods can only acquire two-dimensional grayscale information or macroscopic height information of the surface, and are not sensitive to the microscopic anisotropic texture of CFRP surface fibers, resulting in an inherent defect of insufficient feature characterization ability. The microstructure of CFRP surface fibers is intricate, with small surface undulations and significant texture directionality. Conventional optical imaging is easily affected by surface specular reflection, ambient stray light, and noise from both light and dark areas, making it difficult to effectively capture subtle directional differences in fiber arrangement and accurately distinguish between regular fiber arrangement and slightly disordered microscopic features, resulting in insufficient detection accuracy and robustness.

[0005] Polarized optics possesses microstructural polarization modulation characteristics not found in conventional optics, making it a crucial core technology for accurately characterizing the anisotropic microtextures of CFRP. Polarized light exhibits high sensitivity to the micro-geometry of material surfaces, fine fiber edges, and groove microstructures, overcoming the resolution limitations of conventional visible light imaging and extracting fiber micro-orientation features that conventional testing equipment cannot identify. Simultaneously, the oriented fiber structure of CFRP exhibits stable polarization-selective modulation of reflected light; subtle differences in fiber arrangement direction and regularity can be directly reflected as regular differences in polarization degree and angle, enabling the physical quantitative characterization of fiber anisotropy. Furthermore, polarized imaging effectively suppresses specular reflections and stray light interference from the CFRP surface, demonstrating good imaging stability and feature consistency, providing high signal-to-noise ratio and high reliability raw data for subsequent quantitative modeling of fiber orientation. In summary, polarized optical imaging technology is a necessary technological foundation for the non-destructive and accurate detection of minute layup deviations and micro-fiber disorder defects on the CFRP surface. Summary of the Invention

[0006] This application aims to overcome the technical shortcomings of existing carbon fiber reinforced polymer (CFRP) fiber layup orientation detection technologies, such as destructive testing, low detection efficiency, inability to quantify anisotropic characteristics, and difficulty in achieving batch automated quality control. It provides a CFRP layup orientation verification method and related equipment based on polarization imaging, enabling high-precision, non-destructive, automated detection and quality grading of the CFRP surface fiber layup state.

[0007] To achieve the above objectives, the first aspect of this application provides a method for verifying the layup orientation of carbon fiber composite materials based on polarization imaging, comprising: Acquire multi-polarization angle light intensity images of the carbon fiber composite material surface under test; The multi-polarization angle light intensity image is analyzed pixel by pixel to solve the Stokes vector fundamental parameters, and the surface microscopic normal vector field is reconstructed based on the Stokes vector fundamental parameters. The normalized discrete normal vector is projected onto the unit sphere, and the von Mises-Fischer probability distribution is used to statistically model the spherical orientation data to solve for the global VMF lumped coefficient and the VMF lumped coefficient vectors of each orientation. Using a pre-trained dual neural network model, the global VMF lumped coefficients are mapped to a global roughness scalar, and the directional VMF lumped coefficient vectors are mapped to a full-angle roughness sequence. Outlier determination and abnormal data interpolation correction are performed on the full-angle roughness sequence to obtain a standard roughness sequence; Based on the standard roughness sequence, an angular domain power spectral density function is constructed, and the azimuth angle corresponding to the peak spectral energy is extracted as the actual layup orientation angle. The layup quality defect classification and verification are completed in combination with the preset process threshold.

[0008] Optionally, the step of solving for the global VMF lumped coefficients and the directional VMF lumped coefficient vectors includes: The maximum likelihood estimation method is used to fit and solve the global normal vector data to obtain the globally unique VMF lumped coefficients; A linear directional filter bank with uniform intervals from 0° to 179° is used to extract the directional features of the global normal vector field through directional convolution. The directional VMF lumped coefficients corresponding to each dimension are obtained by fitting the angles one by one, forming a 180-dimensional directional K-value vector.

[0009] Optionally, the dual neural network model includes a single-input single-output fully connected neural network A and a multi-input multi-output fully connected neural network B; The process of constructing the pre-trained dual neural network model includes: The label dataset is the global average roughness of the corresponding region measured by a high-precision probe roughness meter. The mean square error loss function is used to iteratively train network A until convergence, and then the network weights and bias parameters are fixed. Using the probe roughness data corresponding to 180 directions as the true labels, the mean square error loss function is used to iteratively train network B until convergence, and then the network weights and bias parameters are fixed.

[0010] Optionally, the outlier determination and abnormal data interpolation correction of the full-angle roughness sequence includes: Calculate the relative deviation of single-angle roughness relative to the global roughness scalar; Using the global roughness scalar as a constraint benchmark, and combining the average deviation and standard deviation of all sampling points relative to the benchmark roughness, an adaptive outlier threshold is set. A binary decision logic is established. If the relative deviation of the measurement point exceeds the adaptive outlier threshold, it is marked as an abnormal outlier and removed. For the outlier data that has been removed, equal-interval interpolation correction or second-order spline interpolation is performed using adjacent valid measurement points to generate an equal-interval standard roughness sequence without outliers.

[0011] Optionally, constructing the angular domain power spectral density function based on the standard roughness sequence includes: The DC component and macroscopic tilt trend of the standard roughness sequence are removed, and a pure noise-reduced sequence is obtained by introducing a Hanning window function to constrain the boundary. For the equiangular circular sampling signal, the complex spectrum is solved by discrete Fourier transform, and the autocorrelation function in the angle domain is solved by Wiener-Khinchin theorem. A Fourier transform is performed on the angular domain autocorrelation function to eliminate energy redundancy in the high-frequency band and complete energy normalization, thereby constructing a one-sided power spectral density function.

[0012] Optionally, the azimuth angle corresponding to the extracted spectral energy peak is used as the actual layup orientation angle, and combined with a preset process threshold to complete the layup quality defect classification verification, including: The azimuth energy distribution is obtained by integrating the characteristic wavenumber interval, and the azimuth angle corresponding to the energy peak is determined as the true carbon fiber layup orientation angle. The ratio of maximum energy to minimum energy is calculated as the anisotropy coefficient to quantify the orientation of fiber arrangement; Retrieve the preset design ply angle of the component under test and solve for the absolute deviation between the actual ply orientation angle and the design ply angle; Combining the absolute deviation and the anisotropy coefficient, and referring to the pre-stored process allowable angle tolerance threshold and fiber regularity threshold, a four-level standardized layup quality judgment logic is executed.

[0013] To achieve the above objectives, a second aspect of this application provides a carbon fiber composite material layup orientation verification device based on polarization imaging, comprising: The image acquisition module is used to acquire multi-polarization angle light intensity images of the carbon fiber composite material surface under test; The normal vector reconstruction module is used to perform pixel-by-pixel analysis on the multi-polarization angle light intensity image, solve the Stokes vector fundamental parameters, and reconstruct the surface microscopic normal vector field based on the Stokes vector fundamental parameters. The directional statistical modeling module is used to project the normalized discrete normal vector onto the unit sphere, and to perform statistical modeling of the spherical directional data using the von Mises-Fischer probability distribution to solve for the global VMF lumped coefficient and the directional VMF lumped coefficient vector. The roughness mapping module is used to map the global VMF lumped coefficients to a global roughness scalar using a pre-trained dual neural network model, and to map the directional VMF lumped coefficient vectors to a full-angle roughness sequence. The data correction module is used to determine outliers and correct abnormal data interpolation in the full-angle roughness sequence to obtain a standard roughness sequence. The orientation verification module is used to construct an angular domain power spectral density function based on the standard roughness sequence, extract the azimuth angle corresponding to the peak of the spectral energy as the actual layup orientation angle, and complete the graded verification of layup quality defects in combination with a preset process threshold.

[0014] To achieve the above objectives, a third aspect of this application provides an electronic device, including a processor and a memory, wherein a computer program is stored in the memory; The processor is configured to implement the polarization imaging-based carbon fiber composite layup orientation verification method as described above when executing the computer program.

[0015] To achieve the above objectives, a fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the polarization imaging-based carbon fiber composite layup orientation verification method as described above.

[0016] To achieve the above objectives, a fifth aspect of this application provides a carbon fiber composite ply orientation verification system, comprising: A split-focus plane polarization camera is used to perform global standardized imaging of the carbon fiber composite material surface under test, and simultaneously acquire surface reflected light intensity images at multiple polarization angles in a single exposure. A probe-type roughness meter is used to collect true physical roughness data of a corresponding area to construct a neural network training label dataset. And as provided above, a carbon fiber composite material layup orientation verification device based on polarization imaging, wherein the focal plane polarization camera and the probe-type roughness meter are respectively connected to the input end of the verification device.

[0017] This application discloses a method, apparatus, device, storage medium, and system for verifying the layup orientation of carbon fiber composite materials based on polarization imaging. The method includes acquiring multi-polarization angle light intensity images of the carbon fiber composite material surface to be tested; performing pixel-by-pixel analysis on the multi-polarization angle light intensity images to solve for the Stokes vector fundamental parameters, and reconstructing the surface microscopic normal vector field based on the Stokes vector fundamental parameters; projecting the normalized discrete normal vectors onto a unit sphere, and statistically modeling the spherical orientation data using the von Mises-Fischer probability distribution to solve for the global VMF lumped coefficient and the partial-direction VMF lumped coefficient vectors; using a pre-trained dual neural network model to map the global VMF lumped coefficient to a global roughness scalar, and mapping the partial-direction VMF lumped coefficient vector to a full-angle roughness sequence; performing outlier determination and abnormal data interpolation correction on the full-angle roughness sequence to obtain a standard roughness sequence; constructing an angular domain power spectral density function based on the standard roughness sequence, extracting the azimuth angle corresponding to the spectral energy peak as the actual layup orientation angle, and combining it with a preset process threshold to complete the graded verification of layup quality defects.

[0018] This application achieves accurate verification of the orientation state of CFRP fiber layup through polarization optics global feature extraction, pixel-level surface normal field reconstruction, VMF direction statistical modeling, neural network layered morphology parameter reconstruction, and spectral function analysis. Specifically: 1. A single exposure using a split-focus plane polarization camera simultaneously acquires light intensity images from multiple polarization angles, effectively suppressing specular reflections and stray light interference from the CFRP surface. This overcomes the resolution limitations of conventional visible light imaging, providing high signal-to-noise ratio and high reliability raw data for subsequent processing. 2. Based on Stokes parameters, the microscopic normal vector field is reconstructed, and directional statistical modeling is performed using VMF probability distribution. This transforms the abstract spatial aggregation characteristics of the normal vector into quantifiable global and sub-directional concentration coefficients, accurately characterizing the anisotropic texture features of the CFRP surface. 3. Using a pre-trained dual neural network model, a nonlinear mapping relationship between polarization optical statistical parameters and the true value of physical roughness is established, achieving high-precision reconstruction of the global roughness scalar and the full-angle roughness sequence. This solves the problem of traditional one-dimensional roughness detection lacking spatial directional dimension information. 4. Through adaptive outlier threshold determination, abnormal data interpolation correction, and angular domain power spectral density function construction, noise interference is eliminated, and the azimuth angle corresponding to the spectral energy peak is extracted. Combined with preset process thresholds, a four-level standardized classification verification of layup quality defects is completed. The overall solution enables high-precision, non-destructive, automated detection and quantitative evaluation of the CFRP surface fiber layup state, meeting the actual needs of high-end equipment fields for CFRP component manufacturing and industrial mass production quality control. Attached Figure Description

[0019] Figure 1A schematic flowchart illustrating a method for verifying the layup orientation of carbon fiber composite materials based on polarization imaging, provided in an embodiment of this application; Figure 2 A schematic diagram of network A for a carbon fiber composite layup orientation verification method based on polarization imaging provided in an embodiment of this application; Figure 3 A schematic diagram of network B for a carbon fiber composite layup orientation verification method based on polarization imaging provided in an embodiment of this application; Figure 4 This is a flowchart illustrating the carbon fiber composite layup orientation verification method based on polarization imaging provided in this application embodiment. Detailed Implementation

[0020] Explanation of key terms: Target data: refers to the multi-polarization angle light intensity image of the carbon fiber composite material surface to be tested, as well as the original optical and geometric feature data such as the Stokes vector fundamental parameters and the surface micro normal vector field obtained based on the image.

[0021] Training samples: refers to a labeled dataset consisting of the true value of the global average roughness of the corresponding area measured by a high-precision probe roughness meter and the probe measured roughness data in 180 directions, which is used for iterative training of the neural network model.

[0022] Eigenvectors: refer to the global VMF lumped coefficients obtained by statistical modeling of the von Mises-Fischer probability distribution, and the vector of VMF lumped coefficients in the sub-directions composed of 180-dimensional sub-direction K values.

[0023] Model input: refers to the feature data input to the dual neural network model, including the global VMF lumped coefficients of a single input and the 180-dimensional VMF lumped coefficient vector of multiple inputs.

[0024] Inference results: refer to the global roughness scalar and the full-angle roughness sequence obtained from the output of the dual neural network model.

[0025] Prompt message: During the ply quality defect classification and verification process, the system outputs quantitative feedback information regarding the actual ply orientation angle, anisotropy coefficient, and the fourth-level standardized judgment results.

[0026] Constraint information includes boundary conditions used for data correction and quality assessment, such as adaptive outlier thresholds, process allowable angle tolerance thresholds, and fiber regularity thresholds, which are set based on global roughness scalars.

[0027] Multimodal data refers to the fusion application of multi-polarization angle light intensity images (optical modes) acquired by a split-plane polarization camera and physical roughness true value data (physical contact modes) collected by a probe-type roughness meter.

[0028] Edge device / server / terminal: Electronic device that performs the method of this application, which may be deployed in production line edge computing nodes, cloud servers or portable testing terminals.

[0029] Business object / control object: refers to the carbon fiber composite component to be tested and its layup process quality status. The verification results directly affect the production quality control process and defect classification judgment.

[0030] Currently, the industry primarily relies on microscopic sectioning (destructive and inefficient), manual visual inspection (subjective and lacking quantitative standards), or traditional one-dimensional roughness meters and white light interferometers (lacking directional dimension information) for the inspection of CFRP fiber layup orientation. These technologies cannot simultaneously meet the demands of non-destructive testing, high-precision quantification, and batch automated quality control. Conventional optical imaging is susceptible to interference from surface specular reflections, making it difficult to capture subtle differences in the microscopic anisotropic texture of fibers. While polarization optics can extract microscopic orientation features, traditional methods fail to establish an effective nonlinear mapping relationship between polarization parameters and physical roughness, and lack adaptive correction mechanisms for anomalous data and precise orientation identification methods based on spectral analysis.

[0031] The purpose of this invention is to overcome the technical defects of existing CFRP fiber layup orientation detection technology, such as destructive testing, low efficiency, inability to quantify anisotropic characteristics, and difficulty in batch automated testing. This invention provides a CFRP layup orientation verification method based on polarization imaging, which can achieve high-precision and automated detection of the CFRP surface fiber layup state and meet the quality control requirements of CFRP component manufacturing and industrial mass production.

[0032] The surface fibers of CFRP are arranged in a regular and oriented manner. The oriented distribution of fiber bundles will form anisotropic micro-textures on the surface of the board. The micro-textures of fibers in different directions have a differentiated modulation effect on incident polarized light. Based on this physical mechanism, this invention employs a split-focus plane polarization camera to complete a single full-domain imaging, simultaneously acquiring multi-dimensional polarization intensity data. It solves for pixel-level polarization degree and polarization angle parameters based on Stokes polarization theory, reconstructing a high-density microscopic normal vector field on the CFRP surface. After normalizing the discrete normal vectors, it projects them onto a unit sphere, using VMF probability distribution to quantify the spatial orientation aggregation characteristics of the normal vectors and solve for multi-directional multi-dimensional concentration coefficients. A pre-trained neural network model is used to achieve high-precision reconstruction of the global roughness scalar and the full-angle roughness sequence. A second-order anisotropic structure tensor is constructed based on the full-angle roughness sequence, and the texture dominance direction and anisotropic amplitude are calculated through tensor eigenvalue decomposition to accurately map the actual layup orientation of the surface fibers. Combined with designed layup parameters and preset process thresholds, it completes layup orientation deviation calculation, fiber arrangement regularity evaluation, and defect classification verification, achieving non-destructive, quantitative, and automated detection throughout the entire process.

[0033] refer to Figure 4The first embodiment of this application provides a method for verifying the layup orientation of carbon fiber composite materials based on polarization imaging, to solve the technical problems of existing CFRP fiber layup orientation detection techniques mentioned in the background art, such as destructive nature, low efficiency, inability to quantify anisotropic characteristics, and difficulty in batch automated quality control. This method can be executed by a processor, which can be located in a terminal or server. The execution process of the carbon fiber composite material layup orientation verification method based on polarization imaging is as follows: Step S101: Obtain multi-polarization angle light intensity images of the carbon fiber composite material surface to be tested; refer to Figure 1 In one embodiment of this application, the surface of the CFRP sheet to be tested after hot-press curing is pre-treated by cleaning. A lint-free cloth and anhydrous ethanol are used to wipe the surface to thoroughly remove surface dust, resin spillage, attached impurities, oil stains, and other external interferences, ensuring the original texture of the tested area. Imaging parameters such as the working distance, light source illuminance, incident angle, and exposure parameters of the fixed-focus plane polarization camera are used to ensure uniform imaging conditions for all samples. Global standardized imaging is performed on the tested area of ​​the CFRP sheet, simultaneously acquiring surface reflected light intensity images at four polarization angles (0°, 45°, 90°, and 135°) in a single exposure; multiple frames are continuously acquired and pixel-level averaging noise reduction is performed.

[0034] Step S102: Analyze the multi-polarization angle light intensity image pixel by pixel, solve the Stokes vector fundamental parameters, and reconstruct the surface microscopic normal vector field based on the Stokes vector fundamental parameters. Furthermore, the acquired multi-polarization angle light intensity images are analyzed pixel by pixel to extract the light intensity values ​​corresponding to the four polarization angles for each pixel. Solve for the fundamental parameters of the Stokes vector:

[0035]

[0036]

[0037] In the formula, Total light intensity For linear polarization components. Based on the Stokes parameters mentioned above, the degree of polarization (DOP) and polarization angle (AOP) are calculated pixel-by-pixel:

[0038]

[0039] Using the camera imaging plane as the XY plane and the direction perpendicular to the depth of the material as the Z-axis, the pitch angle of the normal vector is solved based on the degree of polarization, and the azimuth angle of the normal vector is solved based on the polarization angle. The microscopic normal vectors of the surface are reconstructed pixel by pixel to form a global dense normal vector field.

[0040]

[0041]

[0042]

[0043]

[0044] In the formula, The pitch angle is the normal vector. The azimuth angle of the normal vector. These are the three-axis components of the normal vector. The normal vectors corresponding to all pixels are normalized to ensure that all vectors are unit-direction vectors, eliminating vector magnitude interference and unifying the statistical calculation benchmark for direction.

[0045] Step S103: Project the normalized discrete normal vector onto the unit sphere, and use the von Mises-Fischer probability distribution to statistically model the spherical orientation data to solve for the global VMF lumped coefficient and the VMF lumped coefficient vectors of the sub-directions. In an optional embodiment of this application, all the obtained normalized microscopic normal vectors are projected onto a unit sphere, and the discrete directional data of the sphere are statistically modeled using the von Mises-Fischer (VMF) probability distribution. The standard probability density function of the VMF is:

[0046] In the formula, x is an arbitrary unit direction vector on a unit sphere, μ is the average principal direction vector of the VMF distribution, and κ is the VMF concentration coefficient. The maximum likelihood estimation method is used to fit and solve the global normal vector data to obtain the globally unique VMF concentration coefficient K. Simultaneously, 180 sets of linear direction filters with uniform intervals from 0° to 179° are used to extract the directional features of the global normal vector field through convolution, and the directional VMF concentration coefficients corresponding to 180 dimensions are obtained by fitting angle by angle, ultimately forming a 180-dimensional directional K-value vector:

[0047] A schematic diagram of the construction using 180 groups of linear directional filters with uniform intervals from 0° to 179°.

[0048] Step S104: Using a pre-trained dual neural network model, the global VMF lumped coefficients are mapped to a global roughness scalar, and the directional VMF lumped coefficient vectors are mapped to a full-angle roughness sequence. refer to Figure 2 In one embodiment of this application, a fully connected neural network A with single input and single output is pre-constructed. The network input is the global VMF lumped coefficient K, and the network output is the global average roughness Ra of the CFRP surface. The ground truth Ra values ​​of the corresponding region measured by a high-precision probe roughness meter are used as the label dataset to construct a training sample set, a validation set, and a test set.

[0049] Neural network A is trained iteratively using the mean squared error loss function, which is expressed as follows:

[0050] In the formula, For neural networks A The mean squared error loss function value; The total number of training samples; For the first i The global roughness prediction value for each sample; For the first i The true global roughness values ​​of each sample were obtained by actual measurement using a high-precision probe roughness tester.

[0051] After training converges, the network weights and bias parameters are fixed, completing the network solidification. During the detection process, the global K value of the test region is input into the solidified neural network A, and the predicted roughness value of the entire test region is output.

[0052] refer to Figure 3 Furthermore, a fully connected neural network with multiple inputs and multiple outputs is pre-constructed. B The network input is the 180-dimensional VMF lumped coefficient vector obtained by solving, and the network output is the surface roughness scalar corresponding to 180 directions from 0° to 179°. The corresponding dataset is constructed using the probe measured roughness data corresponding to the 180 directions as the ground truth labels to complete the network training and solidification.

[0053] Neural Networks B Iterative training is performed using the mean squared error loss function, which is expressed as follows:

[0054] After the network training converges, the parameters are fixed. During detection, the effective 180-dimensional K-value vector output by S3 is input into the network. B Output the roughness sequence across the entire 0°~179° angle, and construct a complete anisotropic topography data sequence:

[0055] Step S105: Outlier determination and abnormal data interpolation correction are performed on the full-angle roughness sequence to obtain the standard roughness sequence; For example, firstly, a mathematical model for outlier detection is constructed. The surface roughness of carbon fiber composites exhibits significant anisotropy; normal measurement data fluctuates smoothly around the global roughness. However, abnormal data generated by detecting impurities, resin agglomerates, and fiber breaks are highly abrupt and do not belong to the inherent texture characteristics of the material. An adaptive statistical detection model is constructed based on the global roughness to accurately eliminate outliers.

[0056] (1) Calculation of relative roughness deviation: To eliminate range interference, a single-angle roughness relative deviation is defined to quantify the degree to which a single measuring point deviates from the macroscopic reference:

[0057] In the formula, This is a dimensionless relative deviation; For the first i Roughness values ​​corresponding to each angular direction; For global baseline roughness, i.e., by neural network A Output global roughness scalar.

[0058] (2) Adaptive outlier threshold solution: The global reference roughness obtained from S104 through neural network A Assuming absolute true values, and combining 180 sets of angular roughness deviation data, we calculate the deviation statistical characteristic quantity to quantify the fluctuation and dispersion of the normal texture on the surface of carbon fiber composite materials:

[0059]

[0060] In the formula, Roughness relative to the reference for all sampling points The average deviation; This represents the standard deviation of the biased samples. Unlike conventional statistical methods without a benchmark, this invention uses measured constant global roughness as a constraint benchmark, combines it with the texture fluctuation characteristics of carbon fiber materials, modifies the 3σ criterion, and sets an adaptive outlier threshold as follows: .

[0061] (3) Outlier determination rules: Establish a binary decision logic: if a test point satisfies the following formula, it is determined to be an outlier:

[0062] Then mark the set of anomaly point indices. Abnormal data is removed to avoid noise interference in subsequent spectrum calculations and layer identification.

[0063] Then, abnormal data interpolation correction is performed. To retain 180 sets of equal-angle sampling structures and ensure the continuity of Fourier transform data, the abnormal data is supplemented by adjacent measurement point interpolation, which conforms to the continuous and gradual roughness characteristics of composite materials.

[0064] (4) Generate a standard roughness sequence with no anomalies and equal intervals: Let the abnormal measurement point be Take adjacent valid measuring points. , Based on the characteristic of equal angular intervals, the interpolation formula is derived as follows:

[0065] In the formula, For abnormal measurement points The roughness value after interpolation correction; and These are the roughness values ​​at the front and rear adjacent effective measuring points, respectively; The angle position corresponding to the abnormal measurement point; and These represent the angular positions of the two adjacent valid measuring points, one before and one after.

[0066] Substitute constant angular interval This simplifies to the specific interpolation formula:

[0067] After correction, a standard roughness sequence with no anomalies and equal intervals is generated:

[0068] In the formula, For the first m Roughness values ​​after interpolation correction at each angle (abnormal measuring point); and The first m -1 and the first m Roughness values ​​for +1 angle (adjacent effective measuring points before and after); This is the standard roughness sequence generated after correction, containing 180 equally spaced angular data.

[0069] For continuous multi-point abnormal operating conditions, second-order spline interpolation is used, and three sets of effective data before and after the abnormal section are selected to fit the curve, ensuring that the corrected data fits the fiber bundle fluctuation law and avoiding interpolation distortion.

[0070] Step S106: Construct an angular domain power spectral density function based on the standard roughness sequence, extract the azimuth angle corresponding to the peak of the spectral energy as the actual layup orientation angle, and complete the layup quality defect classification verification in combination with the preset process threshold. Further, step S106 may include the following process: (1) Signal preprocessing: Remove the DC component and macroscopic tilting trend from the sequence to extract the microscopic roughness morphology of the material:

[0071]

[0072] In the formula, This is the arithmetic mean of the standard roughness sequence, i.e., the DC component; This represents the roughness value corresponding to the i-th angle in the standard roughness sequence. The roughness fluctuation component corresponding to the i-th angle after removing the DC component is used to extract the micro-roughness morphology features of the material.

[0073] To suppress spectral leakage of finite sequences, a Hanning window function is introduced to constrain the boundary:

[0074] In the formula, Windowing process yields a clean, denoised sequence:

[0075] In the formula, This represents the value of the i-th clean, denoised sequence after windowing. This refers to the roughness fluctuation component after removing the DC component; Let be the window function value of the Hanning window function at the i-th sampling point, used to suppress spectral leakage of finite-length sequences.

[0076] (2) Derive the discrete Fourier transform spectrum: For a circularly sampled signal with equal angles, the complex spectrum is solved using the Discrete Fourier Transform, and the complete expression is:

[0077] In the formula, It is a complex spectrum in the angular domain; the amplitude represents the energy of roughness fluctuations, and the phase represents the spatial arrangement of fibers.

[0078] (3) Derive the power spectral density: Solve for the autocorrelation function in the angle domain using the Wiener-Khinchin theorem:

[0079] Performing a Fourier transform on the autocorrelation function yields the two-sided power spectral density:

[0080] The angular domain roughness is a real discrete signal with a spectrum exhibiting conjugate symmetry, resulting in energy redundancy in the high-frequency band. To eliminate this symmetry redundancy and achieve energy normalization, a one-sided power spectral density function is constructed:

[0081] Among them, P The roughness power spectral density in the angular domain; This represents the two-sided power spectral density. The wave number is k=1, which corresponds to the DC component, and k=2,3,…,N / 2+1 corresponds to the effective harmonic components. N = 180, which represents the total number of sampling points.

[0082] (4) Wavenumber calibration and texture physical mapping: Define angular waves to quantify texture density and establish a physical correspondence between spectral parameters and fiber structure:

[0083]

[0084] In the formula, For angular wavenumber, These are the characteristic wavelengths of the texture. High-frequency wavenumbers correspond to the microscopic roughness of monofilament fibers, while low-frequency wavenumbers correspond to the macroscopic wrinkles of the layup.

[0085] (5) Ply orientation identification and determination: The spectral energy of carbon fiber composite materials is concentrated along the fiber orientation, and the azimuth energy distribution is obtained by integrating the characteristic wavenumber intervals.

[0086] In the formula, Azimuth The corresponding total energy of the spectrum; Angular wavenumber; and These are the lower and upper limits of integration for the characteristic wavenumber interval, respectively; is the power spectral density function in the angular domain.

[0087] In the formula, Azimuth The corresponding total spectral energy, and the azimuth angle corresponding to the energy peak, is the true layup orientation angle of the carbon fiber. Anisotropy coefficients are defined to quantify the degree of fiber orientation.

[0088] In the formula, Anisotropy Index is used to quantify the degree of fiber orientation. This represents the maximum spectral energy value in the azimuth energy distribution. This represents the minimum spectral energy value in the azimuth energy distribution.

[0089] The higher the AI ​​value, the stronger the fiber orientation; when the AI ​​value approaches 1, the material is a random and disordered layup.

[0090] (6) Calculation of fiber layup orientation deviation and verification of defect classification To achieve quantitative assessment of carbon fiber composite ply quality, deviation calculation and defect classification are performed based on the identified actual ply orientation angles and design process parameters. The pre-designed ply angles of the component under test are retrieved. The measured optimal layup orientation angle is determined by combining the peak energy of the spectrum. Solve for the absolute deviation of the ply orientation:

[0091] In the formula, This represents the absolute deviation of the layup orientation. The actual layup orientation angle corresponding to the peak spectral energy; The pre-designed ply angle is set for the component to be tested. This deviation represents the degree to which the actual ply angle deviates from the process design angle.

[0092] This invention defines the spectral anisotropy amplitude and uses the ratio of the maximum energy to the minimum energy to characterize the regularity of fiber arrangement:

[0093] Pre-stored process allows for angular tolerance thresholds With fiber regularity threshold Combined with orientation deviation With anisotropic amplitude A four-level standardized layup quality judgment logic is constructed to achieve defect type classification and identification as follows: 1. Qualified status: Meets the requirements and The fiber layup angle of the component surface layer was determined to be accurate, the arrangement was regular, and there were no defects in the manufacturing process. 2. Ply angle skew defect: meets the requirements. and The fiber arrangement was determined to be regular, with only an overall offset in the laying angle and no local fiber disorder issues. 3. Fiber disorder defect: meets the requirements. and The macroscopic layup angle was determined to meet the design requirements, but the component surface had microscopic forming defects such as fiber slippage, local wrinkles, and disordered arrangement. 4. Seriously non-compliant condition: Meets the requirements and The component was determined to have both ply angle deviation and fiber disorder defects, and the molding quality did not meet the usage requirements.

[0094] Based on the above method embodiments, a second aspect of this application provides a carbon fiber composite material layup orientation verification device based on polarization imaging, comprising: The image acquisition module is used to acquire multi-polarization angle light intensity images of the carbon fiber composite material surface under test; The normal vector reconstruction module is used to perform pixel-by-pixel analysis on the multi-polarization angle light intensity image, solve the Stokes vector fundamental parameters, and reconstruct the surface microscopic normal vector field based on the Stokes vector fundamental parameters. The directional statistical modeling module is used to project the normalized discrete normal vector onto the unit sphere, and to perform statistical modeling of the spherical directional data using the von Mises-Fischer probability distribution to solve for the global VMF lumped coefficient and the directional VMF lumped coefficient vector. The roughness mapping module is used to map the global VMF lumped coefficients to a global roughness scalar using a pre-trained dual neural network model, and to map the directional VMF lumped coefficient vectors to a full-angle roughness sequence. The data correction module is used to determine outliers and correct abnormal data interpolation in the full-angle roughness sequence to obtain a standard roughness sequence. The orientation verification module is used to construct an angular domain power spectral density function based on the standard roughness sequence, extract the azimuth angle corresponding to the peak of the spectral energy as the actual layup orientation angle, and complete the graded verification of layup quality defects in combination with a preset process threshold.

[0095] Based on the above method embodiments, a third aspect of this application provides an electronic device, including a processor and a memory, wherein the memory stores a computer program; The processor is configured to implement the polarization imaging-based carbon fiber composite layup orientation verification method as described above when executing the computer program.

[0096] Based on the above method embodiments, the fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the carbon fiber composite material layup orientation verification method based on polarization imaging as provided above.

[0097] Based on the above method embodiments, a fifth aspect of this application provides a carbon fiber composite material layup orientation verification system, comprising: A split-focus plane polarization camera is used to perform global standardized imaging of the carbon fiber composite material surface under test, and simultaneously acquire surface reflected light intensity images at multiple polarization angles in a single exposure. A probe-type roughness meter is used to collect true physical roughness data of a corresponding area to construct a neural network training label dataset. And as provided above, a carbon fiber composite material layup orientation verification device based on polarization imaging, wherein the focal plane polarization camera and the probe-type roughness meter are respectively connected to the input end of the verification device.

[0098] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for verifying the orientation of a carbon fiber composite material ply based on polarimetric imaging, characterized in that, include: Acquire multi-polarization angle light intensity images of the carbon fiber composite material surface under test; The multi-polarization angle light intensity image is analyzed pixel by pixel to solve the Stokes vector fundamental parameters, and the surface microscopic normal vector field is reconstructed based on the Stokes vector fundamental parameters. The normalized discrete normal vector is projected onto the unit sphere, and the von Mises-Fischer probability distribution is used to statistically model the spherical orientation data to solve for the global VMF lumped coefficient and the VMF lumped coefficient vectors of each orientation. Using a pre-trained dual neural network model, the global VMF lumped coefficients are mapped to a global roughness scalar, and the directional VMF lumped coefficient vectors are mapped to a full-angle roughness sequence. Outlier determination and abnormal data interpolation correction are performed on the full-angle roughness sequence to obtain a standard roughness sequence; Based on the standard roughness sequence, an angular domain power spectral density function is constructed, and the azimuth angle corresponding to the peak spectral energy is extracted as the actual layup orientation angle. The layup quality defect classification and verification are completed in combination with the preset process threshold.

2. The carbon fiber composite material layup orientation verification method based on polarization imaging as described in claim 1, characterized in that, The process of solving for the global VMF lumped coefficients and the directional VMF lumped coefficient vectors includes: The maximum likelihood estimation method is used to fit and solve the global normal vector data to obtain the globally unique VMF lumped coefficients; A linear directional filter bank with uniform intervals from 0° to 179° is used to extract the directional features of the global normal vector field through directional convolution. The directional VMF lumped coefficients corresponding to each dimension are obtained by fitting the angles one by one, forming a 180-dimensional directional K-value vector.

3. The carbon fiber composite material layup orientation verification method based on polarization imaging as described in claim 1, characterized in that, The dual neural network model includes a single-input single-output fully connected neural network A and a multi-input multi-output fully connected neural network B; The process of constructing the pre-trained dual neural network model includes: The label dataset is the global average roughness of the corresponding region measured by a high-precision probe roughness meter. The mean square error loss function is used to iteratively train network A until convergence, and then the network weights and bias parameters are fixed. Using the probe roughness data corresponding to 180 directions as the true labels, the mean square error loss function is used to iteratively train network B until convergence, and then the network weights and bias parameters are fixed.

4. The carbon fiber composite material layup orientation verification method based on polarization imaging as described in claim 1, characterized in that, The outlier determination and abnormal data interpolation correction for the full-angle roughness sequence includes: Calculate the relative deviation of single-angle roughness relative to the global roughness scalar; Using the global roughness scalar as a constraint benchmark, and combining the average deviation and standard deviation of all sampling points relative to the benchmark roughness, an adaptive outlier threshold is set. A binary decision logic is established. If the relative deviation of the measurement point exceeds the adaptive outlier threshold, it is marked as an abnormal outlier and removed. For the outlier data that has been removed, equal-interval interpolation correction or second-order spline interpolation is performed using adjacent valid measurement points to generate an equal-interval standard roughness sequence without outliers.

5. The carbon fiber composite material layup orientation verification method based on polarization imaging as described in claim 1, characterized in that, The construction of the angular domain power spectral density function based on the standard roughness sequence includes: The DC component and macroscopic tilt trend of the standard roughness sequence are removed, and a pure noise-reduced sequence is obtained by introducing a Hanning window function to constrain the boundary. For the equiangular circular sampling signal, the complex spectrum is solved by discrete Fourier transform, and the autocorrelation function in the angle domain is solved by Wiener-Khinchin theorem. A Fourier transform is performed on the angular domain autocorrelation function to eliminate energy redundancy in the high-frequency band and complete energy normalization, thereby constructing a one-sided power spectral density function.

6. The carbon fiber composite material layup orientation verification method based on polarization imaging as described in claim 1, characterized in that, The azimuth angle corresponding to the extracted spectral energy peak is used as the actual layup orientation angle, and combined with a preset process threshold, layup quality defect classification verification is completed, including: The azimuth energy distribution is obtained by integrating the characteristic wavenumber interval, and the azimuth angle corresponding to the energy peak is determined as the true carbon fiber layup orientation angle. The ratio of maximum energy to minimum energy is calculated as the anisotropy coefficient to quantify the orientation of fiber arrangement; Retrieve the preset design ply angle of the component under test and solve for the absolute deviation between the actual ply orientation angle and the design ply angle; Combining the absolute deviation and the anisotropy coefficient, and referring to the pre-stored process allowable angle tolerance threshold and fiber regularity threshold, a four-level standardized layup quality judgment logic is executed.

7. A device for verifying the layup orientation of carbon fiber composite materials based on polarization imaging, characterized in that, include: The image acquisition module is used to acquire multi-polarization angle light intensity images of the carbon fiber composite material surface under test; The normal vector reconstruction module is used to perform pixel-by-pixel analysis on the multi-polarization angle light intensity image, solve the Stokes vector fundamental parameters, and reconstruct the surface microscopic normal vector field based on the Stokes vector fundamental parameters. The directional statistical modeling module is used to project the normalized discrete normal vector onto the unit sphere, and to perform statistical modeling of the spherical directional data using the von Mises-Fischer probability distribution to solve for the global VMF lumped coefficient and the directional VMF lumped coefficient vector. The roughness mapping module is used to map the global VMF lumped coefficients to a global roughness scalar using a pre-trained dual neural network model, and to map the directional VMF lumped coefficient vectors to a full-angle roughness sequence. The data correction module is used to determine outliers and correct abnormal data interpolation in the full-angle roughness sequence to obtain a standard roughness sequence. The orientation verification module is used to construct an angular domain power spectral density function based on the standard roughness sequence, extract the azimuth angle corresponding to the peak of the spectral energy as the actual layup orientation angle, and complete the graded verification of layup quality defects in combination with a preset process threshold.

8. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program; The processor is configured to implement the carbon fiber composite layup orientation verification method based on polarization imaging as described in any one of claims 1 to 6 when executing the computer program.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the carbon fiber composite ply orientation verification method based on polarization imaging as described in any one of claims 1 to 6.

10. A carbon fiber composite material layup orientation verification system, characterized in that, include: A split-focus plane polarization camera is used to perform global standardized imaging of the carbon fiber composite material surface under test, and simultaneously acquire surface reflected light intensity images at multiple polarization angles in a single exposure. A probe-type roughness meter is used to collect true physical roughness data of a corresponding area to construct a neural network training label dataset. And the carbon fiber composite material layup orientation verification device based on polarization imaging as described in claim 7, wherein the focal plane polarization camera and the probe-type roughness meter are respectively connected to the input end of the verification device.

Images