Batch detection method and system for production performance of textile fabric

Abstract

The invention provides a batch detection method and system for textile fabric production performance. Wherein a quantitative correlation model of the morphological parameters and the mechanical property parameters of the defects is established based on the morphological parameters and the mechanical property test results of the historical defect samples; according to the composition and the weaving structure of the novel fabric, dynamically adjusting the quantitative correlation model through a transfer learning technology so as to adapt to the fabric mechanical response prediction requirements of different fiber materials and organization structures; synchronously acquiring surface texture features and internal fiber distribution features by using a hyperspectral imaging system, and identifying hidden defect areas which are not found in conventional visual detection; and inputting the spectral feature data, the process parameters of the current production batch and the spatial distribution features into a pre-trained residual network model, carrying out weighted fusion through a feature channel to form a comprehensive quality index, and carrying out rapid classification on the same batch of textile fabrics. According to the technical scheme provided by the invention, the mechanical property prediction precision and the hidden defect detection rate of the textile are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of batch testing technology, and in particular to a batch testing method and system for the production performance of textile fabrics. Background Technology

[0002] With the increasing demand for rapid quality inspection of diverse, small-batch fabrics in the textile industry, especially when producing fabrics containing new fibers or complex weaving structures, existing testing methods struggle to simultaneously assess the correlation between surface quality and internal mechanical properties. In actual production, microstructural defects often cannot be detected through conventional testing methods, and the impact of process parameter fluctuations on defect formation lacks systematic analysis. Therefore, there is an urgent need for an automated testing technology that can integrate multi-dimensional data and adapt to dynamic production conditions.

[0003] Existing solutions employ multi-sensor data fusion technology, combining industrial cameras and infrared sensors to acquire fabric surface images and thermal distribution data, and then using a support vector machine (SVM) model to classify and detect visible defects. This solution utilizes texture analysis algorithms to extract surface defect features and combines thermal imaging data to infer areas of abnormal fiber density.

[0004] While existing solutions can detect some surface and thermodynamic anomalies, they cannot establish a quantitative relationship between fiber density anomalies and the actual mechanical properties of the fabric (such as tear resistance), resulting in the detection results not being able to directly guide process optimization. In addition, the model is sensitive to changes in fiber material; when the fabric composition changes, the sensor parameters need to be recalibrated and the classification model updated, making it difficult to achieve rapid deployment across different fabric types. Summary of the Invention

[0005] This application provides a method and system for batch testing of the production performance of textile fabrics, in order to solve the problems in the prior art.

[0006] Firstly, this application provides a method for batch testing of the production performance of textile fabrics, including: Based on the morphological parameters and mechanical performance test results of historical defect samples, a quantitative correlation model between defect morphological parameters and mechanical performance parameters is established. The quantitative correlation model includes a nonlinear mapping relationship between defect size, shape and fabric tear strength and tensile strength. Based on the composition and weaving structure of the new fabric, the quantitative correlation model is dynamically adjusted using transfer learning technology to adapt to the mechanical response prediction requirements of fabrics with different fiber materials and weave structures. The dynamic adjustment process depends on the generalization ability of the nonlinear mapping relationship. The surface texture features and internal fiber distribution features of the novel fabric are simultaneously acquired using a hyperspectral imaging system. Combined with the mechanical property predictions output by the quantitative correlation model and the nonlinear mapping relationship, hidden defect areas that are not detected by conventional visual inspection are identified. The spectral feature data acquired by the hyperspectral imaging system, the process parameters of the current production batch, and the spatial distribution characteristics of the hidden defect area are input into a pre-trained residual network model. The model is then fused into a comprehensive quality index through weighted feature channels. Based on the comprehensive quality index and the nonlinear mapping relationship, the same batch of textile fabrics is quickly graded.

[0007] Optionally, the step of simultaneously acquiring the surface texture features and internal fiber distribution features of the novel fabric using a hyperspectral imaging system, and combining the predicted mechanical properties output by the quantitative correlation model with the nonlinear mapping relationship to identify hidden defect areas not detected by conventional visual inspection includes: In the visible light band of the hyperspectral imaging system, a high spatial resolution imaging mode is set to obtain the directional gradient distribution map of surface texture features. At the same time, the transmission imaging mode is switched in the near-infrared band to generate a transmission attenuation map of the internal fiber distribution features based on the optical absorption characteristics of the fiber. The directional gradient distribution map is spatially decomposed in the frequency domain to separate the fundamental frequency texture component and the abnormal frequency band residual component that are consistent with the weaving cycle of the textile fabric. The fiber density change rate in the transmission attenuation spectrum is superimposed with the fundamental frequency texture component to generate a joint distribution map that integrates surface and internal features. Based on the mechanical property prediction values ​​output by the quantitative correlation model, the coordinates of the areas to be inspected where the tensile strength prediction value is lower than a set threshold are extracted, and the coordinates of the areas to be inspected are mapped to the corresponding positions of the joint distribution map to generate candidate defect areas with mechanical and spectral correlation. Within the candidate defect region, the deviation between the predicted mechanical performance value and the actual mechanical test value is calculated using the nonlinear mapping relationship. Regions where the deviation exceeds the upper limit of the braided structure tolerance and are not shown in the directional gradient distribution map are marked as latent defect regions.

[0008] Optionally, within the candidate defect region, calculating the deviation between the predicted mechanical performance value and the actual mechanical test value through the nonlinear mapping relationship, and marking the region where the deviation exceeds the upper limit of the braided structure tolerance and is not shown in the directional gradient distribution map as a latent defect region includes: Within the candidate defect area, the predicted tensile strength value output by the quantitative correlation model is extracted, and a uniaxial tensile test is performed on the fabric sample corresponding to the candidate defect area to obtain the actual tensile strength test value. Based on the tensile strength response function in the nonlinear mapping relationship, the degree of deviation between the predicted tensile strength value and the actual tensile strength test value is calculated. Based on the upper limit of tolerance of the weave structure, a maximum allowable threshold for the degree of deviation is set, and the maximum allowable threshold is related to the weave density of the textile fabric and the strength characteristics of the fiber material. The region whose deviation exceeds the maximum allowable threshold and is not visible in the directional gradient distribution map is marked as a latent defect region. The latent defect region satisfies the dual criteria of abnormal mechanical properties and being visually invisible.

[0009] Optionally, the step of establishing a quantitative correlation model between defect morphological parameters and mechanical performance parameters based on the morphological parameters and mechanical performance test results of historical defect samples includes: Extract a set of morphological features containing yarn breakage, weft skew, and stain defects from historical defect samples. Based on the depth curvature distribution and boundary topology of the morphological feature set, define the equivalent ellipse major axis length of the defect size and the irregularity parameter of the defect shape. Uniaxial tensile and trapezoidal tear tests were performed on the fabric samples corresponding to the historical defect samples, and the peak tensile strength and the rate of decrease in tear strength were recorded simultaneously to form mechanical property test results corresponding to the morphological feature set. A quantitative correlation model is established between the equivalent ellipse major axis length, the irregularity parameter and the mechanical performance test results through a dynamic weight factor allocation mechanism. The dynamic weight factor allocation mechanism adaptively adjusts the influence weight of different morphological characteristics on mechanical performance according to the differences in defect types. Based on the quantitative correlation model, a parametric response surface covering a continuous range of values ​​is generated. The parametric response surface constructs a nonlinear mapping relationship between defect size, shape and fabric tear strength and tensile strength through an interpolation function, and embeds the nonlinear mapping relationship into the quantitative correlation model.

[0010] Optionally, the step of dynamically adjusting the quantitative correlation model based on the composition and weave structure of the novel fabric using transfer learning technology to adapt to the fabric mechanical response prediction requirements of different fiber materials and weave structures includes: The fiber chemical properties of the new fabric composition are analyzed to obtain a fiber feature set including parameters such as moisture absorption rate and bending stiffness. At the same time, the warp density and weft yarn intersection distribution of the new fabric are quantitatively extracted to form a weaving structure feature set. The fiber feature set and the weaving structure feature set are input into a pre-constructed transfer feature projection layer. Through a cross-domain feature matching mechanism, the features of the new fabric are spatially mapped to the historical defect sample features in the quantitative correlation model to generate a transfer feature vector with cross-material generalization ability. Based on the migration feature vector, the nonlinear mapping relationship is dynamically constrained, the general mapping parameters that are independent of fiber material in the quantitative correlation model are retained, the target mapping parameters affected by the weaving structure are reset, and an incremental mapping relationship is formed. Based on the incremental mapping relationship and the warp density distribution characteristics in the weaving structure feature set, the quantitative correlation model is dynamically adjusted to reconstruct the response function of the defect size to the tensile strength, so as to adapt to the fabric mechanical response prediction requirements of different fiber materials and weave structures.

[0011] Optionally, the step of dynamically constraining the nonlinear mapping relationship based on the migration feature vector, retaining the general mapping parameters in the quantitative correlation model that are independent of fiber material, resetting the target mapping parameters affected by the weaving structure, and forming an incremental mapping relationship includes: Based on the fiber moisture absorption rate and bending stiffness parameters in the migration feature vector, identify the set of general mapping parameters that are strongly correlated with fiber chemical properties in the nonlinear mapping relationship; Based on the distribution density of weft yarn intersection points in the migration feature vector, locate the target mapping parameter set that is affected by the periodicity of the weaving structure in the nonlinear mapping relationship; Apply parameter freezing constraints to the general mapping parameter set to maintain the numerical invariance of the general mapping parameter set when the fiber material changes, and perform parameter space reconstruction on the target mapping parameter set to generate the reconstructed target mapping parameter set. The parameter freezing constraint results are jointly optimized with the reconstructed target mapping parameter set to generate an incremental mapping relationship that retains fiber material independence and is adapted to the weaving structure.

[0012] Optionally, the step of inputting the spectral feature data acquired by the hyperspectral imaging system, the process parameters of the current production batch, and the spatial distribution characteristics of the latent defect region into a pre-trained residual network model, and then weighting and fusing them into a comprehensive quality index through feature channels, and rapidly classifying the same batch of textile fabrics based on the comprehensive quality index and the nonlinear mapping relationship, includes: The spectral feature data acquired by the hyperspectral imaging system are divided into a visible light feature subset and a near-infrared feature subset according to the spectral band. The process parameters of the current production batch are converted into a set of process influencing factors, which includes the yarn tension fluctuation coefficient and the loom speed uniformity index. The spatial distribution characteristics of the latent defect region are quantified into a defect density distribution map, which reflects the number and degree of aggregation of latent defects per unit area. The visible light feature subset, the near-infrared feature subset, the process influence factor set, and the defect density distribution map are input into a pre-trained residual network model. A comprehensive quality index is generated through a weighted fusion mechanism of multi-source feature channels. Based on the comprehensive quality index and the nonlinear mapping relationship, the same batch of textile fabrics is quickly graded and output.

[0013] Secondly, this application provides a batch testing system for the production performance of textile fabrics, comprising: The modeling module is used to establish a quantitative correlation model between the morphological parameters and mechanical performance parameters of defects based on the morphological parameters and mechanical performance test results of historical defect samples. The quantitative correlation model includes a nonlinear mapping relationship between defect size, shape and fabric tear strength and tensile strength. The adjustment module is used to dynamically adjust the quantitative correlation model based on the composition and weaving structure of the new fabric using transfer learning technology, so as to adapt to the mechanical response prediction requirements of fabrics with different fiber materials and weave structures. The dynamic adjustment process depends on the generalization ability of the nonlinear mapping relationship. The identification module is used to simultaneously acquire the surface texture features and internal fiber distribution features of the new fabric using a hyperspectral imaging system, and combine the mechanical property prediction values ​​output by the quantitative correlation model and the nonlinear mapping relationship to identify hidden defect areas that are not detected by conventional visual inspection. The generation module is used to input the spectral feature data acquired by the hyperspectral imaging system and the process parameters of the current production batch into a pre-trained residual network model, generate a comprehensive quality index through weighted fusion of feature channels, and quickly classify and output the same batch of textile fabrics based on the comprehensive quality index and the nonlinear mapping relationship.

[0014] Thirdly, embodiments of this application provide a computing device, including a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement a batch testing method for the production performance of textile fabrics as described in the first aspect above.

[0015] Fourthly, embodiments of this application provide a computer storage medium storing a computer program, which, when executed by a computer, implements a method for batch testing of the production performance of textile fabrics as described in the first aspect.

[0016] In this embodiment, based on the morphological parameters and mechanical performance test results of historical defect samples, a quantitative correlation model between defect morphological parameters and mechanical performance parameters is established. This quantitative correlation model includes a nonlinear mapping relationship between defect size, shape, and fabric tear strength and tensile strength. According to the composition and weaving structure of the new fabric, the quantitative correlation model is dynamically adjusted using transfer learning technology to adapt to the mechanical response prediction requirements of fabrics with different fiber materials and structures. The dynamic adjustment process relies on the generalization ability of the nonlinear mapping relationship. A hyperspectral imaging system is used to simultaneously acquire the surface texture features and internal fiber distribution features of the new fabric. Combined with the mechanical performance prediction values ​​output by the quantitative correlation model and the nonlinear mapping relationship, hidden defect areas not detected by conventional visual inspection are identified. The spectral feature data acquired by the hyperspectral imaging system, the process parameters of the current production batch, and the spatial distribution features of the hidden defect areas are input into a pre-trained residual network model. These are then weighted and fused into a comprehensive quality index through feature channels. Based on the comprehensive quality index and the nonlinear mapping relationship, the same batch of textile fabrics is rapidly graded.

[0017] The technical solution of this application has the following beneficial effects: By analyzing the morphological parameters and mechanical properties of historical defect samples, a nonlinear mapping relationship is constructed between defect size, shape, and fabric tear strength and tensile strength. This provides a theoretical model basis for predicting fabric mechanical properties and enables quantitative correlation analysis between defect morphology and mechanical properties. Based on transfer learning technology, model parameters are dynamically adjusted according to the composition and weaving structure of new fabrics, preserving common patterns and adapting to differences in fiber materials and structures, thereby improving the model's generalization ability and prediction accuracy for different fabric types. Hyperspectral imaging is used to simultaneously acquire surface texture and internal fiber distribution characteristics. Combined with mechanical property predictions and nonlinear mapping relationships, areas with mechanical anomalies but not visible to the naked eye are screened, solving the problem of missed detection of hidden defects (such as microfiber breakage and internal stress concentration) in traditional testing. By integrating hyperspectral features, process parameters, and hidden defect distribution data, a comprehensive quality index is generated through weighted fusion. Combined with nonlinear mapping relationships, a multi-dimensional collaborative grading of the appearance quality, mechanical properties, and process stability of the same batch of fabrics is achieved.

[0018] Furthermore, based on the hyperspectral imaging system, surface texture direction gradient distribution maps and internal fiber transmission attenuation maps are acquired in different bands. The fundamental frequency texture component and anomalous residual component are separated through spatial frequency domain decomposition, and a joint distribution map is generated by fusing the fiber density change rate. Candidate defect regions are located by combining predicted mechanical properties. The deviation between the predicted and actual mechanical properties is calculated using a nonlinear mapping relationship, and regions exceeding the structural tolerance limit and not visually visible are selected as latent defects. Through the fusion of multimodal features of surface texture and internal fibers, regions with abnormal mechanical properties are accurately located. Combining nonlinear mapping relationships and the condition of visual invisibility, explicit and latent defects are effectively distinguished, significantly improving the detection rate of latent defects such as microscopic damage and internal stress concentration in complex woven fabrics. Simultaneously, misjudgments caused by fluctuations in process parameters are avoided, achieving full-dimensional coverage of textile fabric quality inspection from appearance to mechanical properties.

[0019] These or other aspects of this application will become more apparent in the following description of the embodiments. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart of a batch testing method for the production performance of textile fabrics provided in this application is shown; Figure 2 A schematic diagram of the structure of a batch testing system for the production performance of textile fabrics provided in this application is shown. Figure 3 A schematic diagram of the structure of a computing device provided in this application is shown. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0023] In some of the processes described in the specification, claims, and accompanying drawings of this application, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as 101, 102, etc., are merely used to distinguish different operations and do not themselves represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a chronological order, nor do they limit "first" and "second" to different types.

[0024] This application addresses the complex needs of defect detection and quality assessment in textile fabric production. First, it constructs a quantitative correlation model based on the morphological parameters and mechanical property data of historical defect samples, establishing a nonlinear mapping relationship between defect size, shape, and fabric tear strength and tensile strength. Then, it dynamically adjusts the model parameters using transfer learning technology to adapt to the mechanical response prediction requirements of fabrics with different fiber materials and weaving structures. Next, it utilizes a hyperspectral imaging system to simultaneously acquire surface texture and internal fiber distribution characteristics, combining the predicted mechanical properties with the nonlinear mapping relationship to identify hidden defect areas. Finally, it integrates hyperspectral features, process parameters, and hidden defect distribution data, generating a comprehensive quality index through weighted fusion. Based on the nonlinear mapping relationship, it achieves multi-dimensional rapid grading of the same batch of fabrics, thus forming a fully intelligent solution for the entire process from defect detection to quality grading.

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] Figure 1 This application provides a flowchart of a method for batch testing of the production performance of textile fabrics, as shown in the embodiments. Figure 1 As shown, the method includes: 101. Based on the morphological parameters and mechanical performance test results of historical defect samples, establish a quantitative correlation model between defect morphological parameters and mechanical performance parameters. The quantitative correlation model includes a nonlinear mapping relationship between defect size, shape and fabric tear strength and tensile strength. In this step, the quantitative correlation model refers to a mathematical relationship model established based on the morphological parameters (defect size, shape) and mechanical property parameters (tear strength, tensile strength) of historical defect samples, which is used to describe the nonlinear mapping law between morphological parameters and mechanical properties.

[0027] In this embodiment, morphological parameters of historical defect samples are first obtained using 3D laser scanning technology, including the equivalent major axis length of the defect region (characterizing size) and shape irregularity (calculated through boundary topology). Uniaxial tensile and trapezoidal tear tests are performed on the same batch of samples, recording the peak tensile strength and the rate of decrease in tear strength. Support vector regression (SVR) is used to nonlinearly fit the morphological parameters and mechanical property data. A kernel function maps low-dimensional features to a high-dimensional space to capture complex relationships, and a regularization term is introduced during optimization to prevent overfitting. The resulting quantitative correlation model contains the nonlinear mapping relationship between defect size, shape, and mechanical properties. The optimal hyperparameter combination is determined through cross-validation.

[0028] Suppose a textile company has accumulated 500 sets of linen fabric samples containing yarn breaks and weft skew defects. A 3D laser scanner (accuracy ±0.01mm) is used to obtain the depth curvature distribution of the defect areas. The equivalent ellipse major axis length (range 2-8mm, measurement standard ISO 9073) and shape irregularity (0.1-0.9, based on the Hausdorff distance algorithm for boundary topology) are calculated. Mechanical testing is performed using an INSTRON universal testing machine, executing uniaxial tensile tests (ASTM D5035 standard) and trapezoidal tear tests (ASTM D5587 standard), recording the peak tensile strength (80-200MPa) and tear strength reduction rate (5-25N / s). A Gaussian kernel function support vector regression model is used, with hyperparameters optimized through 5-fold cross-validation (penalty coefficient C=1.2, kernel coefficient γ=0.05). After training, the model prediction error is controlled within 3.8%, generating a quantitative correlation model for real-time production line prediction.

[0029] 102. Based on the composition and weaving structure of the new fabric, the quantitative correlation model is dynamically adjusted using transfer learning technology to adapt to the mechanical response prediction requirements of fabrics with different fiber materials and weave structures. The dynamic adjustment process depends on the generalization ability of the nonlinear mapping relationship. In this step, dynamic adjustment refers to updating the parameters of the quantitative correlation model based on the composition (fiber material) and weaving structure (organic structure) of the new fabric through transfer learning technology, so as to adapt it to the mechanical response prediction requirements of different fabrics.

[0030] In this embodiment, infrared spectroscopy analysis is performed on the fiber composition of the novel fabric to extract chemical property parameters such as moisture absorption rate and bending stiffness. Digital image processing technology is used to quantify the warp density and weft yarn intersection distribution of the weave structure. The features of the novel fabric are mapped to the parameter space of the quantitative correlation model through a feature alignment algorithm, identifying general parameters unrelated to fiber material (such as the influence coefficient of shape on tear strength) and freezing their weights. For parameters related to the weave structure (such as the correlation factor of size on tensile strength), an incremental learning strategy is used to control the parameter update amplitude, and adversarial training is used to enhance the model's generalization ability to adapt to different fabric mechanical response prediction requirements while retaining the generalization characteristics of the original nonlinear mapping relationship.

[0031] For example, continuing the previous example, the company added a new type of fabric blended with bamboo fiber (40% bamboo fiber). The fiber's moisture absorption rate (ASTM D2495 test value of 8.2%) and bending stiffness (3.5 mN·mm² measured in a three-point bending test) are significantly different from pure linen. The warp density of the weave structure is increased to 32 yarns / cm (compared to 28 yarns / cm for linen), and the periodicity index of the weft yarn intersection point is 0.82 (previously 0.75). Using a feature alignment algorithm (maximum mean difference MMD < 0.15), the correlation parameters between shape and tear strength in the quantitative correlation model are frozen (weight 0.48). For the size and tensile strength parameters, adversarial training is used (generator learning rate 0.001, discriminator 0.0005). The initial values ​​of the correlation factors are adjusted from 0.32 to 0.41. After the model adapts to the mechanical response prediction requirements of the new fabric, the prediction error for the tensile strength of the new fabric is reduced to 4.5%.

[0032] 103. Using a hyperspectral imaging system to simultaneously acquire the surface texture features and internal fiber distribution features of the new fabric, and combining the mechanical property prediction values ​​output by the quantitative correlation model and the nonlinear mapping relationship, identify hidden defect areas that are not detected in conventional visual inspection. In this step, the latent defect area refers to the defect area that cannot be detected by conventional visual inspection, which is identified through the combined use of hyperspectral imaging and mechanical modeling. Examples include microfiber fractures or internal stress concentrations.

[0033] In this embodiment, high-resolution imaging is used in the visible light band to obtain a gradient distribution map of the surface texture direction. Fourier transform is used to extract the fundamental frequency texture component (reflecting the normal weaving cycle) and the abnormal residual component (reflecting defects). The near-infrared band is switched to transmission mode, and the fiber density change rate is calculated based on Beer and Lambert's laws to generate a transmission attenuation spectrum. Wavelet packet transform is used to spatially superimpose the fundamental frequency component and the fiber density change rate to generate a joint distribution map. Combining the tensile strength prediction value (coordinates of areas below a threshold) output by the quantitative correlation model, candidate defect areas are screened through coordinate mapping. Finally, a nonlinear mapping relationship is used to calculate the mechanical deviation, and after eliminating visually visible areas, undetected hidden defect areas are identified.

[0034] For example, continuing the previous example, hyperspectral imaging was performed on the bamboo fiber blended fabric. In the visible light band (400-700nm), a linear CCD camera (5μm / pixel resolution) was used to capture the directional gradient distribution map. Principal direction angle analysis showed that the warp direction was 45°±3°, and the area with a gradient amplitude > 0.18 accounted for 12%. Near-infrared transmission imaging (wavelength 1300nm) generated a transmission attenuation spectrum based on a fiber optic spectrometer, and the fiber density change rate was measured to be between -10% and +12%. The fundamental frequency texture component (frequency 0.32mm⁻¹, matching warp density 32 yarns / cm) and the anomalous residual component (amplitude > 0.07) were separated by wavelet packet transform (Daubechies-4 wavelet basis). After superimposing the fiber density change rate, a joint distribution map was generated, locating 6 previously undetected latent defect areas (area > 0.5mm²).

[0035] 104. Input the spectral feature data acquired by the hyperspectral imaging system, the process parameters of the current production batch, and the spatial distribution features of the hidden defect area into a pre-trained residual network model, and fuse them into a comprehensive quality index through weighted feature channels. Based on the comprehensive quality index and the nonlinear mapping relationship, quickly classify the same batch of textile fabrics.

[0036] In this step, the comprehensive quality index refers to a quantitative evaluation index generated by weighting hyperspectral characteristics, process parameters, and latent defect distribution data, which is used for fabric grading.

[0037] In this embodiment, hyperspectral data is divided into visible light (color uniformity) and near-infrared (fiber crystallinity) feature subsets; process parameters (such as yarn tension and loom speed) are converted into standardized process influencing factors; and the spatial distribution characteristics of latent defect regions are converted into a defect density heatmap. A channel attention mechanism is introduced into the pre-trained residual network model to dynamically allocate the weights of each feature subset. After weighted fusion, a comprehensive quality score of 0-100 is generated. Combining the defect size and tensile strength curve in the nonlinear mapping relationship, products are classified as superior, qualified, and substandard.

[0038] For example, continuing the previous example, the hyperspectral data (visible light color uniformity HSV score 92, near-infrared fiber crystallinity index 0.88) and process parameters (yarn tension fluctuation coefficient 0.06, loom speed variation index 0.03) of a batch of bamboo fiber were input into a pre-trained ResNet-50 model. The latent defect distribution was converted into a kernel density estimation heatmap (bandwidth 0.4mm, peak value 0.38). Weights were assigned using a channel attention mechanism: visible light 0.55, near-infrared 0.3, process 0.1, and defects 0.05, resulting in a comprehensive quality index score of 89.7. Combining the defect size and tensile strength curve in the nonlinear mapping relationship (equivalent major axis 5mm corresponds to tensile strength of 190MPa), the batch was determined to be of superior quality (score > 85 and tensile strength > 185MPa), completing the rapid grading of the same batch of textile fabrics.

[0039] Steps 101-104 construct a precise relationship between defect morphology and mechanical properties through a quantitative correlation model, utilize transfer learning to achieve dynamic adaptation of cross-material models, and combine hyperspectral imaging and multi-source data fusion to overcome the bottleneck of latent defect detection. Finally, a comprehensive grading system covering surface quality, internal structure and process stability is formed, which significantly improves detection accuracy and efficiency and meets the industrial needs of multi-variety, small-batch textile fabrics.

[0040] To further improve the detection accuracy of latent defects in textile fabrics and achieve multimodal data collaborative analysis, hyperspectral imaging technology is used to simultaneously acquire surface texture and internal fiber features. Combined with mechanical property prediction and nonlinear mapping relationships, a full-process detection system from feature extraction to defect determination is constructed. In some embodiments, the hyperspectral imaging system is used to simultaneously acquire the surface texture features and internal fiber distribution features of the novel fabric. Combined with the mechanical property prediction values ​​output by the quantitative correlation model and the nonlinear mapping relationships, latent defect areas not detected by conventional visual inspection are identified, including: 201. Set a high spatial resolution imaging mode in the visible light band of the hyperspectral imaging system to obtain the directional gradient distribution map of surface texture features. At the same time, switch to the transmission imaging mode in the near-infrared band to generate a transmission attenuation map of the internal fiber distribution features based on the optical absorption characteristics of the fiber. In step 201, the directional gradient distribution map refers to a visual map of the surface texture direction and gradient intensity distribution of textile fabrics obtained through high spatial resolution imaging in the visible light band, reflecting the yarn arrangement direction and texture continuity. The transmission attenuation map refers to an internal fiber density distribution map generated based on the optical absorption characteristics of fiber materials in the near-infrared band, characterizing the differences in fiber packing density and crystallinity.

[0041] In this embodiment, in the visible light band (400-700nm), a line-scanning hyperspectral camera (5μm / pixel resolution) is used to continuously acquire images along the fabric transport direction. The principal orientation angle (θ) and gradient magnitude (|G|) of each pixel are calculated using the Hessian matrix to generate an orientation gradient distribution map. Simultaneously, in the near-infrared band (900-1700nm), the transmission imaging mode is switched, and the light intensity attenuation rate (I / I0) at different wavelengths is measured using a fiber-coupled spectrometer. Based on Beer and Lambert's law, the fiber density change rate (Δρ=ln(I0 / I) / (μ·d)) is calculated, where μ is the optical absorption coefficient of the fiber material (pre-calibrated value) and d is the fabric thickness (measured in real time by a laser thickness gauge). Finally, a transmission attenuation spectrum is generated.

[0042] 202. Perform spatial frequency domain decomposition on the directional gradient distribution map to separate the fundamental frequency texture component and the abnormal frequency band residual component that are consistent with the weaving cycle of the textile fabric. Then, superimpose the fiber density change rate in the transmission attenuation spectrum with the fundamental frequency texture component to generate a joint distribution map that integrates surface and internal features. In step 202, the fundamental frequency texture component refers to the regular texture frequency component that matches the weaving cycle of the textile fabric, reflecting the spatial periodicity of the normal weaving structure. The abnormal frequency band residual component refers to the texture frequency component that deviates from the fundamental frequency, characterizing surface defects or internal structural anomalies.

[0043] In this embodiment, the directional gradient distribution map is subjected to two-dimensional Gabor filtering (center frequency f0 = 1 / T, where T is the measured value of the weaving period) to extract the fundamental frequency texture component (amplitude > 0.8A_max, where A_max is the maximum amplitude), and the remaining components are classified as abnormal residuals. The transmission attenuation spectrum is decomposed into multi-scale frequency bands using wavelet packet transform (Daubechies-8 wavelet basis) to extract the fiber density change rate (Δρ > ±5%) region. The spatial pyramid matching (SPM) algorithm is used to weight and superimpose the fundamental frequency component and the density change rate at multiple scales (weight: fundamental frequency 0.6, density 0.4) to generate a joint distribution map that integrates surface texture uniformity and internal fiber consistency.

[0044] 203. Based on the mechanical property prediction values ​​output by the quantitative correlation model, extract the coordinates of the areas to be inspected where the tensile strength prediction value is lower than a set threshold, map the coordinates of the areas to be inspected to the corresponding positions of the joint distribution map, and generate candidate defect areas for mechanical and spectral correlation. In step 203, the candidate defect region refers to the set of potential defect spatial coordinates selected by the mechanical performance prediction values, which needs to be further verified to determine whether it is a latent defect.

[0045] The predicted tensile strength (P_ts) is obtained from the quantitative correlation model, and a threshold P_th is set (P_th = 0.85P_design, where P_design is the design strength). Morphological closing operations (structuring element radius r = 5px) are performed on the pixel coordinates where P_ts < P_th to eliminate isolated noise points, generating a mask for the region to be inspected. The mask coordinates are mapped to the corresponding positions in the joint distribution map through affine transformation. Candidate defect regions with mechanical and spectral correlations are then selected based on the pixel confidence score (C > 0.7) of the joint distribution map.

[0046] 204. Within the candidate defect region, the deviation between the predicted mechanical performance value and the actual mechanical test value is calculated using the nonlinear mapping relationship. Regions where the deviation exceeds the upper limit of the braided structure tolerance and are not shown in the directional gradient distribution map are marked as hidden defect regions.

[0047] In step 204, the upper limit of the tolerance for the weave structure refers to the maximum deviation of the mechanical properties of the fabric weaving process from the maximum threshold. If the deviation exceeds the threshold, it is determined to be an unacceptable defect.

[0048] In this embodiment, within the candidate defect region, the deviation (Δ=|P_ts-T_ts| / T_ts) between the predicted mechanical value (P_ts) and the actual uniaxial tensile test value (T_ts) is calculated for each pixel. An upper tolerance limit Δ_max (Δ_max=0.15) is set based on fabric weaving process parameters (such as warp density and weft twist). Simultaneously, the development intensity (G>G_th, G_th=0.2) of the corresponding region in the directional gradient distribution map is analyzed, and regions where Δ>Δ_max and G<G_th are marked as latent defect regions.

[0049] Here is a specific example: Suppose a batch of bamboo fiber / polyester blended fabric needs to be inspected (warp density 32 yarns / cm, weft yarn intersection period 0.28mm). In step 201, visible light imaging (600nm band) generates a directional gradient distribution map with a principal direction angle of 45°±3° and a gradient amplitude >0.15 region accounting for 18%. Near-infrared transmission imaging (1300nm) measures the fiber density change rate ±8% and generates a transmission attenuation spectrum. In step 202, Gabor filtering extracts the fundamental frequency component (frequency 0.31mm⁻¹). Wavelet packet decomposition identifies regions with abnormal density changes (Δρ > 7%), and weighted superposition generates a joint distribution map, locating 5 candidate regions. In step 203, the quantitative model predicts the tensile strength threshold P_th = 153 MPa (design value 180 MPa), and after mapping, 3 candidate defect regions (area > 0.3 mm²) are selected. Through step 204, the actual test value T_ts = 142 MPa (Δ = 0.077), but the directional gradient development intensity G = 0.18 < G_th, which is determined to be a latent defect region.

[0050] Steps 201-204 utilize hyperspectral multi-band collaborative imaging and deep fusion of mechanical and spectral features to overcome the reliance of traditional detection on visible surface defects, accurately locating internal fiber damage and micromechanical anomaly areas; combining nonlinear mapping relationships and process tolerance thresholds, automated judgment of latent defects is achieved, significantly improving the defect detection rate and quality assessment comprehensiveness of complex woven fabrics, and meeting the batch testing needs of high-end textiles.

[0051] To further improve the accuracy of latent defect detection and establish a collaborative judgment mechanism for mechanical properties and visual features, this application achieves accurate marking of latent defects through comparative analysis of quantitative correlation models and measured data, combined with weaving process tolerance thresholds. In some embodiments, within the candidate defect region, the deviation between the predicted mechanical property value and the actual mechanical test value is calculated using the nonlinear mapping relationship. Regions where the deviation exceeds the upper limit of the weaving structure tolerance and are not shown in the directional gradient distribution map are marked as latent defect regions, including: 301. Within the candidate defect area, extract the tensile strength prediction value output by the quantitative correlation model, and perform a uniaxial tensile test on the fabric sample corresponding to the candidate defect area to obtain the actual tensile strength test value. In step 301, the tensile strength prediction value is the theoretical prediction result of the tensile strength of the textile fabric calculated by the quantity correlation model based on the defect morphology parameters.

[0052] Actual tensile strength test value: refers to the true tensile strength data obtained by physical testing of the fabric sample corresponding to the candidate defect area using a uniaxial tensile testing machine.

[0053] In this embodiment, within the candidate defect area, the predicted tensile strength value (P_ts) at the corresponding location is extracted from the output matrix of the quantitative correlation model through spatial coordinate mapping. Simultaneously, fabric samples of standard size (50mm × 50mm) are cut from the same area and subjected to uniaxial tensile testing using an electronic universal testing machine (loading rate 2mm / min, ASTM D5035 standard). The peak tensile force (F_max) at the moment of fracture is recorded, and the actual tensile strength (T_ts = F_max / cross-sectional area) is calculated. To ensure data consistency, a spatial interpolation algorithm is used to convert the test results of discrete samples into a strength distribution map with the same resolution as the predicted value.

[0054] 302. Based on the tensile strength response function in the nonlinear mapping relationship, calculate the degree of deviation between the predicted tensile strength value and the actual tensile strength test value; In step 302, the degree of deviation refers to the proportion of the difference between the predicted tensile strength value and the actual test value, which is used to quantify the accuracy of model prediction and the abnormal strength of defects.

[0055] In this embodiment, based on a predefined tensile strength response function (such as a quadratic polynomial function) in the nonlinear mapping relationship, the residual (Δ=|P_ts-T_ts|) between the predicted value (P_ts) and the actual value (T_ts) is calculated for each pixel. After fitting the residual distribution curve using robust regression (Huber loss function) to eliminate outlier interference, the standardized deviation is calculated (Δ_norm=Δ / (T_ts+ε), ε=1e-5 to prevent division by zero errors). After calculation, the deviation is smoothed using Gaussian filtering (σ=2px) to suppress local noise, outputting a continuously distributed deviation heatmap.

[0056] 303. Based on the upper limit of the tolerance of the weaving structure, set the maximum allowable threshold of the deviation, wherein the maximum allowable threshold is related to the weaving density of the textile fabric and the strength characteristics of the fiber material; In step 303, the maximum allowable threshold refers to the mechanical properties that deviate from the tolerance limit set according to the weaving density (warp / weft density) and tensile strength of the textile fabric. If the deviation exceeds the threshold, it is determined to be an unacceptable defect.

[0057] In this embodiment, the weaving density (e.g., 32 warp yarns / cm) and tensile strength of the fiber material (e.g., 1.2 GPa for bamboo fiber) of the current fabric are obtained through the process database. The maximum allowable threshold Δ_max is calculated using a multi-objective optimization algorithm (NSGA-II). The objective function includes the stability of the weaving structure (the higher the density, the smaller Δ_max) and the fiber strength tolerance (the higher the strength, the larger Δ_max). A fuzzy logic controller (with a trapezoidal distribution of membership function) is introduced to dynamically adjust Δ_max to ensure a strong correlation with the fabric type.

[0058] 304. The region whose deviation exceeds the maximum allowable threshold and is not visible in the directional gradient distribution map is marked as a latent defect region. The latent defect region satisfies the dual judgment conditions of abnormal mechanical properties and being visually invisible.

[0059] In step 304, the dual judgment condition refers to the defect area judgment rule that simultaneously meets the condition that the mechanical properties deviate from the threshold and the visual features are not visible.

[0060] In this embodiment, regions where Δ_norm > Δ_max in the deviation heatmap are marked as mechanically abnormal areas. Simultaneously, the gradient development intensity (G) at the corresponding location is extracted from the directional gradient distribution map, and a visual development threshold G_th = 0.2 (an empirical value) is set. Adjacent abnormal areas are connected using a morphological closing operation (circular structuring element radius r = 3px), and a logical AND operation is performed with low-development areas (G < G_th). Regions that simultaneously satisfy both mechanical abnormality and visual invisibility conditions are eliminated and marked as latent defect areas.

[0061] Here is a specific example: Taking aramid / cotton blended protective clothing fabric as an example (warp density 40 yarns / cm, weft yarn intersection period 0.25mm, aramid content 60%), this fabric is used for the production of high-temperature work protective clothing, and it is necessary to ensure that the internal fibers have no hidden damage in order to avoid deterioration of mechanical properties. Step 301: Extract predicted values ​​P_ts = 210-230 MPa from the candidate defect region (coordinates x = 200-220px, y = 150-170px), and measure T_ts = 195-215 MPa from the cut sample. Interpolate to generate an actual strength distribution map. Step 302: Use Tukey's dual-weight function to fit the residuals and calculate Δ_norm = 0.07-0.18. Generate a heat map showing that the maximum deviation region is located at the coordinate center. Step 303: Based on the aramid tensile strength (3.5 GPa) and high-density weaving characteristics (40 threads / cm), NSGA-II optimization sets Δ_max = 0.10, and the dynamic fuzzy logic is adjusted to 0.11. Step 304: Mark the region (area 0.8 mm²) where Δ_norm > 0.11 and directional gradient development intensity G = 0.15 (< 0.2), and confirm it as a latent defect region generated by internal fracture of aramid fiber.

[0062] Steps 301-304 effectively distinguish between real latent defects and false detection noise by conducting a refined comparative analysis of mechanical prediction and measured data, combined with dynamic threshold setting and multimodal feature logic judgment; and introduce adaptive process parameter optimization and fuzzy logic control to improve the adaptability of detection results to complex weaving structures and fiber materials, thus realizing a technological leap from single mechanical assessment to multi-dimensional collaborative judgment in textile fabric defect detection.

[0063] To address the insufficient accuracy of modeling the correlation between morphological features and mechanical properties in textile fabric defect detection, and to achieve dynamic weight adaptation across defect types, a highly interpretable quantitative correlation model is constructed through three-dimensional morphological quantification, simultaneous testing of multiple mechanical parameters, and nonlinear surface modeling. In some embodiments, the quantitative correlation model between defect morphological parameters and mechanical property parameters, based on the morphological parameters and mechanical property test results of historical defect samples, includes: 401. Extract a set of morphological features containing yarn breakage, weft skew, and stain defects from historical defect samples. Define the equivalent ellipse major axis length of the defect size and the irregularity parameter of the defect shape based on the depth curvature distribution and boundary topology of the morphological feature set. In step 401, the morphological feature set refers to the set of morphological parameters extracted from defect samples such as yarn breakage, weft skew, and stains, including depth curvature distribution (degree of concavity / protrusion of the defect) and boundary topology (geometric complexity of the defect edge). The equivalent ellipse major axis length refers to the actual measured value of the major axis after the defect area is equivalent to an ellipse model, used to quantify the defect size. The irregularity parameter refers to the geometric irregularity index calculated based on the boundary topology, reflecting the degree to which the defect shape deviates from the standard shape.

[0064] In this embodiment, a structured light 3D scanner (accuracy ±0.005mm) is used to acquire point cloud data of the defect area, and a 3D topographic model is generated through Poisson surface reconstruction. The curvature distribution of the reconstructed surface is calculated (the maximum curvature κ_max is used to characterize the depth), and the boundary topological complexity is quantified using a fractal dimension algorithm (box-counting dimension method). The defect area is projected onto a 2D plane after principal component analysis (PCA), and a minimum area enclosing ellipse is fitted. The major axis length is extracted as the equivalent size. Simultaneously, the Fourier descriptor of the boundary contour is calculated, and an irregularity parameter (range 0-1, the larger the value, the more irregular) is generated through energy entropy normalization.

[0065] 402. Perform uniaxial tensile and trapezoidal tear tests on the fabric samples corresponding to the historical defect samples, and simultaneously record the peak tensile strength and the rate of decrease in tear strength to form mechanical property test results corresponding to the morphological feature set. In step 402, the mechanical property test results refer to the set of quantitative mechanical data obtained through uniaxial tensile testing (peak tensile force) and trapezoidal tear testing (strength reduction rate).

[0066] In this embodiment, a bi-column tensile testing machine (ASTM D5035 standard) is used to perform uniaxial tensile tests on the defective specimens. The load is applied at a rate of 5 mm / min until fracture, and the peak tensile strength (F_max) is recorded, along with the strength per unit thickness (σ = F_max / (width × thickness)). For the trapezoidal tear test, a constant speed control (100 mm / min) is used, and a high frame rate camera (1000 fps) captures the crack propagation process, calculating the tear strength reduction rate (ΔF / Δt). To eliminate equipment response delay errors, a Dynamic Time Warping (DTW) algorithm is used to align the mechanical curves with the time series data, generating a standardized dataset of mechanical performance test results.

[0067] 403. A quantitative correlation model is established between the equivalent ellipse major axis length, the irregularity parameter and the mechanical performance test results through a dynamic weight factor allocation mechanism. The dynamic weight factor allocation mechanism adaptively adjusts the influence weight of different morphological features on mechanical performance according to the differences in defect types. In step 403, the dynamic weight factor allocation mechanism refers to a parameter optimization strategy that automatically adjusts the weight of the influence of morphological characteristics on mechanical properties based on the defect type (yarn breakage / weft skew / stain).

[0068] In this embodiment, a random forest classifier (number of decision trees = 200) is constructed to pre-identify defect types (input is depth curvature distribution and irregularity parameters), and the output is a type probability distribution. A dynamic weighting factor allocation mechanism is used to design differentiated weighting strategies for different defect types: yarn breakage emphasizes the equivalent major axis (initial weight 0.7), weft skew emphasizes irregularity (initial weight 0.6), and stains use a balanced weighting (0.5 / 0.5). The weighting coefficients are iteratively optimized using a gradient boosting machine (GBM), with the objective function being the mean squared error (MSE) between the predicted intensity and the actual test value, and the constraint that the total weight sum is always 1, thus establishing a quantitative correlation model.

[0069] 404. Generate a parametric response surface covering a continuous range of values ​​based on the quantitative correlation model. The parametric response surface constructs a nonlinear mapping relationship between defect size, shape and fabric tear strength and tensile strength through an interpolation function, and embeds the nonlinear mapping relationship into the quantitative correlation model.

[0070] In step 404, the parametric response surface refers to a continuous surface that describes the nonlinear mapping relationship in a three-dimensional space consisting of the equivalent major axis length, irregularity parameter and mechanical properties (tear / tensile strength).

[0071] In this embodiment, Kriging interpolation is used to construct the initial response surface. Equivalent major axis (X-axis), irregularity (Y-axis), and mechanical properties (Z-axis) data are input, and spatial correlation is fitted using a semi-variogram. A radial basis function network (RBFN) is introduced to smooth and optimize the interpolation results, with the number of hidden layer nodes equal to the number of training samples, and a Gaussian kernel (bandwidth σ=0.3) as the activation function. The optimized surface is discretized into a high-resolution grid (1000×1000), and a continuous nonlinear mapping relationship is generated through bicubic spline interpolation. Finally, it is stored as a three-dimensional lookup table (LUT) embedded in the quantitative correlation model.

[0072] Here is a specific example: Taking the defect modeling of carbon fiber reinforced bulletproof fabric (warp density 45 yarns / cm, weft shear coating) as an example, in step 401, the yarn breakage defects are scanned in three dimensions (depth curvature κ_max=0.12mm⁻¹, fractal dimension 1.85), with an equivalent ellipse major axis of 6.2mm and an irregularity of 0.78. In step 402, the uniaxial tensile test measures σ=520MPa, trapezoidal tear ΔF / Δt=85N / s, and 3% of outlier data are removed after DTW alignment. In step 403, the random forest classifier identifies a yarn breakage probability of 0.92, and after GBM optimization, the weight equivalent major axis is 0.73 and the irregularity is 0.27. In step 404, Kriging interpolation generates a parametric response surface, and after RBFN optimization, the root mean square error (RMSE) of the surface is reduced to 4.7MPa. The embedded quantitative correlation model supports real-time querying.

[0073] Steps 401-404 significantly improve the physical interpretability of the defect morphology and performance correlation model through three-dimensional morphology quantification and multi-mechanical parameter collaborative analysis; the dynamic weighting mechanism enables accurate modeling across defect types, and the parameter response surface supports nonlinear mapping queries in continuous space, providing high-precision theoretical tools for textile fabric defect detection and process optimization.

[0074] To further improve the adaptability of textile fabric mechanical property prediction models to new materials and processes, this application utilizes fiber chemical property analysis, weaving structure quantification, and transfer learning techniques to achieve dynamic optimization and enhanced cross-material generalization ability of the quantitative correlation model. In some embodiments, the step of dynamically adjusting the quantitative correlation model based on the composition and weaving structure of the novel fabric using transfer learning techniques to adapt to the mechanical response prediction requirements of fabrics with different fiber materials and weave structures includes: 501. Analyze the fiber chemical properties of the composition of the new fabric to obtain a fiber feature set including parameters such as moisture absorption rate and bending stiffness. At the same time, quantitatively extract the warp density and weft yarn intersection distribution of the weaving structure of the new fabric to form a weaving structure feature set. In step 501, fiber chemical property analysis refers to the quantitative testing and analysis of the hygroscopicity (moisture absorption rate) and bending resistance (bending stiffness) of the fiber components of the new fabric. The weaving structure feature set refers to the data set extracted by quantitatively analyzing the warp density (number of yarns per unit length) and weft yarn intersection point distribution (spatial arrangement regularity) in the weaving process of the new fabric.

[0075] In this embodiment, X-ray photoelectron spectroscopy (XPS) was used to analyze the chemical composition of the fiber surface, and dynamic thermomechanical analyzer (DMA) was used to test the fiber bending stiffness (three-point bending method, loading rate 1 mm / min). High-resolution linear array cameras (5 μm / pixel) were used to capture images of the fabric surface, and Hough transform was used to detect the warp and weft yarn intersections. The warp density (number of yarns per centimeter) and the periodicity index of the weft intersections (the percentage of the amplitude of the main peak in the Fourier spectrum) were calculated. The fiber moisture absorption rate was obtained by testing the rate of change of mass in a constant temperature and humidity chamber (ASTM D1776 standard) for 24 hours, ultimately generating a fiber feature set and a weave structure feature set.

[0076] 502. Input the fiber feature set and the weaving structure feature set into a pre-constructed transfer feature projection layer, and use a cross-domain feature matching mechanism to spatially map the new fabric features with the historical defect sample features in the quantitative association model to generate a transfer feature vector with cross-material generalization ability. In step 502, the transfer feature projection layer refers to a neural network layer that maps the features of the new fabric to the feature space of the historical model, used for cross-material feature alignment. The cross-domain feature matching mechanism refers to a technical method that reduces the difference in the distribution of features between the historical model and the new fabric through metric learning.

[0077] In this embodiment, a Generative Adversarial Network (GAN) framework is constructed. The generator concatenates the fiber feature set (moisture absorption rate, bending stiffness) and the weave structure feature set (warp density, periodicity of intersection points) into a high-dimensional vector, which is then projected onto the historical model feature space through a fully connected layer. The discriminator uses Wasserstein distance to measure the similarity of the projected features with the features of historical defect samples. The generator parameters are optimized using the maximum mean difference (MMD) loss function, forcing the new fabric features to align with the distribution of historical features in the latent space, ultimately outputting a transfer feature vector with cross-material generalization ability.

[0078] 503. Based on the migration feature vector, the nonlinear mapping relationship is dynamically constrained, the general mapping parameters that are independent of fiber material in the quantitative correlation model are retained, the target mapping parameters affected by the weaving structure are reset, and an incremental mapping relationship is formed; In step 503, parameter dynamic constraint refers to the technical strategy of selectively freezing or adjusting the parameters of the quantitative correlation model based on the migration feature vector.

[0079] In this embodiment, gradient backpropagation significance analysis (Grad-CAM) is used to identify parameters in the quantitative correlation model that are strongly correlated with fiber material (such as the correlation weight between moisture absorption rate and tear strength). L2 regularization constraints are applied to these parameters to retain universal mapping parameters that are independent of fiber material. For parameters affected by the weaving structure (such as the correlation factor between warp density and tensile strength), a meta-learning (MAML) framework is used for rapid re-initialization. The parameters are iteratively updated based on small sample data (10 sets of samples of the new fabric), ultimately forming an incremental mapping relationship that retains universal rules and adapts to the new structure.

[0080] 504. Based on the incremental mapping relationship and the warp density distribution characteristics in the weaving structure feature set, dynamically adjust the quantitative correlation model and reconstruct the response function of the defect size to the tensile strength to adapt to the fabric mechanical response prediction requirements of different fiber materials and weave structures.

[0081] In step 504, the response function refers to the function that describes the mathematical relationship between defect size and tensile strength, and its parameters are dynamically adjusted according to the incremental mapping relationship and the warp density distribution.

[0082] In this embodiment, based on the incremental mapping relationship, a multi-scale response function based on the attention mechanism is constructed: the defect size (equivalent ellipse major axis) is input into the multi-head self-attention module (number of heads = 8), and the weight allocation under different warp density ranges (such as 20-30 yarns / cm, 30-40 yarns / cm) is calculated; the spatial continuity of the density distribution is captured by the gated recurrent unit (GRU), the slope and intercept of the size and strength curves are dynamically adjusted, the response function is reconstructed, and finally the response function adapted to the new fabric is generated, supporting real-time mechanical performance prediction.

[0083] Here is a specific example: Assuming a batch of aramid / carbon fiber blended stab-resistant fabric is used for bulletproof vest manufacturing, in step 501, XPS measures the aramid moisture absorption rate to be 4.2%, and DMA tests the bending stiffness to be 12.5 mN·mm²; image analysis shows a warp density of 48 yarns / cm and a weft yarn intersection periodicity index of 0.85; in step 502, the GAN projection layer maps the features to the historical Kevlar model space, reducing the MMD loss to 0.12 and generating a transfer feature vector (dimension 128); in step 503, Grad-CAM identifies and freezes the moisture absorption rate-tear strength parameter (weight 0.68), and MAML reinitializes the warp density and tensile strength parameters (from 0.25 to 0.41); in step 504, the multi-head attention module assigns a weight of 0.7 to the high-density region (>45 yarns / cm), and GRU reconstructs the response function, reducing the prediction error by 42% compared to the traditional model.

[0084] Steps 501-504 significantly improve the model's generalization ability to novel fiber materials through chemical property analysis and migration feature space alignment; dynamic parameter constraints and multi-scale response function reconstruction techniques enable precise adaptation of sensitive parameters of the weaving structure, providing reliable technical support for rapid process verification and quality control of high value-added textiles.

[0085] To further improve the decoupling ability of the quantitative correlation model to the influence of fiber chemical properties and weaving structure, and to achieve stable transfer across material models, this application constructs an incremental mapping relationship with strong interpretability through feature-driven parameter classification, dynamic constraints, and joint optimization techniques. In some embodiments, the step of dynamically constraining the nonlinear mapping relationship based on the transfer feature vector, retaining the general mapping parameters in the quantitative correlation model that are independent of fiber material, and resetting the target mapping parameters affected by the weaving structure to form an incremental mapping relationship includes: 601. Based on the fiber moisture absorption rate and bending stiffness parameters in the migration feature vector, identify the set of general mapping parameters that are strongly correlated with the chemical properties of the fiber in the nonlinear mapping relationship; In step 601, the general mapping parameter set specifies the set of parameters in the quantity correlation model that are strongly correlated with the chemical properties of the fiber (moisture absorption rate, bending stiffness), and their values ​​remain stable among different fiber materials.

[0086] In this embodiment, mutual information analysis is used to examine the statistical correlation between the moisture absorption rate and flexural stiffness parameters in the migration feature vector and the parameters of the quantitative correlation model, calculating the mutual information value between the parameters and features (MI > 0.6 indicates a strong correlation). High mutual information parameters are grouped into a general mapping parameter set (such as the correlation weight between shape and tear strength) using hierarchical clustering (Ward's variance minimization algorithm), and the remaining parameters are labeled as the set to be analyzed. The clustering effect of the parameter set is verified using t-SNE dimensionality reduction visualization to ensure the independence of chemically related parameters.

[0087] 602. Based on the distribution density of weft yarn intersection points in the migration feature vector, locate the target mapping parameter set that is affected by the periodicity of the weaving structure in the nonlinear mapping relationship; In step 602, the target mapping parameter set specifies the set of parameters in the correlation model that are significantly affected by the weaving structure characteristics (distribution density of weft yarn intersections), and needs to be dynamically adjusted as the weaving process changes.

[0088] In this embodiment, a periodicity influence factor for the weft yarn intersection point distribution density (spatial Fourier spectrum main peak amplitude) is constructed based on the weft yarn intersection point distribution density (spatial Fourier spectrum main peak amplitude) in the migration feature vector. The quantitative correlation model parameters are divided into a high-sensitivity group (CPI > 0.7) and a low-sensitivity group (CPI < 0.3) according to their CPI sensitivity using spectral clustering (Normalized Cut algorithm). The high-sensitivity group parameters are then screened using LASSO regression (λ = 0.01), and redundant parameters are removed to form a target mapping parameter set (such as the correlation factor between size and tensile strength).

[0089] 603. Apply parameter freezing constraints to the general mapping parameter set to maintain the numerical invariance of the general mapping parameter set when the fiber material changes, and perform parameter space reconstruction on the target mapping parameter set to generate the reconstructed target mapping parameter set. In step 603, parameter freezing constraint refers to the technical means of fixing the values ​​of the general mapping parameter set to prevent the loss of common rules due to changes in fiber material. Parameter space reconstruction refers to the technical process of re-initializing the target mapping parameter set according to the weaving structure characteristics of the new fabric.

[0090] In this embodiment, gradient masking is applied to the general mapping parameter set to shield its gradient update signal during backpropagation, ensuring the invariance of parameter values. For the target mapping parameter set, Bayesian optimization (Gaussian process surrogate model) is used to search for optimal initial values ​​within a preset parameter space (e.g., correlation factor range of 0.1-0.9), with the objective function being the mean square error (MSE) of the mechanical prediction of the new fabric sample. One hundred candidate parameter combinations are generated through Monte Carlo sampling, and the top 5% with the smallest MSE are selected as the reconstructed target mapping parameter set.

[0091] 604. Jointly optimize the parameter freezing constraint results with the reconstructed target mapping parameter set to generate an incremental mapping relationship that retains fiber material independence and is adapted to the weaving structure.

[0092] In step 604, joint optimization refers to the global model tuning process that integrates the frozen parameters and the reconstructed parameters.

[0093] In this embodiment, a dual-channel adversarial training framework is constructed. The generator outputs a joint vector of frozen and reconstructed parameters, and the discriminator distinguishes the similarity between the historical model parameter distribution and the joint vector. Cycle-consistency loss is used to constrain the consistency of the model's prediction results before and after parameter adjustment, while contrastive loss is introduced to enhance the discriminativeness of the parameter set. The AdamW optimizer (learning rate 0.001, weight decay 0.01) is used for iterative optimization until the loss converges, generating an incremental mapping relationship that retains fiber material independence and adapts to the weave structure.

[0094] Here is a specific example: Taking aramid flame-retardant fabric (used in fire suit manufacturing) as an example, this fabric contains aramid fiber (moisture absorption rate 3.8%, bending stiffness 9.6 mN·mm²) blended with basalt fiber. The warp density of the weave structure is 42 yarns / cm, and the periodicity index of the weft yarn intersection distribution density is 0.88. In step 601, mutual information analysis identifies the moisture absorption rate and tear strength parameters (MI=0.72) and classifies them into the general mapping parameter set, while the bending stiffness and shape parameters (MI=0.65) are simultaneously included. Through step 602, spectral clustering... The size and tensile strength parameters (CPI=0.82) were separated into a target set. LASSO regression was used to remove three low-sensitivity parameters to locate the target mapping parameter set. In step 603, Bayesian optimization was used to search for the optimal value of the target parameters (association factor 0.47). Monte Carlo sampling was used to select the reconstructed target mapping parameter set. In step 604, the cyclic loss was reduced to 0.15 after adversarial training. After joint optimization, an incremental mapping relationship was generated, which reduced the prediction error of the tensile strength of aramid fabric by 55% compared with the baseline model.

[0095] Steps 601-604 use parameter classification and dynamic optimization mechanisms to accurately distinguish the influence weights of fiber chemical properties and weaving structure on the model; joint adversarial training ensures the balance between parameter freezing and reconstruction of incremental mapping relationships, significantly improving the model's adaptability to high-performance blended fabrics and providing reliable technical support for rapid process iteration of special textiles.

[0096] To further improve the comprehensiveness and accuracy of textile fabric quality grading and achieve collaborative analysis of multi-source heterogeneous data, this application constructs an intelligent grading system integrating surface quality, internal structure, and production stability through fine spectral feature segmentation, dynamic conversion of process parameters, and defect distribution quantification technology. In some embodiments, the process involves inputting the spectral feature data acquired by the hyperspectral imaging system, the process parameters of the current production batch, and the spatial distribution characteristics of the latent defect region into a pre-trained residual network model. This model is then weighted and fused into a comprehensive quality index through feature channel weighting. Based on this comprehensive quality index and the nonlinear mapping relationship, the same batch of textile fabrics is rapidly graded, including: 701. Divide the spectral feature data acquired by the hyperspectral imaging system into a visible light feature subset and a near-infrared feature subset according to the spectral band; In step 701, the visible light feature subset refers to the data subset collected by the hyperspectral imaging system in the visible light band (400-700nm), reflecting visual characteristics such as the surface color and texture uniformity of the textile fabric. The near-infrared feature subset refers to the data subset collected by the hyperspectral imaging system in the near-infrared band (700-1700nm), characterizing the chemical structural features of the fiber, such as crystallinity and hygroscopicity.

[0097] In this embodiment, the Spectral Angle Mapping (SAM) algorithm is used to group the hyperspectral data by band: the similarity between the spectral curve of each band and the standard visible / near-infrared reference curve is calculated (cosine angle <15° is assigned to the visible subset, >75° is assigned to the near-infrared subset). The visible subset is superpixel segmented (SLIC algorithm, superpixel size 32×32px), and the HSV color histogram of each superpixel (H channel bin number = 36) is extracted to obtain the visible feature subset; the near-infrared subset is filtered using the Continuous Projection Algorithm (SPA) to select feature bands strongly correlated with fiber crystallinity (such as 1450nm, 1550nm), and the absorption peak integral intensity of each pixel is calculated to obtain the near-infrared feature subset.

[0098] 702. Convert the process parameters of the current production batch into a set of process influencing factors, wherein the set of process influencing factors includes the yarn tension fluctuation coefficient and the loom speed uniformity index. In step 702, the process influence factor set refers to the set of standardized influence indicators that convert the process parameters (yarn tension, loom speed) of the production batch into standardized influence indicators, which are used to quantify the impact of process fluctuations on quality.

[0099] In this embodiment, wavelet packet decomposition (db4 wavelet basis, decomposition level = 5) is performed on the yarn tension sensor data (sampling rate 1kHz) to extract the standard deviation of the low-frequency components as the tension fluctuation coefficient. The loom speed data is reconstructed in phase space (delay time τ = 10, embedding dimension m = 3) to calculate the maximum Lyapunov exponent, reflecting speed uniformity. Min-Max normalization is used to map the original parameters to the [0,1] interval, generating a set of process influencing factors. Simultaneously, Pearson correlation analysis (|r| > 0.7) is performed to remove redundant parameters and retain key influencing factors.

[0100] 703. The spatial distribution characteristics of the latent defect region are quantified into a defect density distribution map, which reflects the number and degree of aggregation of latent defects per unit area. In step 703, the defect density distribution map refers to converting the spatial distribution of the latent defect area into a heat map that reflects the degree of defect aggregation. The higher the value, the greater the defect density per unit area.

[0101] In this embodiment, kernel density estimation (Epanechnikov kernel function, bandwidth h=5px) is performed on the coordinates of the latent defect region to calculate the defect density value of each pixel. Morphological opening operation (circular structuring element radius r=3px) is used to eliminate noise points, and the density levels are divided using the watershed algorithm (low density <0.1, medium density 0.1-0.3, high density >0.3). The density levels are mapped to the HSV color space (H=0-120° corresponding to low to high density) to generate a defect density distribution map and register it with the original image space.

[0102] 704. Input the visible light feature subset, the near-infrared feature subset, the process influence factor set, and the defect density distribution map into the pre-trained residual network model, generate a comprehensive quality index through a weighted fusion mechanism of multi-source feature channels, and quickly grade and output the same batch of textile fabrics based on the comprehensive quality index and the nonlinear mapping relationship.

[0103] In step 704, the weighted fusion mechanism of multi-source feature channels refers to the technical strategy of dynamically assigning weights to feature channels from different data sources (spectrum, process, defect).

[0104] In this embodiment, a dual-path attention mechanism is introduced into the pre-trained ResNet-50 model. The spatial attention module (SAM) focuses on areas with high defect density, while the channel attention module (CBAM) dynamically allocates weights for visible light, near-infrared, process, and defect features. A gated recurrent unit (GRU) captures the temporal correlation of features (such as the hysteresis effect between loom speed fluctuations and defect density), outputs a comprehensive quality index (0-100 points), and combines the defect size and tensile strength curves in the nonlinear mapping relationship to set a grading threshold (superior grade > 90 points and tensile strength > 90% of design value).

[0105] Here is a specific example: Taking the quality grading of flame-retardant aramid fire-fighting suit fabric (warp density 40 yarns / cm, weft anti-drip coating) as an example, this fabric must ensure no hidden carbonization defects and meet mechanical performance standards. Through step 701, the visible light band (550nm) superpixel segmentation shows a color uniformity score of 92, and the near-infrared (1450nm) absorption peak intensity reflects a crystallinity index of 0.89. Through step 702, the yarn tension fluctuation coefficient is obtained as 0.08 (fluctuation range ±2.5N), and the loom speed uniformity index is 0.93 (Lyapunov index 0.12). After normalization... The process is converted into a set of process influence factors. In step 703, kernel density estimation identifies three high-density defect areas (density > 0.25), accounting for 0.7% of the total area, and the spatial distribution characteristics are quantified into a defect density distribution map. In step 704, CBAM assigns weights of 0.5 for visible light, 0.3 for near-infrared, 0.15 for process, and 0.05 for defects, with a comprehensive score of 94. Combined with the size and strength curve (a 4mm defect corresponds to a tensile strength of 210MPa, while the design value is 230MPa), it is judged as a qualified product (score > 90 but strength does not reach 95% of the design value).

[0106] Steps 701-704 achieve multi-dimensional collaborative evaluation of textile fabric quality through multi-source feature refinement and dynamic weight fusion; combined with process parameter fluctuation analysis and defect spatial distribution quantification, the sensitivity of grading results to production stability is significantly improved, providing a full-chain solution for the quality control of high-end protective textiles from data collection to intelligent decision-making.

[0107] Figure 2 This application provides a schematic diagram of the structure of a batch testing system for the production performance of textile fabrics, as shown in the embodiment of the present application. Figure 2 As shown, the system includes: Modeling module 21 is used to establish a quantitative correlation model between defect morphological parameters and mechanical performance parameters based on the morphological parameters and mechanical performance test results of historical defect samples. The quantitative correlation model includes a nonlinear mapping relationship between defect size, shape and fabric tear strength and tensile strength. The adjustment module 22 is used to dynamically adjust the quantitative correlation model according to the composition and weaving structure of the new fabric through transfer learning technology, so as to adapt to the mechanical response prediction requirements of fabrics with different fiber materials and weave structures. The dynamic adjustment process depends on the generalization ability of the nonlinear mapping relationship. The identification module 23 is used to simultaneously acquire the surface texture features and internal fiber distribution features of the new fabric using a hyperspectral imaging system, and combine the mechanical property prediction values ​​output by the quantitative correlation model and the nonlinear mapping relationship to identify hidden defect areas that are not detected in conventional visual inspection. The generation module 24 is used to input the spectral feature data acquired by the hyperspectral imaging system and the process parameters of the current production batch into the pre-trained residual network model, generate a comprehensive quality index through weighted fusion of feature channels, and quickly classify and output the same batch of textile fabrics based on the comprehensive quality index and the nonlinear mapping relationship.

[0108] Figure 2 The aforementioned batch testing system for the production performance of textile fabrics can perform... Figure 1 The implementation principle and technical effects of the batch testing method for the production performance of textile fabrics described in the illustrated embodiment will not be repeated here. The specific operation methods of each module and unit in the batch testing system for the production performance of textile fabrics in the above embodiments have been described in detail in the embodiments related to this method, and will not be elaborated upon here.

[0109] In one possible design, Figure 2 The batch testing system for the production performance of textile fabrics shown in the embodiment can be implemented as a computing device, such as... Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32; The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are invoked and executed by the processing component 32.

[0110] The processing component 32 is used for the above Figure 1 The embodiment describes a method for batch testing of the production performance of textile fabrics.

[0111] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above-described method. Alternatively, the processing component may be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above-described method.

[0112] Storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0113] Of course, computing devices may also include other components, such as input / output interfaces, display components, communication components, etc.

[0114] Input / output interfaces provide interfaces between processing components and peripheral interface modules, which can be output devices, input devices, etc.

[0115] The communication components are configured to facilitate wired or wireless communication between computing devices and other devices.

[0116] The computing device can be a physical device or an elastic computing host provided by a cloud computing platform. In this case, the computing device can refer to a cloud server, and the aforementioned processing components, storage components, etc., can be basic server resources rented or purchased from the cloud computing platform.

[0117] This application also provides a computer storage medium storing a computer program, which, when executed by a computer, can perform the above-described functions. Figure 1 The XX method of the illustrated embodiment.

[0118] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0119] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0120] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for batch testing of the production performance of textile fabrics, characterized in that, include: Based on the morphological parameters and mechanical performance test results of historical defect samples, a quantitative correlation model between defect morphological parameters and mechanical performance parameters is established. The quantitative correlation model includes a nonlinear mapping relationship between defect size, shape and fabric tear strength and tensile strength. Based on the composition and weaving structure of the new fabric, the quantitative correlation model is dynamically adjusted using transfer learning technology to adapt to the mechanical response prediction requirements of fabrics with different fiber materials and weave structures. The dynamic adjustment process depends on the generalization ability of the nonlinear mapping relationship. The surface texture features and internal fiber distribution features of the novel fabric are simultaneously acquired using a hyperspectral imaging system. Combined with the mechanical property predictions output by the quantitative correlation model and the nonlinear mapping relationship, hidden defect areas that are not detected by conventional visual inspection are identified. The spectral feature data acquired by the hyperspectral imaging system, the process parameters of the current production batch, and the spatial distribution characteristics of the hidden defect area are input into a pre-trained residual network model. The model is then fused into a comprehensive quality index through weighted feature channels. Based on the comprehensive quality index and the nonlinear mapping relationship, the same batch of textile fabrics is quickly graded.

2. The method according to claim 1, characterized in that, The method involves simultaneously acquiring the surface texture features and internal fiber distribution features of the novel fabric using a hyperspectral imaging system, combining the predicted mechanical properties output by the quantitative correlation model with the nonlinear mapping relationship, to identify hidden defect areas that are not detected by conventional visual inspection, including: In the visible light band of the hyperspectral imaging system, a high spatial resolution imaging mode is set to obtain the directional gradient distribution map of surface texture features. At the same time, the transmission imaging mode is switched in the near-infrared band to generate a transmission attenuation map of the internal fiber distribution features based on the optical absorption characteristics of the fiber. The directional gradient distribution map is spatially decomposed in the frequency domain to separate the fundamental frequency texture component and the abnormal frequency band residual component that are consistent with the weaving cycle of the textile fabric. The fiber density change rate in the transmission attenuation spectrum is superimposed with the fundamental frequency texture component to generate a joint distribution map that integrates surface and internal features. Based on the mechanical property prediction values ​​output by the quantitative correlation model, the coordinates of the areas to be inspected where the tensile strength prediction value is lower than a set threshold are extracted, and the coordinates of the areas to be inspected are mapped to the corresponding positions of the joint distribution map to generate candidate defect areas with mechanical and spectral correlation. Within the candidate defect region, the deviation between the predicted mechanical performance value and the actual mechanical test value is calculated using the nonlinear mapping relationship. Regions where the deviation exceeds the upper limit of the braided structure tolerance and are not shown in the directional gradient distribution map are marked as latent defect regions.

3. The method according to claim 2, characterized in that, Within the candidate defect region, the deviation between the predicted mechanical performance value and the actual mechanical test value is calculated using the nonlinear mapping relationship. Regions where the deviation exceeds the upper tolerance limit of the braided structure and are not shown in the directional gradient distribution map are marked as latent defect regions, including: Within the candidate defect area, the predicted tensile strength value output by the quantitative correlation model is extracted, and a uniaxial tensile test is performed on the fabric sample corresponding to the candidate defect area to obtain the actual tensile strength test value. Based on the tensile strength response function in the nonlinear mapping relationship, the degree of deviation between the predicted tensile strength value and the actual tensile strength test value is calculated. Based on the upper limit of tolerance of the weave structure, a maximum allowable threshold for the degree of deviation is set, and the maximum allowable threshold is related to the weave density of the textile fabric and the strength characteristics of the fiber material. The region whose deviation exceeds the maximum allowable threshold and is not visible in the directional gradient distribution map is marked as a latent defect region. The latent defect region satisfies the dual criteria of abnormal mechanical properties and being visually invisible.

4. The method according to claim 1, characterized in that, The quantitative correlation model between defect morphological parameters and mechanical performance parameters is established based on the morphological parameters and mechanical performance test results of historical defect samples, including: Extract a set of morphological features containing yarn breakage, weft skew, and stain defects from historical defect samples. Based on the depth curvature distribution and boundary topology of the morphological feature set, define the equivalent ellipse major axis length of the defect size and the irregularity parameter of the defect shape. Uniaxial tensile and trapezoidal tear tests were performed on the fabric samples corresponding to the historical defect samples, and the peak tensile strength and the rate of decrease in tear strength were recorded simultaneously to form mechanical property test results corresponding to the morphological feature set. A quantitative correlation model is established between the equivalent ellipse major axis length, the irregularity parameter and the mechanical performance test results through a dynamic weight factor allocation mechanism. The dynamic weight factor allocation mechanism adaptively adjusts the influence weight of different morphological characteristics on mechanical performance according to the differences in defect types. Based on the quantitative correlation model, a parametric response surface covering a continuous range of values ​​is generated. The parametric response surface constructs a nonlinear mapping relationship between defect size, shape and fabric tear strength and tensile strength through an interpolation function, and embeds the nonlinear mapping relationship into the quantitative correlation model.

5. The method according to claim 1, characterized in that, The step of dynamically adjusting the quantitative correlation model based on the composition and weaving structure of the novel fabric using transfer learning technology to adapt to the fabric mechanical response prediction requirements of different fiber materials and weave structures includes: The fiber chemical properties of the new fabric composition are analyzed to obtain a fiber feature set including parameters such as moisture absorption rate and bending stiffness. At the same time, the warp density and weft yarn intersection distribution of the new fabric are quantitatively extracted to form a weaving structure feature set. The fiber feature set and the weaving structure feature set are input into a pre-constructed transfer feature projection layer. Through a cross-domain feature matching mechanism, the features of the new fabric are spatially mapped to the historical defect sample features in the quantitative correlation model to generate a transfer feature vector with cross-material generalization ability. Based on the migration feature vector, the nonlinear mapping relationship is dynamically constrained, the general mapping parameters that are independent of fiber material in the quantitative correlation model are retained, the target mapping parameters affected by the weaving structure are reset, and an incremental mapping relationship is formed. Based on the incremental mapping relationship and the warp density distribution characteristics in the weaving structure feature set, the quantitative correlation model is dynamically adjusted to reconstruct the response function of the defect size to the tensile strength, so as to adapt to the fabric mechanical response prediction requirements of different fiber materials and weave structures.

6. The method according to claim 5, characterized in that, The step of dynamically constraining the nonlinear mapping relationship based on the migration feature vector, retaining the general mapping parameters in the quantitative correlation model that are independent of fiber material, resetting the target mapping parameters affected by the weaving structure, and forming an incremental mapping relationship includes: Based on the fiber moisture absorption rate and bending stiffness parameters in the migration feature vector, identify the set of general mapping parameters that are strongly correlated with fiber chemical properties in the nonlinear mapping relationship; Based on the distribution density of weft yarn intersection points in the migration feature vector, locate the target mapping parameter set that is affected by the periodicity of the weaving structure in the nonlinear mapping relationship; Apply parameter freezing constraints to the general mapping parameter set to maintain the numerical invariance of the general mapping parameter set when the fiber material changes, and perform parameter space reconstruction on the target mapping parameter set to generate the reconstructed target mapping parameter set. The parameter freezing constraint results are jointly optimized with the reconstructed target mapping parameter set to generate an incremental mapping relationship that retains fiber material independence and is adapted to the weaving structure.

7. The method according to claim 1, characterized in that, The process involves inputting the spectral feature data acquired by the hyperspectral imaging system, the process parameters of the current production batch, and the spatial distribution characteristics of the latent defect region into a pre-trained residual network model. This model is then weighted and fused into a comprehensive quality index through feature channel weighting. Based on this comprehensive quality index and the nonlinear mapping relationship, the same batch of textile fabrics is rapidly graded. This includes: The spectral feature data acquired by the hyperspectral imaging system are divided into a visible light feature subset and a near-infrared feature subset according to the spectral band. The process parameters of the current production batch are converted into a set of process influencing factors, which includes the yarn tension fluctuation coefficient and the loom speed uniformity index. The spatial distribution characteristics of the latent defect region are quantified into a defect density distribution map, which reflects the number and degree of aggregation of latent defects per unit area. The visible light feature subset, the near-infrared feature subset, the process influence factor set, and the defect density distribution map are input into a pre-trained residual network model. A comprehensive quality index is generated through a weighted fusion mechanism of multi-source feature channels. Based on the comprehensive quality index and the nonlinear mapping relationship, the same batch of textile fabrics is quickly graded and output.

8. A batch testing system for the production performance of textile fabrics, characterized in that, include: The modeling module is used to establish a quantitative correlation model between the morphological parameters and mechanical performance parameters of defects based on the morphological parameters and mechanical performance test results of historical defect samples. The quantitative correlation model includes a nonlinear mapping relationship between defect size, shape and fabric tear strength and tensile strength. The adjustment module is used to dynamically adjust the quantitative correlation model based on the composition and weaving structure of the new fabric using transfer learning technology, so as to adapt to the mechanical response prediction requirements of fabrics with different fiber materials and weave structures. The dynamic adjustment process depends on the generalization ability of the nonlinear mapping relationship. The identification module is used to simultaneously acquire the surface texture features and internal fiber distribution features of the new fabric using a hyperspectral imaging system, and combine the mechanical property prediction values ​​output by the quantitative correlation model and the nonlinear mapping relationship to identify hidden defect areas that are not detected by conventional visual inspection. The generation module is used to input the spectral feature data acquired by the hyperspectral imaging system and the process parameters of the current production batch into a pre-trained residual network model, generate a comprehensive quality index through weighted fusion of feature channels, and quickly classify and output the same batch of textile fabrics based on the comprehensive quality index and the nonlinear mapping relationship.

9. A computing device, characterized in that, It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement a batch testing method for the production performance of textile fabrics as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that, The device contains a computer program that, when executed by a computer, implements a batch testing method for the production performance of textile fabrics as described in any one of claims 1 to 7.

Images