A system and method for detecting the processing performance of a chemical fiber sampling

Abstract

The application discloses a kind of chemical fiber sampling processing performance detection system and method, belong to chemical fiber detection technical field, comprising: obtaining multi-angle microscopic image sequence and near infrared spectrum data sequence;Multi-angle microscopic image sequence is extracted and stereoscopic matching is generated fiber surface three-dimensional point cloud model to feature point;Near infrared spectrum data sequence is obtained by adaptive complete set empirical mode decomposition to obtain multiple intrinsic mode function component;The curvature change rate of each local region in three-dimensional point cloud model is calculated, and the instantaneous energy distribution entropy of each intrinsic mode function component is calculated;Curvature change rate and instantaneous energy distribution entropy are fused to construct fiber performance feature matrix by multiple scales;The multiple performance category centers are obtained by using artificial fish swarm algorithm with chaos initialization to optimize clustering center of fiber performance feature matrix;The performance category of the chemical fiber sample to be measured is determined according to the distance between each performance category center and the feature vector of the fiber to be measured.

Description

Technical Field

[0001] This invention relates to the field of chemical fiber testing technology, specifically to a chemical fiber sampling and processing performance testing system and method. Background Technology

[0002] The performance of chemical fibers is directly related to subsequent processing techniques and the quality of end products. Efficient and accurate sampling and performance testing are crucial aspects of production control. Currently, common fiber performance testing methods mainly rely on microscopic image analysis or near-infrared spectroscopy. Microscopic image analysis acquires two-dimensional images of the fiber surface and extracts morphological parameters such as contour and texture to assess the fiber's physical state. However, this method can only capture planar projection information of the fiber and cannot reflect the curvature fluctuations of the fiber surface's microscopic three-dimensional morphology, losing sensitivity to details such as tiny protrusions and depressions that affect performance. Near-infrared spectroscopy uses spectral absorption characteristics to deduce the chemical composition and structure of broken fibers. However, the raw spectral data generally suffers from baseline drift, scattering effects, and noise interference. Conventional wavelet transform or traditional empirical mode decomposition is prone to mode aliasing when processing such nonlinear and non-stationary signals, making it difficult to completely separate the noise-dominant component from the signal-dominant component carrying effective information. The extracted spectral energy characteristics lack stability, limiting the accuracy of fiber performance characterization. The two detection methods mentioned above often operate independently, providing only a single dimension of detection information. They lack the means to collaboratively analyze the physical morphology parameters of the fiber surface and the energy distribution information of its internal chemical composition. This makes it impossible for classification models built based on single features to comprehensively map the true performance of the fiber, resulting in significant limitations in detection sensitivity and accuracy. When classifying fiber performance based on extracted features, commonly used clustering methods such as the K-means algorithm are highly sensitive to the selection of initial cluster centers. Randomly set initial centers often lead to clustering results converging to local extrema, resulting in unclear category boundaries and making it difficult to guarantee the consistency of performance judgment for fiber samples from different batches or under varying process conditions. These limitations make existing methods insufficient in terms of sensitivity and accuracy when dealing with the core quality problem in chemical fiber production—fineness anomalies. There is an urgent need for a detection technology that can correlate the three-dimensional surface curvature distribution of fibers with dynamic changes in spectral energy and achieve globally stable optimization of cluster centers. Summary of the Invention

[0003] This invention aims to provide a method and system for detecting the performance of chemical fiber sampling and processing. It constructs a high-discrimination feature matrix by fusing the three-dimensional curvature change rate of the fiber surface and the instantaneous energy distribution entropy of the spectrum at multiple scales, and uses a chaotic initialization artificial fish swarm algorithm to find the global cluster center, so as to overcome the shortcomings of incomplete detection information in single modality and the tendency of conventional clustering methods to get trapped in local optima.

[0004] The objective of this invention can be achieved through the following technical solutions:

[0005] This invention provides a method for testing the performance of chemical fiber sampling, comprising: acquiring a multi-angle microscopic image sequence and a near-infrared spectral data sequence of a chemical fiber sample; extracting feature points and performing stereo matching on the multi-angle microscopic image sequence to generate a three-dimensional point cloud model of the fiber surface, thereby obtaining a high-precision three-dimensional characterization of the microscopic morphology of the fiber surface; performing adaptive complete ensemble empirical mode decomposition on the near-infrared spectral data sequence to obtain multiple intrinsic mode function components, achieving adaptive separation of noise components and effective signal components in the spectral signal; calculating the rate of curvature change of each local region in the three-dimensional point cloud model to quantify the degree of microscopic undulation of the fiber surface, and simultaneously calculating each intrinsic mode function component. The instantaneous energy distribution entropy of the modal function components characterizes the complexity and differences in the energy distribution of chemical components within the fiber. The curvature change rate and the instantaneous energy distribution entropy are fused at multiple scales to construct a fiber performance feature matrix, enabling complementary integration of surface structural features and internal spectral features within a unified feature space. A chaotic-initialized artificial fish swarm algorithm is used to optimize the clustering centers of the fiber performance feature matrix, obtaining multiple performance category centers, thereby improving the global convergence ability and accuracy of the clustering center search. Based on the distance between each performance category center and the feature vector of the fiber to be tested, the performance category of the chemical fiber sample is determined, completing the objective classification of fiber performance.

[0006] As a preferred technical solution of the present invention, the process of extracting feature points and performing stereo matching on the multi-angle microscopic image sequence to generate a three-dimensional point cloud model of the fiber surface includes: selecting a first-angle image and a second-angle image from the multi-angle microscopic image sequence to form a binocular image pair; using a fast and robust feature algorithm to detect candidate feature points in the binocular image pair to generate a first feature point set and a second feature point set; and selecting matching feature point pairs based on the angle difference between the principal direction of the local gradient of each feature point in the first feature point set and the principal direction of the local gradient of the corresponding feature point in the second feature point set, preferably by calculating... The absolute angle difference between the first and second local gradient principal directions is calculated. Combinations with an absolute angle difference less than a preset angle threshold are marked as candidate matching pairs. The optimal matching feature point pair is further selected based on the Hamming distance between feature descriptors. The spatial three-dimensional coordinates corresponding to each matching feature point are calculated based on the pixel coordinate difference between the matching feature point pairs in the first and second angle images. All spatial three-dimensional coordinates corresponding to the matching feature points are aggregated into an initial point cloud. Outlier removal and hole filling are then performed on the initial point cloud to obtain a high-completeness and high-precision three-dimensional point cloud model of the fiber surface. Preferably, outlier removal of the initial point cloud is performed using statistical filtering, and hole filling is performed using radial basis function interpolation. This allows the point cloud model to recover missing microstructures while removing noise outliers, ensuring the reliability of subsequent curvature calculations.

[0007] As a further preferred technical solution of the present invention, the specific process of performing adaptive complete ensemble empirical mode decomposition on the near-infrared spectral data sequence includes: acquiring the original near-infrared spectral data sequence; adding a first white noise sequence and a second white noise sequence to the original near-infrared spectral data sequence, wherein the first white noise sequence and the second white noise sequence are opposites of each other; performing empirical mode decomposition on the spectral data sequence after adding noise to obtain a first set of intrinsic mode function components and a second set of intrinsic mode function components; performing an overall average calculation on the first set of intrinsic mode function components and the second set of intrinsic mode function components to obtain multiple intrinsic mode function components under adaptive complete ensemble empirical mode decomposition; arranging the multiple intrinsic mode function components in descending order of frequency, extracting the first three high-frequency intrinsic mode function components as noise-dominant components, and the remaining intrinsic mode function components as signal-dominant components. By introducing white noise pairs with opposite signs, mode aliasing can be effectively suppressed, making the extracted signal-dominant components more accurately reflect the absorption characteristics caused by the chemical composition of the fiber.

[0008] When calculating the instantaneous energy distribution entropy of each intrinsic mode function component, it is preferable to perform a Hilbert transform on the intrinsic mode function component of each dominant signal component to obtain the analytic signal of each intrinsic mode function component; extract the instantaneous amplitude sequence and instantaneous phase sequence from the analytic signal; calculate the instantaneous energy value corresponding to each sampling point based on the instantaneous amplitude sequence to obtain the instantaneous energy sequence; divide the instantaneous energy sequence into multiple equally spaced energy intervals, and statistically analyze the probability of occurrence of the instantaneous energy value in each energy interval; calculate the Shannon entropy based on each probability of occurrence to obtain the instantaneous energy distribution entropy of each intrinsic mode function component. The instantaneous energy distribution entropy obtained in this way can sensitively reflect the changes in the energy distribution uniformity caused by differences in the internal chemical structure of the fiber.

[0009] As another preferred technical solution of the present invention, the specific process for calculating the curvature change rate of each local region in the three-dimensional point cloud model includes: voxel mesh downsampling of the three-dimensional point cloud model of the fiber surface to obtain a uniformly distributed resampled point cloud; fitting the local surface of each point and its neighboring points in the resampled point cloud using the moving least squares method to obtain the local fitted surface of each point; calculating the Gaussian curvature and mean curvature of each point on the local fitted surface; calculating the principal curvature of each point based on the Gaussian curvature and the mean curvature, and calculating the ratio of the absolute value of the principal curvature difference between adjacent points in the same local region to the radius of the local region to obtain the curvature change rate of the local region. Voxel mesh downsampling ensures the uniformity of the point cloud density, and moving least squares surface fitting improves the smoothness and continuity of the local surface estimation, enabling the curvature change rate to accurately characterize the abrupt changes in the micromorphology and roughness distribution of the fiber surface.

[0010] In the technical solution of this invention, the step of multi-scale fusion of the curvature change rate and the instantaneous energy distribution entropy to construct a fiber performance feature matrix includes: arranging the curvature change rates corresponding to the same chemical fiber sample into a one-dimensional curvature feature vector according to their spatial positions; arranging the instantaneous energy distribution entropy corresponding to the same chemical fiber sample into a one-dimensional entropy feature vector according to the order of the intrinsic mode function components; performing max-min normalization on the one-dimensional curvature feature vector to obtain a normalized curvature vector; performing max-min normalization on the one-dimensional entropy feature vector to obtain a normalized entropy vector; and horizontally concatenating the normalized curvature vector and the normalized entropy vector to form a two-dimensional fiber performance feature matrix, wherein the number of rows in the two-dimensional fiber performance feature matrix is ​​the number of feature dimensions, and the number of columns is the number of samples. Through normalization and concatenation, the differences in dimensionality and numerical scale between curvature features and entropy features are eliminated, enabling multi-source features to collaboratively express the comprehensive performance attributes of the fiber at a unified scale.

[0011] When using a chaotic initialization-based artificial fish swarm algorithm to optimize the cluster centers of the fiber performance feature matrix, the population size, perception distance, movement step size, and crowding factor of the artificial fish swarm are set, along with the number of cluster categories. Initial artificial fish swarm positions are generated using logistic chaotic mapping, with each position corresponding to a set of candidate cluster center matrices. The current fitness value of each artificial fish is calculated, where fitness is the sum of squared distances from each sample in the fiber performance feature matrix to the nearest candidate cluster center. Foraging, swarming, and tail-chasing behaviors are performed on each artificial fish to update its position. Each behavior is iteratively executed until the maximum number of iterations is reached. The position of the artificial fish with the smallest fitness value is taken as the optimal cluster center matrix, and multiple performance category centers are extracted from this optimal matrix. Chaotic initialization enhances the diversity and ergodicity of the initial population. Combined with the collaborative search of multiple behaviors, it effectively avoids cluster centers getting trapped in local optima, allowing the final performance category centers to more realistically reflect the inherent distribution patterns of different performance fibers. Preferably, in behavior updates, the artificial fish randomly selects a trial direction and generates a trial position within its current perception distance. If the fitness value of the trial position is less than the fitness value of the current artificial fish position, it moves one step towards the trial position to complete the foraging behavior; otherwise, it repeats the attempt until the maximum number of attempts is reached, after which it moves randomly. In swarming behavior, the number of companions and the center position within the perception distance are obtained. If the number of companions is not zero and the fitness value of the center position multiplied by the crowding factor is less than the fitness value of the current artificial fish, it moves towards the center position. In tail-chasing behavior, the optimal companion with the smallest fitness value within the perception distance is obtained. If the fitness value of the optimal companion multiplied by the crowding factor is less than the fitness value of the current artificial fish, it moves towards the optimal companion.

[0012] When determining the performance category of a fiber sample, multi-angle microscopic image sequences and near-infrared spectral data sequences of the chemical fiber sample are acquired. The same feature extraction and fusion process as in the training phase is used to generate a feature vector for the fiber. The Euclidean distance between the feature vector and each performance category center is calculated. The performance category center corresponding to the minimum value of all Euclidean distances is taken as the target category center. The performance category represented by the target category center is determined as the performance category of the chemical fiber sample, and the performance determination result is output. This discrimination method is computationally simple, responds quickly, and maintains consistency with the feature space in the training phase, ensuring the stability and reliability of the classification results.

[0013] This invention also includes a chemical fiber sampling and processing performance testing system. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the aforementioned chemical fiber sampling and processing performance testing method. This system organically integrates three-dimensional reconstruction of microscopic images, adaptive near-infrared spectral decomposition, multimodal feature fusion, and chaotic intelligent clustering, enabling automated performance testing and classification of fiber samples, significantly improving detection efficiency and classification accuracy.

[0014] The beneficial effects of this invention are:

[0015] The curvature change rate of a local region on a 3D point cloud model of the fiber surface is calculated. Gaussian curvature and mean curvature are obtained based on voxel downsampling and moving least squares local surface fitting, leading to the curvature change rate characterized by the ratio of principal curvature differences. This feature finely depicts the intensity and spatial distribution of microscopic undulations on the fiber surface, and is more sensitive to capturing morphological defects and structural anomalies compared to traditional 2D texture features. Simultaneously, a Hilbert transform is performed on the dominant signal components in the intrinsic mode function components obtained from adaptive complete ensemble empirical mode decomposition to obtain instantaneous amplitude sequences and construct instantaneous energy sequences. The instantaneous energy distribution entropy is calculated using the statistical energy interval probability distribution. This entropy value describes the dynamic disorder of the vibrational energy of fiber chemical groups over time, avoiding the loss of time-varying information caused by using only static indicators such as average energy or peak value. The normalized curvature change rate vector and the instantaneous energy distribution entropy vector are horizontally concatenated to form a 2D fiber performance feature matrix that simultaneously carries complementary information from both spatial morphology and spectral energy dimensions. This multi-scale fusion method preserves the integrity of each modal feature, allowing fiber samples to still show differences in the fusion matrix even when their single-dimensional features are similar, significantly enhancing the discriminative power of features for minor fluctuations in fiber performance.

[0016] Logistic chaotic mapping is used to initialize the artificial fish swarm, ensuring that the initial candidate cluster centers are ergodic and non-repetitive, thus overcoming the shortcomings of random initialization which can lead to individuals clustering in local regions of the search space. During the optimization process, the artificial fish individuals cooperate in the search through foraging, swarming, and tail-chasing behaviors, using the sum of squared intra-cluster distances as the fitness function. The chaotic initial distribution combined with the parallel trial-and-error mechanism of the artificial fish swarm algorithm effectively avoids the cluster center search getting trapped in local maxima, ultimately converging to the globally optimal cluster center matrix. The obtained performance category centers have low dependence on initial conditions, strong consistency among category centers obtained in different batches, and a certain degree of adaptability to sample feature shifts caused by process fluctuations, ensuring the stability and reliability of the fiber performance category determination. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart of the chemical fiber sampling and processing performance testing method;

[0019] Figure 2 This is a flowchart of the process for generating a 3D point cloud model of a fiber surface;

[0020] Figure 3 This is a flowchart of the instantaneous energy distribution entropy calculation based on adaptive complete set empirical mode decomposition;

[0021] Figure 4 This is a flowchart illustrating the process of fusing fiber surface point cloud features with acoustic emission entropy features to generate a feature matrix.

[0022] Figure 5 This is a flowchart of fiber performance clustering center optimization based on the chaotic initialization artificial fish swarm algorithm;

[0023] Figure 6 This is a flowchart for determining fiber performance categories. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] See Figure 1This invention provides a method for testing the performance of chemical fiber samples, comprising: acquiring multi-angle microscopic image sequences and near-infrared spectral data sequences of chemical fiber samples; extracting feature points and performing stereo matching on the multi-angle microscopic image sequences to generate a three-dimensional point cloud model of the fiber surface; performing adaptive complete ensemble empirical mode decomposition on the near-infrared spectral data sequences to obtain multiple intrinsic mode function components; calculating the rate of curvature change of each local region in the three-dimensional point cloud model and calculating the instantaneous energy distribution entropy of each intrinsic mode function component; fusing the rate of curvature change and the instantaneous energy distribution entropy at multiple scales to construct a fiber performance feature matrix; using a chaotic initialization artificial fish swarm algorithm to optimize the clustering centers of the fiber performance feature matrix to obtain multiple performance category centers; and determining the performance category of the chemical fiber sample to be tested based on the distance between each performance category center and the feature vector of the fiber to be tested.

[0026] In practice, the process of extracting feature points and performing stereo matching on multi-angle microscopic image sequences to generate a three-dimensional point cloud model of the fiber surface is as follows.

[0027] See Figure 2 A first-angle image and a second-angle image are selected from a multi-angle microscopic image sequence of chemical fiber samples to form a binocular image pair. The first-angle image and the second-angle image are obtained by the microscopic image acquisition device by taking pictures of the same fiber region at different angles. There is a preset angle between the shooting optical axis corresponding to the first-angle image and the shooting optical axis corresponding to the second-angle image. The preset angle is set according to the baseline distance and working distance of the microscopic image acquisition device to ensure that there is a sufficient overlapping field of view between the binocular image pair.

[0028] A fast and robust feature detection algorithm is used to detect candidate feature points in a pair of stereo images, generating a first feature point set and a second feature point set. In some embodiments, the fast and robust feature detection algorithm for candidate feature points is implemented as follows: integral images are constructed for the first angle image and the second angle image respectively; the determinant response of the Hessian matrix is ​​calculated on the integral image using box filter templates of different sizes to form a scale space; non-maximum suppression is performed in the scale space, and points with response values ​​greater than other extreme points in the neighborhood are selected as candidate feature points, thus obtaining the first feature point set and the second feature point set. Each first feature point in the first feature point set corresponds to a candidate feature point position in the first angle image, and each second feature point in the second feature point set corresponds to a candidate feature point position in the second angle image.

[0029] The first local gradient principal direction is calculated for each first feature point in the first feature point set, and the second local gradient principal direction is calculated for each second feature point in the second feature point set. In specific implementation, the Haar wavelet response of all pixels is calculated within a circular neighborhood with a radius of 6 times the feature scale of the first feature point, centered on the first feature point. A fan-shaped sliding window with an angle of 60 degrees is rotated around the first feature point, and the Haar wavelet response values ​​within the window are accumulated. The longest vector direction is selected as the first local gradient principal direction. The second local gradient principal direction is calculated in the same way for each second feature point in the second feature point set.

[0030] The algorithm iterates through all combinations of the first and second feature points, calculating the absolute angle difference between the principal directions of the first and second local gradients for each combination. The absolute angle difference is expressed in radians. Combinations with an absolute angle difference less than a preset angle threshold are marked as candidate matching pairs. The preset angle threshold is set to... The preset angle threshold is set to... The basis for this is that the surface texture of fibers has local directional consistency, and the gradient principal direction deviation of correctly matched feature points is usually within a certain small angle range. The corresponding 10-degree angle can filter out most of the mismatches caused by excessive directional deviation due to local distortion or texture repetition on the fiber surface, while retaining enough correct matching feature point pairs.

[0031] In each candidate matching pair, the Hamming distance between the feature descriptors of the first feature point and the second feature point is calculated, and the candidate matching pair with the smallest Hamming distance is selected as the matching feature point pair. In specific implementation, the first feature point corresponds to a 64-dimensional feature descriptor generated by a fast robust feature algorithm, and the second feature point corresponds to a 64-dimensional feature descriptor. The Hamming distance is calculated as follows: for each element in the 64-dimensional feature descriptor, a binarization threshold is applied, set to the median of all element values ​​in the 64-dimensional feature descriptor corresponding to the first feature point; element values ​​greater than the binarization threshold are set to 1, and element values ​​less than or equal to the binarization threshold are set to 0, generating the first binary descriptor; for the 64-dimensional feature descriptor of the second feature point, the median of all element values ​​in the 64-dimensional feature descriptor corresponding to the second feature point is used as the binarization threshold, and the same operation is performed to generate the second binary descriptor. The number of corresponding bits that differ between the first and second binary descriptors is calculated to obtain the Hamming distance. Among all candidate matching pairs, the candidate matching pair with the smallest Hamming distance is determined as the matching feature point pair.

[0032] Based on the pixel coordinate difference between the matched feature point pairs in the first and second angle images, the spatial three-dimensional coordinates corresponding to each matched feature point are calculated. The spatial three-dimensional coordinates are calculated using triangulation formulas:

[0033]

[0034] Where b represents the baseline distance between the optical centers of the first and second angle images in the binocular image pair, f represents the focal length of the microscopic image acquisition device, u1 and v1 represent the horizontal and vertical pixel coordinates of the matched feature point in the first angle image, u0 and v0 represent the horizontal and vertical pixel coordinates of the principal point in the first angle image, and d represents the disparity value between the matched feature point pair in the first and second angle images. u2 represents the horizontal pixel coordinate of the matched feature point in the second-angle image. To match the spatial three-dimensional coordinates corresponding to the feature points, where x, y, and z represent the horizontal, vertical, and depth coordinates in the coordinate system of the microscopic image acquisition device, respectively.

[0035] The spatial 3D coordinates corresponding to all matching feature points are aggregated into an initial point cloud. Each point in the initial point cloud corresponds to a spatial 3D coordinate calculated from a pair of matching feature points. Outlier removal and hole filling are performed on the initial point cloud to obtain a 3D point cloud model of the fiber surface.

[0036] Outlier removal employs a statistical filtering method. The implementation process is as follows: For each spatial 3D coordinate point in the initial point cloud, calculate the average distance from that point to its k nearest neighbors. The value of k ranges from 10 to 50, and can be set to 30. Calculate the mean μ and standard deviation σ of the average distances corresponding to all spatial 3D coordinate points. Set a distance threshold. α is the standard deviation factor, ranging from 1.0 to 3.0, and can be set to 1.5. Setting α to 1.5 is to improve the sensitivity of outlier identification while avoiding the rejection of valid points belonging to the microstructural undulations of the fiber surface. Spatial three-dimensional coordinate points with an average distance greater than a distance threshold are identified as outliers and removed, resulting in a filtered point cloud.

[0037] Hole filling employs radial basis function interpolation. The process of performing radial basis function interpolation on hole regions in the filtered point cloud involves: extracting the set of hole boundary points on the boundary of the hole region. This set contains m boundary points, and the spatial coordinates of the j-th boundary point are denoted as p. j A radial basis function is constructed to fit the surface of the hole region. The radial basis function has the following form: Where p is the spatial coordinate of the point to be interpolated within the hole region. For thin plate spline basis functions, w j Let be the weight coefficients to be solved. Establish a system of linear equations on the boundary point set of the hole, and let . Given the known depth or height value of the j-th boundary point in the set of hole boundary points, we can solve for all w values.j Within the cavity area, interpolation grid points are generated at a preset grid spacing. The spatial coordinates of the interpolation grid points are substituted into the solved radial basis function to calculate the interpolated three-dimensional coordinates. The interpolated three-dimensional coordinate points are added to the point cloud to complete the cavity filling and finally generate a complete three-dimensional point cloud model of the fiber surface.

[0038] In specific implementation, please refer to Figure 3 After obtaining the original near-infrared spectral data sequence, the process of adaptive complete set empirical mode decomposition of the original near-infrared spectral data sequence is as follows.

[0039] The original near-infrared spectral data sequence was obtained by performing near-infrared scanning on chemical fiber samples at continuous time points using a spectral acquisition device. The length of the original near-infrared spectral data sequence is denoted as L, and each sampling point corresponds to a spectral intensity value. A first white noise sequence and a second white noise sequence are added to the original near-infrared spectral data sequence; the first and second white noise sequences are opposites of each other. The first white noise sequence is generated using a Gaussian random number generator with a mean of 0 and a standard deviation of σ. n The white noise sequence has the same length L as the original near-infrared spectral data sequence, σ n Based on the standard deviation σ of the original near-infrared spectral data sequence s Configure it. σ n The rationale for setting the standard deviation of the original near-infrared spectral data sequence to 0.2 times is as follows: In adaptive complete ensemble empirical mode decomposition, the amplitude of added noise needs to balance the mode aliasing reduction effect and the residual noise level. When the noise amplitude is 0.2 times the signal standard deviation, it can effectively reduce mode aliasing in high-frequency mode components, while ensuring that the residual noise component after subsequent overall averaging is below an acceptable threshold. The second white noise sequence is directly obtained by inverting each value of the first white noise sequence; that is, the t-th value in the second white noise sequence is the inverse of the t-th value in the first white noise sequence. The first white noise sequence is added point-by-point to the original near-infrared spectral data sequence to obtain the first noisy spectral data sequence; the second white noise sequence is added point-by-point to the original near-infrared spectral data sequence to obtain the second noisy spectral data sequence.

[0040] Empirical mode decomposition (EMD) is performed on the first noisy spectral data sequence to obtain the first set of intrinsic mode function (EMF) components. EMD is also performed on the second noisy spectral data sequence to obtain the second set of EMF components. The EMD process is as follows: For the input noisy spectral data sequence, all local maxima and local minima are identified; cubic spline interpolation is used to fit all local maxima and local minima to construct upper and lower envelopes; the mean curves of the upper and lower envelopes are calculated, and the mean curves are subtracted from the noisy spectral data sequence to obtain the intermediate signal; the intermediate signal is then checked to see if it satisfies the EMF conditions. The EMF conditions include that the number of local extrema and the number of zero-crossings in the sequence are equal or differ by at most 1, and that at any point... The mean of the upper and lower envelopes is 0. If the intermediate signal does not satisfy the intrinsic mode function (IMF) condition, the above screening process is repeated with the intermediate signal as a new input sequence. If the intermediate signal satisfies the IMF condition, the intermediate signal is taken as an IMF component, and the IMF component is subtracted from the noisy spectral data sequence to obtain the residual sequence. The above process is repeated with the residual sequence as a new input sequence until the residual sequence becomes a monotonic function or a constant, and finally a set of IMF components is obtained. The number of IMF components is determined by the characteristics of the noisy spectral data sequence itself.

[0041] The overall average of the first and second groups of intrinsic mode function (IMF) components is calculated to obtain multiple IMF components under the adaptive complete set empirical mode decomposition. The specific method for overall averaging is as follows: the number of components in the first group and the number of components in the second group are determined respectively, and the smaller value between the two groups is taken as the final number of IMF components K; for the k-th order, The k-th eigenmode function component in the first group of eigenmode function components is added point-by-point to the k-th eigenmode function component in the second group of eigenmode function components, and then divided by 2 to obtain the averaged k-th eigenmode function component. The K eigenmode function components obtained by overall averaging constitute the final result of the adaptive complete ensemble empirical mode decomposition.

[0042] Multiple intrinsic mode function (IMF) components are arranged in descending order of frequency. The frequency of each IMF component is determined by calculating the number of zero-crossings: zero-crossing detection is performed on the IMF components, and the number of times the sign of the IMF component sequence value changes is counted. IMF components with more zero-crossings are considered to have higher frequencies. When the number of zero-crossings is the same, signal energy is used as an auxiliary criterion, with higher energy being ranked earlier. K IMF components are sorted in descending order of the number of zero-crossings to obtain an IMF component sequence arranged from high frequency to low frequency.

[0043] From the sorted intrinsic mode function (IMF) component sequence, the first three high-frequency IMF components are extracted as noise-dominant components. The rationale for extracting the first three high-frequency IMF components as noise-dominant components is as follows: In near-infrared spectroscopy measurements, random noise and measurement disturbances are concentrated in the high-frequency band of the signal. The first three components of the IMF component sequence cover the highest frequency band of the original spectral signal. The signal energy in this band is mainly contributed by noise. Spectral analysis verifies that the cumulative noise energy of the first three high-frequency IMF components accounts for more than the vast majority of the total noise energy; therefore, the first three high-frequency IMF components are marked as noise-dominant components. In the sorted IMF component sequence, all remaining IMF components other than the first three noise-dominant components are considered as signal-dominant components. If the total number of IMF components... If the value is less than or equal to 3, all intrinsic mode function components are considered noise-dominant components and no signal-dominant components are generated. In this case, the original near-infrared spectral data sequence is considered to be heavily contaminated with noise, and no subsequent instantaneous energy distribution entropy calculation is performed. The spectral data is then reacquired.

[0044] For each dominant signal component's eigenmode function (EMF) component, a Hilbert transform is performed. The Hilbert transform is implemented as follows: a Fourier transform is performed on the EMF components of the dominant signal component to obtain their spectra; in the frequency domain, the positive frequency portion of the EMF component's spectrum is multiplied by the imaginary unit. The negative frequency part multiplied by the imaginary unit Perform an inverse Fourier transform to obtain the Hilbert transform of the intrinsic mode function (IMF) components. Construct an analytic signal for the IMF components using the IMF components and their Hilbert transforms. The real part of the analytic signal represents the IMF components themselves, and the imaginary part represents their Hilbert transforms.

[0045] Extract the instantaneous amplitude sequence and instantaneous phase sequence from the analytic signal. For the t-th sampling point in the analytic signal, The instantaneous amplitude is equal to the square root of the sum of the squares of the real and imaginary parts of the analytic signal, i.e. ,in, This represents the t-th sampled value of the eigenmode function component of the dominant component of the signal. This represents the t-th sampled value of the Hilbert transform of the intrinsic mode function components. The t-th value of the instantaneous phase sequence is... .

[0046] The instantaneous energy value corresponding to each sampling point is calculated based on the instantaneous amplitude sequence, resulting in the instantaneous energy sequence. The t-th instantaneous energy value of the instantaneous energy sequence is then calculated. Equal to the t-th value of the instantaneous amplitude sequence The square of, that is For each dominant signal component, an instantaneous energy sequence of length L is obtained for its eigenmode function component.

[0047] The instantaneous energy sequence is divided into multiple equally spaced energy intervals. The division method is as follows: calculate the minimum value E of all instantaneous energy values ​​in the instantaneous energy sequence. min and maximum value E max The number of energy intervals is set to M, and the value of M is set to 10. The rationale for setting M to 10 is to strike a balance between the resolution of the energy distribution and the statistical stability of the estimation. An excessively large M would result in too few sample points in some intervals, leading to an increase in the variance of the probability estimation; an excessively small M would obscure the detailed features of the energy distribution. The width of the equally spaced energy intervals is... The range of the m-th energy interval is ,in, The Mth energy range is , including the upper bound.

[0048] Calculate the probability of an instantaneous energy value occurring within each energy interval. For the m-th energy interval, count the number c of instantaneous energy values ​​in the instantaneous energy sequence that fall within the range of the m-th energy interval. m Calculate the probability of the m-th energy range occurring. .like ,but In subsequent Shannon entropy calculations, the following definition is provided. .

[0049] The Shannon entropy is calculated based on the probability of occurrence, yielding the instantaneous energy distribution entropy for each eigenmode function component. The formula for calculating the instantaneous energy distribution entropy is:

[0050]

[0051] Among them, H ent The instantaneous energy distribution entropy of the eigenmode function components of a dominant signal component is represented by M, where M represents the number of energy intervals, and p m This represents the probability of an instantaneous energy value occurring within the m-th energy interval. This represents a logarithmic operation to the base 2. Applying the above formula to the eigenmode function components of each dominant signal component yields the corresponding number of instantaneous energy distribution entropies. All instantaneous energy distribution entropies form a set of instantaneous energy distribution entropies.

[0052] In specific implementation, please refer to Figure 4 The process of downsampling a three-dimensional point cloud model of the fiber surface using a voxel mesh to obtain a uniformly distributed resampled point cloud is as follows.

[0053] The voxel grid edge length for voxel mesh downsampling is set according to the point cloud density of the 3D point cloud model of the fiber surface. The voxel grid edge length is set to an integer multiple of the median of the average distance between all points and their nearest neighbors in the 3D point cloud model of the fiber surface. Setting the voxel grid edge length to 5 times the median of the average distance between all points and their nearest neighbors in the 3D point cloud model of the fiber surface is based on the following reasoning: when the voxel grid edge length is set to 5 times the median of the average nearest neighbor distance, the number of points in the point cloud can be compressed to about 10% to 30% of the original number of points while preserving the 3D morphological features of the fiber surface. This reduces the computational load of the subsequent moving least squares fitting and avoids distortion of local surface fitting due to excessively sparse point clouds.

[0054] The 3D space containing the fiber surface 3D point cloud model is divided into non-overlapping cubic voxel grids according to the voxel grid's side length. For each non-empty cubic voxel grid, the mean coordinates of all spatial points falling within the cubic voxel grid are calculated to obtain a mean coordinate point, which is used as a point in the resampled point cloud. The mean coordinates of all non-empty cubic voxel grids constitute the resampled point cloud, and the number of points in the resampled point cloud is denoted as M. rs .

[0055] The moving least squares method is used to fit the local surface of each point and its neighborhood points in the resampled point cloud, resulting in the local fitted surface corresponding to each point in the resampled point cloud. For any point p in the resampled point cloud... i , The method for obtaining neighboring points is as follows: using point p i Using the spatial coordinates as the center, define a spherical search region with a search radius R. mls Set to 3 times the side length of the voxel grid. All points in the resampled point cloud falling within the spherical search region are taken as point p. i The number of neighborhood points is denoted as N. i .

[0056] The basis function vectors for the moving least squares method are chosen as quadratic polynomial basis function vectors, which are: Where u and v represent points p i The planar coordinate components in a local coordinate system with the origin as the origin, the local coordinate system is formed by the point p i The direction of the normal vector is taken as the direction of the local coordinate system w-axis and orthogonalized. The weight function of the moving least squares method is a cubic spline weight function for point p. i The j-th neighboring point in the neighborhood of the given points. The weight values ​​w of the cubic spline weight function j The calculation method is as follows: Let d j This represents the distance from the j-th neighboring point to point p. i The Euclidean distance, let ,when hour, ;when hour, At point p i Within the local coordinate system, using the local coordinates and weight values ​​w of neighboring points j By fitting the coefficients of the quadratic polynomial using the weighted least squares method, the point p is obtained. i Local fitting surface S i .

[0057] Calculate the Gaussian curvature and mean curvature of each point in the resampled point cloud on the corresponding locally fitted surface. Obtain the locally fitted surface S. i The first fundamental form coefficient E i F i G i Second fundamental form coefficient L i M i N i Gaussian curvature K i The calculation method is as follows Mean curvature H i The calculation method is as follows According to the Gaussian curvature K i and mean curvature H i Calculate point p i First principal curvature k 1,i Second principal curvature k 2,i The calculation formula is: , .

[0058] The rate of change of curvature of the local region is obtained by calculating the ratio of the absolute value of the principal curvature difference between adjacent points within the same local region to the radius of the local region. The radius of the local region is r. local Set to 3 times the side length of the voxel mesh, based on the radius r of the local region. local When set to three times the side length of the voxel mesh, the average number of points contained within the local spherical neighborhood is between 20 and 50, which meets the sample size required for curvature variation statistics and can effectively capture the spatial variation characteristics of microscopic undulations on the fiber surface. For any point p in the resampled point cloud... i , with point p i Centered on, r local The set of neighborhood points within a spherical neighborhood of radius Ω is denoted as Ω. i set Ω i The number of neighboring points included is denoted as Point p i The corresponding rate of change of curvature in the local region Defined as:

[0059]

[0060] in, p represents the i-th point in the resampled point cloud. i The rate of change of curvature in the corresponding local region; Point p i spherical neighborhood Ω i The number of neighborhood points within Ω; i Indicates point p i Centered on a radius of r local spherical neighborhood; q j Represents the spherical neighborhood Ω i The j-th neighboring point within; κ1,i and κ 2,i They represent point p respectively i The first principal curvature and the second principal curvature; and Representing the neighborhood point q respectively j First principal curvature and second principal curvature; r local This represents the radius of the local region. The rate of change of curvature is calculated for all points in the resampled point cloud, yielding M. rs The curvature change rate value.

[0061] The curvature change rates corresponding to the same chemical fiber samples are arranged into a one-dimensional curvature feature vector according to their spatial location. The arrangement method is as follows: all points in the resampled point cloud are sorted according to the lexicographical order of their spatial coordinates. First, the x-coordinate values ​​are compared, with the point with the smaller x-coordinate value placed first; if the x-coordinate values ​​are the same, the y-coordinate values ​​are compared, with the point with the smaller y-coordinate value placed first; if the y-coordinate values ​​are the same, the z-coordinate values ​​are compared, with the point with the smaller z-coordinate value placed first. The curvature change rates of each point are then calculated according to the sorted point order. Arranged sequentially, forming a structure of length M rs The one-dimensional curvature eigenvector has a dimension of M. rs .

[0062] The instantaneous energy distribution entropy corresponding to the same chemical fiber sample is arranged into a one-dimensional entropy feature vector according to the order of the intrinsic mode function components. The instantaneous energy distribution entropy of all intrinsic mode function components corresponding to the dominant signal component of the chemical fiber sample is obtained, and the number of intrinsic mode function components of the dominant signal component is denoted as K. s According to the order of the intrinsic mode function components from high frequency to low frequency in the adaptive complete set empirical mode decomposition, the corresponding instantaneous energy distribution entropy is arranged sequentially to form a sequence of length K. s The one-dimensional entropy eigenvector has K dimensions. s .

[0063] The one-dimensional curvature eigenvector is subjected to min-max normalization to obtain a normalized curvature vector. The min-max normalization process involves obtaining the maximum value of all elements in the one-dimensional curvature eigenvector. and minimum value For each element value v1 in the one-dimensional curvature eigenvector, Normalized value The calculation method is as follows All normalized values ​​form a normalized curvature vector, and the range of values ​​for each element in the normalized curvature vector is... .

[0064] The one-dimensional entropy eigenvector is subjected to min-max normalization to obtain a normalized entropy vector. The maximum value of all elements in the one-dimensional entropy eigenvector is then obtained. and minimum value For each element value u in the one-dimensional entropy eigenvector m , Normalized value The calculation method is as follows All normalized values ​​form a normalized entropy vector, and the value range of each element in the normalized entropy vector is [0,1].

[0065] The normalized curvature vector and the normalized entropy vector are concatenated laterally to form a two-dimensional fiber performance feature matrix. The concatenation method is as follows: A matrix with dimension M is... rs The normalized curvature vector and its dimension are K. s The normalized entropy vectors are concatenated end-to-end along the feature dimension to form a vector of length M. rs +K s The combined feature column vector, the first row to the Mth row of the combined feature column vector. rs The row corresponds to each element of the normalized curvature vector, the Mth row. rs +1 row to the Mth row rs +K s Each row corresponds to an element of the normalized entropy vector. The combined feature column vector serves as a column of the two-dimensional fiber performance feature matrix, which has M rows. rs +K s , representing the total number of dimensions of fiber performance characteristics; the number of columns in the two-dimensional fiber performance characteristic matrix is ​​the number of samples, and for the current single chemical fiber sample, the number of columns is 1.

[0066] In specific implementation, please refer to Figure 5 The process of using the chaotic initialization artificial fish swarm algorithm to optimize the clustering centers of the fiber performance feature matrix and obtain multiple performance category centers is as follows.

[0067] Define the population size, sensing distance, walking stride, and crowding factor for the artificial fish swarm, and also define the number of clusters. Population size N fish The population size is set to 30 because a population size of 30 allows the artificial fish swarm to fully explore the search space with sufficient individual distribution, while keeping the computational cost per iteration within a reasonable range. Perception distance V visual The value is set to 0.2 times the range of all feature dimensions for each sample in the fiber performance feature matrix. The fiber performance feature matrix consists of fiber performance feature vectors from multiple samples. Each eigenvalue of the fiber performance feature vector has been min-max normalized, and its value range is... Therefore, the length of each dimension of the search space is 1, and the perceptual distance V visual The perceptual distance is set to 0.2. The rationale for setting the perceptual distance to 0.2 times the range of each dimension of the search space is that this perceptual distance ensures that each artificial fish's neighborhood contains an average of approximately 3 to 8 other artificial fish, satisfying the requirements for the number of neighboring partners for swarming and tail-chasing behaviors, while avoiding premature population convergence due to an excessively large neighborhood. Movement step size S step Set to sensing distance V visual 1.5 times that, i.e., S step The step size is set to 0.3. The rationale for setting the step size to 1.5 times the sensing distance is that in the artificial fish swarm algorithm, a larger step size helps to escape local optima more quickly. step For V visual At 1.5 times the size of the current, the movement distance of foraging and herding / chasing behaviors can cover the edge of the perception area, improving global optimization efficiency. (Crowding factor) The congestion factor is set to 0.618. The basis for setting the congestion factor to 0.618 is as follows: The value ranges between 0 and 1. The golden ratio value of 0.618 can reasonably measure the crowding level of the center or optimal companion position in swarming and chasing behaviors, allowing artificial fish to preferentially gather towards the center or optimal companion position when the food concentration is moderate, while avoiding gathering when overcrowded, thus maintaining population diversity. The number of clusters, K. cluster The number of cluster categories K is set according to the actual classification requirements of chemical fiber properties. cluster Consistent with the number of chemical fiber performance categories to be distinguished, the number of cluster categories K cluster The value of K is assigned manually during implementation based on prior knowledge. When prior knowledge is lacking, the number of cluster categories K is... cluster The silhouette coefficient is determined by comparing the number of cluster categories.

[0068] The initial locations of the artificial fish swarm are generated using a logistic chaotic mapping, with each artificial fish's location corresponding to a set of candidate cluster center matrices. The iterative formula for the logistic chaotic mapping is:

[0069]

[0070] Among them, z n z represents the state value of the logistic chaotic map at the nth iteration. n+1 Let z0 represent the state value of the logistic chaotic mapping at the (n+1)th iteration, and λ represent the control parameters of the logistic chaotic mapping. λ is set to 4. The reason for setting λ to 4 is that when λ=4, the logistic mapping is in a completely chaotic state, and the generated chaotic sequence has ergodicity and pseudo-randomness, and can uniformly cover the interval (0,1), which is suitable for population initialization. The initial state value z0 is set to 0.37. The reason for setting the initial state value z0 to 0.37 is that z0 needs to avoid the fixed points 0, 0.25, 0.5, 0.75 and 1. 0.37 is a non-special value within the interval (0,1), which can ensure the effective divergence of the chaotic sequence.

[0071] The specific method for generating the initial artificial fish swarm locations is as follows: iterative generation using logistic chaotic mapping. There are several state values, where D is the state value. feat This represents the number of rows in the fiber performance feature matrix, i.e., the dimension of the fiber performance feature vector. The iteratively generated state values ​​are sequentially divided into N... fish There are groups, each containing Each set of state values ​​is filled into a column of size [number] in row-major order. The candidate cluster center matrix is ​​used, where each row represents the fiber performance feature vector of a candidate cluster center. Since the state values ​​generated by the logistic chaotic mapping are in the interval (0,1), and the values ​​of each dimension of the fiber performance feature vector are all in the range [0,1], the direct assignment of the candidate cluster center matrix is ​​the corresponding chaotic state value. The initial position of each artificial fish is represented by a long vector formed by expanding all elements of a candidate cluster center matrix row by row, with the dimension of the vector being... .

[0072] Calculate the current fitness value for each artificial fish. The fitness value is the sum of squared distances from each sample in the fiber performance feature matrix to the nearest candidate cluster center. For a candidate cluster center matrix corresponding to an artificial fish, the candidate cluster center matrix contains K... cluster There are rows of vectors, each representing a candidate cluster center. For the j-th sample in the fiber performance feature matrix, the j-th column of the fiber performance feature matrix is ​​the fiber performance feature vector extracted from the j-th chemical fiber sample. The relationship between this fiber performance feature vector and K is calculated. clusterThe Euclidean distance between candidate cluster centers is used to determine the nearest candidate cluster center for the j-th sample. The distance to the nearest candidate cluster center is calculated for each sample, and the squares of these distances are summed to obtain the fitness value of the artificial fish.

[0073] In each iteration of the artificial fish swarm algorithm, foraging behavior, swarming behavior, and tail-chasing behavior are performed sequentially for each artificial fish, and the position of the artificial fish is updated.

[0074] The foraging behavior is executed as follows: the perceived distance V at the current location of the artificial fish. visual A random direction is selected within the range to generate a trial position. Each dimension value of the trial position equals the value of the dimension corresponding to the current artificial fish position plus a random increment, where the random increment is determined by the interval... The direction factor is obtained by multiplying a uniformly distributed random number within the range of 0 to 1 by a random direction factor uniformly distributed between 0 and 1. The direction factor is generated by randomly generating a unit vector with the same dimension as the current position and scaling it to the range of the movement step. If the value of any dimension of the test position exceeds... If the range is defined, the dimension value is clamped to the boundary value; values ​​exceeding the upper limit are set to 1, and values ​​below the lower limit are set to 0. The fitness value of the trial position is calculated. If the fitness value of the trial position is less than the fitness value of the current artificial fish position, the current artificial fish moves one step S towards the trial position. step The distance, i.e., the distance the current artificial fish's position vector moves along the direction vector between the trial position and the current position, is S. step The displacement is used to update the artificial fish's position. If the fitness value of the trial position is not less than the fitness value of the current artificial fish position, a new trial direction is randomly selected, and the attempts are repeated until the maximum number of attempts N is reached. try If, after that, no probing location is found that would lower the fitness value, the artificial fish will remain at a sensing distance V. visual Move randomly within one step S step The distance. Maximum number of attempts N try The maximum number of attempts is set to 5 because a moderate number of attempts can strike a balance between reducing ineffective search time and avoiding premature random wandering. Five attempts are sufficient to detect whether there is an improved direction within the local area.

[0075] The swarming behavior is executed as follows: obtain the current sensing distance V of the artificial fish. visual The number of companions within and their central location. Perception distance V. visual "Partners within the population" refers to individuals whose Euclidean distance from the current artificial fish's location is less than V. visual The number of all other artificial fish, denoted as N, is a symmetric number. near .when At this point, the center position of a companion is the one-dimensional mean vector of the position vectors of all companion artificial fish. The fitness value of the center position is calculated using the same method as calculating the fitness value of the artificial fish, after decoding the center position into a candidate cluster center matrix. If the fitness value of the center position is multiplied by the crowding factor... If the fitness value is less than that of the current artificial fish, then the current artificial fish moves one step S towards the center. step The distance is used to update the artificial fish's position. The fitness value of the center position is multiplied by the crowding factor. Not less than the current fitness value of artificial fish, or If the group behavior is not executed, then no group behavior will be performed.

[0076] The execution method for the rear-end collision behavior is as follows: obtain the current sensing distance V of the artificial fish. visual The optimal partner with the lowest intrinsic fitness value. At a perception distance V. visual Among all the artificial fish in the pool, the one with the lowest fitness value is selected as the optimal partner. The fitness value of the optimal partner is then multiplied by the crowding factor. If the fitness value is less than that of the current artificial fish, then the current artificial fish moves one step S towards the best partner. step The distance is used to update the artificial fish's position. The fitness value of the optimal partner is multiplied by the crowding factor. Not less than the fitness value of current artificial fish, or the perception distance V visual If there are no other vehicles nearby, a rear-end collision will not occur.

[0077] In one iteration, each artificial fish sequentially performs foraging, grouping, and tail-chasing behaviors. After each behavior is completed, the artificial fish's position is immediately updated, and subsequent behaviors are performed based on the updated position. The foraging, grouping, and tail-chasing behaviors are iteratively executed until the maximum number of iterations, I, is reached. max Maximum number of iterations I max The basis for setting the maximum number of iterations to 100 is that for an artificial fish population of 30, 100 iterations usually make the fitness value curve tend to be stable and the cluster center position reach a convergent state.

[0078] When the maximum number of iterations I is reached max Then, the location of the artificial fish with the lowest fitness value in the population is selected as the optimal cluster center matrix. Each row of the optimal cluster center matrix corresponds to a fiber performance feature vector obtained through cluster center optimization. The fiber performance feature vector of each row in the optimal cluster center matrix is ​​extracted as multiple performance category centers, and the number of performance category centers is equal to the number of cluster categories K. cluster Each performance category center represents a typical characteristic of a specific type of chemical fiber in the fiber performance characteristic space.

[0079] In practice, the process of determining the performance category of the chemical fiber sample to be tested based on the distance between the center of each performance category and the feature vector of the fiber to be tested is as follows.

[0080] See Figure 6 The process involves acquiring multi-angle microscopic image sequences and near-infrared spectral data sequences of the chemical fiber samples to be tested. The same microscopic image acquisition conditions as those applied to the chemical fiber samples collected during the training phase are used to acquire microscopic images of the chemical fiber samples at different angles, forming a multi-angle microscopic image sequence of the chemical fiber samples to be tested. The same spectral acquisition conditions as those applied to the chemical fiber samples collected during the training phase are then applied to the chemical fiber samples to be tested, and near-infrared scanning is performed at continuous time points to obtain the near-infrared spectral data sequence of the chemical fiber samples to be tested.

[0081] The same feature extraction and fusion process as in the training phase is used to generate the feature vector of the fiber to be tested. The feature extraction and fusion process includes: selecting a first-angle image and a second-angle image from a multi-angle microscopic image sequence of the chemical fiber sample to be tested, forming a binocular image pair; using a fast and robust feature algorithm to detect candidate feature points in the binocular image pair to be tested, generating a first feature point set and a second feature point set; calculating the first local gradient principal direction for each first feature point in the first feature point set, and calculating the second local gradient principal direction for each second feature point in the second feature point set; traversing all combinations of the first and second feature points, calculating the absolute angle difference between the first and second local gradient principal directions in each combination, and marking combinations with an absolute angle difference less than a preset angle threshold as candidate matching pairs; Hamming distance is calculated between the feature descriptors of the first and second feature points in each candidate matching pair. The candidate matching pair with the smallest Hamming distance is selected as the matching feature point pair. Based on the pixel coordinate difference between the matching feature point pairs in the first and second angle images, the spatial three-dimensional coordinates corresponding to each matching feature point are calculated. The spatial three-dimensional coordinates corresponding to all matching feature points are aggregated into an initial point cloud. Outlier removal and hole filling are performed on the initial point cloud to obtain a three-dimensional point cloud model of the fiber surface of the chemical fiber sample to be tested. Adaptive complete ensemble empirical mode decomposition is performed on the near-infrared spectral data sequence of the chemical fiber sample to be tested to obtain multiple intrinsic mode function components. After arranging them in descending order of frequency, the first three high-frequency intrinsic mode functions are extracted. The noise component is used as the dominant noise component, and the remaining intrinsic mode function (EMF) components are used as the dominant signal components. A Hilbert transform is performed on the EMF components of each dominant signal component to obtain the analytic signals of each EMF component. Instantaneous amplitude and phase sequences are extracted from the analytic signals. Instantaneous energy sequences are calculated based on the instantaneous amplitude sequences. These instantaneous energy sequences are divided into multiple equally spaced energy intervals. The probability of occurrence of instantaneous energy values ​​within each energy interval is calculated, and Shannon entropy is calculated based on each probability to obtain the instantaneous energy distribution entropy of each EMF component. A voxel grid downsampling is performed on the three-dimensional point cloud model of the fiber surface of the chemical fiber sample to be tested to obtain a resampled point cloud. The moving least squares method is used to fit the resampled point cloud to each... For each point and its neighboring points, calculate the Gaussian curvature and mean curvature of each point on the locally fitted surface. Calculate the principal curvature of each point based on the Gaussian and mean curvatures, and then calculate the ratio of the absolute value of the principal curvature difference between adjacent points within the same local region to the radius of the local region, thus obtaining the curvature change rate of the local region. Arrange the curvature change rates corresponding to the chemical fiber samples under test into a one-dimensional curvature feature vector according to their spatial location. Arrange the instantaneous energy distribution entropy corresponding to the chemical fiber samples under test into a one-dimensional entropy feature vector according to the order of the intrinsic mode function components. Perform maximum and minimum normalization on the one-dimensional curvature feature vector to obtain a normalized curvature vector, and perform maximum and minimum normalization on the one-dimensional entropy feature vector to obtain a normalized entropy vector.The normalized curvature vector and the normalized entropy vector are concatenated laterally to form the feature vector of the fiber to be tested. The dimension of the feature vector of the fiber to be tested is the same as the number of rows of the fiber performance feature matrix during the training phase, both being D; feat .

[0082] Calculate the Euclidean distance between the feature vector of the fiber to be tested and each performance category center. The performance category centers are the row vectors in the optimal cluster center matrix obtained after optimizing the cluster centers of the fiber performance feature matrix during the training phase using the artificial fish swarm algorithm with chaotic initialization. The number of performance category centers is K. cluster For the k-th performance category center, The center vector of the k-th performance category is denoted as C. k Dimension D feat The feature vector of the fiber to be tested is denoted as F. test Dimension D feat The feature vector of the fiber to be tested With the kth performance category center C k The Euclidean distance between them is calculated as follows:

[0083]

[0084] Among them, D k F represents the feature vector of the fiber to be tested. test With the kth performance category center C k The Euclidean distance between them; D feat This represents the dimension of the fiber feature vector to be tested, and also the dimension of the k-th performance category center vector; F represents the feature vector of the fiber to be tested. feat The value at the d-th feature dimension, ; C represents the center vector of the k-th performance category. k The value at the d-th feature dimension. For K cluster Calculate the Euclidean distance for each performance category center to obtain K. cluster Euclidean distance values .

[0085] The performance class center corresponding to the minimum value among all Euclidean distances is taken as the target class center. In K... cluster Find the minimum value among the Euclidean distance values, and denote the index corresponding to the minimum value as k*, satisfying the condition... The k*th performance category center C k* It was identified as the center of the target category.

[0086] The performance category represented by the target category center is determined as the performance category of the chemical fiber sample to be tested. The correspondence between performance categories and performance category centers is established during the training phase through cluster center optimization and category labeling: for K obtained by the artificial fish swarm algorithm with chaotic initialization... cluster Each performance category center is assigned a performance category label based on its relative position in the fiber performance feature space and the annotation information of the training samples within the corresponding region. The performance category label corresponding to the k*th performance category center is the performance category represented by the target category center. The performance determination result is output, which includes the performance category label to which the tested chemical fiber sample is determined, and the Euclidean distance D between the feature vector of the tested fiber and the target category center. k* .

[0087] In some embodiments, when the minimum Euclidean distance value D k* Greater than a preset distance threshold D th When the chemical fiber sample to be tested is determined not to belong to any known performance category, the performance determination result of "unknown category" is output. The preset distance threshold D... th The setup method is as follows: Obtain the Euclidean distance from all samples in the fiber performance feature matrix during the training phase to their respective nearest performance class centers, and calculate the mean of the Euclidean distances. and standard deviation ,set up The preset distance threshold is set to... The basis for this is the assumption that the Euclidean distance distribution of the training samples follows an approximately normal distribution. It covers approximately 95% of the intra-class distance range. Exceeding this threshold indicates that the sample being tested differs significantly from all known categories and is suitable for classification into an unknown category.

[0088] In one embodiment of the present invention, on a continuous chemical fiber production line, to achieve online monitoring of fiber performance and real-time interlocking control of production parameters, the chemical fiber sampling and performance testing system can be integrated near the exit of the spinning tunnel before the winding process. This embodiment describes how, in this online application scenario, the system works collaboratively with existing control units on the production line (such as the GP speed control system) to achieve real-time evaluation of fiber performance and trigger corresponding production control interlocking reactions. Specifically, an online microscopic image acquisition unit and a near-infrared spectroscopy online probe are installed on the production line. The microscopic image acquisition unit consists of two industrial cameras deployed at a fixed angle, forming a stable binocular vision system that continuously and synchronously captures images of the same section of a uniformly descending fiber bundle, thereby acquiring the multi-angle microscopic image sequence in real time. The near-infrared spectroscopy online probe is fixed to the side of the fiber bundle and performs non-contact scanning of the passing fibers at a set frequency, acquiring the near-infrared spectral data sequence in real time. These images and spectral data streams are transmitted in real time to the processor of the detection system via industrial Ethernet.

[0089] The system processor continuously runs the detection method. For a set of stereo images acquired in real time, a fast and robust feature algorithm is immediately used for feature point detection and matching. Due to fiber movement and field vibration in the online environment, the matching process needs to have higher robustness. In practice, in addition to calculating the angle difference of the principal directions of the local gradient of the feature points for initial screening, the Hamming distance of the feature descriptor is combined to find the optimal match among multiple candidate pairs, and the spatial three-dimensional coordinates of the matching feature points are quickly calculated by triangulation based on the online calibrated camera parameters. These coordinates are aggregated in real time and outliers caused by fiber shaking or ambient light interference are removed by statistical filtering. Then, radial basis function interpolation is used to fill in the small holes, thereby generating a dynamic three-dimensional point cloud model of the fiber surface online. This model corresponds to the instantaneous surface morphology of the fiber segment that just passed the detection point.

[0090] Simultaneously, adaptive complete ensemble empirical mode decomposition is performed in real time on the spectral data sequence within a time window transmitted from the near-infrared spectroscopy online probe. In online processing, to meet real-time requirements, the algorithm for adding opposite-signed white noise pairs and the subsequent overall averaging calculation process can be optimized, but the core steps remain unchanged. The aim is to separate the dominant signal components characterizing the fiber's chemical composition from the raw signal, which may contain equipment electromagnetic interference and environmental noise. Subsequently, a Hilbert transform is performed on each dominant signal component to extract the instantaneous amplitude and calculate the instantaneous energy sequence. By statistically analyzing the distribution probability of energy values ​​across multiple equally spaced intervals, the instantaneous energy distribution entropy of each component is quickly calculated. This process characterizes the dynamic energy distribution characteristics of the fiber's internal chemical structure under the current production state.

[0091] For the real-time generated 3D point cloud model, the system establishes a voxel mesh index in memory and performs downsampling to obtain a uniformly distributed resampled point cloud for subsequent fast surface analysis. For each point in the resampled point cloud, the system searches for neighboring points within a specific spatial scale (i.e., local region radius) determined by the fiber type being produced, and fits the local surface using the moving least squares method. Based on the fitted local surface, the Gaussian curvature and mean curvature of the point are calculated, and its principal curvature is derived. By calculating the absolute value of the principal curvature difference between the point and all its neighbors, and taking the ratio of the average of these differences to the local region radius, the curvature change rate, characterizing the severity of surface undulation near the point, is finally obtained. Arranging the curvature change rates of all points in the current fiber segment point cloud in spatial order constitutes a one-dimensional curvature feature vector. Simultaneously, arranging the instantaneous energy distribution entropy of all calculated dominant signal components in order of order constitutes a one-dimensional entropy feature vector. Before online fusion, the global maximum and minimum values ​​determined by the system during offline training are used to perform max-min normalization on the two real-time feature vectors to eliminate the influence of dimensions and maintain consistency with the training feature space. Finally, the normalized curvature feature vector and entropy feature vector are horizontally concatenated along the feature dimension to form a feature vector of the fiber under test that comprehensively represents the surface morphology and internal chemical properties of the current fiber segment.

[0092] The feature vector of the fiber to be tested will be fed into a pre-trained classification model for performance category determination. The core of this classification model is the performance category centers obtained through a chaotic initialization artificial fish swarm algorithm. During the training phase before system deployment, fiber samples from historical normal production batches, known fineness abnormality batches (e.g., due to GP speed fluctuations), and other performance abnormality batches were used to construct a fiber performance feature matrix containing multiple samples and categories through the same feature extraction and fusion process. Subsequently, logistic chaotic mapping was used to generate ergodic initial artificial fish swarm positions, with each "artificial fish" representing a set of candidate cluster centers. By defining the fitness function as the sum of squared distances from in-class samples to their nearest cluster center, and iteratively executing foraging, swarming, and tail-chasing behaviors, the algorithm ultimately searches for the optimal set of cluster centers that minimizes the fitness value. These cluster centers are labeled with different performance categories, such as "Category 1: Excellent performance," "Category 2: Slight surface roughness," "Category 3: Fluctuations in chemical composition," and "Category 4: High risk of fineness abnormality," forming a stable category determination benchmark.

[0093] During online operation, the system calculates the Euclidean distance between the real-time generated feature vector of the fiber under test and the centers of these pre-stored performance categories. By finding the minimum distance, the current fiber segment is assigned to the closest performance category. This determination result is directly linked to the production control system, forming an interlocking control logic. For example, when the system determines that the fiber performance belongs to "Category 1: Excellent Performance," it is considered that the production state is stable, no intervention signal is triggered, the GP speed runs at the originally set high speed, and the winding head and normal production process are not affected. When the determination result enters "Category 2: Slight Surface Roughness" or "Category 3: Chemical Composition Fluctuation," the system can send a warning signal to the central control room to prompt operators to pay attention to relevant process parameters, but may not immediately trigger a production shutdown. The key is that when the system determines that the fiber performance falls into "Category 4: High Risk of Abnormal Fineness," this determination is highly correlated with the actual fineness of the fiber being about to or already abnormal. At this time, the detection system will immediately send a digital alarm signal to the master station module of the GP interlocking control system.

[0094] Upon receiving the "High Risk of Fiber Density Abnormality" interlock signal from the performance monitoring system, the master module of the GP interlock control system immediately activates the interlock logic. First, the system checks if the GP high-speed operation command received by the metering pump inverter is normal. If the GP high-speed signal is normal, but the monitoring system still alarms, it may indicate a complex process abnormality not directly caused by speed, and the system can escalate the alarm level. A more crucial interlock application scenario is: when the performance monitoring system determines an abnormality, if the GP interlock control system, through its remote I / O module, detects that the actual operating speed of the metering pump (from inverter feedback) is lower than the set GP high-speed threshold, or if the GP high-speed command signal itself is lost for some reason, it immediately confirms this as a fiber density abnormality event requiring emergency intervention. In this case, the system does not wait for traditional fiber breakage or poor forming to occur, but actively intervenes. The interlock control system sends an emergency command to the winding machine's controller, and the winding machine automatically performs a filament cutting action, cutting off the currently risky fiber bundle to prevent abnormal fibers from continuing to wind onto the bobbin and becoming waste. Simultaneously, the system can automatically switch the metering pump of the corresponding spinning station to GP low-speed operation mode and illuminate the on-site indicator light to warn operators that the spinning station has been shut down due to abnormal fiber performance triggered by the interlock protection. This pre-shearing of yarn based on real-time performance judgment controls quality before winding and forming, avoiding subsequent waste yarn disposal costs.

[0095] Furthermore, the interlocking logic can be extended to group control of production rhythm. For example, when multiple adjacent spinning stations on a production line are successively identified as having "high risk of abnormal fineness" by the performance detection system within a short period of time and trigger yarn shearing, causing multiple metering pumps to switch to GP low-speed operation simultaneously or successively, it may cause drastic fluctuations in the inlet pressure of the shared booster pump (screw). To prevent the risk of booster pump stalling caused by this, the interlocking control system can embed a group management algorithm. This algorithm counts the number of spinning stations currently in "GP low-speed" operation in real time. When the number of spinning stations in the same group (such as several spinning stations sharing the same booster pump) in low-speed state reaches a preset upper limit (for example, set to a maximum of 2 depending on the product and booster pump capacity), the system will temporarily suspend subsequent new yarn shearing and GP speed reduction requests triggered by performance abnormalities. For the next spinning station that detects an abnormal risk, the system can take alternative strategies, such as maintaining its GP speed but marking the yarn bobbin as pending, or triggering an alarm of different priority. After a spinning station resumes normal high-speed operation and the low-speed count decreases, the normal interlocking response will resume. This integrated real-time performance monitoring and group control system enables the production system not only to respond to explicit equipment signal failures, but also to predict and adaptively adjust based on the fiber's inherent quality status, achieving a shift in control strategy from "post-fault handling" to "pre-quality prevention." The entire implementation process, from online acquisition of images and spectra, real-time extraction and fusion of 3D and spectral features, to instantaneous performance category determination based on intelligent clustering centers, and then to the deep interaction between the determination results and the GP interlocking control system, triggering a series of actions including fiber shearing, speed switching, and group coordination, constitutes a closed-loop online quality monitoring and production control implementation example.

[0096] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for testing the sampling and processing performance of chemical fibers, characterized in that, The method includes: Acquire multi-angle microscopic image sequences and near-infrared spectral data sequences of chemical fiber samples; Feature point extraction and stereo matching are performed on the multi-angle microscopic image sequence to generate a three-dimensional point cloud model of the fiber surface; Adaptive complete ensemble empirical mode decomposition is performed on the near-infrared spectral data sequence to obtain multiple intrinsic mode function components; Calculate the rate of curvature change of each local region in the three-dimensional point cloud model, and calculate the instantaneous energy distribution entropy of each intrinsic mode function component; The curvature change rate and the instantaneous energy distribution entropy are fused at multiple scales to construct a fiber performance characteristic matrix; The artificial fish swarm algorithm with chaotic initialization is used to optimize the clustering centers of the fiber performance feature matrix, thereby obtaining multiple performance category centers; The performance category of the chemical fiber sample to be tested is determined based on the distance between the center of each performance category and the feature vector of the fiber to be tested.

2. The method for testing the sampling and processing performance of chemical fibers according to claim 1, characterized in that, The specific steps for extracting feature points and performing stereo matching on the multi-angle microscopic image sequence to generate a three-dimensional point cloud model of the fiber surface are as follows: A first-angle image and a second-angle image are selected from the multi-angle microscopic image sequence to form a binocular image pair; A fast and robust feature algorithm is used to detect candidate feature points in the stereo image pair, generating a first feature point set and a second feature point set; Based on the angle difference between the principal local gradient direction of each feature point in the first feature point set and the principal local gradient direction of the corresponding feature point in the second feature point set, matching feature point pairs are selected. Based on the pixel coordinate difference of the matching feature point pairs in the first angle image and the second angle image, calculate the spatial three-dimensional coordinates of each matching feature point; All the spatial three-dimensional coordinates corresponding to the matching feature points are aggregated into an initial point cloud, and outlier points are removed and holes are filled in the initial point cloud to obtain a three-dimensional point cloud model of the fiber surface.

3. The method for testing the sampling and processing performance of chemical fibers according to claim 2, characterized in that, Outlier removal from the initial point cloud was performed using statistical filtering, and hole filling was performed using radial basis function interpolation.

4. The method for testing the sampling and processing performance of chemical fibers according to claim 1, characterized in that, The specific steps for performing adaptive complete ensemble empirical mode decomposition on the near-infrared spectral data sequence to obtain multiple intrinsic mode function components are as follows: Obtain the original near-infrared spectral data sequence, and add a first white noise sequence and a second white noise sequence to the original near-infrared spectral data sequence, wherein the first white noise sequence and the second white noise sequence are opposites of each other; Empirical mode decomposition is performed on the spectral data sequence after noise is added to obtain the first set of intrinsic mode function components and the second set of intrinsic mode function components; The first group of intrinsic mode function components and the second group of intrinsic mode function components are averaged to obtain multiple intrinsic mode function components under the adaptive complete set empirical mode decomposition. The multiple intrinsic mode function components are arranged in descending order of frequency. The first three high-frequency intrinsic mode function components are extracted as noise-dominant components, and the remaining intrinsic mode function components are extracted as signal-dominant components.

5. The method for testing the sampling and processing performance of chemical fibers according to claim 4, characterized in that, The specific steps for calculating the instantaneous energy distribution entropy of each of the intrinsic mode function components are as follows: Perform Hilbert transform on the intrinsic mode function components of each dominant signal component to obtain the analytic signal of each intrinsic mode function component; Extract the instantaneous amplitude sequence and instantaneous phase sequence from the analyzed signal; The instantaneous energy value corresponding to each sampling point is calculated based on the instantaneous amplitude sequence to obtain the instantaneous energy sequence; The instantaneous energy sequence is divided into multiple equally spaced energy intervals, and the probability of the instantaneous energy value occurring in each energy interval is calculated. Shannon entropy is calculated based on the occurrence probabilities of each condition, and the instantaneous energy distribution entropy of each intrinsic mode function component is obtained.

6. The method for testing the sampling and processing performance of chemical fibers according to claim 1, characterized in that, The specific steps for calculating the rate of curvature change of each local region in the three-dimensional point cloud model are as follows: The three-dimensional point cloud model of the fiber surface is downsampled using a voxel grid to obtain a uniformly distributed resampled point cloud; The local surface of each point and its neighboring points in the resampled point cloud is fitted using the moving least squares method to obtain the local fitted surface of each point; Calculate the Gaussian curvature and mean curvature of each point on the locally fitted surface; The principal curvature of each point is calculated based on the Gaussian curvature and the average curvature, and the ratio of the absolute value of the principal curvature difference between adjacent points in the same local region to the radius of the local region is calculated to obtain the curvature change rate of the local region.

7. The method for testing the sampling and processing performance of chemical fibers according to claim 1, characterized in that, The specific steps for constructing the fiber performance feature matrix by multi-scale fusion of the curvature change rate and the instantaneous energy distribution entropy are as follows: The rate of change of curvature corresponding to the same chemical fiber sample is arranged into a one-dimensional curvature feature vector according to its spatial position. The instantaneous energy distribution entropy corresponding to the same chemical fiber sample is arranged into a one-dimensional entropy feature vector according to the order of the intrinsic mode function components; The one-dimensional curvature feature vector is subjected to maximum-minimum normalization to obtain a normalized curvature vector; The one-dimensional entropy feature vector is subjected to maximum-minimum normalization to obtain a normalized entropy vector; The normalized curvature vector and the normalized entropy vector are horizontally concatenated to form a two-dimensional fiber performance feature matrix. The number of rows in the two-dimensional fiber performance feature matrix is ​​the number of feature dimensions, and the number of columns is the number of samples.

8. The method for testing the sampling and processing performance of chemical fibers according to claim 7, characterized in that, The specific steps of using the chaotic initialization artificial fish swarm algorithm to optimize the clustering centers of the fiber performance feature matrix and obtain multiple performance category centers are as follows: Set the population size, sensing distance, movement step length, and crowding factor of the artificial fish swarm, and set the number of cluster categories; The initial artificial fish swarm locations are generated using logistic chaotic mapping, with each location corresponding to a set of candidate cluster center matrices; Calculate the current fitness value for each artificial fish, where the fitness value is the sum of squared distances from each sample in the fiber performance feature matrix to the nearest candidate cluster center; For each artificial fish, perform foraging behavior, grouping behavior, and tail-chasing behavior, and update the location of the artificial fish; The process iteratively executes each action until the maximum number of iterations is reached. The position of the artificial fish with the smallest fitness value is taken as the optimal cluster center matrix, and multiple performance category centers are extracted from the optimal cluster center matrix.

9. The method for testing the sampling and processing performance of chemical fibers according to claim 1, characterized in that, The specific steps for determining the performance category of the chemical fiber sample to be tested based on the distance between each performance category center and the feature vector of the fiber to be tested are as follows: Acquire multi-angle microscopic image sequences and near-infrared spectral data sequences of the chemical fiber sample to be tested, and generate feature vectors of the fiber to be tested using the same feature extraction and fusion process as in the training phase. Calculate the Euclidean distance between the feature vector of the fiber under test and the center of each performance category; The performance class center corresponding to the minimum value among all Euclidean distances is taken as the target class center; The performance category represented by the target category center is determined as the performance category of the chemical fiber sample to be tested, and the performance determination result is output.

10. A chemical fiber sampling and processing performance testing system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the chemical fiber sampling and processing performance testing method according to any one of claims 1 to 9.

Images