Non-woven fabric surface quality detection method, equipment and medium

Abstract

The invention discloses a non-woven fabric surface quality detection method, non-woven fabric surface quality detection equipment and a medium. According to the invention, the detection device is installed on the production line, real-time online detection is carried out in the production process, the device has the functions of immediately reporting the detection condition, immediately giving an alarm when an abnormality occurs, and reminding on-duty personnel to immediately find out and eliminate the problem, and the defect position is recorded and marked to facilitate tracing. By acquiring stress-optical response multi-modal data, stress distribution and deformation information reflecting internal structural characteristics of a material are covered, optical, texture and dynamic behavior characteristics representing a surface state are also included, an input basis is provided for subsequent analysis through standardized preprocessing, a detection blind area under a traditional single data dimension is effectively avoided, and the detection accuracy is improved. And it is ensured that all types of defect information from structural defects, surface pollutants to color difference texture anomalies can be completely captured. Precise identification of slight chromatic aberration and texture disorder which are difficult to distinguish by traditional visual detection is realized.

Description

Technical Field

[0001] This invention belongs to the field of nonwoven fabric testing technology, specifically a method, equipment, and medium for testing the surface quality of nonwoven fabrics. Background Technology

[0002] Nonwoven fabric surface quality inspection is a crucial step in the high-speed, high-precision identification and classification of defects on the surface of nonwoven fabrics during production or in finished products, utilizing automated technologies such as machine vision and laser scanning. This technology captures images of the fabric surface using industrial cameras and then employs advanced image processing algorithms to automatically detect and mark various defects such as holes, stains, uneven thickness, and foreign objects, replacing traditional manual visual inspection. It achieves 100% inspection, effectively overcoming problems such as human eye fatigue and inconsistent standards, and significantly improving inspection efficiency and accuracy. The inspection results not only provide a basis for real-time control of the production process but also ensure the quality stability of the final product, making it an indispensable quality control tool on modern intelligent nonwoven fabric production lines.

[0003] However, traditional detection methods often rely on a single data dimension, which can easily lead to blind spots in detection. Traditional algorithms have limited ability to identify complex defects and are difficult to accurately distinguish between minute contaminants or slight color differences and texture anomalies, resulting in low overall detection efficiency and making it difficult to meet the demand for high-precision, rapid detection of all types of defects. Summary of the Invention

[0004] The purpose of this invention is to provide a method, equipment, and medium for testing the surface quality of nonwoven fabrics in order to solve the problems mentioned above.

[0005] The technical solution adopted in this invention is as follows: a method for detecting the surface quality of nonwoven fabrics, the method comprising the following steps: S1: Install the detection device on the production line to perform real-time online detection during the production process; the production line continuously conveys non-woven fabric, and the non-woven fabric is positioned in real time by a dynamic pneumatic suspension positioning device. The multispectral light source array, together with a 4K high-speed camera, synchronously collects a three-dimensional data matrix and records the collection position coordinates and timestamp. S2: Programmable piezoelectric vibrators on the production line conveyor path apply stress to dynamic samples. A micro pressure sensor array and a high-speed camera synchronously track the moving samples, record the pressure distribution in real time, capture the trajectory of foreign objects, and combine multispectral data to distinguish contaminants and mark the location of defects. S3: Input the initial 3D data matrix from step S1 and the stress-optical response data from step S2 into the depth confidence network, and perform collaborative analysis through sub-modules: The structural defect module is used to identify holes, tears, and nodules based on anomalies in pressure distribution and deformation images. Using a surface contaminant module that combines multispectral fluorescence properties and foreign object displacement trajectories, stains, insects, and metallic impurities are classified. The color difference texture module is used to analyze color difference through the CIE Lab color space, and Fourier transform is used to extract texture periodic features. S4: For the high-risk defect areas identified in step three, start the micro-robotic arm to perform targeted testing; S5: Based on the type and quantity of defects and the results of physical performance tests, products are categorized as follows: Excellent: ≤1 minor defect; Acceptable: ≤3 minor defects; Defects: ≥1 critical defect of three levels.

[0006] In a preferred embodiment, step S2, specifically recording the surface pressure distribution in real time using a micro pressure sensor array, includes: Step S2-1A: Array the miniature pressure sensors into an M×N grid, and collect raw pressure data under dynamic stress in real time, denoted as matrix P. raw (i,j,t), where i=1~M, j=1~N are sensor coordinates, t=1~T is the time series, and T=vibration period × sampling frequency; The formula for calculating the removal of high-frequency noise by applying a moving average filter to the original data is as follows: ; In the formula: i,j represent the row and column coordinate indices of the sensor array: i is the row and j is the column; t represents the time index of the current data frame, corresponding to the moment when the dynamic stress was applied; k is the size of the filtering window, and k = 5 to 10 frames is selected to balance real-time performance and smoothing effect. s represents the index of the historical frame within the sliding window: from t-k+1 to t; Let be the filtered pressure value of sensor (i,j) at time t; Step S2-2A: Perform gridded interpolation on the pressure distribution. Using sensor data as discrete points, a continuous pressure field needs to be generated through interpolation to facilitate the analysis of spatial distribution characteristics. The inverse distance weighted interpolation method is used to map the sensor pressure values ​​to a two-dimensional grid in the detection area. The formula for calculating the pressure value at each grid point is: : In the formula: P(x,y,t) represents the pressure value at coordinate (x,y) in the interpolated continuous pressure field at frame t. (x,y) represents the two-dimensional grid coordinates within the detection area; M and N represent the total number of rows and columns of the sensor array; w ij (x,y) represents the weight value of the (i,j)th sensor for the grid point (x,y); d ij (x,y) represents the Euclidean distance between the (i,j)th sensor and the grid point (x,y); p represents the distance decay exponent; Step S2-3A: Calculate the two-dimensional gradient of the continuous pressure field P(x,y,t) using the salient features of stress concentration points to quantify the severity of pressure changes; The formula for calculating the pressure gradient modulus is: ; In the formula: (x,y,t) is a three-dimensional spatiotemporal coordinate variable used to accurately locate the pressure field characteristics of a specific location on the surface of a nonwoven fabric at a specific time. G(x,y,t) is the magnitude of the pressure gradient at coordinate (x,y) in the t-th frame in a continuous pressure field; The partial derivatives of the pressure field in the x and y directions represent the instantaneous rate of change of the pressure field in space; in Where Δx is the distance between adjacent grid points in the x-direction, and (x+Δx,y,t) and (x−Δx,y,t) are the coordinates of the adjacent grid points on the left and right sides of (x,y,t); Similarly, Δy is used to calculate the partial derivative in the y-direction: ; Step S2-4A: Perform local maximum detection and dynamic fluctuation analysis. In a single frame of the pressure field, identify local maximum points and denote them as the candidate point set S. cand (x, y, t), combined with the dynamic stress of the low-frequency sinusoidal vibration in step S2, calculate the pressure fluctuation amplitude of the candidate point in the time series. Due to the instability of the fiber structure, the fluctuation amplitude in the weak area is significantly higher than that in the normal area; the formula for calculating the standard deviation of pressure fluctuation is: ; In the formula: t represents a specific frame in the time series during dynamic stress detection, used to traverse the pressure data throughout the entire detection cycle, and is a discrete-time variable for summation: from frame 1 to frame T. P(x,y,t) represents the instantaneous pressure value of the grid point at coordinates (x,y) on the surface of the nonwoven fabric at time t. It is the core data of the dynamic pressure field and directly participates in the calculation of the standard deviation of the fluctuation. σ(x,y) represents the standard deviation of pressure fluctuation at grid point (x,y) within the time series T frames; This represents the average pressure value of the grid point (x, y) within the time series T frames; T represents the total number of frames in the time series, covering at least one complete cycle of dynamic stress; Step S2-5A: Based on three indicators—grid point pressure value, pressure gradient magnitude, and dynamic fluctuation standard deviation—set a judgment threshold. Static threshold: μP is the average pressure across the entire field, and σP is the standard deviation of the pressure across the entire field, based on statistics from normal samples. Gradient threshold: G(x,y,t) > G0; G0 is an empirical threshold; Dynamic threshold: σ(x,y)>μ σ +2σ σ， μσ is the mean of the standard deviation of the overall fluctuation, and σ σ Its standard deviation; Points that meet the above three conditions are identified as stress concentration points, and their coordinates (x, y) and corresponding pressure characteristic values ​​are output.

[0007] In a preferred embodiment, step S2, which involves using a high-speed camera to capture the displacement trajectory of surface foreign objects during vibration and combining this with multispectral data to distinguish the type of contaminant, specifically includes the following steps: Step S2-1B: Multispectral Image Preprocessing: Input: Time-series image sequence I acquired by a high-speed multispectral camera λ (x p ,y p ,t); λ is the wavelength, x p ,y p (where t is the image pixel coordinate and t is the time frame); Purpose: To eliminate image noise and register images of different wavelengths to the same spatial coordinate system; Step S2-2B: Segment the foreign object region from the background and extract its pixel coordinates: Step S2-3B: Calculate the displacement trajectory of the foreign object, track the position change of the same foreign object in the time series image, calculate the displacement, and distinguish between static impurities and dynamic foreign objects; Step S2-4B: Perform multispectral reflectance feature extraction. For dynamic foreign object regions, extract their reflectance curves in the multispectral bands as a basis for type differentiation.

[0008] In a preferred embodiment, step S3, identifying holes, tears, and nodule defects using the structural defect module based on pressure distribution anomalies and deformation images, includes the following steps: Step S3-1A: Perform data preprocessing and spatiotemporal registration of the pressure field and deformation field: Step S3-2A: Identify abnormal areas through the statistical characteristics of the pressure field. First, calculate the pressure standard deviation field and calculate the pressure fluctuation standard deviation of each spatial point (x,y) in the time series to reflect pressure stability. Step S3-3A: Calculation through deformation geometry characteristics: Differentiate defect types: Calculate the first spatial derivative of the deformation field to reflect the deformation rate; Step S3-4A: Perform multi-feature fusion and defect classification; Input: Pressure anomaly region ΩP(x,y), deformation gradient ∇U(x,y)=(∇U x ,∇U y ), Gaussian curvature K(x,y); For Ω P For the region (x,y)=1, calculate the following indices: Pressure anomaly level: Dimensionless >1 indicates significant pressure fluctuations; Deformation gradient mode: Dimensionless, reflecting the degree of drastic deformation changes; The curvature symbol sgn(K) represents positive curvature, and -1 represents negative curvature. The defect classification rules are as follows: Classification based on indicator combinations: Hole: A P >1.5 and And sgn(K) = −1; Tear: A P >1.2 and |∣∇U∣>0.1 and sgn(K)=0; Nodules: AP > 1.0 and And sgn(K)=1.

[0009] In a preferred embodiment, step S3, which uses a surface contaminant module to classify stains, insects, and metallic impurities by combining multispectral fluorescence characteristics and foreign object displacement trajectories, specifically includes: Step S3-1B: First, the multispectral fluorescence image and the foreign object motion trajectory data need to be spatiotemporally aligned to ensure consistency in subsequent feature extraction; Step S3-2B: Based on the differences in fluorescence emission characteristics of pollutants at different wavelengths, three key features are extracted to distinguish the composition of substances: Step S3-3B: By analyzing the motion patterns of foreign objects in consecutive frames, two types of dynamic features are extracted to distinguish motion modes: Step S3-4B: Combining the rules of fluorescence characteristics and trajectory characteristics, the three types of pollutants are classified: Stains: Exhibit fluorescence peaks, blue to green light intensity ratio greater than 2.5, average speed less than 50 pixels / second, and trajectory curvature of 0; Mosquitoes: exhibit fluorescence peaks, intensity ratios of 1.2-2.0, speed fluctuations ranging from 200-800 pixels / second, and trajectory curvature greater than 0.05; Metallic impurities: No fluorescence peaks, average velocity exceeding 500 pixels / second, and trajectory curvature less than 0.01.

[0010] In a preferred embodiment, step S3, which uses a color difference texture module to analyze color difference through the CIE Lab color space and extracts texture periodic features using Fourier transform, specifically includes the following steps: Step S3-1C: Extract color difference features based on CIE Lab space to quantify the color difference between the target area and the normal area; first, calculate the brightness difference, that is, the difference between the average L value of the target area and the normal area, reflecting the change in brightness of the area; then calculate the difference between the a channel and the b channel respectively; finally, integrate the differences of the three channels through the comprehensive color difference formula to obtain a comprehensive index, which is used to measure the overall color difference that can be perceived by the human eye. Step S3-2C: Extract the periodic features of the texture through Fourier transform to analyze the regularity of the surface structure. This step targets the periodic texture of the nonwoven fabric surface. First, the L* channel image in Lab space is converted into a grayscale image. Then, a two-dimensional Fourier transform is performed on the grayscale image to convert the texture information in the spatial domain into the spectral distribution in the frequency domain. By identifying these peak frequencies and combining them with the physical resolution of the image, the period length of the texture, i.e., the distance between adjacent textures, is calculated. Under normal circumstances, the texture period should remain stable. If the period fluctuation exceeds 10%, it indicates that the texture structure may be disordered. Step S3-3C: Fuse color difference features with texture periodic features to achieve accurate defect classification; if the texture periodic fluctuation is greater than 10% and the color difference is less than 3, it belongs to texture disorder defects such as broken yarn or wrinkles. At this time, the color does not change significantly but the structural regularity is destroyed; if both are abnormal at the same time, it may be a mixed defect.

[0011] In a preferred embodiment, in step S4, the robotic arm receives the precise coordinates of the defect area and the predicted defect type output by S3, with the predicted type being: metallic impurities; the robotic arm control system, in conjunction with its own kinematic model, plans the optimal path from the initial standby position to the target area, with a positioning accuracy of ±0.02mm.

[0012] In a preferred embodiment, in step S4, the robotic arm automatically switches the end effector based on the defect type predicted in S3: If the impurity is determined to be metallic, a miniature electromagnetic probe is activated. A magnetic field is generated by applying a 5V DC current to detect whether there is a magnetic response in the target area. The response intensity is recorded simultaneously to estimate the volume of the impurity. If the hole / tear is predicted, activate the vacuum adsorption pressure sensor and apply a negative pressure of -30 kPa after contacting the target area. If the sensor detects that the pressure is maintained for less than 0.5 seconds, the hole is confirmed to exist. At the same time, the pressure decay rate is recorded to assess the hole area. If the defect is predicted to be a mixture of uneven dyeing and wrinkles, a micro hardness tester is used to collect hardness values ​​at 3 points in the defect area and the surrounding normal area. By comparing the differences: the normal area HV=85±5, and the defect area HV<70 and fluctuation>15% verifies that the structural integrity is damaged.

[0013] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. In this invention, the detection device is installed on the production line for real-time online detection during the production process. It provides immediate reporting of detection status, instant alarms for abnormalities, and prompts on-duty personnel to identify and resolve problems promptly. It also records and marks defect locations for easy traceability. By acquiring stress-optical response multimodal data, it encompasses stress distribution and deformation information reflecting the internal structural characteristics of the material, as well as optical, textural, and dynamic behavioral features characterizing the surface state. Standardized preprocessing provides the input basis for subsequent analysis, effectively avoiding the detection blind spots of traditional single-data-dimensional methods. This ensures that all types of defects, from structural defects and surface contaminants to color difference and texture anomalies, can be completely captured. It achieves accurate identification of slight color differences and texture disturbances that are difficult to distinguish using traditional visual inspection. This multi-module collaborative intelligent analysis mode not only improves the accuracy of defect classification but also reduces the limitations of a single algorithm through feature complementarity and cross-validation between sub-modules, significantly reducing false positives and false negatives.

[0014] 2. In this invention, the mechanical feature recognition of the structural defect module, the optical-motion feature fusion of the surface contaminant module, and the color-texture quantification analysis of the color difference texture module are organically integrated. By utilizing the strong fitting ability of deep learning to complex nonlinear data, the accurate classification and location of different types of defects can be achieved: the structural defect module enhances the ability to identify hidden structural defects such as holes and tears through the collaborative judgment of stress distribution and deformation images; the surface contaminant module solves the problem of detecting small or low-contrast contaminants by using the spatiotemporal matching of multispectral fluorescence characteristics and foreign object displacement trajectory. Attached Figure Description

[0015] Figure 1 This is a schematic diagram illustrating the process principle of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0017] Example: Reference Figure 1 A method for detecting the surface quality of nonwoven fabrics, the method comprising the following steps: S1: Install the detection device on the production line to perform real-time online detection during the production process; the production line continuously conveys non-woven fabric, and the non-woven fabric is positioned in real time by a dynamic pneumatic suspension positioning device. The multispectral light source array, together with a 4K high-speed camera, synchronously collects a three-dimensional data matrix and records the collection position coordinates and timestamp. S2: Programmable piezoelectric vibrators on the production line conveyor path apply stress to dynamic samples. A micro pressure sensor array and a high-speed camera synchronously track the moving samples, record the pressure distribution in real time, capture the trajectory of foreign objects, and combine multispectral data to distinguish contaminants and mark the location of defects. S3: Input the initial 3D data matrix from step S1 and the stress-optical response data from step S2 into the depth confidence network, and perform collaborative analysis through sub-modules: The structural defect module is used to identify holes, tears, and nodules based on anomalies in pressure distribution and deformation images. Using a surface contaminant module that combines multispectral fluorescence properties and foreign object displacement trajectories, stains, insects, and metallic impurities are classified. The color difference texture module is used to analyze color difference through the CIE Lab color space, and Fourier transform is used to extract texture periodic features. S4: For the high-risk defect areas identified in step three, start the micro-robotic arm to perform targeted testing; S5: Based on the type and quantity of defects and the results of physical performance tests, products are categorized as follows: Excellent: ≤1 minor defect; Acceptable: ≤3 minor defects; Defects: ≥1 critical defect of three levels.

[0018] In step S2, the real-time recording of surface pressure distribution using a micro pressure sensor array specifically includes: Step S2-1A: Array the miniature pressure sensors into an M×N grid (e.g., 16×16=256 sensors, spaced 5mm), and collect raw pressure data under dynamic stress in real time, denoted as matrix P. raw (i,j,t), where i=1~M, j=1~N are sensor coordinates, t=1~T is the time series, T=vibration period × sampling frequency, such as 5Hz vibration sampling for 100 frames, then T=20; The formula for calculating the removal of high-frequency noise by applying a moving average filter to the original data is as follows: ; In the formula: i,j represent the row and column coordinate indices of the sensor array: i is the row and j is the column; t represents the time index of the current data frame, corresponding to a certain moment when the dynamic stress is applied; k is the size of the filtering window, and k = 5 to 10 frames is selected to balance real-time performance and smoothing effect. s represents the index of the historical frame within the sliding window: from t-k+1 to t; Let be the filtered pressure value of sensor (i,j) at time t; Step S2-2A: Perform gridded interpolation on the pressure distribution. Using sensor data as discrete points, a continuous pressure field needs to be generated through interpolation to facilitate the analysis of spatial distribution characteristics. An inverse distance-weighted interpolation method is used to map the sensor pressure values ​​to a two-dimensional grid of the detection area. The formula for calculating the pressure value at each grid point is: : In the formula: P(x,y,t) represents the pressure value at coordinate (x,y) in the interpolated continuous pressure field at frame t. (x,y) represents the two-dimensional grid coordinates (interpolated continuous spatial coordinates) within the detection area; M and N represent the total number of rows and columns of the sensor array (e.g., in a 16×16 array, M=16 and N=16). w ij (x,y) represents the weight value of the (i,j)th sensor for the grid point (x,y). d ij (x,y) represents the Euclidean distance (straight-line distance) between the (i,j)th sensor and the grid point (x,y). p represents the distance decay exponent (the rate at which the control weight decays with increasing distance). Step S2-3A: Calculate the two-dimensional gradient of the continuous pressure field P(x,y,t) using the salient features of stress concentration points to quantify the severity of pressure changes; The formula for calculating the pressure gradient modulus is: ; In the formula: (x,y,t) is a three-dimensional spatiotemporal coordinate variable used to accurately locate the pressure field characteristics of a specific location on the surface of a nonwoven fabric at a specific time. G(x,y,t) is the magnitude of the pressure gradient at coordinate (x,y) in the t-th frame in a continuous pressure field; The partial derivatives of the pressure field in the x and y directions represent the instantaneous rate of change of the pressure field in space; in Where Δx is the spacing between adjacent grid points in the x-direction (e.g., 0.5mm, which is the grid resolution after interpolation in step 2), and (x+Δx,y,t) and (x−Δx,y,t) are the coordinates of adjacent grid points on the left and right sides of (x,y,t).

[0019] Similarly, Δy is used to calculate the partial derivative in the y-direction: ; Step S2-4A: Perform local maximum detection and dynamic fluctuation analysis. In a single frame of the pressure field, identify local maximum points (points with pressure values ​​higher than 8 neighboring grid points) and denote them as the candidate point set S. cand (x, y, t), combined with the dynamic stress of the low-frequency sinusoidal vibration in step S2, calculate the pressure fluctuation amplitude of the candidate point in the time series. Due to the instability of the fiber structure, the fluctuation amplitude in the weak area is significantly higher than that in the normal area; the formula for calculating the standard deviation of pressure fluctuation is: ; In the formula: t represents a frame in the time series during dynamic stress detection. It is used to traverse the pressure data throughout the entire detection cycle and is a discrete-time variable for summation (from frame 1 to frame T). P(x,y,t) represents the instantaneous pressure value of the grid point at coordinates (x,y) on the surface of the nonwoven fabric at time t. It is the core data of the dynamic pressure field and directly participates in the calculation of the standard deviation of the fluctuation.

[0020] σ(x,y) represents the standard deviation of the pressure fluctuation of grid point (x,y) within the time series T frames (reflecting dynamic stability); This represents the average pressure value of the grid point (x, y) within the time series T frames; T represents the total number of frames in the time series, covering at least one complete cycle of dynamic stress, such as one cycle of sinusoidal vibration; Step S2-5A: Based on three indicators—grid point pressure value, pressure gradient magnitude, and dynamic fluctuation standard deviation—set a judgment threshold. Static threshold: μP is the average pressure across the entire field, and σP is the standard deviation of the pressure across the entire field, based on statistics from normal samples. Gradient threshold: G(x,y,t) > G0; G0 is an empirical threshold, such as 1.5 times the maximum gradient in the normal region). Dynamic threshold: σ(x,y)>μ σ +2σ σ(μσ is the mean of the standard deviation of the overall fluctuation, σ) σ (It is its standard deviation).

[0021] Points that meet the above three conditions are identified as stress concentration points, and their coordinates (x, y) and corresponding pressure characteristic values ​​are output.

[0022] In step S2, the hidden defect development is achieved through dynamic fluctuation analysis. The microcrack area has a decreased structural stiffness due to fiber breakage. Under vibration stress, the standard deviation of pressure fluctuation is 2 to 3 times higher than that of the normal area, thus realizing the development of "micro-defects that cannot be identified by static detection". Positioning accuracy: Through interpolation of a 16×16 sensor array and a 100×100 grid, the positioning error of stress concentration points is ≤0.1mm, which meets the detection requirements for non-woven fabric surface defects (minimum detectable microcrack width 0.05mm).

[0023] In step S2, using a high-speed camera to capture the displacement trajectory of surface foreign objects during vibration, and combining multispectral data to distinguish the type of pollutant, specifically includes the following steps: Step S2-1B: Multispectral Image Preprocessing: Input: Time-series image sequence I acquired by a high-speed multispectral camera λ (x p ,y p ,t)(λ is the wavelength, x p ,y p (where t is the image pixel coordinate and t is the time frame). Objective: To eliminate image noise (camera sensor noise, ambient light interference) and register images of different wavelengths to the same spatial coordinate system.

[0024] Preprocessing specifically includes: Gaussian filtering denoising: Spatial domain smoothing is performed on the single-wavelength image Iλ(xp,yp,t) to suppress high-frequency noise. The denoising calculation formula is as follows: In the formula: Iλ(xp,yp,t) represents the grayscale value of the original multispectral image. Wherein, λ is the wavelength of light (unit: nm, value range 400~900nm, including visible light and near-infrared bands); x p ,y p Image pixel coordinates (unit: pixels, origin is the top left corner of the image, x to the right) p Positive direction, downward is y p positive direction); t is the time frame index (unit: frame, synchronized with the sampling frequency of the high-speed camera, such as t=1 corresponding to the 0.001st second when the sampling frequency is 1000 frames / second).

[0025] G(m,n) represents a two-dimensional Gaussian kernel function. Where σ = 1.5~2.0 pixels and k = 3~5 pixels, controlling the filtering strength. m and n are the relative pixel coordinates within the kernel function window (e.g., in a 31×31 kernel, the values ​​of m and n range from -15 to 15). σ is the standard deviation of the Gaussian kernel (unit: pixels, value 1.5~2.0, controls the smoothness of the filter; the larger the σ, the stronger the noise reduction but the more details are lost). k is half the kernel window size (unit: pixels; a value of 15 corresponds to a 31×31 window, determined based on the minimum size of the foreign object being ≥5 pixels).

[0026] I λ,filt (x p ,y p ,t): The grayscale value of the filtered image, which is the result of the convolution of the original image with a Gaussian kernel, used to eliminate sensor noise (such as dark current noise) and ambient light interference.

[0027] Preprocessing specifically includes multispectral image registration: Using the visible light band image at λ=550nm as a reference, spatial offset of images at other wavelengths is corrected through affine transformation. The calculation formula is as follows: ; Where (xp′, yp′) are the registered coordinates, a, b, c, d, e, f are affine transformation parameters, obtained through pre-calibration using a checkerboard calibration plate, a, e are scaling factors, b, d are rotation / shearing factors, and c, f are translation amounts, ensuring that the spatial deviation between multispectral images is ≤0.5 pixels.

[0028] Step S2-2B: Segment the foreign object region from the background and extract its pixel coordinates: First, based on the grayscale difference between the nonwoven fabric surface and the foreign object, calculate the local threshold: ;in: For (x) p ,y p The mean gray level within a 31×31 pixel window centered at σ(x) p ,y p ,t) is the standard deviation of grayscale within the window; k=1.2~1.5 (adjusted according to the background color of the non-woven fabric: k=1.2 for white non-woven fabric and k=1.5 for dark fabric).

[0029] Calculate a local threshold and perform connected component analysis on the segmented binary image (1 for foreign object, 0 for background) to extract the center pixel coordinates of the foreign object's bounding rectangle: (x obj (i,t),y obj (i,t)), where i=1,2,⋯,N; N is the number of foreign objects in the current frame.

[0030] Step S2-3B: Calculate the displacement trajectory of foreign objects, track the positional changes of the same foreign object in the time series image, calculate the displacement amount, and distinguish between static impurities (such as background blemishes) and dynamic foreign objects (such as falling fiber lint). The specific steps are as follows: A: Set the image pixel coordinates (x) obj (i,t),y obj (i,t) is converted to the physical coordinates of the nonwoven fabric surface (X). obj (i,t),Y obj (i,t)): ; ; In the formula: (x0, y0) is the image origin (pixel coordinates of the intersection of the camera optical axis and the sample stage, pre-calibrated value); s x ,s y =0.02~0.05mm / pixel (camera resolution, high-speed camera takes s) x =s y =0.03mm / pixel).

[0031] B: For the i-th foreign object, calculate the physical displacement of adjacent frames (t and t−1): Criterion: If D(i,t)>0.1mm (dynamic threshold, adjusted according to camera frame rate, 0.1mm at 500fps), it is determined as "dynamic foreign object" (trajectory needs to be tracked); otherwise it is "static impurity" (excluded).

[0032] Step S2-4B: Perform multispectral reflectance feature extraction. For dynamic foreign matter areas, extract their reflectance curves in the multispectral bands as a basis for type differentiation (inorganic particles vs. organic stains).

[0033] In this step, a multispectral camera simultaneously acquires an image Iw(λ,t) of a standard white board (reflectance Rw(λ) = 99%), and the reflectance of the foreign object region is calculated using the following formula. ; In the formula Let be the average gray value of the i-th foreign object at wavelength λ and frame t (take the average gray value of all pixels in the connected region); Feature wavelength selection: Three key wavelengths were selected: λ1=450nm (blue light, absorption peak of organic molecules), λ2=550nm (green light, general reflectance standard), and λ3=850nm (near-infrared, reflection peak of inorganic particles), resulting in the feature vector: ; .

[0034] Step S2-5B: Classify pollutant types by distinguishing inorganic particles (such as dust and metal shavings) from organic stains (such as oil and lint) based on reflectance feature vectors and spectral similarity matching. Calculating the reflectivity vector of foreign objects by pollutant type classification With two types of standard sample spectral libraries (inorganic particles) Organic stains The included angle θ: In the formula: θ is the spectral angle, which measures the similarity of the spectral shape between the foreign object reflectance feature vector and the standard sample spectral library vector. The smaller the angle, the closer the spectral curve shape is. The reflectance feature vector of the foreign object represents a three-dimensional vector composed of the reflectance of the foreign object under three characteristic wavelengths, reflecting the reflectance characteristics of the foreign object to light of a specific wavelength. The standard sample spectral library vector represents the reflectance feature vector of two types of pollutant standard samples (inorganic particles and organic stains) collected in advance, which serves as the benchmark template for type classification; Let L2 be the L2 norm of the foreign object reflectivity vector, and let be the square root of the sum of squares of the vector components. The length of the foreign object reflectivity vector is normalized to eliminate the influence of absolute differences in light intensity. Let L2 norm be the standard spectral library vector, representing the square root of the sum of squares of the components of the standard vector. Its physical meaning is... The action is symmetric, and the standard vector is normalized in length to ensure that the inner product result reflects only the difference in direction (spectral shape), rather than the vector length (absolute value of reflectance).

[0035] In step S3, the use of the structural defect module to identify holes, tears, and nodule defects based on abnormal pressure distribution and deformation images includes the following steps: Step S3-1A: Perform data preprocessing and spatiotemporal registration of the pressure field and deformation field: I. Spatial Coordinate Registration Geometric transformations are used to adjust the physical coordinates of the deformation field data to be consistent with those of the pressure field. The specific process includes: Scaling and rotation correction: Based on the stretching characteristics of the nonwoven fabric in the width direction and the camera shooting angle, adjust the ratio of the deformation field (such as converting the image pixel size to the actual physical size) and the rotation angle (to eliminate perspective deviation caused by camera tilt) to ensure that the horizontal and vertical spatial scales are consistent with the pressure field.

[0036] Translation compensation: Using a pre-set checkerboard calibration plate (with a grid pattern of known spacing printed on the surface), the positional offset of the pressure sensor and the camera in the horizontal (width direction) and vertical (length direction) directions is measured, and the deformation field is translated as a whole to align with the origin of the pressure field coordinates (usually set as the center point of the nonwoven fabric surface).

[0037] Precision control: All transformation parameters (such as scaling ratio, rotation angle, translation amount) are determined through experimental calibration to ensure that the spatial position deviation after registration does not exceed 0.1 mm (far smaller than the minimum defect size, avoiding misjudgment of features due to misalignment).

[0038] II. Time Synchronization Correction Eliminating the minute delay between the pressure sensor and the DIC camera sampling time ensures that the pressure and deformation data correspond at the same moment. The specific process includes: Delay measurement: A laser synchronous triggering device is used (the laser signal triggers the sensor and camera to sample simultaneously) to record the timestamp difference between pressure data and deformation data (usually in the microsecond range, such as pressure data lagging behind deformation data by 0.5 milliseconds).

[0039] Sequence alignment: Based on the measured delay value, the time series of the pressure field or deformation field is shifted as a whole (e.g., the timestamp of the pressure data is uniformly increased by 0.5 milliseconds) to ensure that the two are completely synchronized on the time axis and avoid feature misalignment caused by dynamic processes (e.g., high-speed movement of nonwoven fabric).

[0040] Step S3-2A: Identify abnormal regions through the statistical characteristics of the pressure field. First, calculate the pressure standard deviation field, calculating the pressure fluctuation standard deviation of each spatial point (x, y) over the time series to reflect pressure stability. The calculation formula is: ; In the formula: This represents the average pressure over time, in kPa. T is the total number of sampling frames (e.g., T=500, corresponding to a sampling duration of 1s@500Hz). σP(x,y) is the standard deviation of pressure (unit: kPa). In the normal region, σP < 5 kPa. In the defective region, σP increases significantly due to structural instability (σP > 15 kPa at the hole and > 10 kPa at the nodule).

[0041] A dual-threshold segmentation is constructed based on σP(x,y) and the mean pressure Pˉ(x,y): ; In the formula: Ω P (x,y) is the mask for the pressure anomaly region (1 indicates anomaly, 0 indicates normal). τ σ =8kPa is the standard deviation threshold, determined using the 95% confidence interval method: μ σ For the normal region σ P (mean) (The average pressure is the reference pressure, which is the average pressure value of the defect-free area). τ P =10kPa is the average pressure deviation threshold at the puncture site. nodule ; Step S3-3A: Calculate and differentiate defect types using deformation geometry features. The first spatial derivative of the deformation field is calculated to reflect the rate of deformation (the gradient at the tear point increases sharply due to fiber breakage): ; ; The numerical calculation uses the central difference method: ; In the formula: U x (x,y) represents the lateral displacement at coordinate point (x,y) on the surface of the nonwoven fabric. x represents the horizontal spatial coordinate, and y represents the vertical spatial coordinate; Δx represents the spatial sampling interval, that is, the lateral distance between two adjacent displacement measurement points. U x (x+Δx,y) and U x (x−Δx,y) represents the lateral displacement values ​​at intervals Δx to the right and left of the current point (x,y); This represents the change in lateral displacement per unit length, dimensionless, with a normal range <0.02 (gradual displacement change) and a tear range >0.1. Step S3-4A: Perform multi-feature fusion and defect classification: Input: Pressure anomaly region ΩP(x,y), deformation gradient ∇U(x,y)=(∇U x ,∇U y ), Gaussian curvature K(x,y); For Ω P For the region (x,y)=1, calculate the following indices: Pressure anomaly level: Dimensionless >1 indicates significant pressure fluctuations; Deformation gradient mode: Dimensionless, reflecting the degree of drastic deformation changes; The curvature symbol sgn(K) represents positive curvature, and -1 represents negative curvature. The defect classification rules are as follows: Classification based on indicator combinations: Hole: A P>1.5 and And sgn(K) = −1 (sudden pressure drop + concave deformation); Tear: A P >1.2 and |∣∇U∣>0.1 and sgn(K)=0 (pressure fluctuation + high deformation gradient + in-plane tearing); Nodules: AP > 1.0 and And sgn(K)=1 (slightly higher pressure + convex deformation) In step S3, the classification of stains, insects, and metallic impurities using the surface contaminant module, combined with multispectral fluorescence characteristics and foreign object displacement trajectories, specifically includes: Step S3-1B: First, the multispectral fluorescence image and the foreign object motion trajectory data need to be spatiotemporally aligned to ensure consistency in subsequent feature extraction.

[0042] Spectral image registration: For fluorescence images with different excitation wavelengths, by identifying the fixed texture features on the surface of the nonwoven fabric, the spatial offset caused by the optical system is eliminated, so that the pixel positions of the images of each wavelength are completely corresponding.

[0043] Trajectory coordinate mapping: The trajectory of the moving object captured by the high-speed camera is transformed into the fluorescence image coordinate system. By using pre-calibrated camera parameters, each coordinate point in the trajectory is converted into its corresponding position in the fluorescence image, ensuring that the trajectory and the fluorescence signal originate from the same physical region.

[0044] Step S3-2B: Based on the differences in fluorescence emission characteristics of pollutants at different wavelengths, three key features are extracted to distinguish the composition of substances: Fluorescence peak wavelength: The wavelength with the highest intensity is found by scanning the fluorescence intensity in the ultraviolet to visible light band of 365-600 nm. For example, stains usually show a peak at around 430 nm, the peak of chitin on the surface of mosquitoes is about 480 nm, while metallic impurities do not show obvious peaks because they have no fluorescent properties.

[0045] Fluorescence intensity ratio: Calculate the ratio of fluorescence intensity at two specific wavelengths, such as 450 nm blue light to 532 nm green light. The blue light intensity of stains is significantly higher than that of green light, with a ratio greater than 2.5. The blue and green light intensities of mosquitoes are relatively close (ratio 1.2-2.0), while metallic impurities have no fluorescence, resulting in an intensity ratio of 0.

[0046] Fluorescence spectral width: The wavelength range at which the intensity on both sides of the fluorescence peak drops to half. Stains have a narrow spectral width (less than 50 nm), mosquitoes have a wide spectral width, and there is no spectral width data for metallic impurities.

[0047] Step S3-3B: By analyzing the motion patterns of foreign objects in consecutive frames, two types of dynamic features are extracted to distinguish motion modes: Average velocity: The total displacement of foreign objects over a short period of time is statistically analyzed to calculate the average velocity. Metal impurities, due to their high density, move synchronously with the substrate, resulting in a stable and relatively fast velocity; insects, due to their active crawling or wing flapping, exhibit large velocity fluctuations; while stains, adhering to the substrate surface, have an extremely low velocity.

[0048] Trajectory curvature: Calculates the average curvature of the trajectory, reflecting the frequency of change in the direction of motion. Metal impurities are pulled by the substrate, and their trajectories are nearly straight (curvature less than 0.01); mosquitoes, due to random turning, have significantly curved trajectories (curvature greater than 0.05); stains are stationary, and their trajectory curvature is 0.

[0049] Step S3-4B: Combining the rules of fluorescence characteristics and trajectory characteristics, the three types of pollutants are classified: Stains: exhibit fluorescence peaks (420-440 nm), blue to green light intensity ratio greater than 2.5, average speed less than 50 pixels / second, and trajectory curvature of 0.

[0050] Mosquitoes: exhibit fluorescence peaks (470-490 nm), intensity ratios of 1.2-2.0, velocity fluctuations of 200-800 pixels / second, and trajectory curvatures greater than 0.05.

[0051] Metallic impurities: No fluorescence peak (intensity ratio of 0), average velocity exceeding 500 pixels / second, and trajectory curvature less than 0.01.

[0052] Through the layer-by-layer screening and combination of the above features, the pollutant type, center location coordinates and classification confidence level are finally output, providing a basis for subsequent removal or alarm processing.

[0053] In step S3, the color difference texture module is used to analyze the color difference through the CIE Lab color space, and the texture periodic features are extracted by combining Fourier transform.

[0054] In step S3, the color difference texture module is used to analyze color difference through the CIE Lab color space, and the texture periodic features are extracted by combining Fourier transform. Specifically, this includes the following steps: Step S3-1C: Extract color difference features based on the CIE Lab color space to quantify the color difference between the target area and the normal area. First, calculate the brightness difference, which is the difference between the average L value of the target area (suspected defect area) and the normal area, reflecting the brightness change of the area; then, calculate the difference between the a channel and the b channel respectively. For example, if the target area is reddish, the a difference will be positive; if it is yellowish, the b difference will be positive; finally, integrate the differences of these three channels through a comprehensive color difference formula to obtain a comprehensive index, which is used to measure the overall color difference that the human eye can perceive. When this comprehensive index is greater than 3, it indicates that the color difference has reached the level that the human eye can distinguish, and it can be preliminarily judged as a potential defect.

[0055] Step S3-2C: Extract the periodic features of the texture using Fourier transform to analyze the regularity of the surface structure. This step targets the periodic texture of the nonwoven fabric surface. First, the L* channel image in Lab space is converted into a grayscale image. Then, a two-dimensional Fourier transform is performed on the grayscale image to convert the spatial domain texture information into a frequency domain spectral distribution. A series of peaks will appear in the spectrum, corresponding to the periodic frequencies of the texture—for example, if there are 5 transverse lines per millimeter on the nonwoven fabric surface, obvious peaks will appear at the corresponding frequencies in the spectrum. By identifying these peak frequencies and combining them with the physical resolution of the image, the period length of the texture, i.e., the distance between adjacent lines, can be calculated. Under normal circumstances, the texture period should remain stable. If the period fluctuation exceeds 10%, it indicates that the texture structure may be disordered.

[0056] Step S3-3C: The color difference feature and texture periodicity feature are fused to achieve accurate defect classification. For example, when the overall color difference is greater than 5 and the texture periodicity fluctuation is less than 2%, it can be identified as an uneven dyeing defect because the color difference is significant but the texture structure is not destroyed. If the texture periodicity fluctuation is greater than 10% and the color difference is less than 3, it belongs to texture disorder defects such as yarn breakage or wrinkles, where the color does not change significantly but the structural regularity is destroyed. If both are abnormal, it may be a mixed defect. Through this multi-feature fusion method, the system can distinguish different types of surface defects, providing a basis for subsequent quality judgment. The entire process ensures the accuracy and robustness of defect identification through dual verification of color and structural features.

[0057] In step S4, the robotic arm receives the precise coordinates of the defect area (based on the physical coordinate system transmitted by S2, in millimeters) and the defect type prediction result output by S3, such as "coordinates (X=125.3mm, Y=48.7mm), prediction type: metallic impurity". The robotic arm control system, combined with its own kinematic model (such as a 6-DOF joint structure), plans the optimal path from the initial standby position (usually located at the edge of the detection platform) to the target area. The path planning needs to avoid the substrate surface of non-defect areas (to avoid secondary damage), and the positioning accuracy is controlled within ±0.02mm (to ensure that the test point coincides with the defect center).

[0058] Based on the defect type predicted by S3, the robotic arm automatically switches to the end effector: a detachable tool head. If the target area is identified as a "metallic impurity", a miniature electromagnetic probe is activated. A magnetic field is generated by applying a 5V DC current to detect whether there is a magnetic response in the target area (metallic impurities will cause the probe inductance value to change by more than 5%). The response intensity (unit: mH) is recorded simultaneously to estimate the impurity volume. If the prediction is "hole / tear", activate the vacuum adsorption pressure sensor, apply a negative pressure of -30kPa after contacting the target area. If the sensor detects that the pressure is maintained for less than 0.5 seconds (i.e., air leakage), the hole is confirmed to exist. At the same time, record the pressure decay rate (kPa / s) to assess the hole area. If the defect is predicted to be a mixture of uneven dyeing and wrinkles, a micro hardness tester (with a loading force of 100mN) is used to collect hardness values ​​(unit: HV) at three points in both the defect area and the surrounding normal area. The structural integrity is verified by comparing the differences (HV = 85±5 in the normal area, and HV < 70 with fluctuation > 15% in the defect area). An apparatus and medium include: at least one processor; and a memory communicatively connected to the at least one processor; wherein: the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the above-described method for detecting the surface quality of nonwoven fabrics.

[0059] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the surface quality detection method for nonwoven fabrics as described above.

[0060] From the above, we can conclude that: In this application, stress-optical response multimodal data is acquired simultaneously, encompassing stress distribution and deformation information reflecting the internal structural characteristics of materials, as well as optical, textural, and dynamic behavioral features characterizing surface states. Standardized preprocessing provides a clean and comprehensive input foundation for subsequent analysis, effectively avoiding detection blind spots inherent in traditional single-data-dimension methods. This ensures that information on all types of defects, from structural defects and surface contaminants to color difference and texture anomalies, is fully captured. In the analysis and decision-making stage, step S3 uses a deep belief network as its core framework to construct a sub-module collaborative analysis mechanism. This mechanism organically integrates the mechanical feature identification of the structural defect module, the optical-motion feature fusion of the surface contaminant module, and the color-texture quantification analysis of the color difference and texture module. Leveraging the strong fitting ability of deep learning to complex nonlinear data, it achieves accurate classification and localization of different types of defects. The structural defect module enhances the ability to identify hidden structural defects such as holes and tears by coordinating the judgment of stress distribution and deformation images; the surface contaminant module solves the problem of detecting tiny or low-contrast contaminants by using multispectral fluorescence characteristics and spatiotemporal matching of foreign object displacement trajectories; and the color difference and texture module achieves accurate identification of slight color differences and texture disorder that are difficult to distinguish by traditional visual inspection through quantitative analysis of CIE Lab color space and Fourier texture period.

[0061] This multi-module collaborative intelligent analysis model not only improves the accuracy of defect classification but also reduces the limitations of a single algorithm through feature complementarity and cross-validation between sub-modules, significantly reducing false positives and false negatives. Simultaneously, the high-quality data preprocessing in step S2 and the deep learning parallel computing architecture in step S3 optimize the analysis process while ensuring detection accuracy, avoiding redundant calculations and providing precise guidance for the targeted testing in the subsequent step S4. Ultimately, this forms a complete closed loop from data acquisition and intelligent analysis to physical verification, providing efficient and reliable technical support for quality control in the nonwoven fabric production process, helping to improve product qualification rates and reduce manual inspection costs.

[0062] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the term "comprising" or any other variations thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0063] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention 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 the present invention.

Claims

1. A method for detecting the surface quality of nonwoven fabrics, characterized in that: The method includes the following steps: S1: Install the detection device on the production line to perform real-time online detection during the production process; the production line continuously conveys non-woven fabric, and the non-woven fabric is positioned in real time by a dynamic pneumatic suspension positioning device. The multispectral light source array, together with a 4K high-speed camera, synchronously collects a three-dimensional data matrix and records the collection position coordinates and timestamp. S2: Programmable piezoelectric vibrators on the production line conveyor path apply stress to dynamic samples. A micro pressure sensor array and a high-speed camera synchronously track the moving samples, record the pressure distribution in real time, capture the trajectory of foreign objects, and combine multispectral data to distinguish contaminants and mark the location of defects. S3: Input the initial 3D data matrix from step S1 and the stress-optical response data from step S2 into the depth confidence network, and perform collaborative analysis through sub-modules: The structural defect module is used to identify holes, tears, and nodules based on anomalies in pressure distribution and deformation images. Using a surface contaminant module that combines multispectral fluorescence properties and foreign object displacement trajectories, stains, insects, and metallic impurities are classified. The color difference texture module is used to analyze color difference through the CIE Lab color space, and Fourier transform is used to extract texture periodic features. S4: For the high-risk defect areas identified in step three, start the micro-robotic arm to perform targeted testing; S5: Based on the type and quantity of defects and the results of physical performance tests, products are categorized as follows: Excellent: ≤1 minor defect; Acceptable: ≤3 minor defects; Defects: ≥1 critical defect of three levels.

2. The surface quality testing method for nonwoven fabrics as described in claim 1, characterized in that: In step S2, the real-time recording of surface pressure distribution using a micro pressure sensor array specifically includes: Step S2-1A: Array the miniature pressure sensors into an M×N grid, and collect raw pressure data under dynamic stress in real time, denoted as matrix P. raw (i,j,t), where i=1~M, j=1~N are sensor coordinates, t=1~T is the time series, and T=vibration period × sampling frequency; The formula for calculating the removal of high-frequency noise by applying a moving average filter to the original data is as follows: ； In the formula: i,j represent the row and column coordinate indices of the sensor array: i is the row and j is the column; t represents the time index of the current data frame, corresponding to the moment when the dynamic stress was applied; k is the size of the filtering window, and k = 5 to 10 frames is selected to balance real-time performance and smoothing effect. s represents the index of the historical frame within the sliding window: from t-k+1 to t; Let be the filtered pressure value of sensor (i,j) at time t; Step S2-2A: Perform gridded interpolation on the pressure distribution. Using sensor data as discrete points, a continuous pressure field needs to be generated through interpolation to facilitate the analysis of spatial distribution characteristics. The inverse distance weighted interpolation method is used to map the sensor pressure values ​​to a two-dimensional grid in the detection area. The formula for calculating the pressure value at each grid point is: : In the formula: P(x,y,t) represents the pressure value at coordinate (x,y) in the interpolated continuous pressure field at frame t. (x,y) represents the two-dimensional grid coordinates within the detection area; M and N represent the total number of rows and columns of the sensor array; w ij (x,y) represents the weight value of the (i,j)th sensor for the grid point (x,y); d ij (x,y) represents the Euclidean distance between the (i,j)th sensor and the grid point (x,y); p represents the distance decay exponent; Step S2-3A: Calculate the two-dimensional gradient of the continuous pressure field P(x,y,t) using the salient features of stress concentration points to quantify the severity of pressure changes; The formula for calculating the pressure gradient modulus is: ； In the formula: (x,y,t) is a three-dimensional spatiotemporal coordinate variable used to accurately locate the pressure field characteristics of a specific location on the surface of a nonwoven fabric at a specific time. G(x,y,t) is the magnitude of the pressure gradient at coordinate (x,y) in the t-th frame in a continuous pressure field; The partial derivatives of the pressure field in the x and y directions represent the instantaneous rate of change of the pressure field in space; in Where Δx is the distance between adjacent grid points in the x-direction, and (x+Δx,y,t) and (x−Δx,y,t) are the coordinates of the adjacent grid points on the left and right sides of (x,y,t); Similarly, Δy is used to calculate the partial derivative in the y-direction: ； Step S2-4A: Perform local maximum detection and dynamic fluctuation analysis. In a single frame of the pressure field, identify local maximum points and denote them as the candidate point set S. cand (x, y, t), combined with the dynamic stress of the low-frequency sinusoidal vibration in step S2, calculate the pressure fluctuation amplitude of the candidate point in the time series. Due to the instability of the fiber structure, the fluctuation amplitude in the weak area is significantly higher than that in the normal area; the formula for calculating the standard deviation of pressure fluctuation is: ； In the formula: t represents a specific frame in the time series during dynamic stress detection, used to traverse the pressure data throughout the entire detection cycle, and is a discrete-time variable for summation: from frame 1 to frame T. P(x,y,t) represents the instantaneous pressure value of the grid point at coordinates (x,y) on the surface of the nonwoven fabric at time t. It is the core data of the dynamic pressure field and directly participates in the calculation of the standard deviation of the fluctuation. σ(x,y) represents the standard deviation of pressure fluctuation at grid point (x,y) within the time series T frames; This represents the average pressure value of the grid point (x, y) within the time series T frames; T represents the total number of frames in the time series, covering at least one complete cycle of dynamic stress; Step S2-5A: Based on three indicators—grid point pressure value, pressure gradient magnitude, and dynamic fluctuation standard deviation—set a judgment threshold. Static threshold: μP is the average pressure across the entire field, and σP is the standard deviation of the pressure across the entire field, based on statistics from normal samples. Gradient threshold: G(x,y,t) > G0; G0 is an empirical threshold; Dynamic threshold: σ(x,y)>μ σ +2σ σ， μσ is the mean of the standard deviation of the overall fluctuation, σ σ Its standard deviation; Points that meet the above three conditions are identified as stress concentration points, and their coordinates (x, y) and corresponding pressure characteristic values ​​are output.

3. The surface quality testing method for nonwoven fabrics as described in claim 1, characterized in that: In step S2, using a high-speed camera to capture the displacement trajectory of surface foreign objects during vibration, and combining multispectral data to distinguish the type of pollutant, specifically includes the following steps: Step S2-1B: Multispectral Image Preprocessing: Input: Time-series image sequence I acquired by a high-speed multispectral camera λ (x p ,y p ,t); λ is the wavelength, x p ,y p (where t is the image pixel coordinate and t is the time frame); Purpose: To eliminate image noise and register images of different wavelengths to the same spatial coordinate system; Step S2-2B: Segment the foreign object region from the background and extract its pixel coordinates: Step S2-3B: Calculate the displacement trajectory of the foreign object, track the position change of the same foreign object in the time series image, calculate the displacement, and distinguish between static impurities and dynamic foreign objects; Step S2-4B: Perform multispectral reflectance feature extraction. For dynamic foreign object regions, extract their reflectance curves in the multispectral bands as a basis for type differentiation.

4. The surface quality testing method for nonwoven fabrics as described in claim 1, characterized in that: In step S3, the use of the structural defect module to identify holes, tears, and nodule defects based on abnormal pressure distribution and deformation images includes the following steps: Step S3-1A: Perform data preprocessing and spatiotemporal registration of the pressure field and deformation field: Step S3-2A: Identify abnormal areas through the statistical characteristics of the pressure field. First, calculate the pressure standard deviation field and calculate the pressure fluctuation standard deviation of each spatial point (x,y) in the time series to reflect pressure stability. Step S3-3A: Calculation through deformation geometry characteristics: Differentiate defect types: Calculate the first spatial derivative of the deformation field to reflect the deformation rate; Step S3-4A: Perform multi-feature fusion and defect classification; Input: Pressure anomaly region ΩP(x,y), deformation gradient ∇U(x,y)=(∇U x ,∇U y ), Gaussian curvature K(x,y); For Ω P For the region (x,y)=1, calculate the following indices: Pressure anomaly level: Dimensionless >1 indicates significant pressure fluctuations; Deformation gradient mode: Dimensionless, reflecting the degree of drastic deformation changes; The curvature symbol sgn(K) represents positive curvature, and -1 represents negative curvature. The defect classification rules are as follows: Classification based on indicator combinations: Hole: A P >1.5 and And sgn(K) = −1; Tear: A P >1.2 and |∣∇U∣>0.1 and sgn(K)=0; Nodules: AP > 1.0 and And sgn(K)=1.

5. The surface quality testing method for nonwoven fabrics as described in claim 1, characterized in that: In step S3, the classification of stains, insects, and metallic impurities using the surface contaminant module, combined with multispectral fluorescence characteristics and foreign object displacement trajectories, specifically includes: Step S3-1B: First, the multispectral fluorescence image and the foreign object motion trajectory data need to be spatiotemporally aligned to ensure consistency in subsequent feature extraction; Step S3-2B: Based on the differences in fluorescence emission characteristics of pollutants at different wavelengths, three key features are extracted to distinguish the composition of substances: Step S3-3B: By analyzing the motion patterns of foreign objects in consecutive frames, two types of dynamic features are extracted to distinguish motion modes: Step S3-4B: Combining the rules of fluorescence characteristics and trajectory characteristics, the three types of pollutants are classified: Stains: Exhibit fluorescence peaks, blue to green light intensity ratio greater than 2.5, average speed less than 50 pixels / second, and trajectory curvature of 0; Mosquitoes: exhibit fluorescence peaks, intensity ratios of 1.2-2.0, speed fluctuations ranging from 200-800 pixels / second, and trajectory curvature greater than 0.05; Metallic impurities: No fluorescence peaks, average velocity exceeding 500 pixels / second, and trajectory curvature less than 0.

01.

6. The surface quality testing method for nonwoven fabrics as described in claim 1, characterized in that: In step S3, the color difference texture module is used to analyze color difference through the CIE Lab color space, and the texture periodic features are extracted by combining Fourier transform. Specifically, this includes the following steps: Step S3-1C: Extract color difference features based on CIE Lab space to quantify the color difference between the target area and the normal area; first, calculate the brightness difference, that is, the difference between the average L value of the target area and the normal area, reflecting the change in brightness of the area; then calculate the difference between the a channel and the b channel respectively; finally, integrate the differences of the three channels through the comprehensive color difference formula to obtain a comprehensive index, which is used to measure the overall color difference that can be perceived by the human eye. Step S3-2C: Extract the periodic features of the texture through Fourier transform to analyze the regularity of the surface structure. This step targets the periodic texture of the nonwoven fabric surface. First, the L* channel image in Lab space is converted into a grayscale image. Then, a two-dimensional Fourier transform is performed on the grayscale image to convert the texture information in the spatial domain into the spectral distribution in the frequency domain. By identifying these peak frequencies and combining them with the physical resolution of the image, the period length of the texture, i.e., the distance between adjacent textures, is calculated. Under normal circumstances, the texture period should remain stable. If the period fluctuation exceeds 10%, it indicates that the texture structure may be disordered. Step S3-3C: Fuse color difference features with texture periodic features to achieve accurate defect classification; if the texture periodic fluctuation is greater than 10% and the color difference is less than 3, it belongs to texture disorder defects such as broken yarn or wrinkles. At this time, the color does not change significantly but the structural regularity is destroyed; if both are abnormal at the same time, it may be a mixed defect.

7. The surface quality testing method for nonwoven fabrics as described in claim 1, characterized in that: In step S4, the robotic arm receives the precise coordinates of the defect area and the predicted defect type from the output of S3. The predicted type is: metal impurity. The robotic arm control system, combined with its own kinematic model, plans the optimal path from the initial standby position to the target area, with a positioning accuracy of ±0.02mm.

8. The surface quality testing method for nonwoven fabrics as described in claim 1, characterized in that: In step S4, based on the defect type predicted in S3, the robotic arm automatically switches the end effector: If the impurity is determined to be metallic, a miniature electromagnetic probe is activated. A magnetic field is generated by applying a 5V DC current to detect whether there is a magnetic response in the target area. The response intensity is recorded simultaneously to estimate the volume of the impurity. If the hole / tear is predicted, activate the vacuum adsorption pressure sensor and apply a negative pressure of -30 kPa after contacting the target area. If the sensor detects that the pressure is maintained for less than 0.5 seconds, the hole is confirmed to exist. At the same time, the pressure decay rate is recorded to assess the hole area. If the defect is predicted to be a mixture of uneven dyeing and wrinkles, a micro hardness tester is used to collect hardness values ​​at 3 points in the defect area and 3 points in the surrounding normal area. By comparing the differences: the normal area HV=85±5, and the defect area HV<70, and the fluctuation>15%, it is verified that the structural integrity is damaged.

9. A device and a medium, characterized in that: include: At least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor to enable the at least one processor to perform the surface quality testing method for nonwoven fabrics as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the surface quality detection method for nonwoven fabrics as described in any one of claims 1 to 8.

Images