A method, system, equipment, and storage medium for detecting surface defects in rolling mill rolls.

Abstract

This invention discloses a method, system, equipment, and storage medium for detecting surface defects in rolling mill rolls, relating to the field of machine vision inspection technology. The method includes: first, establishing six-channel visual features of the cylindrical surface of the rolling mill roll through simultaneous acquisition of data via four-channel polarization linear array imaging, infrared temperature measurement, and an encoder; then, establishing fatigue vulnerability probability by combining contact load, bending stress, thermal stress, frictional shear stress, and cumulative cyclic damage; subsequently, generating geometric masks and relative depths for microcracks, pitting, and spalling using the fatigue vulnerability probability and preferred crack propagation direction; forming virtual defect samples through a network generated under mask-constrained conditions; and finally, completing defect identification using a seven-channel lightweight segmentation network. This invention improves the identification capability of microcracks, pitting, and spalling, while balancing detection accuracy, sample expansion effect, and on-site deployment efficiency.

Description

Technical Field

[0001] This invention relates to the field of machine vision inspection technology, and in particular to a method, system, equipment and storage medium for detecting surface defects in rolls. Background Technology

[0002] With the development of steel rolling equipment towards high speed, continuous operation, and intelligence, the surface condition of rolls has become a crucial factor affecting strip surface quality, rolling stability, and roll service life. Under the combined effects of cyclic contact loads, bending loads, frictional traction, and localized thermal cycles, rolls are prone to surface defects such as microcracks, pitting, and spalling. Traditional roll inspection mainly relies on manual observation during shutdown, magnetic particle testing, eddy current testing, and surface profile measurement. While these methods can confirm some defects, their efficiency and online applicability are limited. With the development of industrial cameras, linear array imaging, polarization vision, and deep learning technologies, machine vision-based roll surface inspection is increasingly being applied to continuous surface image acquisition, texture enhancement, defect localization, and category recognition, especially… Convolutional neural networks and semantic segmentation networks can learn the edge, texture, and shape features of defects from a large number of samples. Linear scan cameras, combined with rotary encoders, can also convert the rotational motion of the roll into a continuous cylindrical surface unfolding image, thus providing a technical basis for online, non-contact detection of roll surface defects. However, the roll surface has characteristics such as large curvature, strong reflection, dense grinding marks, and possible oil film adhesion. Microcracks in a single grayscale or color image are easily confused with normal grinding marks, specular reflection, and local contamination. If the cylindrical unfolding image lacks stable zero-phase and physical scale calibration, it will be difficult to correspond the defect positions in different detection cycles and fail to provide sufficient information related to roll fatigue damage. It is also difficult to distinguish between early weak defects in high-risk areas and reflective interference in low-risk areas. Summary of the Invention

[0003] In view of the aforementioned existing problems, the present invention is proposed.

[0004] Therefore, the present invention provides a method, system, device and storage medium for detecting surface defects of rolls, which solves the problem that the defect locations in different detection cycles of the prior art are difficult to correspond and that the information is not sufficiently related to roll fatigue damage.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for detecting surface defects in rolling mill rolls, comprising, Collect roll parameters and machine vision data, establish the coordinates of the roll cylindrical surface and a unified cylindrical mesh, and map the machine vision data to the unified cylindrical mesh to further generate six-channel roll scalar features and imaging quality masks; Based on the roll parameters, the roll stress is calculated and mechanical fatigue damage index, roll surface temperature gradient and machine vision surface perturbation features are generated. The fatigue vulnerability probability and the preferred propagation direction of roll surface cracks are then obtained. Classify surface defects, determine the location of defect initiation points based on fatigue vulnerability probability and generate defect geometric masks, and calculate the maximum relative depth for each type of defect to form a relative depth map; A normal roll surface background library is established based on the six-channel roll scalar features, and real defect patches are collected. Training data is formed based on the real defect patches. The generator and discriminator are trained using the training data to output a defect appearance generation network. A virtual defect training set is formed based on the defect appearance generation network. A lightweight roll defect segmentation network is constructed and detection labels are defined. Defect training data is established to train the lightweight roll defect segmentation network. Based on the trained defect segmentation network, roll surface defects are identified and output.

[0006] As a preferred embodiment of the roll surface defect detection method of the present invention, the steps of collecting roll parameters and machine vision data, establishing roll cylindrical surface coordinates and a unified cylindrical mesh, mapping machine vision data to the unified cylindrical mesh, and further generating six-channel roll scalar features and imaging quality masks are specifically as follows: acquiring the geometric parameters, material parameters, rolling process parameters and machine vision data of the target roll, and establishing data association according to the same roll number and the same acquisition time; The geometric parameters include the reference radius and the effective length of the roll body; The material parameters correspond to the rolls and the strip being rolled. The roll material parameters include the elastic modulus, Poisson's ratio, yield strength, coefficient of linear expansion, fatigue strength at the reference number of cycles, and fatigue index. The strip material parameters include the elastic modulus and Poisson's ratio. The rolling process parameters include the current total rolling force, current reduction, inlet tension, outlet tension, roll surface linear velocity, strip linear velocity, and roll rotation speed; The machine vision data uses a four-channel polarization linear array camera that is synchronously triggered by a rotary encoder to acquire original polarization images with polarization directions of 0°, 45°, 90°, and 135° respectively. At the same time, an infrared linear array camera that is synchronously triggered with the rotary encoder is used to acquire the original image of the roll surface temperature. The collected machine vision data is corrected; Define the circumferential and axial coordinates of the roll surface to establish the cylindrical surface coordinates of the roll. Set a unique zero-phase mark at the end of the roll and define the circumferential position when the encoder zero-phase pulse arrives as the 0 position of the circumferential coordinate. Record the number of effective trigger pulses actually generated by the encoder per revolution and calculate the roll rotation angle corresponding to each effective trigger pulse. Calculate the corresponding circumferential coordinates through the roll rotation angle. Calculate the axial pixel physical resolution by setting a calibration plate. Record the pixel number corresponding to the left end of the roll. Multiply the difference between the axial pixel and the pixel number corresponding to the left end of the roll by the axial pixel physical resolution to obtain the axial coordinate. Define the target minimum stable crack width and the minimum crack width coverage pixels. Obtain a uniform cylindrical grid spacing based on the ratio of the target minimum stable crack width and the minimum crack width coverage pixels. Calculate the number of circumferential grids using the roll circumference and the uniform cylindrical grid spacing, and obtain the coordinates of each circumferential grid. Simultaneously, calculate the number of axial grids using the effective roll length and the uniform cylindrical grid spacing, and obtain the axial grid coordinates. Combine the circumferential grid coordinates and the axial grid coordinates to form a cylindrical grid point. Obtain the continuous coordinates of each cylindrical grid point in the machine vision data image. Based on the correspondence between cylindrical grid points and continuous coordinates, bicubic interpolation is used to map the corrected machine vision data image onto a unified cylindrical grid; The total intensity component and the linear polarization difference component are calculated based on the polarization images of the four polarization directions, and the low illumination confidence and linear polarization degree are constructed based on the total intensity component and the linear polarization difference component. The linear polarization direction angle is calculated based on the linear polarization differential component, and the linear polarization direction is continuously encoded. Take the logarithm of the total intensity and then subtract the low-frequency background to extract local reflection texture features; Two roll defect size detection scales are defined. The corresponding pixel scales are calculated according to the roll defect size detection scales. The circumferential gradient and axial gradient after Gaussian kernel smoothing are calculated for the two pixel scales respectively. The edge intensity under the corresponding scale is obtained. The larger edge intensity value between the two pixel scales is selected as the multi-scale edge intensity. At each pixel scale, a structure tensor is constructed and two feature values ​​of the structure tensor are extracted. The texture anisotropy at the corresponding pixel scale is obtained by the ratio of the difference between the two feature values ​​to the sum. The larger of the texture anisotropy at the two pixel scales is selected as the multi-scale anisotropy. The local reflection texture features, linear polarization degree, multi-scale edge intensity, and multi-scale anisotropy are used to form the roll scalar features to be standardized. The 5% and 95% quantiles of each item in the roll scalar features are extracted respectively. Each item in the roll scalar features to be standardized is standardized according to the quantiles. At the same time, the linear polarization direction code is mapped to 0-1. The mapped linear polarization direction code and the standardized roll scalar features are connected and combined to form a new six-channel roll scalar feature. An imaging quality mask is generated by combining low-light markers, channel saturation markers, and encoder trigger integrity markers.

[0007] As a preferred embodiment of the roll surface defect detection method of the present invention, the step of calculating roll stress based on roll parameters and generating mechanical fatigue damage index, roll surface temperature gradient and machine vision surface perturbation features, and further obtaining fatigue vulnerability probability and preferred expansion direction of roll surface cracks specifically involves: converting axial coordinates into dimensionless positions through the effective length of the roll body, and constructing axial linear load distribution using zero-mean quadratic basis functions. The equivalent elastic modulus between the roll and the strip is calculated based on the material parameters, and the reference radius of the roll is used as the equivalent radius of curvature. The Hertz contact half-width and maximum contact pressure at the axial position are calculated by combining the equivalent elastic modulus and the equivalent radius of curvature. The maximum bending stress on the roll surface at the axial position is obtained by calculating the bending moment at the axial position and the moment of inertia of the roll's circular cross-section. The ratio of the bending moment multiplied by the effective radius of the roll to the moment of inertia is then used to obtain the maximum bending stress on the roll surface at the axial position. The median temperature in the circumferential direction is extracted for each axial position based on the machine vision data image mapped to a unified cylindrical mesh, and the local temperature difference between the current point and the median temperature is calculated. The constraint thermal stress caused by the local temperature difference is then calculated based on the local temperature difference. The difference between the exit tension and the inlet tension is calculated, and the ratio of the difference to the total rolling force is taken to obtain the equivalent friction coefficient. The slip direction coefficient is calculated based on the strip linear velocity and the roll surface linear velocity. The product of the maximum contact pressure, the equivalent friction coefficient, and the slip direction coefficient is taken as the roll surface shear stress. The surface normal stress in the circumferential and axial directions is calculated based on the maximum contact pressure, constraint thermal stress and maximum bending stress on the roll surface. The equivalent stress on the roll surface is calculated under two-dimensional plane stress state. Then the cumulative number of rotations of the target roll since the most recent grinding is read, and the mechanical fatigue damage index is calculated. The temperature gradients in the circumferential and axial directions of the roll surface are calculated separately and synthesized into the roll surface temperature gradient. The local reflective texture features are filtered by Gaussian kernel to extract the local normal texture background. The absolute value of the difference between the normalized local reflective texture and the local normal texture background is calculated and then the machine vision surface perturbation features are calculated. The machine vision surface perturbation features are corrected by using an imaging quality mask. The mechanical fatigue damage index, the roll surface temperature gradient, and the corrected machine vision surface perturbation features are weighted after quantile-stable normalization to obtain the comprehensive fatigue driving amount of the roll. The sigmoid function is used to map the comprehensive fatigue driving amount to a fatigue vulnerability probability in the range of 0-1. Based on the principal stress direction relationship under two-dimensional stress state, the maximum principal stress direction angle is calculated, and the preferred propagation direction of the crack on the roll surface is calculated based on the maximum principal stress direction angle.

[0008] As a preferred embodiment of the roll surface defect detection method of the present invention, the step of classifying surface defect types, determining the defect initiation point location based on fatigue vulnerability probability and generating a defect geometric mask, and calculating the maximum relative depth for each type of defect to form a relative depth map specifically involves: classifying the defects to be generated on the roll surface into three categories: microcracks, pitting, and spalling, assigning category labels to each, and counting the actual occurrence of the three types of defects to determine the generation ratio of the three types of defects in the virtual sample; The fatigue vulnerability probability is divided into fixed-size blocks according to a uniform cylindrical grid, and the average fatigue vulnerability degree is calculated for each block; Within the image block, the probability of defect initiation points is obtained by weighted sampling based on the fatigue vulnerability probability of each pixel. A minimum physical spacing is set for different initiation points, and the defect initiation points are selected based on their probabilities. Generate geometric masks based on the three types of defects; For each type of defect, define the maximum relative depth range; For microcrack defects, read the microcrack geometry mask and extract the microcrack path and crack width. Calculate the shortest actual distance from the pixel in the cylindrical mesh to the crack center path, and select the minimum value. At the same time, record the position of the crack path closest to the pixel and the corresponding crack width. Combine the minimum value and the corresponding crack width to calculate the relative depth of the microcrack defect. For pitting defects, the pitting geometry mask is read and the pitting center, major axis, minor axis, shape index, reference direction, and low-frequency perturbation are extracted. For pixels in the pitting area, their cylindrical coordinates are converted to local rotation coordinates with the pitting center as the origin according to the reference direction. Based on the relative position of the pixel along the major and minor axes of the pitting, its normalized radial position is calculated, and the relative depth of the pitting is calculated based on the normalized radial position. For peeling defects, read the peeling geometry mask, calculate the shortest Euclidean distance from each pixel to the outer boundary of the peeling inside the peeling mask, select the largest boundary distance from the current peeling defect, normalize the boundary distance of each pixel by the largest boundary distance, and generate the peeling relative depth. The relative depths of the three types of defects are combined to form a relative depth map.

[0009] As a preferred embodiment of the roll surface defect detection method of the present invention, the step of establishing a normal roll surface background library based on the six-channel roll scalar features and collecting real defect patches, forming training data based on the real defect patches, training the generator and discriminator with the training data to output a defect appearance generation network, and forming a virtual defect training set based on the defect appearance generation network specifically involves: establishing a normal roll surface background library based on the six-channel standardized roll surface features; Simultaneously collect real defect maps and label the real defect areas and defect categories; The real defect patch is used as the target image that the generative network needs to learn. Based on the real defect mask, the defect area in the patch is deleted. Then, the existing image inpainting algorithm is used to fill in the deleted area based on the normal grinding marks, local brightness and polarization features around the defect, so as to obtain a defect-free background patch corresponding to the original real defect patch. For each real defect sample, the distance from each pixel inside the defect region to the mask boundary is calculated according to its geometric mask. The positions with larger distances are marked as relatively deep, and the positions with closer distances to the boundary are marked as relatively shallow, thus obtaining a relative depth guide map that matches the real defect mask. Read the fatigue vulnerability probability corresponding to the actual defect location on the roll and generate training data; The generator network uses the U-Net conditional generator, which combines the training data according to the channel direction to form the generator input. The encoding part of the U-Net generator extracts the comprehensive features of normal background and defect conditions step by step, and the decoding part restores the image resolution step by step. It also preserves the original normal roll texture through the skip connection between the encoding layer and the decoding layer, and finally outputs a six-channel candidate defect image with the same size as the normal background. The discriminator uses the PatchGAN local discriminator. The discriminator receives the defect image and its corresponding defect mask and category information, divides the image into multiple local regions, and determines whether each local region is a real defect appearance or a generated defect appearance. The discriminator trains the generator and outputs the defect appearance generation network. The corresponding defect appearance is generated based on the defect appearance generation network, and candidate six-channel defect images are output to form a virtual defect training set.

[0010] As a preferred embodiment of the roll surface defect detection method of the present invention, the step of constructing a lightweight roll defect segmentation network and defining detection labels, and establishing defect training data to train the lightweight roll defect segmentation network specifically involves: combining virtual defect samples, real defect samples, and real normal samples to form defect training data and labeling them; for each sample, connecting the six-channel standardized roll surface features and fatigue vulnerability probability according to the channel direction to form a seven-channel network input. A lightweight roll defect segmentation network was constructed using a MobileNetV3-Small encoder and a U-Net decoder, and the detection labels were defined as four categories, including normal, microcracks, pitting, and spalling. The defect segmentation network is trained using a combination of multi-class cross-entropy loss and Dice loss. Different training weights are set for different defect categories. The AdamW optimizer is used to update the network parameters, and the existing learning rate decay method is used to gradually reduce the learning rate. If the loss does not decrease for several consecutive training rounds, the training is stopped, the defect segmentation network parameters are output, and the final lightweight roll defect segmentation network is formed.

[0011] As a preferred embodiment of the roll surface defect detection method of the present invention, the step of identifying and outputting roll surface defects based on the trained defect segmentation network specifically involves: extracting six-channel roll surface features and fatigue vulnerability probability from the collected target roll machine vision data image, and then outputting the final lightweight roll defect segmentation network to obtain the roll surface defect detection result; and generating a roll surface defect detection map and a detection report based on the roll surface defect detection result.

[0012] Secondly, the present invention provides a roll surface defect detection system, comprising, The data acquisition and feature extraction module is used to acquire roll parameters and machine vision data, establish the coordinates of the roll cylindrical surface and a unified cylindrical mesh, and map the machine vision data to the unified cylindrical mesh, and further generate six-channel roll scalar features and imaging quality masks. The roll parameter analysis module is used to calculate roll stress based on roll parameters and generate mechanical fatigue damage index, roll surface temperature gradient and machine vision surface perturbation features, and further obtain fatigue vulnerability probability and preferred propagation direction of roll surface cracks. The roll defect construction module is used to classify surface defect types, determine the location of defect initiation points based on fatigue vulnerability probability and generate defect geometric masks, and calculate the maximum relative depth for each type of defect to form a relative depth map; The virtual data generation module is used to establish a normal roll surface background library based on the six-channel roll scalar features and collect real defect patches. Training data is formed based on the real defect patches. The generator and discriminator are trained using the training data to output a defect appearance generation network and a virtual defect training set is formed based on the defect appearance generation network. The detection network construction and output module is used to construct a lightweight roll defect segmentation network and define detection labels, establish defect training data to train the lightweight roll defect segmentation network, and identify and output roll surface defects based on the trained defect segmentation network.

[0013] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the roll surface defect detection method as described in the first aspect of the present invention.

[0014] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein, when the computer program is executed by a processor, it implements any step of the roll surface defect detection method as described in the first aspect of the present invention.

[0015] The beneficial effects of this invention are as follows: This invention achieves co-location characterization of multi-source data on the roll surface through encoder synchronous polarization imaging, infrared thermometry, and cylindrical mesh mapping, reducing the interference of reflection and unfolding seams on defect features. By fusing contact load, thermal gradient, cyclic damage, and visual perturbation to generate fatigue vulnerability probability, it achieves correlation between defect appearance and service conditions. By constraining geometric masks and relative depths based on risk location and crack direction, it forms defect samples with accurate labels and reasonable physical morphology. By using mask-constrained generation to maintain the normal roll background, and by using real, virtual, and normal samples to train a lightweight segmentation network, it improves the identification ability of microcracks, pitting, and spalling, while taking into account detection accuracy, sample expansion effect, and field deployment efficiency. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the roll surface defect detection method in Example 1.

[0018] Figure 2 A schematic diagram of the physical fatigue vulnerability field on the surface of the roll.

[0019] Figure 3 This is a schematic diagram of virtual defect sample generation and lightweight segmentation network training.

[0020] Figure 4 This is a structural diagram of the roll surface defect detection system in Example 1. Detailed Implementation

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0023] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0024] Example 1, referring to Figures 1-4 This is the first embodiment of the present invention, which provides a method for detecting surface defects in rolls, comprising the following steps: S1. Collect roll parameters and machine vision data, establish the coordinates of the roll cylindrical surface and a unified cylindrical mesh, and map the machine vision data to the unified cylindrical mesh to further generate six-channel roll scalar features and imaging quality mask; S1.1 Obtain the geometric parameters, material parameters, rolling process parameters and machine vision data of the target roll, and establish data association according to the same roll number and the same acquisition time; Specifically, the geometric parameters include the reference radius and the effective length of the roll body. The reference radius is obtained by uniformly selecting multiple positions along the axial direction of the target roll body, measuring the actual diameter at each position using a laser diameter gauge, and taking the median of the multiple measured diameters as the reference diameter of the target roll. The reference radius is then calculated based on the reference diameter. The effective length of the roll body is obtained from the roll design drawings or mill equipment files. The material parameters correspond to the rolls and the strip being rolled. The roll material parameters are obtained from the corresponding material tests, including the roll material's elastic modulus, Poisson's ratio, yield strength, coefficient of linear expansion, fatigue strength at the reference number of cycles, and fatigue index. The strip material parameters are obtained from the material archive, including the strip's elastic modulus and Poisson's ratio. The rolling process parameters are read from the mill controller, including the current total rolling force, current reduction, inlet tension, outlet tension, roll surface linear velocity, strip linear velocity, and roll speed. Simultaneously, the cumulative number of rotations since the last time the roll was ground or replaced is calculated based on the roll speed. Machine vision data is obtained using a four-channel polarization linear array camera synchronously triggered by a rotary encoder. The camera's imaging lines are arranged along the roll's axial direction, ensuring that different circumferential positions of the roll pass through the same illumination area and observation position sequentially during its rotation. The four polarization channels acquire raw polarization images with polarization directions of 0°, 45°, 90°, and 135°, respectively. Simultaneously, an infrared linear array camera synchronously triggered by the rotary encoder acquires raw images of the roll's surface temperature. Before formally acquiring roll images, the light source is turned off and the camera is blocked, and dark-field images of the four polarization channels are acquired. The optimal number of dark-field image acquisitions is 200. Averaging these 200 images reduces random readout noise to approximately [a fraction of] the noise of a single acquisition. Meanwhile, the calibration time remains within an acceptable range. Dark field references are calculated for each polarization direction and axial pixel. That is, the axial pixels of each polarization channel image are summed in all acquired dark field images and the average value is calculated to obtain the dark field reference for each axial pixel. Then, a non-polarized uniform diffuse reflection standard plate is placed at the roll imaging position. The same exposure time and gain as the formal inspection are maintained, and white field images are acquired. The number of white field image acquisitions is also 200 times. The white field reference for each axial pixel is calculated. The acquired machine vision data is corrected. Specifically, the unified white field reference response of all effective pixels in the four polarization channels is first calculated. That is, the difference between the white field reference and the dark field reference of each axial pixel is calculated. The difference is summed in all axial pixels and all polarization channels and then averaged to obtain the unified white field reference response. The total number of axial pixels is the number of pixels along the axis of the line scan camera. The gain of the original line image of each polarization channel is corrected by the unified white field reference response. Gain correction is specifically expressed as:

[0025] in polarization direction The original line image of the j-th axial pixel at the k-th encoder trigger number. and polarization direction The white and dark reference values ​​of the j-th axial pixel. To standardize the white field reference response, This is the camera's maximum digital grayscale value. This represents the polarization intensity after gain correction, with a value ranging from 0 to 1. S1.2. Define the circumferential and axial coordinates of the roll surface. Establish the coordinates of the roll cylindrical surface. The circumferential coordinates need to be between 0 and the roll circumference, and the axial coordinates need to be between 0 and the effective length of the roll. Set a unique zero-phase mark at the end of the roll, and define the circumferential position when the encoder zero-phase pulse arrives as the 0 position of the circumferential coordinates. Record the number of effective trigger pulses actually generated per encoder revolution and calculate the roll rotation angle corresponding to each effective trigger pulse. Calculate the corresponding circumferential coordinates through the roll rotation angle. Calculate the axial pixel physical resolution by setting a calibration plate. Specifically, set a calibration plate with a known scribe line spacing at the roll imaging position, obtain the actual axial distance between two reference points on the calibration plate, and extract the image pixel number of the corresponding reference point. Obtain the axial pixel physical resolution by dividing the actual axial distance between the two reference points by the difference between the image pixel numbers. Record the pixel number corresponding to the left end of the roll. Obtain the axial coordinate by multiplying the difference between the axial pixel number and the pixel number corresponding to the left end of the roll by the axial pixel physical resolution. S1.3 Define the target minimum stable detection crack width and the minimum crack width coverage pixels. Obtain a uniform cylindrical grid spacing based on the ratio of the target minimum stable detection crack width and the minimum crack width coverage pixels. The target minimum stable detection crack width is preferably 0.10 mm. At this width, roll defects can be stably identified without overemphasizing extremely small targets, ensuring applicability. The minimum crack width coverage pixels are preferably 4. If the crack width only covers 1 pixel, it is difficult to distinguish that pixel anomaly from camera dead pixels, random noise, or interpolation errors. If it covers 2 pixels, although the basic sampling requirements are met, it is still difficult to distinguish the inside and sides of the crack. Simultaneous representation, when covering 4 pixels, can usually form crack interior response, left boundary response, right boundary response, and at least one transition pixel, which helps to extract edge and anisotropic features. Therefore, the preferred value for the uniform cylindrical grid spacing is 0.025mm. The number of grids in the circumferential direction is calculated by the circumference of the roll and the uniform cylindrical grid spacing, and the coordinates of each grid in the circumferential direction are obtained. At the same time, the number of grids in the axial direction is calculated by the effective length of the roll and the uniform cylindrical grid spacing, and the coordinates of the axial grid are obtained. The circumferential grid coordinates and the axial grid coordinates are combined to form a cylindrical grid point, and the continuous coordinates of each cylindrical grid point in the machine vision data image are obtained. Specifically, for each cylindrical grid point, its corresponding continuous coordinates in the machine vision data image are represented as follows:

[0026] in For the nth circumferential direction grid coordinates, Let m be the coordinate of the m-th axis grid. This refers to the actual number of valid trigger pulses generated per revolution of the encoder. The reference radius of the roll. The physical resolution of the axial pixels. This is the pixel number corresponding to the left end of the roller body. and Continuous coordinates; Based on the correspondence between cylindrical grid points and continuous coordinates, bicubic interpolation is used to map the corrected machine vision data image onto a unified cylindrical grid, including the corrected polarization image and the original temperature map. It should be noted that when mapping the image, a periodic boundary needs to be set in the circumferential direction to avoid defects located near the zero-phase seam of the cylindrical unfolded diagram being truncated. S1.4 Calculate the total intensity component and the linear polarization difference component based on the polarization images of the four polarization directions. Specifically, the total intensity component is obtained by dividing the sum of the polarization intensities of corresponding pixels in the polarization images of the four polarization directions by 2. The linear polarization difference component includes the intensity difference between the 0° and 90° polarization directions and the intensity difference between the 45° and 135° polarization directions. Based on the total intensity component and the linear polarization difference component, construct the low-light confidence score and the degree of linear polarization. Specifically, the low-light confidence score is obtained by dividing the difference between the total intensity component and the low-light threshold by the difference between the high-light threshold and the low-light threshold. The difference in values ​​is truncated within the range of 0 to 1. The preferred values ​​for the high illumination threshold and the low illumination threshold are 0.15 and 0.03, respectively. The high illumination threshold is used to determine the intensity level at which the polarization feature transitions from partially reliable to fully reliable. 0.15 represents the intensity starting point at which the polarization estimation reaches a stable signal-to-noise ratio. The low illumination threshold is used to determine whether the polarization information of the current pixel has been dominated by dark current and readout noise. 0.03 makes the probability of pure dark field noise exceeding this threshold sufficiently low. The degree of linear polarization is obtained by taking the square root of the sum of the squares of the two linear polarization difference components and the ratio of the total intensity component. The linear polarization direction angle is calculated based on the linear polarization difference components, and the linear polarization direction is continuously encoded. Specifically, since the linear polarization direction has a periodicity of 180°, directly using the linear polarization direction angle will result in numerical jumps at the periodic boundaries. Therefore, a double-angle continuous encoding is adopted, which includes two forms: cosine and sinine. The two forms of double angle are expressed as follows:

[0027]

[0028] in This is an approximate representation of a double angle in cosine form. The approximate representation of double angles in sin form, all within the range of -1 to 1, For low illumination confidence level, and For linearly polarized differential components, As the stability constant, it is preferred to take ; Based on the Stokes parameter relationship for linear polarization, under ideal conditions where the linear polarization amplitude is non-zero and numerical stability terms are not considered... and These are strictly equal to cos² and sin², respectively. By multiplying by the low-illuminance confidence level and adding a stability constant to the denominator, the directional encoding decays towards zero in the low-illuminance or low-polarity region, resulting in the actual obtained... and For double-angle continuous encoding with confidence amplitude, when the effective polarization signal is obvious, it is approximately equal to cos double angle and sin double angle respectively; The total intensity of the roll surface can be approximated as the product of the slowly changing illumination term and the local surface reflection texture. The logarithm of the total intensity is taken, and then the low-frequency background is subtracted to extract the local reflection texture features. The low-frequency background is obtained by convolving the logarithm of the total intensity with a two-dimensional Gaussian kernel. The physical background scale of the Gaussian kernel is preferably 0.50 mm. When the uniform cylindrical grid spacing is 0.025 mm, the pixel standard deviation corresponding to the Gaussian kernel is 20. This scale is larger than the width of the early microcracks, which can remove the slow changes of oil film and illumination, while retaining the local texture anomalies caused by microcracks and pitting. Two roll defect size detection scales are defined, preferably 0.05mm and 0.15mm, respectively. 0.05mm is half the target minimum crack width, and 0.15mm is 1.5 times the target minimum crack width. These scales are used to enhance the edges of cracks with larger widths, pitting boundaries, and spalling boundaries. The corresponding pixel scale is calculated according to the roll defect size detection scale, where 0.05mm corresponds to a pixel scale of 2 pixels and 0.15mm corresponds to a pixel scale of 6 pixels. The circumferential gradient and axial gradient after Gaussian kernel smoothing are calculated for the two pixel scales, and the edge strength at the corresponding scale is obtained. The larger edge strength value between the two pixel scales is selected as the multi-scale edge strength. At each pixel scale, a structure tensor is constructed and two feature values ​​of the structure tensor are extracted. The first feature value is greater than or equal to the second feature value and both feature values ​​are greater than or equal to 0. The texture anisotropy at the corresponding pixel scale is obtained by the ratio of the difference between the two feature values ​​to the sum. The larger of the texture anisotropy at the two pixel scales is selected as the multi-scale anisotropy. The local reflection texture features, linear polarization degree, multi-scale edge strength, and multi-scale anisotropy are used to form the roll scalar features to be standardized. The 5% and 95% quantiles of each item in the roll scalar features are extracted respectively. Specifically, after the target roll has been ground for the most recent time, a reference scan is performed on the entire roll body. The absence of obvious cracks, pitting, and peeling on the current surface is confirmed by manual inspection or non-destructive testing. This scan is recorded as the normal reference scan. The 5% and 95% quantiles of each item are calculated in the normal reference scan. For the axial direction, an axial local statistical half-width needs to be set, and the quantiles are unified within the axial local statistical half-width range. Each item in the roll scalar features to be standardized is standardized according to the quantiles. The standardized roll scalar features need to be truncated in the range of 0 to 1. At the same time, the linear polarization direction code is mapped to 0-1. The mapped linear polarization direction code and the standardized roll scalar features are connected and combined to form a new six-channel roll scalar feature. S1.5. An imaging quality mask is generated by combining the low-light indicator, channel saturation indicator, and encoder trigger integrity indicator. Specifically, the low-light indicator is either 0 or 1. It is set to 1 when the total intensity feature is greater than or equal to the low-light threshold, and 0 otherwise. The channel saturation indicator is either 0 or 1. It is set to 1 when the polarization intensity of the largest pixel in the polarization channel is greater than the saturation threshold, and 0 otherwise. The saturation threshold is preferably 0.98. When the camera is close to full scale, it may experience nonlinear response, pixel charge overflow, or quantization compression. A 2% full-scale margin is reserved. Pixels exceeding this value are no longer used to accurately calculate the ratio between the four polarization channels. The encoder trigger integrity indicator is either 0 or 1, corresponding to the k-th scan line. If the encoder counts continuously, the indicator is 1; otherwise, it is 0. The imaging quality mask is obtained by multiplying the low-light indicator, channel saturation indicator, and encoder trigger integrity indicator. The imaging quality mask is 1 only when all three are 1.

[0029] S2. Calculate roll stress based on roll parameters and generate mechanical fatigue damage index, roll surface temperature gradient and machine vision surface perturbation features, and further obtain fatigue vulnerability probability and preferred roll surface crack propagation direction; S2.1. The axial coordinates are converted into dimensionless positions by using the effective length of the roller body, and the axial linear load distribution is constructed using a zero-mean quadratic basis function. Specifically, the axial linear load distribution is represented as follows:

[0030] in Let V be the normal load per unit length at axial position v, and F be the current total rolling force. The effective length of the roller body. The axial load shape factor is determined through pressure calibration tests after the rolls are installed in the mill. , which is the dimensionless position of the axial coordinate; Since the axial coordinate is between 0 and Within range The average value is ,therefore Furthermore, the axial load satisfies The axial distributed load is consistent with the total rolling force measured by the rolling mill; The equivalent elastic modulus between the roll and the strip is calculated based on the material parameters, and the reference radius of the roll is used as the equivalent radius of curvature. The Hertz contact half-width and maximum contact pressure at the axial position are calculated by combining the equivalent elastic modulus and the equivalent radius of curvature. Specifically, the Hertz contact half-width and maximum contact pressure are expressed as follows:

[0031] in For Hertzian contact half-width, For maximum contact pressure, The equivalent radius of curvature, It is the equivalent elastic modulus; To obtain a complete contact pressure distribution, a local contact coordinate x is established in the rolling direction. When x is 0, it represents the center of the contact area; when x is 0, it represents the center of the contact area. At that time, the contact pressure is ,when When the contact pressure is 0, within a contact half-width of 2 Hertz, the total integral value of the contact pressure is equal to The contact load obtained by integrating along the rolling direction is consistent with the axial linear load; The maximum bending stress on the roll surface at the axial position is obtained by calculating the bending moment at the axial position and the moment of inertia of the roll's circular cross-section. The ratio of the bending moment multiplied by the effective radius of the roll to the moment of inertia is then used to obtain the maximum bending stress on the roll surface at the axial position. Bending moment at axial position Represented as:

[0032] in Total rolling force For axial integration variables; Moment of inertia of the circular cross section of the roll Represented as:

[0033] in The effective diameter of the roll; The median temperature in the circumferential direction is extracted for each axial position based on the machine vision data image mapped to a unified cylindrical mesh, and the local temperature difference between the current point and the median temperature is calculated. The constraint thermal stress caused by the local temperature difference is then calculated based on the local temperature difference. Constrained thermal stress Represented as:

[0034] in The elastic modulus of the roll material. The coefficient of linear expansion of the roll material. Poisson's ratio, For local temperature differences, The surface thermal constraint coefficient is preferably 0.65. If it is 1, it means that the surface area cannot expand freely at all. If it is 0, it means that it can expand freely. The local surface of the roll is constrained by the surrounding material but is not completely fixed. Therefore, 0.65 is selected. The difference between the exit tension and the inlet tension is calculated, and the ratio of this difference to the total rolling force is taken to obtain the equivalent friction coefficient. After obtaining the ratio, it needs to be truncated between the lower limit and the upper limit of the friction coefficient. The preferred values ​​for the lower limit and the upper limit are 0.02 and 0.20, respectively. The lower limit of 0.02 is used to avoid the friction effect being completely eliminated due to tension measurement noise, and the upper limit of 0.20 is used to avoid excessive shear stress that does not conform to the lubrication rolling conditions caused by abnormal tension signals. The slip direction coefficient is calculated based on the strip linear velocity and the roll surface linear velocity. Specifically, the slip direction coefficient is obtained by calculating the ratio of the difference between the strip linear velocity and the roll surface linear velocity to the absolute value of the difference. The product of the maximum contact pressure, the equivalent friction coefficient, and the slip direction coefficient is taken as the roll surface shear stress. The equivalent friction coefficient and the slip direction coefficient are dimensionless values ​​when calculated. S2.2 Calculate the circumferential and axial surface normal stresses based on the maximum contact pressure, constraint thermal stress, and maximum bending stress on the roll surface. The circumferential normal stress is the difference between the constraint thermal stress and the maximum contact pressure, and the axial surface normal stress is the sum of the constraint thermal stress and the maximum bending stress on the roll surface. Under two-dimensional plane stress conditions, calculate the equivalent stress on the roll surface. Then, read the cumulative number of rotations of the target roll since the most recent grinding, where the cumulative number of rotations is obtained by integrating the roll speed. Calculate the mechanical fatigue damage index based on the Basquin fatigue relation and the linear cyclic damage concept. Specifically, the equivalent stress on the roll surface Represented as:

[0035] in The surface normal stress is in the circumferential direction. For axial surface normal stress, This refers to the shear stress on the roll surface. The surface von Mises equivalent stress is derived from the distortion energy intensity theory under two-dimensional plane stress conditions. It is used to unify normal stress, bending stress, thermal stress and frictional shear stress into an equivalent load scale. Mechanical fatigue damage index Represented as:

[0036] in To accumulate the number of revolutions, The number of cycles is obtained from the roll's factory parameters for reference. For fatigue strength, The fatigue index. As the stability constant, it is preferred to take ; S2.3 Calculate the circumferential and axial temperature gradients of the roll surface respectively, and synthesize them into the roll surface temperature gradient. Use a Gaussian kernel to filter the local reflective texture features to extract the local normal texture background. Calculate the absolute value of the difference between the normalized local reflective texture and the local normal texture background, and then calculate the machine vision surface perturbation features. Specifically, the machine vision surface perturbation features are obtained by calculating the ratio of the absolute value of the difference to the absolute median difference of the local reflective texture features. During the calculation, the absolute median difference needs to be converted into a robust scale similar to the standard deviation through a coefficient. To avoid low-quality image areas affecting the fatigue vulnerability field, an imaging quality mask is used to correct the machine vision surface perturbation features. S2.4 The mechanical fatigue damage index, the temperature gradient of the roll surface, and the corrected machine vision surface perturbation features are weighted after quantile-stable normalization to obtain the comprehensive fatigue driving amount of the roll. The sum of all weights is , and the preferred values ​​are 0.60, 0.25, and 0.15, respectively. The mechanical fatigue damage has the highest weight because cyclic contact load and bending load are the main driving factors for the formation of roll fatigue defects. The temperature gradient has the second highest weight, which is used to reflect local thermal shock. The machine vision perturbation has the lowest weight, which is only used to compensate for surface roughness changes, micro scratches, and early texture anomalies that cannot be directly described by the analytical mechanical model. The sigmoid function is used to map the comprehensive fatigue driving amount to a fatigue vulnerability probability in the range of 0-1. S2.5. Based on the principal stress direction relationship under two-dimensional stress state, calculate the maximum principal stress direction angle, where the maximum principal stress direction angle refers to the angle between the maximum principal stress direction and the roll circumferential direction. The calculation is performed using the four-quadrant arctangent function. The input is the shear stress on the roll surface and the difference between the surface normal stress in the circumferential and axial directions. Calculate the preferred crack propagation direction on the roll surface based on the maximum principal stress direction angle. The preferred crack propagation direction on the roll surface consists of two components: the first component is the negative sin value of the maximum principal stress direction angle, corresponding to the roll circumferential direction; the second component is the cos value of the maximum principal stress direction angle, corresponding to the roll axial direction.

[0037] S3. Classify the surface defects, determine the location of the defect initiation point based on the fatigue vulnerability probability and generate a defect geometric mask, and calculate the maximum relative depth for each type of defect to form a relative depth map. S3.1 Divide the defects to be generated on the roll surface into three categories: microcracks, pitting, and spalling, and assign category labels to each. Count the actual number of occurrences of the three types of defects to determine the generation ratio of the three types of defects in the virtual sample. The actual number of occurrences can be counted from real roll maintenance records, grinding verification records, or non-destructive testing records. In order to avoid the fact that a certain type of defect cannot be generated at all when there are few real samples, set a basic prior quantity for the number of the three types of defects so that each type of defect has a non-zero generation probability. Before sufficient real defect statistics have been accumulated, it is preferable to set the initial generation ratios of microcracks, pitting, and spalling to 60%, 25%, and 15%, respectively. Microcracks account for the highest proportion because they have small areas, low contrast, and are easily confused with grinding marks or local reflections, making them the most likely defects to be missed in machine vision inspection. Pitting and spalling usually have large areas and obvious boundary changes, making them relatively easy for the detection model to identify. Therefore, their initial proportions are relatively lower. As the number of confirmed inspection samples increases, the actual number of the three types of defects is recounted, and the category generation ratios are updated so that the virtual defect category distribution gradually approaches the actual defect distribution of the target roll, rather than maintaining a fixed proportion in the long term. The fatigue vulnerability probability is divided into fixed-size blocks according to a uniform cylindrical grid. The block size is consistent with the input size of the subsequent defect appearance generation network and segmentation network, preferably 256×256 pixels. When the uniform cylindrical grid spacing is 0.025 mm, the actual surface area of ​​the roll corresponding to one block is 6.4 mm×6.4 mm. This range can accommodate short cracks, pitting, and local spalling, while ensuring that the proportion of small defects in the whole image is not too low due to the block area being too large. The average fatigue vulnerability degree is calculated for each block by averaging the fatigue vulnerability probabilities in the block. A higher average fatigue vulnerability degree indicates that the area is more prone to fatigue defects under the combined effects of contact load, bending stress, thermal stress, frictional shear stress, and machine vision surface perturbation. Therefore, more defects are allowed to be generated in the block. When the average fatigue vulnerability degree is low, only a few defects are generated. To prevent excessive density of defects in a single block, the number of defects in a block is preferably limited to 1 to 4. S3.2 Within the block, the probability of a defect initiation point is obtained by weighted sampling based on the fatigue vulnerability probability of each pixel. The higher the fatigue vulnerability probability, the higher the probability of the location being selected as a defect initiation point. To ensure that multiple defect instances have distinguishable geometric boundaries, a minimum physical distance is set for different initiation points. The physical distance is converted into the number of pixels through a uniform cylindrical grid spacing. The defect initiation point is selected based on the probability of the defect initiation point. If the distance between the newly selected initiation position and the existing initiation position is less than the minimum distance, the position is abandoned and resampling is performed. It should be noted that the fatigue vulnerability probability is used to increase the possibility of high-risk positions being selected, rather than simply setting all local maximum values ​​as defects. This can maintain the consistency between the defect position and the roll stress state, and also retain a certain sample diversity, avoiding the generation of samples being concentrated in a few risk peak positions. S3.3 Generate geometric masks according to the three types of defects. Specifically, for the initiation location of a microcrack, this location is taken as the crack initiation point, and the next path point is selected from the adjacent pixels of the current path point. The selection of candidate path points considers fatigue vulnerability probability, physical crack direction, path continuity, and related perturbations. The physical crack direction represents the physical direction deviation cost, which is calculated by the candidate unit direction vector of the candidate path point and the current position and the preferred crack propagation direction on the roll surface, and is expressed as 1 minus the square of the inner product of the candidate unit direction vector and the preferred crack propagation direction on the roll surface. Path continuity represents the path turning cost, which is calculated by the candidate unit direction vector and the current position. The candidate unit direction vector is calculated as 1 minus the inner product of two candidate unit direction vectors, divided by 2. If the current position is the crack initiation point, a random variable is obtained by sampling from an independent standard normal distribution. The initial extension sign of the crack initiation point is determined based on the random variable and multiplied by the preferred extension direction of the crack on the roll surface at the crack initiation point. If the random variable is greater than or equal to 0, the initial extension sign is 1; otherwise, it is -1. The related disturbance represents the related disturbance cost, which is obtained by generating independent standard normal noise of the same size as the current patch and then performing Gaussian kernel filtering, and normalizing the filtering result. The fatigue vulnerability probability, physical crack direction, path continuity, and related disturbance are weighted. The cost of selecting candidate path points is obtained, where the fatigue vulnerability probability is negative and incremented by 1 during weighting. The preferred weighting values ​​are 0.45, 0.30, 0.20, and 0.05, with fatigue vulnerability having the highest weight to ensure the crack path primarily conforms to the fatigue vulnerability distribution on the roll surface. The physical direction weight is secondary to ensure the path extension direction matches the stress result. Continuity weight is used to avoid jagged paths, and disturbance weight is the lowest to prevent random patterns from disrupting physical laws. Candidate path points are selected based on the minimum cost to form the crack path. The process stops when the cumulative actual crack length reaches a preset value, there are no valid candidate points, or the block boundary is reached. To determine the crack width, the following steps are first performed: The standard normal random variable is input into the sigmoid function and multiplied by the fatigue vulnerability probability and the difference between the maximum and minimum crack widths. The minimum crack width is then added to obtain the crack width at that location. The maximum and minimum crack widths are preferably 0.20 mm and 0.10 mm, respectively. The minimum crack width is larger than the physical scale of a single pixel in the standard image, which can form a weak crack that can be learned by the detection network. The maximum crack width is used to cover cracks that have expanded to a certain extent but have not yet formed a large area of ​​peeling. Pixels that are less than half the crack width from the center path are set as crack regions according to the crack width. The crack regions are used to generate microcrack geometric masks. For the initiation location of pitting corrosion, this location is taken as the pitting corrosion center, and the preferred crack propagation direction at this location is read and used as the reference direction for the geometric major axis of the pitting corrosion. Using this direction does not equate pitting corrosion with cracks, but rather considers that slippage on the roll surface, tangential traction, and principal stress directions will cause local material damage to propagate along a certain direction, making actual pitting corrosion usually not perfectly circular. The major and minor semi-axes of the pitting corrosion region are calculated separately and converted to pixel scale using a uniform cylindrical grid spacing. The major semi-axe is obtained by multiplying the sigmoid value of a standard normal random variable by the difference between the maximum and minimum major semi-axes, plus the minimum major semi-axe. The preferred values ​​for the maximum and minimum major semi-axes are 0.60 mm and 0.10 mm, respectively. The minimum major semi-axe corresponds to 4 pixels, which avoids generating sub-pixel pitting corrosion that cannot form an effective area. The maximum major semi-axe is 0.60 mm. mm can cover early pitting and pitting areas that have already expanded to a certain extent. The minor axis is obtained by multiplying the sigmoid value of a standard normal random variable by 0.4, adding 0.6, and finally multiplying by the major axis. Therefore, pitting is not fixed to be a circle, but varies between a circle and an ellipse with a certain directionality. A standard normal random variable is regenerated, and the sigmoid value of twice the standard normal random variable is added by 2 to obtain the pitting shape index. Based on the major and minor axes, a rotated hyperellipse is used to construct the main body of the pitting. An amplitude-controlled hyperellipse is added to the boundary of the hyperellipse. The limited low-frequency perturbation causes the boundary to undulate slowly rather than forming a completely regular mathematical curve. The low-frequency perturbation also needs to be calculated using standard normal random variables. First, the local boundary angle of the boundary point of the pitting region is obtained. Then, three sets of standard normal random variables are generated, each containing two. The unnormalized angular perturbation is constructed by combining the local boundary angle and the three sets of standard normal random variables. The unnormalized angular perturbation is normalized to obtain the angular perturbation. After the pitting region is generated, the pixels inside the rotated superellipse are set as the pitting region to generate the pitting geometric mask. Specifically, unnormalized angular perturbation Represented as:

[0038] in For local boundary angles, and Let h be a standard normal random variable in the h-th group; For the initiation location of the spalling category, the spalling region is constructed by combining multiple pitting sub-regions with short cracks. Spalling on the roll surface is typically formed by the initiation of microcracks, the propagation of local pitting, and the interconnection of multiple damaged areas. While directly using regular large circles or ellipses to represent spalling is simple, it fails to reflect the geometric characteristics of spalling as multiple local fatigue damages gradually connect. A standard normal random variable is generated centered on the spalling initiation location. The number of pitting sub-regions is obtained by downsampling based on three times the sigmoid value of the standard normal random variable and adding 2. The location of each sub-region is still prioritized. Located in a position with a high probability of fatigue damage and larger in size than ordinary early pitting corrosion, but with a certain overlap between different sub-regions, thus forming an irregular material detachment body. In order to make the size of the detached sub-region larger than that of ordinary pitting corrosion, the range of the major semi-axis of the sub-region is determined to be 0.15mm and 0.80mm. Subsequently, one or two short cracks are generated between the pitting sub-regions. The short cracks preferably extend in the direction of propagation and in the local high fatigue damage area to simulate the connection of fatigue cracks between multiple local damage areas. The union of all pitting sub-regions and short crack regions is obtained to obtain the detachment geometry mask for the detachment region. S3.4. For each type of defect, define the maximum relative depth range. For microcracks, the maximum relative depth is preferably limited to between 0.002 mm and 0.020 mm. For pitting, the maximum relative depth is preferably limited to between 0.010 mm and 0.100 mm. For spalling, the maximum relative depth is preferably limited to between 0.050 mm and 0.300 mm. It should be noted that the maximum relative depth range should be determined primarily based on the historical grinding records of the target roll, the surface profilometer measurement results, or the defect cross-section data after maintenance. In specific implementation, the measured depths of defects of the same type are statistically analyzed, and the lower quantile depth is used as the lower limit and the higher quantile depth is used as the upper limit, so as to avoid individual abnormally deep defects determining the depth range of the entire virtual sample. When sufficient measured depth data has not yet been accumulated, the defined range can be used as the initial implementation range. The depth range of microcracks is the smallest because it mainly manifests as narrow and shallow surface grooves. Pitting is a local material depression, and its depth is usually greater than that of early microcracks. Spalling involves a large area of ​​material detachment, so a larger depth range is set. For microcrack defects, the microcrack geometry mask is read and the microcrack path and crack width are extracted. The shortest actual distance from the pixel in the cylindrical mesh to the crack center path is calculated. That is, the distance from the pixel to each crack path point or adjacent path segment is calculated in turn, and the minimum value is selected. At the same time, the position of the crack path closest to the pixel and the corresponding crack width are recorded. The relative depth of the microcrack defect is calculated by combining the minimum value and the corresponding crack width. Relative depth of microcrack defects Represented as:

[0039] in Let x be the relative depth of the k-th microcrack defect at pixel x. This represents the maximum relative depth of the corresponding defect. The minimum distance, The crack width at the nearest path location. As a microcrack mask, As the stability constant, it is preferred to take ; For pitting defects, the pitting geometry mask is read and the pitting center, major axis, minor axis, shape index, reference direction, and low-frequency perturbation are extracted. For pixels in the pitting area, their cylindrical coordinates are converted to local rotation coordinates with the pitting center as the origin according to the reference direction. Based on the relative position of the pixel along the major and minor axes of the pitting, its normalized radial position is calculated, and the relative depth of the pitting is calculated based on the normalized radial position. Normalized radial position Represented as:

[0040] in and For pitting the long and short semi-axis, and For local rotation coordinates, For shape parameters, For angular perturbation, The roughness coefficient for pitting edge is preferably 0.10. Increasing this value further... This will make the boundaries more irregular, but if it exceeds 0.20, it is easy to generate local fractures or sharp protrusions. If it is less than 0.05, the pitting corrosion is close to the ideal ellipse, and the morphological diversity is insufficient. As the stability constant, it is preferred to take ; relative depth of pitting Represented as:

[0041] in For pitting geometry mask, The pitting morphology coefficient is preferably set to 1. At this value, the pitting depth decreases approximately linearly and continuously from the center to the boundary, thus avoiding the unfounded artificial formation of sharp or flat bottom shapes. Indicates a positive value; For peeling defects, read the peeling geometry mask, calculate the shortest Euclidean distance from each pixel to the outer boundary of the peeling inside the peeling mask, select the largest boundary distance from the current peeling defect, normalize the boundary distance of each pixel by the largest boundary distance, and generate the peeling relative depth. relative depth of peeling Represented as:

[0042] in This is the shortest Euclidean distance. The maximum boundary distance, To peel off the geometric mask, As the stability constant, it is preferred to take , The morphology coefficient of the peeling bottom is preferably 0.7. The normalized distance increases rapidly inside the peeling, so that a large area of ​​the internal region has a relatively deep depth, which can form a smooth bottom after the sheet material falls off. The relative depth maps are composed of the relative depths of the three types of defects. It should be noted that for each pixel in the cylindrical mesh, all defect instances are checked in turn. If the pixel does not belong to any defect instance, that is, the mask of all instances at the pixel is 0, then the pixel is set as a normal roll surface. If the pixel belongs to only one defect instance, then the mask, category and relative depth of the defect are directly assigned to the pixel. If the pixel belongs to two or more defect instances at the same time, then the absolute value of the relative depth of each instance at the pixel is compared, and the instance with the deepest depression is selected as the dominant instance. If two overlapping instances have the same relative depth at a certain pixel, then the priority is determined according to the damage coverage of spalling-pitting-microcrack. It should also be noted that each time a standard normal random variable is generated, it is obtained by resampling.

[0043] S4. Based on the six-channel scalar features of the roll, establish a normal roll surface background library and collect real defect patches. Based on the real defect patches, form training data. Use the training data to train the generator and discriminator to output a defect appearance generation network and form a virtual defect training set based on the defect appearance generation network. S4.1. Based on the surface characteristics of the six-channel standardized roll, select images whose imaging quality masks meet the requirements and are confirmed by manual inspection or non-destructive testing to be free of cracks, pitting, and peeling, and establish a normal roll surface background library. The normal background library should cover the left, middle and right sides of the roll body, and cover the initial stage after grinding, the normal service period and the state with normal oil film. The background samples of different axial areas are set because there may be different grinding texture density, surface reflection and normal wear degree in the middle and ends of the roll. If the background library only contains images of the middle of the roll body, the generator is likely to mistakenly copy the middle texture to the ends of the roll body. A certain number of real defect blocks confirmed by inspection, grinding verification or non-destructive testing are collected simultaneously, and the real defect areas and defect categories are marked. The role of real defect data is not to directly provide a large number of model training samples, but to determine the actual appearance changes of microcracks, pitting and spalling in the six-channel features. Specifically, a normal region adjacent to each real defect is selected, and the differences between the defect region and the adjacent normal region in local reflection texture, polarization degree, polarization direction encoding, edge intensity, and texture anisotropic channel are compared. Through the median statistics of multiple real defect samples, the typical channel response directions of various defects are obtained. For example, microcracks usually produce obvious responses in edge intensity and linear anisotropic channel, pitting corrosion causes changes in local reflection, polarization degree, and closed boundary, while spalling is characterized by a large range of reflection changes and complex boundaries. The real defect patch is used as the target image that the generative network needs to learn. Based on the real defect mask, the defect area in the patch is deleted. Then, the existing image inpainting algorithm is used to fill in the deleted area based on the normal grinding marks, local brightness and polarization features around the defect, so as to obtain a defect-free background patch corresponding to the original real defect patch. Image restoration preferably uses the PatchMatch texture restoration algorithm, which can search for similar normal texture blocks around the defect and fill the defect area with similar textures. It is suitable for the grinding texture background with strong repetition on the roll surface. For microcracks with small area, a fast image restoration algorithm based on neighborhood diffusion can also be used to fill them. For each real defect sample, the distance from each pixel inside the defect region to the mask boundary is calculated according to its geometric mask. The positions with larger distances are marked as relatively deep, and the positions with closer distances to the boundary are marked as relatively shallow, thus obtaining a relative depth guide map that matches the real defect mask. Read the fatigue vulnerability probability corresponding to the roll position where the real defect is located. If complete process data is missing when the real defect image is acquired, the fatigue vulnerability probability can be calculated according to the adjacent working conditions of the roll. Alternatively, the sample can be used only for appearance authenticity training without participating in fatigue vulnerability condition training, and training data can be formed. Each set of training data includes: the repaired six-channel normal background, the real defect mask, the defect category, the relative depth guide map, the fatigue vulnerability probability, and the original six-channel real defect image. S4.2 The generator network uses the U-Net conditional generator, which combines the training data according to the channel direction to form the generator input. The six-channel normal background provides the generator with the local reflection texture, polarization degree, polarization direction, edge intensity and texture anisotropy of the real roll surface. The defect mask tells the generator which pixel positions to generate defects. The defect category is used to distinguish microcracks, pitting and spalling. The relative depth guide map is used to distinguish the defect center, slope and edge. The fatigue vulnerability probability is used to provide the generator with the fatigue risk background of the defect location. The encoding part of the U-Net generator extracts the comprehensive features of the normal background and defect conditions step by step. The decoding part restores the image resolution step by step and retains the original normal roll texture through the skip connection between the encoding layer and the decoding layer. Finally, the output is a six-channel candidate defect map with the same size as the normal background. It should be noted that, in order to prevent the generator from changing the normal area unrelated to the defect, a transition area is extended outward to the defect mask to accommodate the bright edges, dark edges and polarization transitions generated near the crack, pitting and peeling edges. The transition width is calculated according to the uniform cylindrical grid spacing, preferably corresponding to 2 to 3 pixels. When synthesizing the final image, only the output of the generator in the defect mask and transition area is used. Pixels outside this area directly copy the original six-channel normal background. S4.3 The discriminator adopts the PatchGAN local discriminator. The discriminator receives the defect image and its corresponding defect mask and category information, divides the image into multiple local regions, and determines whether each local region is a real defect appearance or a generated defect appearance. The discriminator trains the generator and outputs the defect appearance generation network. When training the generator, the training objectives are as follows: First, the generated defect region should closely resemble the real defect in terms of local texture, reflection changes, polarization changes, and edge structure. Second, the generated result should be similar to the corresponding real defect image within the defect region. Third, the region outside the defect mask should be consistent with the normal background after repair. Finally, the generated result near the defect boundary should have a continuous transition to avoid obvious pasting boundaries. All training objectives can be achieved directly using the adversarial loss and pixel reconstruction loss commonly used in generative adversarial networks. During training, the data is divided into training and validation sets according to the roll number. Adjacent patches collected from the same roll should not be included in the training and validation sets at the same time to avoid overestimating the generation effect due to the high similarity of background textures. Training is stopped when the validation set result no longer improves after several consecutive training rounds, and the generator with the best validation result is saved. S4.4 Generate corresponding defect appearances based on the defect appearance generation network and output candidate six-channel defect images to form a virtual defect training set. Only the generation results in the defect area and narrow transition area are retained, while other areas continue to use the original normal background. For microcracks, the main formations are dark lines distributed along the crack center path, local bright and dark edges, and linear anisotropic changes. For pitting, the main formations are pit reflections, polarization changes, and closed boundaries that change continuously between the center and the edge. For exfoliation, the main formations are irregular reflection changes with large areas, internal rough textures, and complex edges.

[0044] S5. Construct a lightweight roll defect segmentation network and define detection labels. Establish defect training data to train the lightweight roll defect segmentation network. Based on the trained defect segmentation network, identify roll surface defects and output them. S5.1. Combine virtual defect samples, real defect samples, and real normal samples to form defect training data and label them. For each sample, connect the six-channel standardized roll surface features and fatigue vulnerability probability according to the channel direction to form a seven-channel network input. A lightweight roll defect segmentation network was constructed using a MobileNetV3-Small encoder and a U-Net decoder. Four detection labels were defined: normal, microcracks, pitting, and spalling. Specifically, the number of input channels in the first layer of the MobileNetV3-Small network was increased from the conventional image channels to seven, receiving six channels of roll surface features and one fatigue vulnerability probability channel. The remaining inverse residual module, depthwise separable convolution, and activation function all adopted the MobileNetV3-Small structure. The encoder progressively reduced the feature map resolution and increased the number of feature channels to extract features from local... The multi-scale features from crack edge to pitting and peeling are used to obtain the overall morphology. The shallow coding features mainly retain microcracks, grinding marks and local polarization changes, while the deep coding features mainly reflect the overall relationship between the defect category and the surrounding background. The decoder adopts the existing U-Net structure, restores the spatial resolution by upsampling at each level, and connects the upsampled features at each level with the shallow features of the same resolution in the encoder. The decoder finally outputs four channels, corresponding to normal background, microcracks, pitting and peeling, respectively. After normalizing the four channels, the probability of each pixel belonging to the four target categories is obtained, and the category with the highest probability is used as the initial segmentation category. S5.2. The defect segmentation network is trained based on multi-class cross-entropy loss and Dice loss. Multi-class cross-entropy loss is used to make the predicted class of each effective pixel close to the real class. Dice loss is used to alleviate the class imbalance caused by the number of defect pixels being far less than that of normal background pixels, and to improve the overall overlap between the small crack area and the real label. Different training weights are set for different defect categories. Microcracks have small areas and are easy to miss, so they should be given higher weights. Pitting corrosion has the next higher weight. Peeling areas have large areas and obvious visual features, so they can be given relatively lower weights. The class weights can be determined according to the inverse proportion or square root inverse proportion of the number of pixels of each category in the training set, avoiding complete reliance on human experience. The AdamW optimizer is used to update the network parameters, and the existing learning rate decay method is used to gradually reduce the learning rate. If the loss does not decrease for several consecutive training rounds, the training is stopped, the defect segmentation network parameters are output, and the final lightweight roll defect segmentation network is formed. It should be noted that the training loss is only calculated for pixel locations where the imaging quality mask is valid. For invalid pixels, neither the classification loss nor the Dice loss is calculated.

[0045] S5.3 After extracting the six-channel roll surface features and fatigue vulnerability probability from the collected machine vision data image of the target roll, the final lightweight roll defect segmentation network is output to obtain the roll surface defect detection results. Based on the roll surface defect detection results, the roll surface defect detection map and detection report are generated and output.

[0046] This embodiment also provides a roll surface defect detection system, including: The data acquisition and feature extraction module is used to acquire roll parameters and machine vision data, establish the coordinates of the roll cylindrical surface and a unified cylindrical mesh, and map the machine vision data to the unified cylindrical mesh, and further generate six-channel roll scalar features and imaging quality masks. The roll parameter analysis module is used to calculate roll stress based on roll parameters and generate mechanical fatigue damage index, roll surface temperature gradient and machine vision surface perturbation features, and further obtain fatigue vulnerability probability and preferred propagation direction of roll surface cracks. The roll defect construction module is used to classify surface defect types, determine the location of defect initiation points based on fatigue vulnerability probability and generate defect geometric masks, and calculate the maximum relative depth for each type of defect to form a relative depth map; The virtual data generation module is used to establish a normal roll surface background library based on the six-channel roll scalar features and collect real defect patches. Training data is formed based on the real defect patches. The generator and discriminator are trained using the training data to output a defect appearance generation network and a virtual defect training set is formed based on the defect appearance generation network. The detection network construction and output module is used to construct a lightweight roll defect segmentation network and define detection labels, establish defect training data to train the lightweight roll defect segmentation network, and identify and output roll surface defects based on the trained defect segmentation network.

[0047] This embodiment also provides a computer device applicable to the roll surface defect detection method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the roll surface defect detection method proposed in the above embodiment.

[0048] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0049] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the roll surface defect detection method proposed in the above embodiments. The storage medium 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 Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0050] In summary, this invention achieves co-location characterization of multi-source data on the roll surface through encoder-synchronized polarization imaging, infrared thermography, and cylindrical mesh mapping, reducing the interference of reflection and unfolded seams on defect features. By fusing contact load, thermal gradient, cyclic damage, and visual perturbation to generate fatigue vulnerability probability, it correlates defect appearance with service conditions. By constraining geometric masks and relative depths based on risk locations and crack directions, it forms defect samples with accurate labels and reasonable physical morphology. By using mask-constrained generation to maintain a normal roll background, and by training a lightweight segmentation network using real, virtual, and normal samples, it improves the identification capabilities of microcracks, pitting, and spalling, balancing detection accuracy, sample expansion effect, and field deployment efficiency.

[0051] It should be noted that 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting surface defects in rolling mill rolls, characterized in that: include, Collect roll parameters and machine vision data, establish the coordinates of the roll cylindrical surface and a unified cylindrical mesh, and map the machine vision data to the unified cylindrical mesh to further generate six-channel roll scalar features and imaging quality masks; Based on the roll parameters, the roll stress is calculated and mechanical fatigue damage index, roll surface temperature gradient and machine vision surface perturbation features are generated. The fatigue vulnerability probability and the preferred propagation direction of roll surface cracks are then obtained. Classify surface defects, determine the location of defect initiation points based on fatigue vulnerability probability and generate defect geometric masks, and calculate the maximum relative depth for each type of defect to form a relative depth map; A normal roll surface background library is established based on the six-channel roll scalar features, and real defect patches are collected. Training data is formed based on the real defect patches. The generator and discriminator are trained using the training data to output a defect appearance generation network. A virtual defect training set is formed based on the defect appearance generation network. A lightweight roll defect segmentation network is constructed and detection labels are defined. Defect training data is established to train the lightweight roll defect segmentation network. Based on the trained defect segmentation network, roll surface defects are identified and output.

2. The method for detecting surface defects of a roll as described in claim 1, characterized in that: The process of acquiring roll parameters and machine vision data, establishing roll cylindrical surface coordinates and a unified cylindrical mesh, mapping machine vision data to the unified cylindrical mesh, and further generating six-channel roll scalar features and imaging quality masks involves: acquiring the geometric parameters, material parameters, rolling process parameters, and machine vision data of the target roll, and establishing data associations according to the same roll number and the same acquisition time; The geometric parameters include the reference radius and the effective length of the roll body; The material parameters correspond to the rolls and the strip being rolled. The roll material parameters include the elastic modulus, Poisson's ratio, yield strength, coefficient of linear expansion, fatigue strength at the reference number of cycles, and fatigue index. The strip material parameters include the elastic modulus and Poisson's ratio. The rolling process parameters include the current total rolling force, current reduction, inlet tension, outlet tension, roll surface linear velocity, strip linear velocity, and roll rotation speed; The machine vision data uses a four-channel polarization linear array camera that is synchronously triggered by a rotary encoder to acquire original polarization images with polarization directions of 0°, 45°, 90°, and 135° respectively. At the same time, an infrared linear array camera that is synchronously triggered with the rotary encoder is used to acquire the original image of the roll surface temperature. The collected machine vision data is corrected; Define the circumferential and axial coordinates of the roll surface to establish the cylindrical surface coordinates of the roll. Set a unique zero-phase mark at the end of the roll and define the circumferential position when the encoder zero-phase pulse arrives as the 0 position of the circumferential coordinate. Record the number of effective trigger pulses actually generated by the encoder per revolution and calculate the roll rotation angle corresponding to each effective trigger pulse. Calculate the corresponding circumferential coordinates through the roll rotation angle. Calculate the axial pixel physical resolution by setting a calibration plate. Record the pixel number corresponding to the left end of the roll. Multiply the difference between the axial pixel and the pixel number corresponding to the left end of the roll by the axial pixel physical resolution to obtain the axial coordinate. Define the target minimum stable crack width and the minimum crack width coverage pixels. Obtain a uniform cylindrical grid spacing based on the ratio of the target minimum stable crack width and the minimum crack width coverage pixels. Calculate the number of circumferential grids using the roll circumference and the uniform cylindrical grid spacing, and obtain the coordinates of each circumferential grid. Simultaneously, calculate the number of axial grids using the effective roll length and the uniform cylindrical grid spacing, and obtain the axial grid coordinates. Combine the circumferential grid coordinates and the axial grid coordinates to form a cylindrical grid point. Obtain the continuous coordinates of each cylindrical grid point in the machine vision data image. Based on the correspondence between cylindrical grid points and continuous coordinates, bicubic interpolation is used to map the corrected machine vision data image onto a unified cylindrical grid; The total intensity component and the linear polarization difference component are calculated based on the polarization images of the four polarization directions, and the low illumination confidence and linear polarization degree are constructed based on the total intensity component and the linear polarization difference component. The linear polarization direction angle is calculated based on the linear polarization differential component, and the linear polarization direction is continuously encoded. Take the logarithm of the total intensity and then subtract the low-frequency background to extract local reflection texture features; Two roll defect size detection scales are defined. The corresponding pixel scales are calculated according to the roll defect size detection scales. The circumferential gradient and axial gradient after Gaussian kernel smoothing are calculated for the two pixel scales respectively. The edge intensity under the corresponding scale is obtained. The larger edge intensity value between the two pixel scales is selected as the multi-scale edge intensity. At each pixel scale, a structure tensor is constructed and two feature values ​​of the structure tensor are extracted. The texture anisotropy at the corresponding pixel scale is obtained by the ratio of the difference between the two feature values ​​to the sum. The larger of the texture anisotropy at the two pixel scales is selected as the multi-scale anisotropy. The local reflection texture features, linear polarization degree, multi-scale edge intensity, and multi-scale anisotropy are used to form the roll scalar features to be standardized. The 5% and 95% quantiles of each item in the roll scalar features are extracted respectively. Each item in the roll scalar features to be standardized is standardized according to the quantiles. At the same time, the linear polarization direction code is mapped to 0-1. The mapped linear polarization direction code and the standardized roll scalar features are connected and combined to form a new six-channel roll scalar feature. An imaging quality mask is generated by combining low-light markers, channel saturation markers, and encoder trigger integrity markers.

3. The method for detecting surface defects of rolls as described in claim 2, characterized in that: The process of calculating roll stress based on roll parameters and generating mechanical fatigue damage index, roll surface temperature gradient and machine vision surface perturbation features, and further obtaining fatigue vulnerability probability and preferred roll surface crack propagation direction is as follows: the axial coordinate is converted into a dimensionless position through the effective length of the roll body, and the axial linear load distribution is constructed using a zero-mean quadratic basis function. The equivalent elastic modulus between the roll and the strip is calculated based on the material parameters, and the reference radius of the roll is used as the equivalent radius of curvature. The Hertz contact half-width and maximum contact pressure at the axial position are calculated by combining the equivalent elastic modulus and the equivalent radius of curvature. The maximum bending stress on the roll surface at the axial position is obtained by calculating the bending moment at the axial position and the moment of inertia of the roll's circular cross-section. The ratio of the bending moment multiplied by the effective radius of the roll to the moment of inertia is then used to obtain the maximum bending stress on the roll surface at the axial position. The median temperature in the circumferential direction is extracted for each axial position based on the machine vision data image mapped to a unified cylindrical mesh, and the local temperature difference between the current point and the median temperature is calculated. The constraint thermal stress caused by the local temperature difference is then calculated based on the local temperature difference. The difference between the exit tension and the inlet tension is calculated, and the ratio of the difference to the total rolling force is taken to obtain the equivalent friction coefficient. The slip direction coefficient is calculated based on the strip linear velocity and the roll surface linear velocity. The product of the maximum contact pressure, the equivalent friction coefficient, and the slip direction coefficient is taken as the roll surface shear stress. The surface normal stress in the circumferential and axial directions is calculated based on the maximum contact pressure, constraint thermal stress and maximum bending stress on the roll surface. The equivalent stress on the roll surface is calculated under two-dimensional plane stress state. Then the cumulative number of rotations of the target roll since the most recent grinding is read, and the mechanical fatigue damage index is calculated. The temperature gradients in the circumferential and axial directions of the roll surface are calculated separately and synthesized into the roll surface temperature gradient. The local reflective texture features are filtered by Gaussian kernel to extract the local normal texture background. The absolute value of the difference between the normalized local reflective texture and the local normal texture background is calculated and then the machine vision surface perturbation features are calculated. The machine vision surface perturbation features are corrected by using an imaging quality mask. The mechanical fatigue damage index, the roll surface temperature gradient, and the corrected machine vision surface perturbation features are weighted after quantile-stable normalization to obtain the comprehensive fatigue driving amount of the roll. The sigmoid function is used to map the comprehensive fatigue driving amount to a fatigue vulnerability probability in the range of 0-1. Based on the principal stress direction relationship under two-dimensional stress state, the maximum principal stress direction angle is calculated, and the preferred propagation direction of the crack on the roll surface is calculated based on the maximum principal stress direction angle.

4. The method for detecting surface defects of rolls as described in claim 3, characterized in that: The process of classifying surface defects, determining the location of defect initiation points based on fatigue vulnerability probability, generating defect geometric masks, and calculating the maximum relative depth for each type of defect to form a relative depth map is as follows: the defects to be generated on the roll surface are divided into three categories: microcracks, pitting, and spalling, and each category is assigned a label. The actual occurrence of the three types of defects is counted to determine the generation ratio of the three types of defects in the virtual sample. The fatigue vulnerability probability is divided into fixed-size blocks according to a uniform cylindrical grid, and the average fatigue vulnerability degree is calculated for each block; Within the image block, the probability of defect initiation points is obtained by weighted sampling based on the fatigue vulnerability probability of each pixel. A minimum physical spacing is set for different initiation points, and the defect initiation points are selected based on their probabilities. Generate geometric masks based on the three types of defects; For each type of defect, define the maximum relative depth range; For microcrack defects, read the microcrack geometry mask and extract the microcrack path and crack width. Calculate the shortest actual distance from the pixel in the cylindrical mesh to the crack center path, and select the minimum value. At the same time, record the position of the crack path closest to the pixel and the corresponding crack width. Combine the minimum value and the corresponding crack width to calculate the relative depth of the microcrack defect. For pitting defects, the pitting geometry mask is read and the pitting center, major axis, minor axis, shape index, reference direction, and low-frequency perturbation are extracted. For pixels in the pitting area, their cylindrical coordinates are converted to local rotation coordinates with the pitting center as the origin according to the reference direction. Based on the relative position of the pixel along the major and minor axes of the pitting, its normalized radial position is calculated, and the relative depth of the pitting is calculated based on the normalized radial position. For peeling defects, read the peeling geometry mask, calculate the shortest Euclidean distance from each pixel to the outer boundary of the peeling inside the peeling mask, select the largest boundary distance from the current peeling defect, normalize the boundary distance of each pixel by the largest boundary distance, and generate the peeling relative depth. The relative depths of the three types of defects are combined to form a relative depth map.

5. The method for detecting surface defects of a roll as described in claim 4, characterized in that: The process of establishing a normal roll surface background library based on the six-channel roll scalar features and collecting real defect patches, forming training data based on the real defect patches, training the generator and discriminator with the training data to output a defect appearance generation network, and forming a virtual defect training set based on the defect appearance generation network is as follows: establishing a normal roll surface background library based on the six-channel standardized roll surface features. Simultaneously collect real defect maps and label the real defect areas and defect categories; The real defect patch is used as the target image that the generator network needs to learn. Based on the real defect mask, the defect area in the patch is deleted. Then, the existing image inpainting algorithm is used to fill in the deleted area based on the normal grinding marks, local brightness and polarization features around the defect, so as to obtain a defect-free background patch corresponding to the original real defect patch. For each real defect sample, the distance from each pixel inside the defect region to the mask boundary is calculated according to its geometric mask. The positions with larger distances are marked as relatively deep, and the positions with closer distances to the boundary are marked as relatively shallow, thus obtaining a relative depth guide map that matches the real defect mask. Read the fatigue vulnerability probability corresponding to the actual defect location on the roll and generate training data; The generator network uses the U-Net conditional generator, which combines the training data according to the channel direction to form the generator input. The encoding part of the U-Net generator extracts the comprehensive features of normal background and defect conditions step by step, and the decoding part restores the image resolution step by step. It also preserves the original normal roll texture through the skip connection between the encoding layer and the decoding layer, and finally outputs a six-channel candidate defect image with the same size as the normal background. The discriminator uses the PatchGAN local discriminator. The discriminator receives the defect image and its corresponding defect mask and category information, divides the image into multiple local regions, and determines whether each local region is a real defect appearance or a generated defect appearance. The discriminator trains the generator and outputs the defect appearance generation network. The corresponding defect appearance is generated based on the defect appearance generation network, and candidate six-channel defect images are output to form a virtual defect training set.

6. The method for detecting surface defects of a roll as described in claim 5, characterized in that: The construction of a lightweight roll defect segmentation network and the definition of detection labels, and the establishment of defect training data to train the lightweight roll defect segmentation network are specifically as follows: virtual defect samples, real defect samples and real normal samples are combined to form defect training data and labeled. For each sample, the six-channel standardized roll surface features and fatigue vulnerability probability are connected according to the channel direction to form a seven-channel network input. A lightweight roll defect segmentation network was constructed using a MobileNetV3-Small encoder and a U-Net decoder, and the detection labels were defined as four categories, including normal, microcracks, pitting, and spalling. The defect segmentation network is trained using a combination of multi-class cross-entropy loss and Dice loss. Different training weights are set for different defect categories. The AdamW optimizer is used to update the network parameters, and the existing learning rate decay method is used to gradually reduce the learning rate. If the loss does not decrease for several consecutive training rounds, the training is stopped, the defect segmentation network parameters are output, and the final lightweight roll defect segmentation network is formed.

7. The method for detecting surface defects of a roll as described in claim 6, characterized in that: The specific steps for identifying and outputting roll surface defects based on the trained defect segmentation network are as follows: extracting six-channel roll surface features and fatigue vulnerability probabilities from the collected machine vision data image of the target roll, and then outputting the final lightweight roll defect segmentation network to obtain the roll surface defect detection results. Based on the roll surface defect detection results, a roll surface defect detection map and a detection report are generated and output.

8. A roll surface defect detection system, based on the roll surface defect detection method according to any one of claims 1 to 7, characterized in that: include, The data acquisition and feature extraction module is used to acquire roll parameters and machine vision data, establish the coordinates of the roll cylindrical surface and a unified cylindrical mesh, and map the machine vision data to the unified cylindrical mesh, and further generate six-channel roll scalar features and imaging quality masks. The roll parameter analysis module is used to calculate roll stress based on roll parameters and generate mechanical fatigue damage index, roll surface temperature gradient and machine vision surface perturbation features, and further obtain fatigue vulnerability probability and preferred propagation direction of roll surface cracks. The roll defect construction module is used to classify surface defect types, determine the location of defect initiation points based on fatigue vulnerability probability and generate defect geometric masks, and calculate the maximum relative depth for each type of defect to form a relative depth map; The virtual data generation module is used to establish a normal roll surface background library based on the six-channel roll scalar features and collect real defect patches. Training data is formed based on the real defect patches. The generator and discriminator are trained using the training data to output a defect appearance generation network and a virtual defect training set is formed based on the defect appearance generation network. The detection network construction and output module is used to construct a lightweight roll defect segmentation network and define detection labels, establish defect training data to train the lightweight roll defect segmentation network, and identify and output roll surface defects based on the trained defect segmentation network.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the roll surface defect detection method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the roll surface defect detection method according to any one of claims 1 to 7.

Images