A shearing behavior dynamic correction method and system based on wear state identification

Abstract

The present application relates to the technical field of image recognition, in particular to a shearing behavior dynamic correction method and system based on wear state recognition; a digital image sequence of a blade edge region is acquired; an image feature separation network is constructed to perform parallel feature extraction on the preprocessed digital image sequence; the continuity feature of the periodic texture pattern of the blade edge surface is utilized to identify the pixel-level high-frequency texture discontinuous region and the edge gradient direction field to obtain a target defect probability graph; the projection shadow region of the convex edge and the low-frequency light halo are identified by using the inverse light imaging model to obtain an interference artifact probability graph; a spatial mutual exclusion constraint of the target defect probability graph and the interference artifact probability graph is established in the pixel space to generate a defect binary mask; multi-dimensional texture feature calculation is performed on the corresponding region of the defect binary mask to construct a surface state feature vector; the present application discriminates the defect mode category based on the surface state feature vector to generate corresponding shearing correction parameters.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, specifically to a method and system for dynamic correction of shearing behavior based on wear state recognition. Background Technology

[0002] Computer vision-based industrial defect detection technology has been widely applied in the operation and maintenance monitoring of metal shearing equipment. As a core actuator, the wear condition of the shearing blade, such as microscopic chipping or excessive wear, directly determines the processing quality of the sheet metal. Traditional vision inspection systems typically use high-resolution industrial cameras to acquire images of the cutting edge, and then use image enhancement, edge detection, or convolutional neural network algorithms to identify changes in the geometric shape of the cutting edge, thereby providing feedback to control process parameters such as shearing gap or speed.

[0003] However, in actual industrial production sites, the working environment of shearing equipment is extremely harsh, and the surface of the cutting tools inevitably becomes covered with foreign matter such as cutting fluid film, metal shavings, and dust agglomerates. In existing two-dimensional grayscale image processing technologies, these exogenous adhesions and endogenous wear defects have a very high degree of similarity in visual characteristics, often manifesting as dark spots with sudden drops in local grayscale values ​​or irregular texture abrupt changes.

[0004] The core problem with existing technologies is that, under complex industrial lighting conditions, conventional image processing algorithms cannot effectively distinguish between material missing defects and material accumulation artifacts with similar texture features based on two-dimensional digital images. This results in insufficient specificity in wear condition recognition and makes it impossible to generate accurate dynamic correction strategies for shearing behavior.

[0005] To address this, a dynamic correction method and system for shear behavior based on wear state identification is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a dynamic correction method and system for shearing behavior based on wear state recognition. By constructing an image feature separation network to identify the probability map of target defects and the probability map of interference artifacts, a binary mask of defects is generated, and a surface state feature vector is constructed; finally, the corresponding shearing correction parameters are generated.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A dynamic correction method for shear behavior based on wear state identification includes:

[0009] Obtain the digital image sequence of the cutting edge region and perform image enhancement and illumination equalization preprocessing;

[0010] An image feature separation network is constructed to extract features in parallel from the preprocessed digital image sequence. Among them, the high-frequency texture analysis branch uses the continuous features of the periodic texture pattern on the cutting edge surface to identify the pixel-level high-frequency texture discontinuity region and edge gradient direction field, and segment out the target defect probability map representing the material missing. The photometric shadow analysis branch uses the backlighting imaging model to identify the projected shadow region of the protrusion edge and the low-frequency light halo, and segment out the interference artifact probability map representing the material accumulation.

[0011] Construct a pixel-level probabilistic inference model and perform mask fusion and thinning operations; establish spatial mutual exclusion constraints between the target defect probability map and the interference artifact probability map in the pixel space to generate a defect binary mask;

[0012] Multidimensional texture feature calculation is performed on the region corresponding to the defect binary mask to construct a surface state feature vector; based on the distribution of the surface state feature vector in the multidimensional feature space, the defect pattern category is determined and the corresponding shearing correction parameters are generated.

[0013] The identification process of the high-frequency texture analysis branch includes:

[0014] Multi-scale frequency domain filter bank convolution operation is performed on digital image sequences to extract the texture orientation field of the cutting edge surface. The region where the texture streamline is interrupted by a pulse function and the high-frequency energy decreases abruptly is marked as a potential structural damage area by utilizing the continuity prior features of the periodic texture pattern.

[0015] Within the potential structural damage area, a gray-level gradient vector field is calculated, and a radial reference vector field with the region's geometric centroid as the origin is constructed. A pixel-by-pixel dot product operation is performed on the gray-level gradient vector field and the radial reference vector field. If and only if the dot product result of the closed region edge is positive, indicating that the image gradient direction is away from the region's geometric center, and the gray-level distribution inside the region exhibits local minimum characteristics showing sharp step edges, it is determined that there is a negative abrupt change in image depth, and a target defect probability map containing clear boundary information is output.

[0016] The identification process of the photometric shading analysis branch includes:

[0017] For non-rigid interference artifacts, calculate the dark channel prior map of local image blocks, identify regions in the dark channel with gray values ​​higher than a preset benchmark and edges accompanied by Gaussian blur features, and mark them as fluid artifact shadow masks.

[0018] To address structured interference artifacts, a set of pixels with brightness values ​​exceeding the saturation limit in the image is extracted through threshold segmentation and identified as specular highlight points. A backlighting imaging model is constructed based on the known incident direction vector of the light source, and the specular highlight points are used as positioning anchor points to search for the projected shadow region on the opposite side. The gradient direction in the neighborhood of the highlight point is calculated. If the gradient direction points from the surrounding area to the geometric center of the region, it is determined to be a non-defective positive accumulation. Using morphological geodesic dilatation, the identified specular highlight points and projected shadow regions are used as common seed points to perform restricted growth on the gradient map to completely cover the rigid foreign object region, generating a rigid artifact mask.

[0019] The fluid artifact mask and the rigid artifact mask are fused at the pixel level to output an interference artifact probability map containing high-confidence artifact distribution information.

[0020] The pixel-level probabilistic inference model is implemented through a fully connected conditional random field model, which includes a global energy index composed of a weighted sum of data observation terms and pairwise correlation terms.

[0021] The data observations are determined by the target defect probability map and the interference artifact probability map, and represent the initial confidence cost of classification based solely on the current pixel features;

[0022] The pairwise association terms introduce a consistency constraint factor for edge gradient direction. The calculation logic of the consistency constraint factor is as follows: for any adjacent pixel pair, if the color difference between them is lower than a preset threshold and the gradient directions are opposite, it indicates that a topological contradiction of concavity and convexity reversal has occurred. Then, a label penalty is applied between the pixel pairs to prohibit them from being simultaneously marked as defect patterns of the same category.

[0023] The mean field approximation algorithm is used to iteratively optimize the global energy index, so that the probability map of interference artifacts with inward gradient and the probability map of target defects with outward gradient form a mutually exclusive competitive relationship in space. Pixels that are visually within the coverage of artifact projection are marked as untrusted regions and logically filtered out to generate a defect binary mask.

[0024] The surface state feature vector is a hybrid feature vector containing geometric dimension, statistical texture dimension and frequency domain energy dimension;

[0025] Specifically, this includes: the maximum depth estimate of the target defect region calculated based on monocular depth estimation, the aspect ratio of the minimum bounding rectangle of the target defect region, the weighted sum of texture energy and inverse gap moment calculated based on the gray-level co-occurrence matrix, the fractal dimension of the defect edge calculated using the box-counting dimension method, and the energy proportion of the dominant frequency component in the Fourier spectrum of the target defect region image; after normalization processing, the surface state feature vector is obtained.

[0026] The process of determining defect pattern categories based on surface state feature vectors includes:

[0027] A decision boundary model based on multidimensional feature space partitioning is constructed, and the surface state feature vector is projected into the multidimensional feature space. The defect pattern category is determined according to the decision region into which the vector falls.

[0028] If the feature exhibits significant directional anisotropy in geometric aspect ratio and the texture spectrum energy is concentrated in the low-frequency band of the specified direction, indicating that the texture extends in accordance with the processing direction, it is determined to be a linear loss mode.

[0029] If the feature is that the texture entropy exhibits a nonlinear discrete distribution, and is accompanied by the drift of the texture frequency domain feature in the specified frequency band, indicating that there is micro-texture migration on the surface due to thermal effects, then it is determined to be a thermal adhesion mode.

[0030] If the features exhibit a step-like abrupt change in depth features and the edge fractal dimension characterizes the contour with complex geometric topological irregularities, then it is determined to be a brittle fracture mode.

[0031] Output the normalized probability distribution vector of the pattern category to determine the defect pattern category.

[0032] The process of generating shear correction parameter data includes:

[0033] The surface state feature vector sequence within the historical time window is called to construct a time-series trend prediction model, and control strategy data is generated by combining the defect mode category of the current frame.

[0034] When the loss pattern is determined to be linear and the depth component of the feature vector shows a linear growth trend in time, the growth slope is calculated and the corresponding gap step compensation data is generated.

[0035] When the thermal adhesion mode is identified and the temporal change rate of the texture entropy feature increases exponentially, it indicates accelerated thermal accumulation and generates thermal balance correction data that reduces the speed and increases the cooling amount.

[0036] When the fracture mode is determined to be brittle and the depth feature undergoes a pulse-like abrupt change within a single frame, it indicates that the device has been subjected to an impact load and generates shutdown alarm data.

[0037] The calculated data is organized into shear correction parameters and transmitted to the actuator to adjust the process parameters.

[0038] A dynamic correction system for shear behavior based on wear state identification includes:

[0039] The data acquisition module acquires digital image sequences of the cutting edge region and performs image enhancement and illumination equalization preprocessing.

[0040] The feature recognition module constructs an image feature separation network to perform parallel feature extraction on the preprocessed digital image sequence. Among them, the high-frequency texture analysis branch uses the continuous features of the periodic texture pattern on the cutting edge surface to identify pixel-level high-frequency texture discontinuities and edge gradient direction fields, and segments out the probability map of target defects representing material loss. The photometric shadow analysis branch uses the backlighting imaging model to identify the projected shadow area of ​​the protrusion edge and the low-frequency light halo, and segments out the probability map of interference artifacts representing material accumulation.

[0041] The mask generation module constructs a pixel-level probabilistic inference model and performs mask fusion and thinning operations; it establishes spatial mutual exclusion constraints between the target defect probability map and the interference artifact probability map in the pixel space and generates a binary defect mask.

[0042] The correction analysis module performs multidimensional texture feature calculation on the region corresponding to the defect binary mask to construct a surface state feature vector; based on the distribution of the surface state feature vector in the multidimensional feature space, it determines the defect pattern category and generates the corresponding shearing correction parameters.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] 1. This invention effectively solves the problem of traditional edge detection algorithms being sensitive to shallow scratches and surface stains by combining frequency domain texture analysis and spatial domain topology verification. The Gabor filter has a strong selectivity for periodic knife marks and can keenly capture the tiny breaks in the knife marks. Based on the gradient direction constraint with a positive dot product, the concave features are strictly limited from a physical topology perspective, eliminating the interference of color spots on flat surfaces. This design improves the detection rate of real wear under complex lighting conditions and significantly reduces the false alarm rate caused by surface watermarks or rust, providing a reliable data foundation for subsequent accurate correction.

[0045] 2. This invention innovatively introduces dark channel prior and backlight imaging model into industrial defect detection, and constructs a specific identification mechanism for foreign objects with different physical forms; for fluid oil stains, it uses their optical absorption characteristics for segmentation; for rigid debris, it uses the optical geometric features of "highlight-shadow" pairing for locking; in particular, it introduces inward gradient as the criterion for protrusions, which distinguishes protruding foreign objects from dented wear from a physical essence.

[0046] 3. This invention achieves deep logical fusion of multi-source information through a fully connected conditional random field model. In particular, it introduces consistency constraints on edge gradient directions, cleverly transforming the topologically exclusive relationship between concavities and convexities in physical space into energy constraints in the algorithm model. This not only solves the problem of segmenting blurred edge regions, but more importantly, it implements a logical filtering based on physical properties. It utilizes significant artifacts to actively suppress subtle pseudo-wearing, ensuring that the generated binary mask has extremely high physical realism. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating a dynamic correction method for shear behavior based on wear state identification according to the present invention.

[0048] Figure 2 This is a logical schematic diagram of a dynamic correction method for shear behavior based on wear state identification according to the present invention.

[0049] Figure 3 This is a schematic diagram of the structure of a dynamic correction system for shear behavior based on wear state recognition according to the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Example 1:

[0052] This invention proposes a dynamic correction method for shear behavior based on wear state identification. The process of the method is as follows: Figure 1 As shown, the data logic is as follows: Figure 2 As shown, it specifically includes:

[0053] Obtain the digital image sequence of the cutting edge region and perform image enhancement and illumination equalization preprocessing;

[0054] An image feature separation network is constructed to extract features in parallel from the preprocessed digital image sequence. Among them, the high-frequency texture analysis branch uses the continuous features of the periodic texture pattern on the cutting edge surface to identify the pixel-level high-frequency texture discontinuity region and edge gradient direction field, and segment out the target defect probability map representing the material missing. The photometric shadow analysis branch uses the backlighting imaging model to identify the projected shadow region of the protrusion edge and the low-frequency light halo, and segment out the interference artifact probability map representing the material accumulation.

[0055] Construct a pixel-level probabilistic inference model and perform mask fusion and thinning operations; establish spatial mutual exclusion constraints between the target defect probability map and the interference artifact probability map in the pixel space to generate a defect binary mask;

[0056] Multidimensional texture feature calculation is performed on the region corresponding to the defect binary mask to construct a surface state feature vector; based on the distribution of the surface state feature vector in the multidimensional feature space, the defect pattern category is determined and the corresponding shearing correction parameters are generated.

[0057] The identification process of the high-frequency texture analysis branch includes:

[0058] Multi-scale frequency domain filter bank convolution operation is performed on digital image sequences to extract the texture orientation field of the cutting edge surface. The region where the texture streamline is interrupted by a pulse function and the high-frequency energy decreases abruptly is marked as a potential structural damage area by utilizing the continuity prior features of the periodic texture pattern.

[0059] Within the potential structural damage area, a gray-level gradient vector field is calculated, and a radial reference vector field with the region's geometric centroid as the origin is constructed. A pixel-by-pixel dot product operation is performed on the gray-level gradient vector field and the radial reference vector field. If and only if the dot product result of the closed region edge is positive, indicating that the image gradient direction is away from the region's geometric center, and the gray-level distribution inside the region exhibits local minimum characteristics showing sharp step edges, it is determined that there is a negative abrupt change in image depth, and a target defect probability map containing clear boundary information is output.

[0060] Data processing flow and algorithm details of the high-frequency texture analysis branch:

[0061] Multi-scale frequency domain filtering: A set of Gabor filters with different directions and scales are used to convolve the preprocessed image; the response value of each pixel is output to construct the texture orientation field; normal knife marks show a continuous strong response, while worn areas show a sudden decay of the response.

[0062] Potential region marking: Detect texture streamlines. If the energy response at a certain point shows a pulse function interruption, that is, a sudden change in the derivative along the tool mark direction, and the high-frequency energy drops abruptly, that is, the high-frequency component is lost, then it is marked as a potential structural damage area.

[0063] Gradient and topology verification: Calculate the gray-level gradient vector field within the potential region; construct a radial reference vector field, with the direction defined as pointing from the geometric centroid of the region to the edge pixels (from inside to outside); perform dot product operation.

[0064] Judgment logic: If the dot product operation is positive, it indicates that the gradient direction is consistent with the radial direction, i.e., moving outward from the center. Combining this with optical principles, the center of the pit is dark while the edges are bright, confirming that the gradient indeed flows from the inside out. Simultaneously, the second derivative within the region is calculated; if the eigenvalue is large, it indicates the presence of sharp edges. If all the above conditions are met, the target defect probability map is output.

[0065] A multi-scale frequency domain filter bank convolution operation is performed on the digital image sequence. Specifically, a Gabor filter bank with 8 directions (0°, 22.5°, 45°, 67.5°, 90°, 112.5°, 135°, 157.5°) and 3 spatial scales (σ=2, 4, 8 pixels) is constructed, resulting in a total of 24 convolution kernels. The preprocessed image is convolved with each Gabor convolution kernel to obtain 24 response feature maps. For each pixel, the direction corresponding to the maximum response value among the 24 response values ​​is selected as the dominant texture direction for that pixel, and this maximum response value is recorded as the texture intensity. The resulting texture direction and texture intensity constitute a texture orientation field.

[0066] Next, utilizing the continuity prior features of the periodic texture pattern, the texture direction and texture intensity of adjacent pixels are traversed. Normal knife marks are characterized by continuous extension of the texture direction and stable maintenance of the texture intensity; when the texture direction suddenly changes or the texture intensity shows a pulse-like decrease, defined as the attenuation of the texture intensity of the current pixel relative to the median of the neighborhood exceeding 50% and lasting for more than 3 consecutive pixels, this area is marked as a potential structural damage area.

[0067] Further refined verification is performed within the potential structural damage area. First, the potential damage area is analyzed using the 8-connectivity principle, with each independent connected region considered a candidate defect region. For each candidate region, its geometric centroid as a binary region is calculated. Then, the gray-level gradient vector field is calculated at all pixels within the region, while simultaneously constructing a radial reference vector field. A pixel-by-pixel dot product operation is performed on the pixels at the region's edges. A region is considered a true defect region only if more than 70% of the dot product results in the closed edge pixels of the candidate region are greater than 0, indicating that the gradient direction generally deviates from the centroid. Simultaneously, the gray-level values ​​within the region exhibit local minima (i.e., the average gray-level value within the region is significantly lower than the average gray-level value of its neighbors, with a difference of at least 20%). This result is then output to the target defect probability map.

[0068] This invention effectively solves the problem of traditional edge detection algorithms being sensitive to shallow scratches and surface stains by combining frequency domain texture analysis and spatial domain topology verification. The Gabor filter has a strong selectivity for periodic knife marks and can keenly capture the tiny breaks in the knife marks. Furthermore, based on the gradient direction constraint with a positive dot product, the concave features are strictly limited from a physical topology perspective, eliminating interference from color spots on flat surfaces. This design improves the detection rate of real wear by more than 15% under complex lighting conditions and significantly reduces the false alarm rate caused by surface watermarks or rust, providing a reliable data foundation for subsequent accurate correction.

[0069] The identification process of the photometric shading analysis branch includes:

[0070] For non-rigid interference artifacts, calculate the dark channel prior map of local image blocks, identify regions in the dark channel with gray values ​​higher than a preset benchmark and edges accompanied by Gaussian blur features, and mark them as fluid artifact shadow masks.

[0071] To address structured interference artifacts, a set of pixels with brightness values ​​exceeding the saturation limit in the image is extracted through threshold segmentation and identified as specular highlight points. A backlighting imaging model is constructed based on the known incident direction vector of the light source, and the specular highlight points are used as positioning anchor points to search for the projected shadow region on the opposite side. The gradient direction in the neighborhood of the highlight point is calculated. If the gradient direction points from the surrounding area to the geometric center of the region, it is determined to be a non-defective positive accumulation. Using morphological geodesic dilatation, the identified specular highlight points and projected shadow regions are used as common seed points to perform restricted growth on the gradient map to completely cover the rigid foreign object region, generating a rigid artifact mask.

[0072] The fluid artifact mask and the rigid artifact mask are fused at the pixel level to output an interference artifact probability map containing high-confidence artifact distribution information.

[0073] Data processing flow and algorithm details of the photometric shading analysis branch:

[0074] Non-rigid fluid recognition: Oil stains have strong light absorption and the edges have a smooth transition due to surface tension; calculate the dark channel prior of the image; if the gray value of a certain area is significantly higher than that of the surrounding area, and the edge gradient histogram conforms to a Gaussian distribution, it is marked as a fluid artifact mask.

[0075] Structured rigidity recognition: Pixels with saturated brightness (close to 255) are extracted using Otsu threshold segmentation. These pixels can be judged by setting a threshold and used as specular reflection highlights.

[0076] Backlighting model: Set the light source vector and search for dark areas (shadows) in the opposite position based on the position of the highlight point.

[0077] Gradient verification: Calculate the gradient direction within the neighborhood of the highlight point; since the center of the protrusion is the brightest, the gradient points from the surrounding dark areas to the center bright area, that is, from the surrounding areas to the geometric center of the region, converging inwards. If this condition is met, it is determined to be a protrusion.

[0078] Region growth: Using highlights and shadows as seeds, morphological geodesic expansion is used to fill the entire foreign object region under the guidance of gradient map, generating a rigid artifact mask.

[0079] Perform pixel-level union operation on the fluid mask and the rigid mask, and output the interference artifact probability map after normalization.

[0080] This invention innovatively introduces dark channel priors and backlighting imaging models into industrial defect detection, constructing a specific identification mechanism for foreign objects with different physical forms. For fluid oil stains, its optical absorption characteristics are used for segmentation; for rigid debris, the optical geometric features of paired "highlight-shadow" are used for locking; and in particular, an inward gradient is introduced as a criterion for protrusions, physically distinguishing protruding foreign objects from dented wear. This design can accurately identify more than 98% of surface attachments, greatly reducing the detection background and preventing equipment shutdowns due to foreign object interference.

[0081] Furthermore, the construction process of the backlighting imaging model based on the shearing scene includes: establishing a unified geometric space that includes the camera coordinate system, the light source coordinate system, and the blade plane coordinate system;

[0082] During the initialization phase, the spatial position vector between the camera optical center and the fixed light source, as well as the angle between the camera optical axis and the cutting edge plane, are determined using a calibration plate.

[0083] A virtual hemispherical illumination projection model is constructed, and the specular reflection highlight is regarded as the vertex of the protrusion. Based on the principle of geometric optics, the protrusion is set as a quasi-protruding surface, and the theoretical projection area of ​​the surface on the cutting edge plane is calculated using vector calculation.

[0084] Calculate the overlap between the actual projected shadow region and the theoretical projection region in the image; if the overlap is higher than a preset threshold and the gradient direction conforms to the inward convergence feature, then the region is confirmed to be the shadow generated by a rigid foreign object.

[0085] Specifically, during the installation and commissioning phase of the shearing machine, a standard checkerboard calibration board is used to determine the camera's intrinsic and extrinsic parameters. Given the fixed physical position of the internal lighting source (strip LED) relative to the blade holder, the incident light vector is calculated. During operation, when the photometric shadow branch detects a highlight, a tiny spherical protrusion is assumed to exist at that location. Based on the incident light vector and the assumed height of the protrusion, the theoretical shadow should appear at the image coordinates; that is, based on the vertex coordinates of the protrusion, the light source vector, and the camera projection matrix, the area where the shadow falls on the cutting edge plane is calculated using geometric projection relationships.

[0086] Specifically, the system presets an empirical protrusion height, such as the maximum allowable burr height of the tool, like 0.2mm, or estimates the height based on the pixel diameter of the highlight area. Based on the camera calibration parameters and the incident angle of the light source, the theoretical shadow length is calculated according to geometric projection, and an elliptical theoretical shadow search area is generated on the image plane along the direction of light projection.

[0087] The system searches for the actual shadow at that location in the image. If the shape and direction of the actual shadow highly coincide with the theoretical calculation (IoU>0.7) and the gradient is inward, it is determined to be an artifact.

[0088] Traditional shadow detection often relies solely on image grayscale, making it easily confused with rust or dark spots on the tool surface. This design incorporates physical geometric constraints into the algorithm, utilizing the fixed lighting environment of the shearing machine to construct a rigid physical model. This physical optics-based verification mechanism enables the system to achieve nearly 100% accuracy in recognizing metal debris (rigid protrusions), solving the problem of distinguishing between rust and debris in two-dimensional images.

[0089] The pixel-level probabilistic inference model is implemented through a fully connected conditional random field model, which includes a global energy index composed of a weighted sum of data observation terms and pairwise correlation terms.

[0090] The data observations are determined by the target defect probability map and the interference artifact probability map, and represent the initial confidence cost of classification based solely on the current pixel features;

[0091] The pairwise association terms introduce a consistency constraint factor for edge gradient direction. The calculation logic of the consistency constraint factor is as follows: for any adjacent pixel pair, if the color difference between them is lower than a preset threshold and the gradient directions are opposite, it indicates that a topological contradiction of concavity and convexity reversal has occurred. Then, a label penalty is applied between the pixel pairs to prohibit them from being simultaneously marked as defect patterns of the same category.

[0092] The mean field approximation algorithm is used to iteratively optimize the global energy index, so that the probability map of interference artifacts with inward gradient and the probability map of target defects with outward gradient form a mutually exclusive competitive relationship in space. Pixels that are visually within the coverage of artifact projection are marked as untrusted regions and logically filtered out to generate a defect binary mask.

[0093] Architecture and logical details of the pixel-level probabilistic inference model:

[0094] A fully connected conditional random field model is used to construct a global energy assessment system aimed at minimizing the classification uncertainty of the system. The core idea of ​​this system is to transform the image segmentation problem into an energy minimization problem. Its global energy index is composed of two weighted parts: data observation terms and pairwise correlation terms.

[0095] First, the data observations are directly anchored to the target defect probability map and interference artifact probability map output by the digital image sequence. This quantifies the initial confidence cost of classifying a pixel as genuine wear or interference artifact solely based on its own visual features such as grayscale and texture. In short, it represents the neural network's independent judgment tendency for each pixel.

[0096] Secondly, in addition to the conventional constraints of color similarity and spatial proximity, the pairwise association term innovatively introduces a consistency constraint mechanism for edge gradient direction; it calculates the gray-level gradient vector relationship between adjacent pixel pairs within the local neighborhood. If the color difference between adjacent pixels is detected to be below a preset threshold, i.e., they are visually similar in color, but their gradient directions are opposite—one exhibits an inward concave feature, and the other an outward convex feature—the system determines that there is a physical topological contradiction of concavity-convexity reversal in this region. In this case, the model imposes a high label blocking cost on the pairwise association term, forcibly prohibiting these two pixels from being labeled as the same category, thereby using physical laws to break the visual confusion.

[0097] Finally, during the inference process, the mean-field approximation algorithm is used to iteratively optimize the aforementioned global energy index. Through the energy minimization competition process, the interference artifact region with inward gradient characteristics and the target defect region with outward gradient characteristics form a strictly mutually exclusive competitive relationship in space.

[0098] Finally, by utilizing the high-confidence regions of interference artifacts, potential defect signals that are visually located within the artifact projection coverage area are actively suppressed through spatial propagation, marked as untrusted regions and logically filtered out, thereby generating a physically pure binary defect mask.

[0099] This invention achieves deep logical fusion of multi-source information through a fully connected conditional random field model. In particular, it introduces consistency constraints on edge gradient directions, cleverly transforming the topologically exclusive relationship between concave and convex areas in physical space into energy constraints in the algorithm model. This not only solves the problem of segmenting blurred edge regions, but more importantly, it implements a logical filtering based on physical properties. It utilizes significant artifacts to actively suppress subtle pseudo-wearing features, ensuring that the generated binary mask has extremely high physical realism.

[0100] Furthermore, the process of anchoring the data observation to the target defect probability map and interference artifact probability map output by the digital image sequence specifically adopts a negative log-likelihood mapping strategy based on adaptive class balancing.

[0101] First, a negative natural logarithmic transformation is performed on the probability values ​​output by the neural network, mapping high-confidence probability values ​​to low-value energy costs and low-confidence probability values ​​to high-value energy costs, thereby constructing an observation potential that conforms to the principle of energy minimization.

[0102] Based on this, a category balance weight coefficient is introduced to dynamically weight the observed potential energy. The calculation logic of the weight coefficient is as follows: the pixel proportion of the target defect area in the current frame image is statistically analyzed in real time, and an inverse proportional weighting mechanism is adopted. When the pixel proportion of the target defect is lower, a larger weight coefficient is assigned to it, thereby amplifying the signal response intensity of small area defects in the global energy index and overcoming the problem of missed detection of small targets caused by extreme imbalance of positive and negative samples.

[0103] When constructing data observations, the original probability values ​​are not used directly, but are instead converted into energy domains.

[0104] In practice, the first step is to calculate the percentage of initially identified defect pixels in the current frame relative to the entire image. Considering the extremely small size of wear areas in industrial scenes (e.g., less than 0.5%), a gain coefficient inversely proportional to this percentage is automatically calculated. For example, if the wear percentage is very low, the system automatically increases the energy weight of the defect category while keeping the weights of the background and artifacts unchanged. Subsequently, a negative logarithmic function is used to process the probability map, ensuring that pixels with a high probability of being defects have extremely low energy values. The final data observation is a combined representation of the weighted defect energy and artifact energy.

[0105] This invention effectively solves the common problem of small target obscuration in industrial defect detection by introducing negative log-likelihood mapping and adaptive class balancing weights. In conventional algorithms, tiny chips are often dominated by background signals due to their small pixel count. This invention nonlinearly amplifies the probability difference through logarithmic domain transformation and uses an inverse weighting mechanism to give small targets higher energy weight, significantly enhancing the system's sensitivity to capturing micron-level chips and ensuring effective detection in the early stages of defects.

[0106] The surface state feature vector is a hybrid feature vector containing geometric dimension, statistical texture dimension and frequency domain energy dimension;

[0107] Specifically, this includes: the maximum depth estimate of the target defect region calculated based on monocular depth estimation, the aspect ratio of the minimum bounding rectangle of the target defect region, the weighted sum of texture energy and inverse gap moment calculated based on the gray-level co-occurrence matrix, the fractal dimension of the defect edge calculated using the box-counting dimension method, and the energy proportion of the dominant frequency component in the Fourier spectrum of the target defect region image; after normalization processing, the surface state feature vector is obtained.

[0108] Among them, the maximum depth estimate of the target defect region is used to quantify the severity of the defect; the aspect ratio of the minimum bounding rectangle of the target defect region is used to distinguish between the narrow linear defect features and the circular sheet-like defect features; the weighted sum of texture energy and inverse gap moment characterizes the roughness and gray-level uniformity of the defect surface; the fractal dimension of the defect edge calculated using the box-counting dimension method is used to quantify the degree of complex jaggedness of the defect contour; and the energy proportion of the dominant frequency component in the Fourier spectrum of the target defect region image is used to characterize the directional intensity of the texture.

[0109] Specifically, in the process of weighted summation of texture energy and inverse difference moment calculated based on gray-level co-occurrence matrix, the weight coefficients of the weighted summation can be obtained by principal component analysis (PCA) to analyze the contribution of the sample set. In this embodiment, the weight of texture energy is 0.6 and the weight of inverse difference moment is 0.4.

[0110] Standard samples are acquired using a calibrated stereo camera, or the 3D topography of the cutting edge surface is measured temporarily in an industrial environment using a structured light scanner to establish a paired training set. Subsequently, a self-supervised monocular deep learning model (e.g., MiDaS framework or Monodepth2) is pre-trained and fine-tuned to adapt to the specific lighting conditions and tool material. During operation, the fine-tuned model is used to perform depth inference on real-time acquired single-frame grayscale images to obtain a coarse-resolution depth map. Then, boundary-preserving upsampling techniques such as guided filtering are used to upsample the depth map to its original resolution using the boundary information of the original grayscale image. Finally, the maximum depth value is extracted within the region specified by the defect mask.

[0111] The hybrid feature vector constructed in this invention encompasses three dimensions: macroscopic geometry, microscopic texture, and frequency domain topology, enabling a comprehensive digital description of wear conditions. The introduction of fractal dimensions quantifies the jagged edges of the chipping, while Fourier spectra quantify the directional damage to the texture. This multi-dimensional feature representation reflects the physical essence of wear more effectively than single area or depth indicators, providing highly separable data support for the subsequent accurate identification of complex wear mechanisms.

[0112] The process of determining defect pattern categories based on surface state feature vectors includes:

[0113] A decision boundary model based on multidimensional feature space partitioning is constructed, and the surface state feature vector is projected into the multidimensional feature space. The defect pattern category is determined according to the decision region into which the vector falls.

[0114] If the feature exhibits significant directional anisotropy in geometric aspect ratio and the texture spectrum energy is concentrated in the low-frequency band of the specified direction, indicating that the texture extends in accordance with the processing direction, it is determined to be a linear loss mode and falls into the first decision region.

[0115] If the feature is characterized by a nonlinear discrete distribution of texture entropy, accompanied by a drift of texture frequency domain features in a specified frequency band, indicating that there is micro-texture migration on the surface due to thermal effects, then it is determined to be a thermal adhesion mode; and falls into the second decision region.

[0116] If the feature exhibits a step-like abrupt change in depth features, and the edge fractal dimension represents a contour with complex geometric and topological irregularities, then it is determined to be a brittle fracture mode and falls into the third decision region.

[0117] Output the normalized probability distribution vector of the pattern category to determine the defect pattern category.

[0118] Identification process and logic: Decision boundary models are constructed using support vector machines or random forests.

[0119] Linear loss mode: If the vector falls into the first region, it means that the texture still follows the processing direction, but it has become darker.

[0120] Thermal adhesion mode: If the vector falls into the second region, it indicates that material migration and coating have occurred on the surface.

[0121] Brittle fracture mode: If the vector falls into the third region, it indicates that a chipping has occurred.

[0122] Output: Pattern category probability distribution vector, for example: [linear: 0.1, adhesion: 0.1, fracture: 0.8].

[0123] The designated frequency band refers to a significant attenuation of high-frequency component energy and a center frequency shift towards the mid-to-low frequency band. Specifically, the proportion of high-frequency energy above the cutoff frequency in the Fourier spectrum is calculated. When this proportion decreases beyond a preset threshold, texture smearing due to thermal effects is determined to have occurred. The preset threshold is set to 10% of the average pixel grayscale value within a local window, or it can be dynamically calculated based on the current illumination intensity using an adaptive threshold algorithm.

[0124] This invention represents a leap from quantitative detection to qualitative identification. By dividing the multi-dimensional feature space, it not only identifies when a tool is damaged, but also how it was damaged; it accurately distinguishes between thermal adhesion and brittle fracture, avoiding secondary damage that may be caused by a single correction strategy. For example, continuing to apply pressure to a chipped tool can lead to a safety accident, significantly improving the decision-making level of the intelligent maintenance system.

[0125] Furthermore, the decision boundary model adopts a cascaded architecture of a one-class support vector machine (One-ClassSVM) and a multi-class support vector machine based on kernel density estimation;

[0126] The first-level classifier uses a One-Class SVM, which is trained using only normal wear samples to quickly detect abnormal samples that deviate from the normal distribution, i.e. brittle fracture patterns.

[0127] The second-level classifier uses a multi-class SVM to further subdivide the samples not detected by the first level. It uses the radial basis kernel function to map the feature vectors to a high-dimensional space and divides the samples into linear loss mode and thermal adhesion mode by maximizing the classification margin.

[0128] At the same time, a confidence gating mechanism is introduced, which outputs a definite mode category only when the classification probability exceeds 0.85; otherwise, it outputs an undetermined mode and retains the correction parameters of the previous frame.

[0129] When identifying defect patterns, the feature vectors are first input into a One-Class SVM, which is trained only on normal wear data and defines a hypersphere boundary. If the feature vector falls outside the sphere, it is considered an anomaly, possibly indicating chipping or severe adhesion. Subsequently, these anomalous samples enter a second-stage RBF-SVM classifier, which further subdivides them based on the specific distribution of features, such as high fractal dimension indicating chipping and high texture entropy indicating adhesion. Simultaneously, a confidence threshold of 0.85 is set. If the classifier output probability does not meet this threshold, the system considers the classification unreliable and does not adjust the parameters to avoid erroneous actions.

[0130] This invention addresses the problem of scarce abnormal samples by employing a cascaded architecture. Industrial accident samples are rare, making it easy for multi-classifiers to overfit when trained directly. A One-Class SVM acts as the first level of defense, ensuring high sensitivity to unknown anomalies; the second-level classifier guarantees accuracy in classifying known faults. A confidence-gating mechanism further enhances the system's robustness, preventing blind decision-making when data is ambiguous.

[0131] The process of generating shear correction parameter data includes:

[0132] The surface state feature vector sequence within the historical time window is called to construct a time-series trend prediction model, and control strategy data is generated by combining the defect mode category of the current frame.

[0133] When the loss pattern is determined to be linear and the depth component of the feature vector shows a linear growth trend in time, the growth slope is calculated and the corresponding gap step compensation data is generated.

[0134] When the thermal adhesion mode is identified and the temporal change rate of the texture entropy feature increases exponentially, it indicates accelerated thermal accumulation and generates thermal balance correction data that reduces the speed and increases the cooling amount.

[0135] When the fracture mode is determined to be brittle and the depth feature undergoes a pulse-like abrupt change within a single frame, it indicates that the device has been subjected to an impact load and generates shutdown alarm data.

[0136] The calculated data is organized into shear correction parameters and transmitted to the actuator to adjust the process parameters.

[0137] First, the surface state feature vector sequence stored in the database within the historical time window is called to construct a time-series trend prediction model in order to capture the dynamic pattern of wear state changes over time.

[0138] Subsequently, based on the defect mode category determined in the current frame, a differentiated control strategy is generated:

[0139] First, for the linear wear mode, if the system detects that the physical depth component in the feature vector exhibits a stable linear growth trend over time, it indicates that the tool is in a normal progressive wear stage. The slope of this growth trend is calculated, and corresponding clearance step compensation data is generated accordingly. This drives the actuator to perform micron-level fine-tuning of the shear clearance to counteract the clearance expansion effect caused by edge wear.

[0140] Secondly, regarding the thermal adhesion mode, if the system detects a sharp exponential increase in the temporal change rate of the texture entropy feature, it indicates that thermal accumulation in the shear region is accelerating, and material adhesion is imminent. Determining the need for thermal equilibrium intervention, thermal equilibrium correction data is generated, instructing the actuator to reduce the shear rate to decrease frictional heat generation, while simultaneously increasing the amount of cooling medium injected, thereby suppressing the spread of adhesive wear.

[0141] Finally, regarding the brittle fracture mode, if the system detects a pulse-like abrupt change in depth characteristics within a single frame, it indicates that the tool has been subjected to an impact load exceeding expectations or that chipping has occurred. The system determines that immediate load relief is necessary to protect the equipment, and then generates load relief data, instructing the actuator to rapidly increase the shear gap and decrease the compression depth, releasing stress through physical yielding to prevent further expansion of the chipped area or tool breakage.

[0142] All control commands calculated above are packaged and organized into a cut correction parameter data package, which is transmitted to the actuator via the industrial fieldbus to complete the real-time adjustment of process parameters.

[0143] This invention introduces a temporal evolution analysis mechanism, enabling the correction strategy to possess dynamic adaptability and predictability. Instead of relying solely on the current instantaneous state for feedback, it incorporates differentiated control methods based on historical evolution trends: precise compensation for gradual wear to maintain machining accuracy, proactive cooling of thermal effects to prevent adhesion failures, and immediate load removal from sudden impacts to ensure equipment safety. This hierarchical closed-loop control strategy maximizes the balance between machining quality, tool life, and production safety, demonstrating the adaptive decision-making capability of intelligent manufacturing systems.

[0144] Furthermore, the construction of the time-series trend prediction model to capture the dynamic pattern of wear status changes over time specifically adopts a bidirectional long short-term memory network that integrates a self-attention mechanism.

[0145] The input layer of this model receives a sequence of historical surface state feature vectors of length T.

[0146] The feature extraction layer contains two LSTM units in opposite directions, which capture the positive evolution trend and the negative dependency of wear features respectively, and output the hidden layer state vector.

[0147] The attention layer redistributes the importance of features at different time steps by calculating the weighting coefficients of the hidden layer state vectors, giving higher attention weight to features that have recently undergone mutations.

[0148] The output layer uses a fully connected layer to regress and predict the feature vector at the next time step, calculates the deviation between the predicted vector and a set threshold, and outputs the first derivative information of the wear rate.

[0149] Specifically, maintain a historical feature queue of length T=50. Construct a Bi-LSTM-Attention model with an input dimension of 5, and the Bi-LSTM layer contains 64 hidden units;

[0150] The attention layer calculates the weights for each time step, assigning high weights to time points where abrupt changes have occurred. For example, when predicting the wear depth at the next time step, if the model finds that the depth of the last 5 frames increases quadratically, the attention mechanism will focus on these 5 frames and predict that the next frame may exceed the safety threshold.

[0151] The output layer, combined with a fully connected network, directly outputs the predicted depth and wear rate for the next frame. If the wear rate exceeds a linear threshold, the system prematurely determines that it has entered a period of rapid wear.

[0152] The model training employs a self-supervised learning strategy, using the feature vectors of T consecutive historical frames as input and the true feature vector of the (T+1)th frame as the label. The loss function is the mean squared error (MSE) loss. In the initial stage of system operation, pre-training is performed using standard wear-and-tear lifecycle data collected offline, and online fine-tuning is performed during operation using sliding window data.

[0153] This invention employs a Bi-LSTM combined with an attention mechanism, enabling it to simultaneously utilize historical long-term trends and recent abrupt changes. Compared to simple linear regression or Kalman filtering, this model can more accurately capture the nonlinear wear evolution patterns, such as the inflection point when wear enters a severe stage. This allows the system to not only perform post-event corrections but also pre-event predictions, adjusting parameters in advance before wear leads to defective products, thus achieving true predictive maintenance.

[0154] Example 2:

[0155] This invention proposes a dynamic correction system for shear behavior based on wear state recognition, the structure of which is as follows: Figure 3 As shown, it includes:

[0156] The data acquisition module acquires digital image sequences of the cutting edge region and performs image enhancement and illumination equalization preprocessing.

[0157] The feature recognition module constructs an image feature separation network to perform parallel feature extraction on the preprocessed digital image sequence. Among them, the high-frequency texture analysis branch uses the continuous features of the periodic texture pattern on the cutting edge surface to identify pixel-level high-frequency texture discontinuities and edge gradient direction fields, and segments out the probability map of target defects representing material loss. The photometric shadow analysis branch uses the backlighting imaging model to identify the projected shadow area of ​​the protrusion edge and the low-frequency light halo, and segments out the probability map of interference artifacts representing material accumulation.

[0158] Periodic texture patterns are a prerequisite assumption for the high-frequency texture analysis branch. This assumption holds true under the following conditions: the tool is in its early to mid-life (e.g., wear does not exceed 60% of its normal lifespan), at which point, although there are minor defects, the macroscopic texture of the surface remains regular; the object being processed is a conventional metal sheet (steel, aluminum alloy, stainless steel), and the cutting speed is within the recommended process parameters.

[0159] The mask generation module constructs a pixel-level probabilistic inference model and performs mask fusion and thinning operations; it establishes spatial mutual exclusion constraints between the target defect probability map and the interference artifact probability map in the pixel space and generates a binary defect mask.

[0160] The correction analysis module performs multidimensional texture feature calculation on the region corresponding to the defect binary mask to construct a surface state feature vector; based on the distribution of the surface state feature vector in the multidimensional feature space, it determines the defect pattern category and generates the corresponding shearing correction parameters.

[0161] In this embodiment, the image feature separation network can be implemented using a multi-branch structure based on a convolutional neural network. Specifically, the single-frame or multi-frame blade edge images, after preprocessing by image enhancement and illumination equalization, are stacked by channel and input into a shared convolutional feature extraction trunk to obtain an intermediate feature map that maintains the same spatial resolution as the input image.

[0162] Subsequently, based on this intermediate feature map, a high-frequency texture analysis branch and a photometric shadow analysis branch are set up respectively. The high-frequency texture analysis branch enhances the response to periodic knife marks through a series of direction-selective convolution kernels and outputs a high-frequency response feature map with the same size as the input image. The photometric shadow analysis branch responds to oil film and highlight areas through convolution operators targeting dark channels and highlight areas and outputs a photometric feature map with the same size as the input image.

[0163] After normalization, the two feature maps are used as initial estimates for the target defect probability map and the interference artifact probability map, respectively, in the subsequent pixel-level probability inference model. Network parameters can be obtained through offline training using collected labeled blade image samples and loaded as fixed parameters during system runtime. This implementation is merely one specific method for image feature separation networks; those skilled in the art can make conventional adjustments to the structure, such as the number of convolutional layers and channels, without affecting the essential content of this invention.

[0164] This embodiment provides a dynamic correction system for shearing behavior based on wear state recognition, which mainly consists of three physical entities: an image acquisition terminal, an edge computing processing node, and an execution control unit, connected through an industrial fieldbus.

[0165] Image acquisition terminal: Includes a high-resolution global shutter industrial camera, such as the Basler ace series, with a resolution of 5 megapixels, mounted on the side of the shearing machine tool holder, with the optical axis perpendicular to the cutting edge plane.

[0166] It is equipped with a combined light source system, including a coaxial light source for capturing textures and a side strip light source for building backlit models, which is triggered by the controller to flicker.

[0167] Edge computing processing nodes: These employ embedded AI computing boxes (such as the NVIDIA Jetson AGX Orin) with built-in high-performance GPUs for deploying deep learning networks and CRF inference algorithms. This node is responsible for all computational tasks, from image processing to control commands.

[0168] The execution control unit includes a PLC controller, such as a Siemens S7-1500, and a servo drive system, which directly controls the eccentric shaft motor of the shearing machine to adjust the clearance and controls the main drive motor to adjust the speed.

[0169] The edge computing processing nodes deploy a software system based on a microservices architecture, which specifically includes the following functions:

[0170] Feature separation and extraction: Receive the RAW format image transmitted from the camera, process it through ISP and perform illumination equalization, and then input it into the image feature separation network.

[0171] High-frequency texture branch: Outputs the texture gradient field of the cutting edge area to identify potential wear.

[0172] Photometric Shadow Branch: Combines light source calibration parameters to output the shadow probability of potential foreign objects.

[0173] Output: Generate two floating-point probability maps: a target defect probability map and an interference artifact probability map.

[0174] Adaptive probabilistic inference: This solves the problem of missing minor wear caused by imbalanced positive and negative samples. Its internal operating logic is as follows:

[0175] Pixel percentage statistics: Real-time statistics are collected on the number of pixels with a probability value exceeding 0.5 in the current frame, and the percentage of these pixels relative to the total cutting edge area is calculated. In industrial settings, actual wear is often extremely small.

[0176] Adaptive weight calculation: Dynamically calculates category balance weight coefficients, including the actual wear percentage, baseline gain, and smoothing term. When the wear area is extremely small, the weight coefficients will automatically increase (e.g., from 1.0 to 5.0).

[0177] Negative log-likelihood mapping: Instead of directly comparing the original probabilities, the system converts them into data observations in the energy domain. Through this step, those tiny, easily ignored wear signals are given a very high presence in the energy function, forcing subsequent models to preserve these regions.

[0178] Pixel-level probabilistic inference model optimization: By incorporating pairwise correlation terms with gradient consistency constraints, the global energy is iteratively minimized using a mean-field approximation algorithm. This generates a physically pure defect binary mask, successfully preserving micrometer-level chipping features while eliminating large-area oil contamination interference.

[0179] Correction Strategy Generation: Receives a binary mask, calculates feature vectors, and identifies wear modes, including linear loss, thermal adhesion, and brittle fracture. Timing Analysis: Maintains a circular buffer to store feature vectors from the past 100 cycles. Instruction Generation: If linear loss is identified and the depth increases linearly, the system generates specific pulse compensation instructions through protocol conversion to drive the servo motor to fine-tune the tool gap.

[0180] All calculated control command data are encapsulated into a structure containing the following fields:

[0181] Parameter type identifier: 1 byte, 0-linear loss compensation, 1-thermal balance correction, 2-load unloading alarm;

[0182] Gap adjustment amount: 2-byte signed integer, unit 0.01mm, range -50 to +50;

[0183] Speed ​​adjustment percentage: 2-byte unsigned integer, ranging from 0% to 200%;

[0184] Coolant flow rate adjustment percentage: 2-byte unsigned integer, range 0–200%;

[0185] Timestamp: 4-byte Unix timestamp;

[0186] Confidence score: 1 byte, 0-100.

[0187] The PLC controller is transmitted via Ethernet using the industry-standard Modbus TCP protocol with a communication baud rate of 9600bps and an address offset of 0x0100.

[0188] In actual testing on a metal sheet shearing production line, the system demonstrated the following performance advantages:

[0189] Improved detection rate of minute defects: Thanks to the adaptive weighting and negative logarithmic mapping strategy, the system has significantly improved the detection rate of micro-chips with a diameter of less than 0.05 mm.

[0190] Strong anti-interference capability: effectively shields more than 99% of oil and debris interference on the tool surface, reducing the false alarm rate to below 0.1%.

[0191] Closed-loop correction accuracy: It achieves micron-level dynamic gap adjustment, which keeps the burr height of the sheared section of the plate stable within 0.02mm for a long time, and extends the tool life by about 30%.

[0192] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for dynamic correction of shearing behavior based on wear state recognition, characterized in that, The method comprises the following steps: acquiring a digital image sequence of the blade edge region, performing image enhancement and illumination equalization preprocessing; constructing an image feature separation network to perform parallel feature extraction on the preprocessed digital image sequence; wherein, a high-frequency texture analysis branch utilizes the continuity feature of the periodic texture pattern on the blade edge surface to identify the high-frequency texture discontinuous region and the edge gradient direction field at the pixel level, and segment out a target defect probability map representing material loss; a photometric shadow analysis branch utilizes an inverse illumination imaging model to identify the projection shadow region of the protrusion edge and the low-frequency illumination halo, and segment out an interference artifact probability map representing material accumulation; constructing a pixel-level probability inference model to perform mask fusion and refinement operations; establishing a spatial mutual exclusion constraint between the target defect probability map and the interference artifact probability map in the pixel space to generate a defect binary mask; performing multi-dimensional texture feature calculation on the region corresponding to the defect binary mask to construct a surface state feature vector; based on the distribution of the surface state feature vector in the multi-dimensional feature space, the defect mode category is determined, and the corresponding shear correction parameter is generated; the process of determining the defect mode category based on the surface state feature vector comprises: constructing a decision boundary model based on multi-dimensional feature space division, projecting the surface state feature vector into the multi-dimensional feature space, and determining the defect mode category according to the decision region where the vector falls; if the feature shows that the geometric aspect ratio has directional anisotropy, and the texture spectrum energy is concentrated in the low-frequency band of the specified direction, indicating that the texture extends along the processing direction, it is determined as a linear loss mode; if the feature shows that the texture entropy presents a nonlinear discrete distribution, and the texture frequency domain feature drifts in the specified frequency band, indicating that the surface exists micro-texture migration caused by thermal effect, it is determined as a thermal adhesion mode; if the feature shows that the depth feature presents a stepwise mutation, and the edge fractal dimension represents a complex geometric topological irregularity of the profile, it is determined as a brittle fracture mode; output the normalized mode category probability distribution vector to determine the defect mode category.

2. The shear behavior dynamic correction method based on wear state recognition according to claim 1, wherein: the identification process of the high-frequency texture analysis branch comprises: performing multi-scale frequency domain filter bank convolution operation on the digital image sequence to extract the texture orientation field of the blade edge surface, and using the continuity prior feature of the periodic texture pattern to mark the region where the texture flow line occurs pulse function interruption and the high-frequency energy occurs stepwise decline as a potential structure damage area; calculate the gray scale gradient vector field in the potential structure damage area, and construct a radial reference vector field with the area geometric center as the origin; perform pixel-by-pixel dot product operation on the gray scale gradient vector field and the radial reference vector field; only when the dot product result of the closed region edge is positive, it is shown that the image gradient direction deviates from the area geometric center, and the local minimum value characteristic of the area internal gray scale distribution shows a sharp step edge, it is determined that there is a negative mutation of image depth, and the target defect probability map containing clear boundary information is output.

3. The shear behavior dynamic correction method based on wear state recognition according to claim 1, wherein: The identification process of the photometric shadow analysis branch includes: For non-rigid interference artifacts, the dark channel prior map of the local image block is calculated, the area with gray value higher than the preset reference in the dark channel and the edge with Gaussian blur characteristics is identified, and the fluid artifact sub-mask is marked; For structured interference artifacts, the pixel set with brightness value exceeding the saturation limit in the image is extracted by threshold segmentation, and the specular highlight points are determined; based on the known light source incident direction vector, an inverse lighting imaging model is constructed, the specular highlight points are used as the positioning anchor points, and the projection shadow area on the opposite side is searched; the gradient direction in the neighborhood of the highlight point is calculated, and if the gradient direction is pointed to the geometric center of the area from all around, it is determined as a non-defective positive accumulation; using morphological geodesic dilation operation, the determined specular highlight points and the projection shadow area are used as common seed points, and the restricted growth is carried out on the gradient map to completely cover the rigid foreign matter area, and the rigid artifact sub-mask is generated; The fluid artifact sub-mask and the rigid artifact sub-mask are fused at the pixel level, and the interference artifact probability map containing high-confidence artifact distribution information is output.

4. The shear behavior dynamic correction method based on wear state recognition according to claim 1, characterized in that: The pixel-level probability inference model is realized by a fully connected conditional random field model, including a global energy index composed of a data observation item and a pair correlation item weighted summation; The data observation item is determined by the target defect probability map and the interference artifact probability map, representing the initial confidence cost of classification only according to the current pixel characteristics; The pair correlation item introduces a consistency constraint factor of edge gradient direction; the calculation logic of the consistency constraint factor is: for any adjacent pixel pair, if the color difference between the adjacent pixel pair is lower than the preset threshold and the gradient directions are opposite, which represents a topological contradiction of concave-convex inversion, a label penalty is applied between the pixel pair to prohibit being labeled as the same defect mode; The average field approximation algorithm is used to iteratively optimize the global energy index, so that the interference artifact probability map with inward gradient and the target defect probability map with outward gradient form a mutually exclusive competitive relationship in space, the pixels located in the projection coverage range of the artifact in vision are labeled as untrustworthy areas and are logically filtered out, and the defect binary mask is generated.

5. The shear behavior dynamic correction method based on wear state recognition according to claim 1, characterized in that: The surface state feature vector is a mixed feature vector including geometric dimensions, statistical texture dimensions and frequency energy dimensions; Specifically, it includes: the maximum depth estimation value of the target defect area calculated based on monocular depth estimation, the aspect ratio of the minimum bounding rectangle of the target defect area, the weighted sum of texture energy and inverse gap metric calculated based on the gray level co-occurrence matrix, the defect edge fractal dimension calculated by the box counting method, and the energy proportion of the main frequency component in the Fourier spectrum of the target defect area image; after normalization, the surface state feature vector is obtained.

6. The shear behavior dynamic correction method based on wear state recognition according to claim 1, characterized in that: The process of generating shear correction parameter data includes: The surface state feature vector sequence in the call history time window is called to construct a time trend prediction model, and a control strategy data is generated in combination with the current frame defect mode category; When it is determined that the linear loss mode and the depth component of the feature vector presents a linear growth trend in time sequence, the growth slope is calculated and corresponding gap step compensation data is generated; When it is determined that the thermal adhesion mode and the time sequence change rate of the texture entropy feature presents an exponential rise, it represents that the heat accumulation is accelerated, and the thermal balance correction data of reducing speed and increasing cooling amount is generated; When it is determined that the brittle fracture mode and the depth feature has a pulse mutation in a single frame, it represents that it is impacted by the load, and shutdown alarm data is generated; The calculated data is arranged as a shear correction parameter and transmitted to an actuator to adjust the process parameters.

7. A shear behavior dynamic correction system based on wear state recognition, characterized by, It comprises: A data acquisition module acquires a digital image sequence of the blade edge region, and performs image enhancement and illumination equalization preprocessing; A feature recognition module constructs an image feature separation network to perform parallel feature extraction on the preprocessed digital image sequence; wherein, a high-frequency texture analysis branch utilizes the continuity feature of the periodic texture pattern of the blade edge surface to identify the high-frequency texture discontinuous region and the edge gradient direction field at the pixel level, and segment out the target defect probability graph representing material loss; a photometric shadow analysis branch utilizes an inverse illumination imaging model to identify the projection shadow region of the protrusion edge and the low-frequency illumination halo, and segment out the interference artifact probability graph representing material accumulation; A mask generation module constructs a pixel-level probability inference model to perform mask fusion and thinning operations; a spatial mutual exclusion constraint between the target defect probability graph and the interference artifact probability graph is established in the pixel space to generate a defect binary mask; A correction analysis module performs multi-dimensional texture feature calculation on the region corresponding to the defect binary mask to construct a surface state feature vector; based on the distribution of the surface state feature vector in the multi-dimensional feature space, the defect mode category is determined, and corresponding shear correction parameters are generated; The process of determining the defect mode category based on the surface state feature vector includes: A decision boundary model based on multi-dimensional feature space division is constructed, the surface state feature vector is projected into the multi-dimensional feature space, and the defect mode category is determined according to the decision region where the vector falls; If the feature has directional anisotropy in geometric aspect ratio, and the texture spectrum energy is concentrated in the low frequency band of the specified direction, it represents that the texture extends along the processing direction, and it is determined as linear loss mode; If the feature presents a nonlinear discrete distribution of texture entropy, and the texture frequency domain feature drifts in the specified frequency band, it represents that the micro texture migrates due to thermal effect, and it is determined as thermal adhesion mode; If the feature presents a step mutation of depth feature, and the edge fractal dimension represents that the profile has complex geometric topological irregularity, it is determined as brittle fracture mode; A normalized mode category probability distribution vector is output to determine the defect mode category.

Images