High-resolution spectrum rapid analysis method for fabric fiber components

Abstract

The invention relates to the technical field of spectral analysis, in particular to a fabric fiber component high-resolution spectral rapid analysis method which comprises the following steps: S1, acquiring three-dimensional shape data of a fabric surface through laser scanning, and identifying a fiber arrangement direction and a surface curvature characteristic based on the three-dimensional shape data; s2, dynamically adjusting the incident angle and the focusing depth of a spectrum probe according to the fiber arrangement direction and the surface curvature characteristics identified in S1, and selecting a corresponding spectrum excitation mode; s3, performing feature extraction on the multi-mode spectral data acquired in the S2, and fusing feature vectors of different spectral modes through a deep learning algorithm; s4, analyzing the fused feature vector based on a transfer learning model, and outputting a fiber component type and a proportion thereof; and S5, associating the component data output in the step S4 with space coordinates, and generating a component distribution visualization map. By dynamically adjusting the spectrum probe, fusing multi-mode spectrum data and the like, the analysis precision and speed are remarkably improved.

Description

Technical Field

[0001] The invention relates to the technical field of spectral analysis, and in particular to a high-resolution spectral rapid analysis method for fabric fiber components. Background Art

[0002] In the key links of the textile industry such as raw material quality inspection, production process optimization and product performance evaluation, rapid and accurate analysis of fabric fiber composition is the core requirement to ensure product quality and compliance. As textile materials develop towards multifunctionality and compositeness, the complexity of blended fiber components has increased significantly, and traditional laboratory testing methods can no longer meet the strict requirements of industrial sites for real-time and reliability. The limitations of existing technologies in practical applications are mainly reflected in the following aspects: The microscopic surface morphology of textile fibers is highly complex, and structural features such as yarn buckling, fiber crossing, and fabric wrinkles cause serious interference to optical detection. When a fixed spectral probe is used to detect a fabric surface with a complex three-dimensional morphology, the scattering effect of the incident light on the fiber grooves and raised areas is significantly different, causing signal distortion of the characteristic absorption peak. For example, in a blended fabric containing cotton and synthetic fibers, the natural grooves on the cotton fiber surface and the smooth surface of the polyester fiber produce very different modulation effects on the reflected light at the same incident angle. Traditional spectrometers lack a dynamic adjustment mechanism and are unable to capture the effective spectral characteristics of both types of fibers in a single scan, resulting in increased errors in component analysis. In addition, factors such as environmental vibration, temperature and humidity fluctuations in the industrial site further aggravate the noise interference of the spectral signal, making it difficult to guarantee the reliability of existing static detection solutions.

[0003] The spectral feature aliasing problem caused by multi-component fiber mixing seriously restricts the accuracy of component analysis. When the fabric contains three or more fiber materials (such as cotton / polyester / spandex blends), the characteristic absorption peaks of different materials overlap in the infrared spectrum, while the Raman spectrum is easily interfered by the fluorescence background. Existing single-mode spectral analysis methods usually rely on manual experience to select specific bands for analysis, but do not combine the fiber spatial distribution information to decouple the mixed signal, resulting in a significant decrease in the recognition rate of minor components (accounting for less than 10%). For example, in the recycled fiber sorting scenario, the presence of trace foreign fibers may change the mechanical properties of the final product, and traditional methods cannot effectively separate overlapping peak features, which often results in missed detection of key components.

[0004] The stringent requirements of industrial production lines for detection efficiency are in sharp contradiction with the processing speed of existing technologies. Online quality inspection of textiles usually requires that the single-point analysis time be controlled within 5 seconds. In order to eliminate the influence of surface morphology, traditional methods require mechanical polishing and other pre-treatment of samples, which seriously restricts the detection efficiency. In addition, when the existing transfer learning algorithm is applied across devices and batches, it does not consider the modulation effect of the fiber arrangement direction on the spectral characteristics, resulting in insufficient model generalization ability and the need for frequent recalibration of parameters, which cannot meet the stability requirements in continuous production scenarios. In fabric functional testing (such as waterproof and breathable performance evaluation), the contradiction between the need for rapid analysis of component gradient distribution and the existing single-point detection mode is becoming increasingly prominent.

[0005] The lack of spatial resolution of component distribution leads to a single dimension of quality control. Fabric functionality often depends on the microscopic distribution characteristics of fiber components. For example, moisture absorption and perspiration performance requires a gradient distribution of hydrophilic fibers in a specific area. Traditional interpolation algorithms do not take into account the anisotropic characteristics of fiber orientation, and are prone to produce spatial artifacts when reconstructing component distribution maps, making it impossible to accurately locate heterochromatic fiber aggregation areas or impurity enrichment points. In high-end textile testing, this defect may lead to misjudgment of product batch pass rate, resulting in significant economic losses. Summary of the invention

[0006] Based on the above purpose, the present invention provides a method for rapid high-resolution spectral analysis of fabric fiber components, comprising the following steps: S1: acquiring three-dimensional topographic data of the fabric surface by laser scanning, and identifying the fiber arrangement direction and surface curvature characteristics based on the three-dimensional topographic data; S2: According to the fiber arrangement direction and surface curvature characteristics identified by S1, the incident angle and focusing depth of the spectral probe are dynamically adjusted, and the corresponding spectral excitation mode is selected; S3: Extract features from the multi-mode spectral data collected by S2 and fuse the feature vectors of different spectral modes through a deep learning algorithm; S4: Analyze the fused feature vector based on the transfer learning model and output the fiber component type and its proportion; S5: Associate the component data output by S4 with the spatial coordinates to generate a component distribution visualization map.

[0007] Preferably, the S1 specifically includes: S1-1: using laser emitters distributed in a ring array to project multi-angle interference fringes onto the surface of the fabric, and collecting deformation images of the interference fringes through a charge coupled device, wherein the wavelength of the laser emitters is selected from the visible light band; S1-2: performing a phase unwrapping operation on the deformed image, wherein the phase gradient field is solved by using a least squares algorithm, and the three-dimensional height coordinates are converted based on the laser wavelength and the spatial geometric parameters of the projection device; S1-3: Construct a neural network model based on deep separable convolution to process three-dimensional point cloud data. The encoding stage of the model uses a three-dimensional convolution kernel for feature extraction, and the decoding stage uses transposed convolution to achieve resolution recovery. The final output layer uses a normalized exponential function to calculate the probability value of each voxel point belonging to the direction of the main axis of the fiber. The normalized exponential function is a Softmax function, and its input is an unnormalized logarithmic probability vector; S1-4: Calculate the local curvature parameter based on the probability distribution diagram, wherein the curvature parameter is obtained by calculating the ratio of the second-order derivative of the height distribution function along the fiber direction to the sum of the squares of the first-order derivative.

[0008] Preferably, the method for dynamically adjusting the spectrum probe in S2 is: S2-1: Calculate the best incident angle according to the fiber orientation data, wherein the calculation of the incident angle is based on Snell's law, wherein the surface normal deflection angle is derived from the three-dimensional morphology data of S1, and the fiber material refractive index is from a pre-stored database; S2-2: Establish a light flux feedback control system, monitor the reflected light intensity distribution through a photodiode array, and when the uniformity of the light intensity distribution does not reach a preset range, trigger the robotic arm to adjust the probe pitch angle, and the adjustment amount is calculated by a control algorithm including proportional, integral and differential terms, wherein the proportional coefficient, integral time constant and differential gain parameters are calibrated through a system identification experiment; S2-3: Select the spectral excitation mode according to the local curvature characteristics, and start the wide-field infrared spectroscopy mode when the curvature is less than the set threshold, and the set threshold is determined according to the minimum focusing depth of the optical fiber probe; switch to the pulsed laser Raman mode when multiple layers of fiber overlap are detected, and the pulse width is adaptively adjusted according to the interlayer distance; activate the confocal microscopy mode in highly wrinkled areas, and the confocal pinhole aperture is inversely proportional to the current focusing depth.

[0009] Preferably, the feature extraction method in S3 includes: S3-1: Performing continuous wavelet transform on the infrared spectrum, using Mexican hat wavelet basis function to extract absorption peak features at different scales, wherein the Mexican hat wavelet basis function is in the form of a second-order Gaussian derivative function, and its width parameter is dynamically adjusted according to the grating resolution of the spectrometer; S3-2: performing background suppression processing on the Raman spectrum, estimating the background signal by an iterative reweighting algorithm, wherein the reweighting coefficient is dynamically updated according to the signal-to-noise ratio of each band of the spectrum, and the signal-to-noise ratio is calculated by a moving window method to obtain the ratio of the local mean to the standard deviation; S3-3: A gated attention mechanism is used to fuse different spectral features. The attention weight is obtained by performing a nonlinear transformation on the concatenated feature vector through a trainable parameter matrix, where the nonlinear transformation function is a Sigmoid function, and its output value range is limited between 0 and 1.

[0010] Preferably, the transfer learning model construction method in S4 is: S4-1: In the pre-training stage, a capsule network architecture is used to establish the association between primary features and high-level semantics through a dynamic routing algorithm. The number of iterations of the dynamic routing is determined according to the number of dimensions of the feature vector. S4-2: The domain adaptation process uses a kernel function matching strategy to minimize the spectral distribution difference between the source domain and the target domain. The kernel function is a Gaussian kernel, and its bandwidth parameter is optimized by the maximum likelihood estimation method; S4-3: In the online learning phase, key parameters are identified and regularization constraints are imposed by calculating the square mean of parameter gradients. The calculation of the square mean of gradients uses a sliding window method to retain historical gradient information.

[0011] Preferably, the visualization method in S5 includes: S5-1: constructing a spatial coordinate system by fusing inertial sensor data and optical encoder readings through a Kalman filter, wherein the process noise covariance matrix of the Kalman filter is adaptively adjusted according to the motion speed of the probe; S5-2: The component distribution map is reconstructed using an anisotropic interpolation algorithm, wherein the variation function parameters of the interpolation algorithm are dynamically set according to the main direction angle of the fiber orientation distribution, wherein the range parameter along the fiber orientation is set to 1.5 to 2 times that of the vertical direction.

[0012] Preferably, the specific rule for selecting the spectral excitation mode in S2-3 is: When the standard deviation of the surface normal deflection angle of the local area is less than a first threshold, activating the Fourier transform infrared spectroscopy mode, wherein the first threshold is determined according to the minimum incident angle tolerance of the spectral probe; When it is detected that the difference in the refractive index of materials at adjacent sampling points exceeds a second threshold, switching to the surface enhanced Raman spectroscopy mode, the second threshold being 1.2 times the standard deviation of the refractive index of each material in the standard fiber refractive index database; When the three-dimensional topography data indicates the existence of interlayer gaps, the time-gated fluorescence spectroscopy mode is enabled, and the time-gating width is proportional to the square root of the interlayer distance.

[0013] Preferably, the gated attention mechanism implementation process of S3-3 includes: splicing the infrared spectrum feature vector and the Raman spectrum feature vector along the channel dimension, and performing L2 norm normalization processing on the feature vector before splicing; Generate an initial attention score through a fully connected layer, where the weight matrix dimension of the fully connected layer is (M+N)×K, where M is the infrared feature dimension, N is the Raman feature dimension, and K is the number of neurons in the middle layer; Performing a nonlinear activation function transformation on the initial attention score to obtain a normalized weight coefficient, wherein the slope parameter of the activation function is dynamically adjusted according to the surface curvature value of the current sampling point; The two spectral features are weighted and summed according to the weight coefficient to generate a fusion feature vector, where the sum of the weight coefficient of the Raman feature and the weight coefficient of the infrared feature is always equal to 1.

[0014] Preferably, the domain adaptation process of S4-2 specifically includes: Calculating the maximum mean difference between source domain samples and target domain samples in a reproducing kernel Hilbert space, where the reproducing kernel Hilbert space is constructed by Gaussian kernel function mapping; Optimizing the feature transformation matrix by a gradient descent algorithm to minimize the difference between domains, wherein the learning rate of the gradient descent decays exponentially according to the current iteration number; An orthogonality constraint is introduced into the loss function to prevent the feature space from degenerating, and the strength coefficient of the orthogonality constraint is positively correlated with the domain difference metric.

[0015] Preferably, the anisotropic interpolation algorithm implementation process of S5-2 includes: Determining a main direction of spatial interpolation according to the fiber orientation distribution data, wherein the main direction is obtained by calculating a weighted average of local fiber orientation vectors; Constructing a variation function model including a directional weight factor, wherein the directional weight factor is in direct proportion to the cosine value of the main direction angle; Calculate the estimated value of the component concentration at the unsampled point by solving the Kriging equations, wherein the solution of the equations is iteratively calculated using the conjugate gradient method; An edge-preserving filter is performed on the interpolation result to eliminate spatial artifacts, and the kernel function size of the filter is dynamically adjusted according to the average value of the spacing between the nearest neighbor sampling points.

[0016] Beneficial effects of the present invention: By dynamically adjusting the spectral probe, fusing multimodal spectral data, optimizing the analysis process through transfer learning, and visualizing spatial coordinates, the method significantly improves the accuracy and speed of analysis, and can accurately identify fiber components and their distribution in complex fabric surface structures and multi-layer fibers. Through this series of precise operations, it can provide more detailed and real-time component analysis results, improving the efficiency of fabric quality monitoring and material research and development. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 is a flow chart of the steps of the method of the present invention; Figure 2 is a flow chart of the steps of method S1 of the present invention; Figure 3 4 is a flow chart of the steps of method S2 of the present invention. DETAILED DESCRIPTION

[0019] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. At the same time, it is explained here that in order to make the embodiments more detailed, the following embodiments are the best and preferred embodiments, and those skilled in the art may also adopt other alternatives to implement some known technologies; and the accompanying drawings are only for more specific description of the embodiments, and are not intended to specifically limit the present invention.

[0020] See also Figure 1-Figure 3 The embodiment of the present invention provides a high-resolution spectral rapid analysis method for fabric fiber components. In step S1, the three-dimensional morphological data of the fabric surface is first obtained by laser scanning technology. This process can accurately capture the subtle morphological features of the fabric surface, such as the arrangement direction of the fibers and the surface curvature, and establish an accurate three-dimensional surface model. Through this method, the necessary geometric basic data can be provided for the subsequent steps to ensure that the subsequent analysis can fully consider the actual physical properties of the fabric, such as the fiber orientation and surface undulations.

[0021] Based on S1, step S2 dynamically adjusts the incident angle and focusing depth of the spectral probe according to the fiber arrangement direction and surface curvature characteristics identified in the three-dimensional morphology data. This process is very critical because the fabric surface often presents complex morphological changes. Accurate adjustment of the spectral probe can ensure the effectiveness and accuracy of spectral acquisition under different fiber directions and surface curvatures. By selecting a suitable spectral excitation mode (such as infrared, Raman or laser mode), the representativeness and accuracy of the collected data can be further improved, thus laying a solid foundation for subsequent data processing and analysis.

[0022] Based on the multi-mode spectral data collected in S2, the S3 step uses a deep learning algorithm to extract and fuse the features of different spectral modes. This stage uses fine feature extraction to not only extract key absorption peaks and other spectral features from infrared and Raman spectra, but also uses deep learning methods to automatically discover the associations between different modes and fuse them into a comprehensive feature vector. This fusion helps to more accurately describe the fiber composition of the fabric and ensure the meticulousness and accuracy of the component analysis.

[0023] Step S4 uses the transfer learning model to parse the fused feature vector and output the fiber component type and its proportion. The introduction of transfer learning enables the model to perform effective reasoning through existing knowledge (source domain) even when there is less training data. Through transfer learning, the model can adapt to different spectral data sources and achieve higher accuracy and generalization capabilities. In addition, transfer learning can reduce the reliance on a large amount of labeled data, saving time and cost.

[0024] Finally, step S5 combines the composition data output from S4 with the spatial coordinates to generate a visualization map of the composition distribution. This step can intuitively display the spatial distribution of different fiber components in the fabric through visualization, helping researchers and engineers understand the composition structure of the fabric. The association of spatial coordinates with composition data not only improves the interpretability of the analysis results, but also assists in more refined optimization during production and quality control.

[0025] These steps are closely linked to form an efficient and accurate spectral analysis system. Each step has its unique contribution, ensuring the efficiency and accuracy of spectral data acquisition, feature extraction, component analysis and result visualization. For example, dynamic adjustment of the probe angle and focusing depth improves the accuracy of spectral data acquisition, and deep learning algorithms and transfer learning models enhance the ability to process complex data. The resulting component distribution map not only provides researchers with detailed component information, but can also be used for quality control in actual production. Through this method, rapid and accurate analysis of fabric fiber components can be achieved, greatly improving the efficiency of fabric production and material research and development.

[0026] In a possible implementation, at the initial stage of S1, laser emitters distributed in a circular array are used to project multi-angle interference fringes onto the fabric surface. This design can collect deformation images of the fabric surface in multiple directions and collect these deformation images through a charge coupled device (CCD). The selected visible light band laser emitter ensures high optical resolution and helps to obtain clear interference fringe images. The collection of multi-angle interference fringes can provide deformation information of the fabric surface at different angles, laying the foundation for subsequent phase unwrapping and three-dimensional reconstruction.

[0027] After obtaining the deformation image, the S1-2 step accurately calculates the three-dimensional morphology of the fabric surface through phase unwrapping operation. This process uses the least squares algorithm to solve the phase gradient field, thereby eliminating the phase ambiguity in the deformation image. The three-dimensional coordinates are converted by the laser wavelength and the spatial geometric parameters of the projection device to ensure the accuracy and precision of the morphological data. This stage provides the basic high-precision coordinate information for the subsequent three-dimensional point cloud data processing, ensuring the fine expression of the surface morphology.

[0028] In step S1-3, a neural network model based on deep separable convolution is used to process the three-dimensional point cloud data. The encoding stage of this neural network model uses a three-dimensional convolution kernel for feature extraction, which can efficiently extract representative features from complex three-dimensional point cloud data. In the decoding stage, the transposed convolution technique is used for resolution recovery to ensure that the output three-dimensional morphology map has sufficient details and accuracy. The probability value of each voxel point is processed by the Softmax normalized exponential function, so that the network can determine whether each point belongs to the direction of the main axis of the fiber, thereby obtaining the directional information of the fibers on the surface of the fabric. In this stage, the deep learning ability of the neural network significantly improves the parsing ability of the three-dimensional morphology data, providing a reliable basis for the accurate identification of the fiber arrangement direction.

[0029] Step S1-4 calculates the local curvature parameters based on the probability distribution diagram obtained in the previous step. These curvature parameters are obtained by calculating the ratio of the second-order derivative of the height distribution function along the fiber direction to the sum of the squares of the first-order derivative. The local curvature parameters reflect the changing characteristics of the fabric surface curvature, which can further help identify the bending of the fiber and the details of the surface undulation, thereby enhancing the understanding of the fabric morphological characteristics.

[0030] In the process from S1-1 to S1-2, the combination of laser interferometry technology and phase unwrapping algorithm can accurately obtain the three-dimensional morphological information of the fabric surface and ensure the accuracy of the data. In the S1-3 stage, the application of the deep separable convolutional neural network model improves the ability to process complex three-dimensional point cloud data, and realizes the probability analysis of fiber direction through the Softmax function, making the data analysis process more intelligent and efficient. In the local curvature calculation stage of S1-4, combined with the height distribution characteristics of the fiber direction, the ability to analyze the microstructure of the fabric surface is further enhanced.

[0031] Through these closely connected steps, the entire process not only improves the accuracy and resolution of the three-dimensional morphology analysis of the fabric, but also provides more accurate surface data for subsequent spectral analysis, so that the rapid analysis of the fabric fiber composition can obtain high-quality results in a shorter time, significantly improving the analysis efficiency and accuracy.

[0032] In a possible implementation, in S2-1, the orientation of the fibers on the fabric surface is first calculated by using the three-dimensional morphology data obtained in step S1. This fiber orientation information is used to determine the optimal incident angle of the spectral probe to ensure that the excitation light can efficiently irradiate the target area. The calculation of the incident angle is based on Snell's law, which involves the calculation of the surface normal deflection angle, which is derived from the three-dimensional morphology data of step S1, and the refractive index of the fiber material is extracted from a pre-stored database. In this way, the incident angle of light can be accurately matched to the geometric shape of the fiber surface, thereby improving the intensity and quality of the spectral signal.

[0033] In step S2-2, a light flux feedback control system is used to monitor the intensity distribution of the reflected light using a photodiode array. If the intensity distribution of the reflected light fails to meet the preset uniformity requirements, the system will automatically trigger the robotic arm to adjust the pitch angle of the spectral probe. At this time, the adjustment amount is calculated using a control algorithm that includes proportional, integral, and differential terms. The proportional coefficient, integral time constant, and differential gain parameters of the control algorithm are calibrated through system identification experiments to ensure the stability and accuracy of the adjustment process. The feedback mechanism ensures that the probe is always in the best working state by monitoring the light flux in real time, avoiding analysis errors caused by uneven illumination.

[0034] Step S2-3 selects different spectral excitation modes according to the local curvature characteristics to further optimize the effect of spectral analysis. Specifically, when the local curvature is less than the set threshold, the wide-field infrared spectroscopy mode is activated, which is suitable for relatively flat areas; when the local curvature is large, it switches to the pulsed laser Raman mode, which can effectively handle the situation of overlapping multiple layers of fibers, and the pulse width will be adaptively adjusted according to the distance between layers to provide more accurate spectral data. In addition, when a highly wrinkled area is detected, the confocal microscopy mode is activated, and the confocal pinhole aperture is adjusted in inverse proportion to the current focusing depth to improve the image resolution and detail capture capabilities to adapt to different surface structures of the fabric. Through this dynamic adjustment, the fabric surface in different areas can obtain the most suitable spectral excitation mode, thereby providing high-resolution and highly adaptable spectral data.

[0035] This series of steps ensures that the spectral probe can dynamically adapt to the morphological characteristics of the fibers on the fabric surface, significantly improving the accuracy and efficiency of spectral analysis. The incident angle calculation in S2-1 combined with the three-dimensional morphology data of step S1 can ensure that the spectral probe is accurately aligned with the fiber surface and optimize the lighting effect. The light flux feedback control system in S2-2 ensures uniform reflected light intensity by real-time monitoring and automatic adjustment of the position of the spectral probe, thereby improving the stability and reliability of data acquisition. S2-3 selects different spectral excitation modes according to different local curvatures, allowing the system to adapt to a variety of fabric surface morphologies, including flat, wrinkled, overlapping and other complex structures, thereby providing higher resolution and more accurate results for spectral analysis.

[0036] Through this multi-level, dynamically adjusted mechanism, the applicability and flexibility of the high-resolution spectral rapid analysis method of fabric fiber components are improved as a whole, providing a strong technical guarantee for the accurate analysis of fabric fiber components.

[0037] In one possible implementation, in S3-1, a continuous wavelet transform is first performed on the infrared spectral data, and a Mexican hat wavelet basis function is used for feature extraction. The Mexican hat wavelet is a second-order Gaussian derivative function that can effectively extract absorption peak features at different scales. Specifically, the wavelet transform decomposes signals of different frequencies through multi-scale analysis, thereby accurately locating the absorption peak in the infrared spectrum. The width parameter of the Mexican hat wavelet is dynamically adjusted according to the grating resolution of the spectrometer, so that the method can adapt to the resolution requirements of different instruments and improve the accuracy of the absorption peak features. Through this transformation, key chemical composition information can be effectively extracted from complex spectral data.

[0038] In step S3-2, background signal suppression processing is performed on the Raman spectrum. Background signals often affect the analysis results of Raman spectra, so an iterative reweighting algorithm is needed to estimate and suppress these background signals. The reweighting coefficient is dynamically updated based on the signal-to-noise ratio of each band of the spectrum. This signal-to-noise ratio is estimated by calculating the ratio of the local mean to the standard deviation using the moving window method. By dynamically adjusting the reweighting coefficient, the background noise can be accurately removed and the true Raman characteristic peaks can be retained, making the Raman spectrum data clearer and more reliable, and accurately reflecting the molecular structure and composition of the fabric fibers.

[0039] Step S3-3 uses a gated attention mechanism to fuse different spectral features. The gated attention mechanism performs a nonlinear transformation on the spliced ​​infrared spectrum and Raman spectrum feature vectors by training a learnable parameter matrix. The transformation function uses a Sigmoid function, whose output value range is limited to between 0 and 1, and can dynamically adjust the importance of different features. In this way, the model can automatically weight and fuse these features according to the contribution of different spectral data to obtain the optimal feature representation. The gated attention mechanism not only enhances the model's sensitivity to key signals, but also effectively removes redundant or unimportant signals, further improving the accuracy of the analysis.

[0040] The connection and fusion of these steps significantly improve the accuracy and efficiency of spectral data processing. The wavelet transform in S3-1 provides multi-scale absorption peak extraction capabilities for infrared spectroscopy, effectively dealing with different types of spectral peaks and signal fluctuations. S3-2 removes background noise in the Raman spectrum by dynamically updating the background suppression processing of the reweighting coefficient, making the Raman characteristic signal more prominent and providing a clear signal basis for subsequent analysis. The gated attention mechanism in S3-3 further optimizes the feature fusion process and enhances the model's ability to process complex data by intelligently weighting and selecting useful features.

[0041] In summary, through the implementation of these steps, the most representative and important features can be extracted from infrared and Raman spectra, which improves the accuracy and robustness of data analysis, thereby enabling more accurate identification of the composition of fabric fibers.

[0042] In one possible implementation, in S4-1, a capsule network architecture is used in the pre-training stage, and the network establishes the relationship between primary features and high-level semantics through a dynamic routing algorithm. The characteristic of the capsule network is that it can capture complex spatial relationships and directional information, which is particularly important for the analysis of spectral data. The dynamic routing algorithm enables capsule nodes at different levels to dynamically select and transmit information, thereby forming a more robust feature representation. In this step, the number of iterations of the dynamic routing is determined according to the number of dimensions of the feature vector, and higher-dimensional feature vectors usually require more iterations to better learn and map complex relationships in high-dimensional data. Through the capsule network, the spatial relationships and structural features in the spectral data can be effectively captured, and the expression ability of the model in subsequent stages can be improved.

[0043] In S4-2, the domain adaptation process of transfer learning adopts a kernel function matching strategy, which aims to minimize the difference in spectral distribution between the source domain (i.e., training set) and the target domain (i.e., test set or new sample). Since there may be a deviation in the distribution between the source domain and the target domain, directly applying the pre-trained model to the target domain may result in lower accuracy. To solve this problem, the Gaussian kernel function is used for distribution matching. The Gaussian kernel function can nonlinearly map the features of the source domain and the target domain in a high-dimensional space, minimizing the distribution difference. Its bandwidth parameters are optimized by the maximum likelihood estimation method to ensure that the parameters of the kernel function can reflect the difference between the source domain and the target domain to the greatest extent. In this way, the model can achieve better generalization ability between cross-domain data.

[0044] In the online learning phase of S4-3, key parameters are identified by calculating the mean squared gradient of the parameters, and regularization constraints are imposed on these parameters. The mean squared gradient is calculated using the sliding window method, which smoothes the parameter update process by retaining historical gradient information. In this way, the model can keep a memory of historical data during the learning process and avoid overfitting the current samples. In addition, regularization constraints can effectively prevent the overfitting of the model and ensure its robustness and generalization ability on new data. The goal of this phase is to enable the model to continuously adapt to new spectral data and optimize its parameters, making the analysis of spectral data more accurate and efficient.

[0045] The combination of these steps enables the model to efficiently perform transfer learning, solves the domain differences between different spectral data, and optimizes the performance of the model in the continuous learning process. Specifically, the capsule network architecture can handle complex spectral features and spatial relationships, the domain adaptation process optimizes the adaptability of the source domain and the target domain through the Gaussian kernel function, and the online learning stage ensures the stability and generalization ability of the model through the square mean gradient and regularization constraints.

[0046] Suppose we want to quickly analyze the fiber composition of fabrics. First, through the pre-training stage, a capsule network is used to learn a large amount of spectral data to establish the association between primary features and high-level semantics. For example, in infrared spectral data, the primary feature may be the position information of the absorption peak, while the high-level semantics corresponds to the classification label of the fiber component. Through the dynamic routing algorithm, the model can effectively capture this relationship. Then, in the domain adaptation stage, it is assumed that the source domain is the spectral data of a certain type of fabric, and the target domain is another type of fabric. Through kernel function matching and maximum likelihood estimation, the difference in spectral distribution between the source domain and the target domain is minimized, thereby improving the analysis accuracy of the target domain data. In the online learning stage, as new samples arrive, the model will use historical gradient information for optimization, identify which parameters have an important impact on the final result, and prevent overfitting through regularization. Finally, the model can adapt to new spectral data and perform efficient and accurate analysis.

[0047] In summary, through the connection and implementation of these steps, the adaptability and accuracy of spectral analysis can be effectively improved, ensuring the effectiveness and efficiency of this method in the composition analysis of different types of fabric fibers.

[0048] In a possible implementation, in S5-1, a spatial coordinate system is constructed by fusing the data of the inertial sensor and the optical encoder through a Kalman filter. The inertial sensor provides acceleration and angular velocity information of the probe, while the optical encoder provides displacement data. The role of the Kalman filter is to fuse this information, compensate for the noise and uncertainty of the data, and thus more accurately locate the position and direction of the probe. In particular, the process noise covariance matrix of the Kalman filter will be adaptively adjusted according to the movement speed of the probe. As the probe speed changes, the Kalman filter can dynamically adjust the noise weight to improve the positioning accuracy after fusion. This step ensures that the establishment of the spatial coordinate system has a high accuracy and provides a reliable spatial basis for subsequent spectral data analysis.

[0049] In S5-2, an anisotropic interpolation algorithm is used to reconstruct the composition distribution map of the fabric. The anisotropic interpolation algorithm sets interpolation parameters in different directions so that the data has different degrees of smoothness in different directions. Specifically, the parameters of the variation function are dynamically set according to the main direction angle of the fiber orientation distribution. This means that the interpolation algorithm will optimize the interpolation process according to the direction of the fiber orientation, so that the range parameter along the fiber orientation is set to 1.5 to 2 times that of the vertical direction. This setting allows the spectral changes along the fiber direction to be better reflected, while the changes in the vertical direction are appropriately smoothed. In this way, the distribution map of the fabric fiber components in space can be reconstructed more accurately, making the final component analysis more accurate and clear.

[0050] Combined with the Kalman filter and the anisotropic interpolation algorithm, these steps can effectively process and analyze spectral data and ultimately obtain an accurate component distribution map. The Kalman filter provides high-precision probe positioning information by adaptively adjusting the process noise covariance matrix, providing a solid foundation for subsequent interpolation. The anisotropic interpolation algorithm can optimize the reconstruction effect of the component distribution map according to the fiber's orientation distribution characteristics by dynamically adjusting the interpolation parameters. Through the coordination of these two steps, a more realistic and detailed fabric fiber component distribution map can be obtained, which improves the spatial resolution and accuracy of spectral analysis, and ultimately improves the reliability and practicality of the fabric fiber component analysis method.

[0051] Suppose we are carrying out a spectral analysis task of fabric composition. First, in the Kalman filter fusion stage, we use the inertial sensor installed on the probe to measure the acceleration and angular velocity of the probe, and obtain its position data through the optical encoder. By fusing these data, the Kalman filter accurately calculates the position and orientation of the probe on the fabric surface and tracks the motion state of the probe in real time. Then, in the reconstruction stage of the component distribution map, the anisotropic interpolation algorithm is used to reconstruct the fiber component distribution map. During the interpolation process, the system dynamically sets the interpolation variogram parameters according to the fiber orientation distribution. For example, if the main orientation of the fiber has a small angle with the vertical direction, the system will set a larger range parameter accordingly, so that the changes along the fiber direction can be better restored. Finally, the output component distribution map can clearly show the changes in fabric composition in different areas, greatly improving the accuracy of the analysis.

[0052] In summary, through the precise positioning of the Kalman filter and the optimized reconstruction of the anisotropic interpolation algorithm, high-resolution spatial visualization of fabric fiber composition analysis can be achieved, ensuring that this method can efficiently and accurately generate and analyze component distribution maps.

[0053] In one possible implementation, step S2-3 improves the accuracy and efficiency of the analysis by selecting a suitable spectral excitation mode. Specifically, under different material surface characteristics and three-dimensional morphology conditions, different spectral modes are activated according to specific rules, including Fourier transform infrared spectroscopy (FTIR), surface enhanced Raman spectroscopy (SERS), and time-gated fluorescence spectroscopy (TCFS) modes.

[0054] Specifically, in step S2-3, when the spectral probe moves on the fabric surface, the surface normal angle of the area is first collected. The change in the surface normal angle reflects the morphology and flatness of the fabric surface. If the standard deviation of the surface normal deflection angle of the local area is detected to be less than the set first threshold, it means that the surface of the area is relatively flat and uniform, and the Fourier transform infrared spectroscopy (FTIR) mode is activated for analysis.

[0055] The setting of the threshold is usually determined based on the minimum incident angle tolerance of the spectral probe, ensuring that the contact angle between the probe and the surface is maintained within a small tolerance range to avoid affecting the accuracy of the spectral data due to the deviation of the incident angle.

[0056] By activating the FTIR mode, the spectral data of the fabric in this flat area can be obtained, so as to accurately analyze the fiber composition, molecular structure and other information. The FTIR mode is suitable for spectral analysis of uniform and flat surfaces, providing high-resolution chemical composition information.

[0057] Furthermore, when adjacent sampling points show a large difference in material refractive index during spectral analysis, the system switches to surface enhanced Raman spectroscopy (SERS) mode. The difference in refractive index can reflect the presence of different materials or fibers. At this time, the second threshold is set to 1.2 times the standard deviation of the refractive index of each material in the standard fiber refractive index database. Through this refractive index difference, the system can automatically identify the differences between materials, thereby optimizing the excitation mode of spectral analysis and switching to SERS mode. The SERS mode can effectively improve the sensitivity to tiny local differences and is particularly suitable for high-sensitivity analysis of interfaces or heterogeneous regions of different materials. Through the surface enhancement effect, SERS can provide stronger signals in low concentrations or small areas, thereby accurately analyzing different components.

[0058] In step S2-3, when the spectral probe moves on the fabric surface, the system will also combine the three-dimensional morphology data to determine whether there is an interlayer gap on the fabric surface. When the interlayer gap is detected, the system will automatically enable the time-gated fluorescence spectroscopy (TCFS) mode. The setting of the time gate width is directly proportional to the square root of the interlayer distance, which means that as the interlayer distance increases, the gate time will be appropriately extended, thereby accurately capturing the time characteristics of the fluorescence signal.

[0059] Enabling time-gated fluorescence spectroscopy can effectively suppress background noise and improve the timeliness and specificity of the signal. When there is a gap between layers, the fluorescence signal may be subject to greater scattering or attenuation. Time gating technology can help accurately identify and enhance the target signal, thereby achieving more accurate analysis.

[0060] The selection and switching of these three excitation modes are based on different surface morphologies, refractive index differences and three-dimensional structure data, ensuring that the most suitable spectral mode can be automatically selected under variable fabric surface conditions. The activation conditions of each mode are targeted and dynamically adjusted according to actual conditions, improving the efficiency and accuracy of spectral analysis.

[0061] In the specific implementation process, the probe first touches the fabric surface. During the preliminary scanning process, the system measured that the standard deviation of the surface normal was less than the preset first threshold, so the Fourier transform infrared spectroscopy mode was activated to obtain the chemical composition information of the fiber. Then, when scanning different areas, it was detected that the refractive index difference between adjacent sampling points was greater than the second threshold, and the system immediately switched to the surface enhanced Raman spectroscopy mode to enhance the analysis ability of small regional differences. Finally, when encountering interlayer gaps, the system judged and enabled the time-gated fluorescence spectroscopy mode based on the three-dimensional morphology data, further optimizing the extraction and analysis process of the fluorescence signal.

[0062] Through the organic combination of these steps, the final spectral analysis results not only have high resolution, but also can perform accurate component analysis under various complex surface conditions, greatly improving the adaptability and accuracy of fabric fiber composition analysis methods.

[0063] In one possible implementation, step S3-3 implements a gated attention mechanism to optimize the fusion process of infrared and Raman spectroscopy data. In the specific implementation process, a series of steps and technical means are adopted to finally generate a fused feature vector through feature vector concatenation, attention score calculation, nonlinear activation function transformation and weighted summation.

[0064] Specifically, in step S3-3, it is first necessary to extract feature vectors for the infrared spectrum and the Raman spectrum respectively. In order to further fuse, the feature vectors of the two spectra need to be spliced ​​in the channel dimension. Before splicing, in order to eliminate the dimensional differences in different spectral feature dimensions, the feature vectors of the infrared spectrum and the Raman spectrum are L2-norm normalized respectively. The purpose of L2-norm normalization is to ensure that the two feature vectors have the same scale before splicing, so as to avoid affecting the subsequent fusion process due to scale differences.

[0065] L2 norm normalization enables spectral data from different sources to be fused on the same scale, thus ensuring the effectiveness and accuracy of data fusion. In addition, the concatenated feature vector will contain key information of both infrared and Raman spectra, providing sufficient information for subsequent analysis.

[0066] Based on the concatenated feature vector, the initial attention score is then generated through the fully connected layer. The input of the fully connected layer is the concatenated feature vector, and the output is the initial attention score. The weight matrix dimension of this fully connected layer is (M+N)×K, where M is the dimension of the infrared spectrum feature, N is the dimension of the Raman spectrum feature, and K is the number of neurons in the middle layer. The role of this layer is to calculate the importance of each feature in the final fusion by weighted combination of infrared and Raman spectrum features.

[0067] By generating the initial attention score through the fully connected layer, the importance of each feature can be calculated without losing information, providing a basis for the subsequent weighted summation. This is the core of the gated attention mechanism, which can dynamically adjust the weight of spectral features to improve the effect of feature fusion.

[0068] The initial attention score is transformed by a nonlinear activation function to obtain a normalized weight coefficient. The activation function includes but is not limited to ReLU (linear rectification function) and Sigmoid function. Under different surface curvature conditions, the slope of the activation function will be adjusted to control the range of attention scores, making feature weighting more accurate and flexible.

[0069] Dynamically adjusting the slope parameter of the activation function allows the system to adapt to different surface morphologies. Different weighting methods may be required for surfaces with different curvatures. Dynamic adjustment can ensure that the model processes spectral features more accurately, thereby improving the accuracy and robustness of the analysis.

[0070] After obtaining the normalized weight coefficients, these coefficients are used to perform weighted summation on the features of the infrared spectrum and the Raman spectrum, and finally a fused feature vector is generated. It should be noted that the sum of the weight coefficients of the Raman feature and the infrared feature is always equal to 1, which ensures the balance of weights in the fusion process of the two features. Through this weighted summation, the system can combine the two spectral information, make full use of their respective advantages, and generate a comprehensive feature vector for subsequent analysis.

[0071] The weighted summed fusion feature vector can integrate the information of infrared spectrum and Raman spectrum, thus improving the data representation ability. The weighting coefficient ensures that the fusion of the two spectral features has an appropriate balance under different sampling conditions, making the analysis results more accurate and reliable.

[0072] These four steps form a complete gated attention mechanism. Through the process of L2 norm normalization, fully connected layer, nonlinear activation function and weighted summation, the system can adaptively adjust the weights of each spectral feature according to its importance, thereby achieving efficient feature fusion. This mechanism can handle different types of spectral data and flexibly adjust according to the actual surface morphology and curvature information, thus improving the accuracy and adaptability of spectral analysis.

[0073] In the specific implementation process, feature vectors are first extracted through infrared spectroscopy and Raman spectroscopy respectively. On this basis, the system concatenates the two feature vectors and performs L2 norm normalization. Next, the initial attention score is generated using the fully connected layer and nonlinearly transformed through the activation function. Assuming that the current sample has a high curvature value, the slope of the activation function will be dynamically adjusted according to this curvature value. Finally, a fused feature vector is generated through weighted summation and used for subsequent analysis and recognition tasks. Through this process, the system can more accurately reflect the characteristics of the fabric fiber composition, thereby improving the analysis efficiency and the accuracy of the results.

[0074] In one possible implementation, the domain adaptation process in step S4-2 aims to solve the distribution deviation problem between the source domain (such as standard laboratory sampling data) and the target domain (such as actual factory site sampling data) to ensure the generalization ability of the model between different domains. This process combines the three key technologies of maximum mean difference (MMD) metric, gradient descent optimization, and orthogonality constraint to build a stable and efficient cross-domain feature alignment mechanism.

[0075] First, the source domain samples With the target domain sample Mapped to a reproducing kernel Hilbert space (RKHS), which is composed of the Gaussian kernel function Constructed, it can effectively capture nonlinear relationships. In this high-dimensional space, the "maximum mean difference" is used as the distance measure between the source domain and the target domain sample distribution:

[0076] in, and are the number of samples in the source domain and the target domain, respectively. It is the characteristic function after Gaussian kernel mapping, which maps the samples from the original space to RKHS. The MMD metric can quantify the distribution difference between the source domain and the target domain under unsupervised conditions, avoiding the training difficulties caused by missing labels, and is one of the mainstream methods in current domain adaptation tasks.

[0077] Furthermore, a feature transformation matrix is ​​defined , the original spectral features Mapping to a new embedding space , to minimize the MMD loss. The matrix is ​​iteratively optimized using the gradient descent algorithm, and the objective function is:

[0078] The learning rate With the current iteration number Perform exponential decay, such as:

[0079] The exponentially decaying learning rate strategy can converge quickly in the early stage and avoid oscillation in the later stage, which helps to find a more stable optimal mapping matrix. At the same time, through continuous optimization by gradient descent, the difference between domains is effectively reduced, thereby improving the expressiveness of the model in the target domain.

[0080] In order to prevent the feature transformation matrix from degenerating (such as highly correlated features or dimension collapse after transformation), an orthogonality constraint is introduced into the total loss function:

[0081] This constraint mapping matrix column vectors remain approximately orthogonal, maintaining the independence between features. It is positively correlated with the MMD metric value. That is, when the difference between the source domain and the target domain is greater, the system automatically strengthens the orthogonality constraint and improves the model stability.

[0082] This design can effectively avoid information redundancy and dimensionality collapse in the mapped features, improve the diversity and effectiveness of feature expression, and thus improve the discrimination ability of the classifier in the target domain.

[0083] In the specific implementation process, the source domain is the spectral data collected by high-precision instruments in the laboratory, and the target domain is the data collected by sensors on the factory site. Due to the different equipment precision and environment, the distribution of the target domain is significantly different from that of the source domain.

[0084] Specific: After the joint features of infrared and Raman spectra are extracted separately, they enter RKHS through Gaussian kernel mapping; Use MMD to measure the distribution difference between the source domain and the target domain; Defining the transformation matrix , using a learning rate of 0.01 and a decay coefficient of The gradient descent algorithm is used to optimize it; orthogonal constraints are introduced (such as ), ensuring feature diversity; The final trained model improves recognition accuracy in the target domain by about 12% and can quickly adapt to other factory environment data with similar distribution.

[0085] This method achieves efficient migration and adaptation between different data domains in fabric fiber spectral analysis through a systematic and mathematically explainable process, thereby improving the versatility, accuracy and practical deployment capabilities of the model.

[0086] In a possible implementation, the anisotropic interpolation algorithm implementation process of step S5-2 is intended to solve the problem of spatial interpolation result errors caused by uneven fiber distribution directions or different local orientations. This process improves interpolation accuracy and ensures the accuracy of component concentration estimation by accurately calculating local fiber orientations and constructing a variogram model containing orientation information.

[0087] First, the main direction in the local area is determined by analyzing the fiber orientation distribution data. The specific method is to calculate the weighted average of the local fiber orientation vector, which can be obtained by the following formula:

[0088] in, Indicates The direction vector of each fiber.

[0089] Indicates The weight of an individual fiber, usually related to the density or strength of the fiber.

[0090] Indicates the total number of fibers.

[0091] This step ensures that the influence of fiber orientation is taken into account during the interpolation process, avoiding interpolation deviation caused by uneven fiber orientation distribution, thereby improving the accuracy of the interpolation results.

[0092] In the interpolation process, the fiber directionality needs to be modeled through the variation function model. The model adds a directional weight factor to adjust the degree of spatial variation in different directions. The directional weight factor is directly proportional to the cosine value of the main direction angle, that is: ; in is the angle between the current direction and the principal direction. This design reflects the anisotropic nature of the fiber, with less variability in the principal direction and greater variability in the direction perpendicular to the principal direction.

[0093] This design can more accurately capture the effect of fiber orientation on component distribution, avoiding the traditional isotropic interpolation method that ignores the influence of fiber orientation, thereby improving the accuracy of spatial interpolation, especially in fiber samples with obvious directional distribution.

[0094] Kriging interpolation is a widely used spatial interpolation technique that calculates the estimated concentration of components at unsampled points by solving the Kriging equations. Specifically, the Kriging equations are: ; in is the covariance matrix between sample points, is the concentration of the component to be estimated, is the known component concentration data. In this method, the solution of the equations is iteratively calculated using the conjugate gradient method to solve the problem of computational efficiency and memory consumption. The conjugate gradient method optimizes computational efficiency and performs well in processing large-scale data sets, especially in high-resolution spectral analysis, which can effectively reduce computing resource consumption and increase interpolation speed.

[0095] After the interpolation is completed, in order to eliminate the artifacts (such as edge effects) that may appear during the interpolation process, the result is processed by edge-preserving filtering. The kernel function size of the filter is dynamically adjusted according to the mean value of the spacing between the nearest neighbor sampling points, that is: ; in is the size of the kernel function, is the Euclidean distance between sample points.

[0096] Dynamically adjusting the size of the filter kernel function can effectively retain edge features, avoiding the over-smoothing problem that may be caused by the traditional fixed kernel function size, making the interpolation results more accurate and delicate, especially in the edge area.

[0097] Assume that in a fiber composition analysis, high-resolution data obtained by Fourier transform infrared spectroscopy is used and this method is used for spatial interpolation. In some fiber areas, due to the low density of spectral data acquisition, artifacts and discontinuities appear in the interpolation results.

[0098] In the specific implementation, the main direction is first obtained by calculating the direction of each fiber in the spectral data; then, the variation function model is constructed and the directional weight factor is introduced to further improve the directional accuracy of the interpolation. The solution of the Kriging equation quickly obtains the component estimation value of the unsampled point through the conjugate gradient method, and performs edge-preserving filtering on the result to eliminate artifacts. Finally, the interpolation result shows that the component distribution of the fiber area is smoother and consistent with the actual direction, and the interpolation accuracy is improved by 15% compared with the traditional method.

[0099] This method solves the interpolation error caused by uneven fiber orientation distribution through systematic steps and optimization, ensures the accuracy and efficiency of component concentration estimation, and is suitable for the actual needs of high-resolution spectral rapid analysis.

[0100] The present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention. In order to make the public have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention, but those skilled in the art can fully understand the present invention without the description of these details. In addition, in order to avoid unnecessary confusion about the essence of the present invention, well-known methods, processes, procedures, components and circuits are not described in detail.

[0101] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A high-resolution spectral rapid analysis method for fabric fiber components, characterized in that: The following steps are involved: S1: acquiring three-dimensional topographic data of the fabric surface by laser scanning, and identifying the fiber arrangement direction and surface curvature characteristics based on the three-dimensional topographic data; S2: According to the fiber arrangement direction and surface curvature characteristics identified by S1, the incident angle and focusing depth of the spectral probe are dynamically adjusted, and the corresponding spectral excitation mode is selected; S3: Extract features from the multi-mode spectral data collected by S2 and fuse the feature vectors of different spectral modes through a deep learning algorithm; S4: Analyze the fused feature vector based on the transfer learning model and output the fiber component type and its proportion; S5: Associate the component data output by S4 with the spatial coordinates to generate a component distribution visualization map.

2. The high-resolution spectral rapid analysis method of fabric fiber components according to claim 1 is characterized in that: The S1 specifically includes: S1-1: using laser emitters distributed in a circular array to project multi-angle interference fringes onto the surface of the fabric, and collecting deformation images of the interference fringes through a charge coupled device, wherein the wavelength of the laser emitters is selected from the visible light band; S1-2: performing a phase unwrapping operation on the deformed image, wherein the phase gradient field is solved by using a least squares algorithm, and the three-dimensional height coordinates are converted based on the laser wavelength and the spatial geometric parameters of the projection device; S1-3: Construct a neural network model based on deep separable convolution to process three-dimensional point cloud data. The encoding stage of the model uses a three-dimensional convolution kernel for feature extraction, and the decoding stage uses transposed convolution to achieve resolution recovery. The final output layer uses a normalized exponential function to calculate the probability value of each voxel point belonging to the direction of the main axis of the fiber. The normalized exponential function is a Softmax function, and its input is an unnormalized logarithmic probability vector; S1-4: Calculate the local curvature parameter based on the probability distribution diagram, wherein the curvature parameter is obtained by calculating the ratio of the second-order derivative of the height distribution function along the fiber direction to the sum of the squares of the first-order derivative.

3. The high-resolution spectral rapid analysis method of fabric fiber components according to claim 2 is characterized in that: The method for dynamically adjusting the spectrum probe in S2 is: S2-1: Calculate the best incident angle according to the fiber orientation data, wherein the calculation of the incident angle is based on Snell's law, wherein the surface normal deflection angle is derived from the three-dimensional morphology data of S1, and the fiber material refractive index is from a pre-stored database; S2-2: Establish a luminous flux feedback control system, monitor the reflected light intensity distribution through a photodiode array, and when the uniformity of the light intensity distribution does not reach a preset range, trigger the mechanical arm to adjust the probe pitch angle, and the adjustment amount of the probe pitch angle is calculated by a control algorithm including a proportional term, an integral term, and a differential term, wherein the proportional coefficient, the integral time constant, and the differential gain parameters are calibrated through a system identification experiment; S2-3: Select the spectral excitation mode according to the local curvature characteristics, start the wide-field infrared spectroscopy mode when the curvature is less than the set threshold, and the set threshold is determined according to the minimum focusing depth of the optical fiber probe; switch to the pulsed laser Raman mode when multiple layers of fiber overlap are detected, and the pulse width in the pulsed laser Raman mode is adaptively adjusted according to the interlayer distance; In highly wrinkled areas the confocal microscopy mode is activated, in which the confocal pinhole aperture is inversely proportional to the current focus depth.

4. The high-resolution spectral rapid analysis method of fabric fiber components according to claim 3 is characterized in that: The feature extraction method in S3 includes: S3-1: Performing continuous wavelet transform on the infrared spectrum, using Mexican hat wavelet basis function to extract absorption peak features at different scales, wherein the Mexican hat wavelet basis function is in the form of a second-order Gaussian derivative function, and its width parameter is dynamically adjusted according to the grating resolution of the spectrometer; S3-2: performing background suppression processing on the Raman spectrum, estimating the background signal by an iterative reweighting algorithm, wherein the reweighting coefficient in the iterative reweighting algorithm is dynamically updated according to the signal-to-noise ratio of each band of the spectrum, and the signal-to-noise ratio is calculated by a moving window method to obtain the ratio of the local mean to the standard deviation; S3-3: A gated attention mechanism is used to fuse different spectral features. The attention weight in the gated attention mechanism is obtained by performing a nonlinear transformation on the concatenated feature vector through a trainable parameter matrix, wherein the nonlinear transformation function is a Sigmoid function, and its output value range is limited between 0 and 1.

5. The high-resolution spectral rapid analysis method of fabric fiber components according to claim 4 is characterized in that: The transfer learning model construction method in S4 is: S4-1: In the pre-training stage, a capsule network architecture is used to establish the association between primary features and high-level semantics through a dynamic routing algorithm. The number of iterations of the dynamic routing is determined according to the number of dimensions of the feature vector. S4-2: The domain adaptation process uses a kernel function matching strategy to minimize the spectral distribution difference between the source domain and the target domain. The kernel function is a Gaussian kernel, and its bandwidth parameter is optimized by the maximum likelihood estimation method; S4-3: In the online learning phase, key parameters are identified and regularization constraints are imposed by calculating the square mean of parameter gradients. The calculation of the square mean of gradients uses a sliding window method to retain historical gradient information.

6. The high-resolution spectral rapid analysis method for fabric fiber components according to claim 5, characterized in that: The visualization method in S5 includes: S5-1: constructing a spatial coordinate system by fusing inertial sensor data and optical encoder readings through a Kalman filter, wherein the process noise covariance matrix of the Kalman filter is adaptively adjusted according to the motion speed of the probe; S5-2: The component distribution map is reconstructed using an anisotropic interpolation algorithm, wherein the variation function parameters of the interpolation algorithm are dynamically set according to the main direction angle of the fiber orientation distribution, wherein the range parameter along the fiber orientation is set to 1.5 to 2 times that of the vertical direction.

7. The high-resolution spectral rapid analysis method for fabric fiber components according to claim 3 is characterized in that: The specific rules for selecting the spectral excitation mode in S2-3 are: When the standard deviation of the deflection angle of the surface normal of the local area is less than a first threshold, activating the Fourier transform infrared spectroscopy mode, wherein the first threshold is determined according to the minimum incident angle tolerance of the spectral probe; When it is detected that the difference in the refractive index of materials at adjacent sampling points exceeds a second threshold, switching to the surface enhanced Raman spectroscopy mode, the second threshold being 1.2 times the standard deviation of the refractive index of each material in the standard fiber refractive index database; When the three-dimensional topography data indicates the existence of interlayer gaps, the time-gated fluorescence spectroscopy mode is enabled, and the time-gating width is proportional to the square root of the interlayer distance.

8. The high-resolution spectral rapid analysis method for fabric fiber components according to claim 4, characterized in that: The implementation process of the gated attention mechanism of S3-3 includes: splicing the infrared spectrum feature vector and the Raman spectrum feature vector along the channel dimension, and performing L2 norm normalization processing on the feature vector before splicing; Generate an initial attention score through a fully connected layer, where the weight matrix dimension of the fully connected layer is (M+N)×K, where M is the infrared feature dimension, N is the Raman feature dimension, and K is the number of neurons in the middle layer; Performing a nonlinear activation function transformation on the initial attention score to obtain a normalized weight coefficient, wherein the slope parameter of the activation function is dynamically adjusted according to the surface curvature value of the current sampling point; The two spectral features are weighted and summed according to the weight coefficient to generate a fusion feature vector, where the sum of the weight coefficient of the Raman feature and the weight coefficient of the infrared feature is always equal to 1.

9. The high-resolution spectral rapid analysis method for fabric fiber components according to claim 5, characterized in that: The domain adaptation process of S4-2 specifically includes: Calculating the maximum mean difference between source domain samples and target domain samples in a reproducing kernel Hilbert space, where the reproducing kernel Hilbert space is constructed by Gaussian kernel function mapping; Optimizing the feature transformation matrix by a gradient descent algorithm to minimize the difference between domains, wherein the learning rate of the gradient descent decays exponentially according to the current iteration number; An orthogonality constraint is introduced into the loss function to prevent the feature space from degenerating, and the strength coefficient of the orthogonality constraint is positively correlated with the domain difference metric.

10. The high-resolution spectral rapid analysis method for fabric fiber components according to claim 6, characterized in that: The anisotropic interpolation algorithm implementation process of S5-2 includes: Determining a main direction of spatial interpolation according to the fiber orientation distribution data, wherein the main direction is obtained by calculating a weighted average of local fiber orientation vectors; Constructing a variation function model including a directional weight factor, wherein the directional weight factor is in direct proportion to the cosine value of the main direction angle; Calculate the estimated value of the component concentration at the unsampled point by solving the Kriging equations, wherein the solution of the equations is iteratively calculated using the conjugate gradient method; An edge-preserving filtering process is performed on the interpolation result to eliminate spatial artifacts, and the kernel function size of the filter is dynamically adjusted according to the average value of the spacing between the nearest neighbor sampling points.

Images