A metal plate surface defect detection method fusing multi-source sensing data

Abstract

The present application relates to the technical field of image processing, in particular to a metal plate surface defect detection method fusing multi-source sensing data, comprising collecting multi-source data of a metal plate surface to be detected; extracting characteristic parameters of surface rolling stripes; dividing the metal plate surface into sub-regions, and establishing a multi-modal characteristic fingerprint library; reconstructing the height field distribution of the metal plate surface according to three-dimensional topographic data of the metal plate; projecting the multi-source data onto the reconstructed three-dimensional curved surface, and establishing a corresponding relationship; extracting rolling stripe characteristics of a current scanning region, and matching the rolling stripe characteristics with the multi-modal characteristic fingerprint library to calculate the relative difference of stripe phases and correct the relative position offset of the multi-source data; collecting registered eddy current electromagnetic data to obtain multi-source characteristic data; establishing a defect correlation criterion to obtain a characteristic parameter response mode of a defect type; matching the multi-source characteristic data with the defect correlation criterion to confirm a real defect; and marking as a suspected defect and / or a non-defect characteristic when not matched.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a method for detecting surface defects in metal plates by fusing multi-source sensor data. Background Technology

[0002] Cold-rolled thin steel sheets are widely used in new energy vehicles, home appliances, and high-end equipment manufacturing. However, they are prone to shape defects during production, transportation, and warehousing, directly affecting the accuracy of subsequent stamping, welding, and assembly. Traditional inspection methods mainly rely on manual visual inspection or a single sensor, which suffers from low inspection accuracy, high false negative rate, and inability to adapt to high-speed production lines.

[0003] Currently, domestic steel companies have attempted to introduce machine vision into the defect identification process. However, existing solutions mostly use a single data source, relying solely on industrial cameras for two-dimensional image analysis, making it difficult to accurately identify minute defects and three-dimensional morphological defects. Furthermore, on-site interference factors such as steel plate surface reflection, oil stains, and speed fluctuations can severely affect the stability of visual inspection, leading to a persistently high false positive rate. While simple laser displacement sensors can acquire height information, they are insensitive to surface texture defects and have limited scanning speed.

[0004] Current technologies lack effective means to deeply integrate image data with mechanical measurement data, making it impossible to establish a correlation model between defect features and material properties, and hindering intelligent defect classification and severity assessment. This results in detection results failing to directly guide the automatic adjustment of subsequent correction process parameters, impacting the overall intelligence level and production efficiency of the production line. Therefore, achieving accurate identification and real-time feedback of microscopic defects in cold-rolled thin steel sheets has become an urgent problem to be solved.

[0005] To address this, a method for detecting surface defects in metal plates that integrates multi-source sensor data is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a method for detecting surface defects in metal plates by fusing multi-source sensor data. By utilizing the inherent rolling stripe features of the metal plate to establish a multimodal fingerprint database, calculating the stripe phase difference to correct the spatial offset between visual and laser data, high-precision dynamic registration without calibration objects is achieved. At the same time, a defect correlation model is constructed by combining the response patterns of visual, morphological, and eddy current data. Through multi-dimensional feature verification, true and false defects are effectively distinguished, solving the problems of difficult multi-source data fusion and high false alarm rate.

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

[0008] A method for detecting surface defects in metal plates by integrating multi-source sensor data, comprising:

[0009] Collect multi-source data of the surface of the metal plate to be inspected; extract the feature parameters of the rolling stripes on the surface of the metal plate; divide the surface of the metal plate into sub-regions, and establish a multimodal feature fingerprint database containing the feature parameters of the rolling stripes for each sub-region;

[0010] Based on the real-time measured three-dimensional topography data of the metal plate, the height field distribution on the surface of the metal plate is reconstructed; the multi-source data is projected onto the reconstructed three-dimensional curved surface to establish the correspondence between pixels and the physical position of the plate surface; the rolling stripe features of the current scanning area are extracted and matched with the multi-modal feature fingerprint database respectively; the relative position offset of the multi-source data is corrected by calculating the relative difference of the stripe phase.

[0011] After registration, eddy current electromagnetic data at the same location are collected and impedance features are extracted. Combined with the grayscale and depth features of the multi-source data extracted in the defect area, multi-source feature data is obtained.

[0012] Establish a defect correlation criterion to obtain the characteristic parameter response pattern of the defect type; match the multi-source characteristic data with the defect correlation criterion, and when the combination pattern of the characteristic data matches the response pattern of the defect type, it is confirmed as a real defect; when they do not match, they are marked as suspected defects and / or non-defect features.

[0013] Preferably, the multi-source data includes visual image data and laser topography data; the feature parameters include the stripe direction angle and spatial period extracted from the visual image data, and the surface undulation period and amplitude extracted from the laser topography data;

[0014] The step of dividing the surface of the metal plate into sub-regions includes:

[0015] A two-dimensional Fourier transform is performed on the visual image data to identify the dominant spatial frequency peak corresponding to the rolling stripes in the frequency domain. The direction angle and spatial period of the stripes are obtained through inverse transform and peak localization. The surface contour curve is extracted from the laser morphology data along the direction perpendicular to the rolling direction. The periodic components and amplitudes of the surface undulations are extracted through one-dimensional Fourier analysis. The deviations in the direction angle and spatial period of the rolling stripes extracted from the visual image data and laser morphology data are verified, and the regions that meet the conditions are determined as effective registration regions. Based on the surface quality distribution law of the effective registration regions during the cold rolling process, the surface of the metal plate is divided into a head transition region, a tail transition region, an edge region, and a central stable region. Different sub-region sizes are set for different functional regions. The direction angle, spatial period, and undulation amplitude of the rolling stripes are extracted and stored for each sub-region to establish a multimodal feature fingerprint of the sub-region.

[0016] Preferably, establishing the correspondence between pixels and the physical positions on the board surface includes:

[0017] By scanning and measuring the surface of a metal plate, discrete height measurement point data is obtained. Two-dimensional spatial interpolation is performed on the height measurement point data to reconstruct the continuous height field distribution on the metal plate surface. An iterative process is initialized, and perspective projection geometry is used to calculate the initial position coordinates of each pixel on the plate surface in the visual image. Based on the initial position coordinates, the height value of the corresponding position is retrieved from the reconstructed height field. The retrieved height value is used to correct the three-dimensional coordinates of the corresponding pixel on the plate surface. The query and correction steps are repeated until the change in the three-dimensional coordinates calculated between two iterations is less than a preset convergence threshold. The converged three-dimensional coordinates are used as the final physical position of the plate surface.

[0018] Preferably, the step of correcting the relative positional offset of multi-source data includes:

[0019] Rolling stripe extraction is performed on the visual image data of the current scanning area, and the spatial phase distribution of the rolling stripes is calculated using Hilbert transform. The extracted stripe features are then matched with a pre-established multimodal feature fingerprint database to determine the sub-region identifier corresponding to the current area. The same rolling stripe extraction and phase calculation are performed on the laser topography data of the current scanning area. The phase difference between the stripe phase in the visual image data and the stripe phase in the laser topography data is calculated. Based on the known spatial period of the rolling stripes, the relative position offset between the visual image data and the laser topography data is calculated using the linear relationship between the phase difference and spatial displacement. The spatial coordinates of the visual image data and the laser topography data are then translated to correct the relative position offset.

[0020] Preferably, the establishment of the defect correlation criterion includes:

[0021] Standard samples containing known defect types, including scratch defects and indentation defects, are collected. Multi-source data is acquired from the standard samples. Gray-level contrast, edge gradient, and shape features are extracted from visual image data; depth anomalies, width, edge slope, and roughness variations are extracted from laser morphology data; and impedance variation amplitude and phase shift are extracted from eddy current electromagnetic data. Statistical analysis is performed on the standard sample data for each defect type, calculating the mean, variance, and distribution range of each feature parameter. Response patterns for different defect types are analyzed to determine the correlation patterns between feature parameters. For scratch defects, the correlation pattern is that visual gray-level contrast, laser depth anomalies, and eddy current impedance variation amplitudes all show anomalies. For indentation defects, the correlation pattern is that laser depth anomalies are significant, but the eddy current impedance variation is lower than the impedance variation caused by scratches. A multi-dimensional feature space for defect types is established, defining reasonable value ranges for feature parameters and constraints between parameters.

[0022] Preferably, matching multi-source feature data with defect correlation criteria includes:

[0023] For the same registered board position, obtain multi-source feature data; calculate the similarity between the multi-source feature data and the feature space of each defect type in the defect correlation criterion; verify whether the multi-source feature data satisfies the parameter constraint relationship of a certain defect type; when the multi-source feature data conforms to the feature space of the defect type and satisfies the corresponding parameter constraint relationship, the position is determined as a real defect of the defect type; when the feature vector is greater than the first distance threshold with the feature space of all known defect types, and / or does not satisfy the parameter constraint relationship of any defect type, the position is marked as a suspected defect and / or a non-defect feature.

[0024] Preferably, it also includes an adaptive partitioning step based on the plate deformation field:

[0025] Based on the reconstructed height field distribution on the metal plate surface, the second-order partial derivative of the height field is calculated to obtain the surface curvature at the plate position. A curvature threshold is set, and regions with surface curvature greater than the curvature threshold are marked as high curvature regions, while regions with surface curvature less than the curvature threshold are marked as low curvature regions. The sub-region size of the head transition zone is set for the high curvature region, and the sub-region size of the central stabilization zone is set for the low curvature region.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] 1. This invention establishes a mapping relationship between "multi-source data and real physical location" in metal plate surface defect detection by deeply fusing visual images, laser 3D topography, and eddy current electromagnetic signals. The proposed method for iterative reconstruction of the 3D height field and convergence calculation of pixel-corresponding physical locations achieves spatial consistency between visual image coordinates, laser depth data, and electromagnetic signal sampling points, ensuring that multi-source data are strictly aligned to the same plate coordinate system. The phase difference calculation of rolling stripe features corrects the positional offset of the multi-source data, overcoming the error accumulation caused by illumination changes, plate vibration, and scanning delay in traditional methods. By jointly analyzing visual grayscale features, laser depth anomalies, and eddy current impedance changes under a unified coordinate system, the accuracy of defect localization and the distinguishability of different defect types are improved, making the detection results more robust to plate surface fluctuations, uneven forming, and measurement noise.

[0028] 2. This invention introduces multimodal parameters such as rolling stripe direction angle, spatial period, and undulation amplitude into metal plate inspection, and establishes a feature fingerprint database at the sub-region level, thereby realizing differentiated detection strategies for different functional regions (head transition area, edge area, and central stable area). This invention utilizes the objective laws governing regional quality distribution differences in the rolling process, extracting periodic texture features from visual and topographic data through two-dimensional and one-dimensional Fourier analysis, and verifying the consistency of direction angle / period deviations between visual image data and laser topographic data, making region division more reliable. By using different sub-region sizes in high-curvature and low-curvature regions, this invention adaptively considers the influence of the plate deformation field, making stripe feature extraction, registration, and defect analysis more stable. The multimodal fingerprint database enables the system to quickly identify the position of the current scanning area in the entire plate and select the matching model, avoiding threshold failure, misjudgment, or missed detection caused by regional differences in traditional methods. This achieves higher resolution and more accurate detection that conforms to the actual rolling process, improving the overall reliability of detection on complex curved surfaces or processing transition areas.

[0029] 3. This invention systematically analyzes the visual, laser, and electromagnetic response characteristics of scratch and indentation defects to construct defect correlation criteria and a multi-dimensional feature space, achieving physical correlation discrimination between multi-source features. The model of this invention not only considers traditional visual / morphological features such as grayscale contrast, depth anomalies, and width, but also introduces eddy current impedance and phase shift parameters, giving defect judgment stronger physical meaning and type differentiation capabilities. By modeling the mean, variance, and feature distribution of typical defect samples, the "feature response pattern" of each defect in multi-source data is identified. This invention ensures that a true defect can only be confirmed when the multi-source feature data simultaneously satisfy the feature space location and parameter constraints. When the feature vector's distance from all defect types exceeds a threshold, automatic filtering is performed to avoid misclassifying non-defect features such as texture changes, stripe fluctuations, and surface reflections as defects. This improves the accuracy of defect classification and the filtering capability of non-defect areas, making the system more suitable for high-noise, multi-interference environments in actual industrial production. Attached Figure Description

[0030] Figure 1 A flowchart of a method for detecting surface defects in metal plates that integrates multi-source sensor data is provided by the present invention.

[0031] Figure 2 This is a flowchart of dividing the surface of a metal plate into sub-regions provided in an embodiment of the present invention;

[0032] Figure 3 This is a flowchart for multi-source data relative position offset correction provided in an embodiment of the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0034] Example 1:

[0035] Please see Figure 1 This invention provides a method for detecting surface defects in metal plates by integrating multi-source sensor data. The technical solution is as follows: Multi-source data of the surface of the metal plate to be inspected is collected; feature parameters of rolling stripes on the surface of the metal plate are extracted; the surface of the metal plate is divided into sub-regions, and a multi-modal feature fingerprint library containing rolling stripe feature parameters is established for each sub-region; the height field distribution of the metal plate surface is reconstructed based on real-time measured three-dimensional topography data; the multi-source data is projected onto the reconstructed three-dimensional surface, and a correspondence between pixels and the physical position of the plate surface is established; the rolling stripe features of the currently scanned area are extracted and matched with the multi-modal feature fingerprint library respectively, and the relative positional offset of the multi-source data is corrected by calculating the relative difference of the stripe phase; eddy current electromagnetic data at the same registered position are collected, and impedance features are extracted, combined with the grayscale features and depth features of the multi-source data extracted in the defect area to obtain multi-source feature data; a defect correlation criterion is established to obtain the feature parameter response pattern of the defect type; the multi-source feature data is matched with the defect correlation criterion, and when the combination pattern of the feature data matches the response pattern of the defect type, it is confirmed as a real defect; when they do not match, they are marked as suspected defects and / or non-defect features.

[0036] Furthermore, the multi-source data includes visual image data and laser topography data; the feature parameters include the stripe direction angle and spatial period extracted from the visual image data, and the surface undulation period and amplitude extracted from the laser topography data;

[0037] The step of dividing the surface of the metal plate into sub-regions includes:

[0038] A two-dimensional Fourier transform is performed on the visual image data to identify the dominant spatial frequency peak corresponding to the rolling stripes in the frequency domain. The direction angle and spatial period of the stripes are obtained through inverse transform and peak localization. The surface contour curve is extracted from the laser morphology data along the direction perpendicular to the rolling direction. The periodic components and amplitudes of the surface undulations are extracted through one-dimensional Fourier analysis. The deviations in the direction angle and spatial period of the rolling stripes extracted from the visual image data and laser morphology data are verified, and the regions that meet the conditions are determined as effective registration regions. Based on the surface quality distribution law of the effective registration regions during the cold rolling process, the surface of the metal plate is divided into a head transition region, a tail transition region, an edge region, and a central stable region. Different sub-region sizes are set for different functional regions. The direction angle, spatial period, and undulation amplitude of the rolling stripes are extracted and stored for each sub-region to establish a multimodal feature fingerprint of the sub-region. The orientation angle deviation is calculated as the difference between the rolling stripe orientation angle in the visual image data and the rolling stripe orientation angle in the laser topography data, and the absolute value of this difference is taken as the orientation angle deviation data. The spatial period deviation is calculated as the difference between the spatial period in the visual image data and the surface undulation period in the laser topography data, and the absolute value of this difference is taken as the spatial period deviation data. It is determined whether the deviation is within the allowable range. Regions within the allowable range are identified as valid registration regions, while regions outside the allowable range are identified as defect regions. See details... Figure 2 .

[0039] Surface quality distribution patterns refer to the specific distribution patterns of surface texture and morphology characteristics exhibited by metal sheets during the cold rolling process due to differences in stress state, temperature field, and roll wear at different locations on the sheet surface. Specifically, this pattern manifests as follows: the head and tail transition zones of the metal sheet typically exhibit instability in the rolling state, resulting in rapid changes in surface texture; the edge regions, due to stress concentration or sheet shape control, often experience significant deformation or texture distortion; while the central stable zone exhibits relatively uniform surface quality and stable texture characteristics. Based on this inherent process distribution pattern, the surface of the metal sheet is logically divided into these four functional regions to allow for differentiated treatment strategies tailored to the texture stability characteristics of different regions.

[0040] The process of constructing a multimodal feature fingerprint database is based on the extraction and structured storage of inherent rolling process features on the surface of metal plates.

[0041] First, visual image data and laser morphology data of the surface of the metal plate to be inspected are collected. For the visual image data, two-dimensional frequency domain analysis is performed, specifically using two-dimensional Fourier transform processing. The dominant spatial frequency peak representing the rolling stripes is identified in the frequency domain spectrum. Through inverse transform and localization analysis of this peak, two key feature parameters are obtained: stripe direction angle (characterizing the extension direction of the texture) and spatial period (characterizing the density of the texture). Simultaneously, for the laser morphology data, a surface contour curve is extracted along a path perpendicular to the rolling direction. One-dimensional Fourier analysis is performed on this contour curve to extract the periodic components of the surface undulations, thereby obtaining two more key feature parameters: surface undulation period and amplitude (characterizing the depth of the texture). Finally, the system divides the surface of the metal plate into several sub-regions. For each sub-region, the stripe direction angle and spatial period extracted from the visual image data, and the surface undulation period and amplitude extracted from the laser morphology data are correlated and stored. This set of parameters constitutes the "multimodal feature fingerprint" of the sub-region; the database stores the set of the above four feature parameters corresponding to each sub-region.

[0042] The process of setting different sub-region sizes is an adaptive adjustment process based on the deformation characteristics of the sheet metal, aiming to balance detection accuracy and computational efficiency. First, based on the reconstructed height field distribution data of the metal sheet surface, the second-order partial derivatives of this height field at various locations are calculated, thereby quantifying the surface curvature at different points on the sheet. Then, the calculated surface curvature is compared with a preset curvature threshold. Regions with surface curvature greater than the curvature threshold are marked as high-curvature regions. These regions typically correspond to the aforementioned head transition zone, tail transition zone, or edge region, where surface features change drastically. For these regions, a smaller sub-region size is set to improve the spatial resolution of feature extraction, ensuring accurate capture of rapidly changing rolling stripe features. Regions with surface curvature less than the curvature threshold are marked as low-curvature regions. These regions typically correspond to the central stable region, where surface features are smooth and uniform. For these regions, a larger sub-region size is set to reduce data processing and improve overall detection efficiency. Finally, sub-regions of different sizes collectively cover the entire surface of the metal sheet.

[0043] In this embodiment, a multimodal feature fingerprint library for metal sheet rolling texture is constructed by fusing two types of multi-source data: visual images and laser morphology, enabling high-precision identification of complex surface defects. Core parameters such as stripe direction angle, spatial period, and surface undulation period and amplitude are extracted through two-dimensional and one-dimensional Fourier analysis. Based on the surface quality distribution pattern during cold rolling, the sheet surface is divided into functional regions such as head, tail, edge, and central stability. Furthermore, the size of the sub-regions is adaptively adjusted according to the surface curvature, allowing for higher-resolution feature representation in areas of drastic texture changes. This method accurately registers multi-source features on a three-dimensional curved surface, forming a structured feature fingerprint library, providing a stable regional benchmark for subsequent defect detection. Overall, it achieves refined modeling of metal sheet surface texture, improves the reliability of multi-source data fusion, and lays a high-confidence data foundation for distinguishing between real and false defects.

[0044] Furthermore, establishing the correspondence between pixels and their physical locations on the board includes:

[0045] By scanning and measuring the surface of a metal plate, discrete height measurement point data is obtained. Two-dimensional spatial interpolation is performed on the height measurement point data to reconstruct the continuous height field distribution on the metal plate surface. An iterative process is initialized, and perspective projection geometry is used to calculate the initial position coordinates of each pixel on the plate surface in the visual image. Based on the initial position coordinates, the height value of the corresponding position is retrieved from the reconstructed height field. The retrieved height value is used to correct the three-dimensional coordinates of the pixel on the plate surface. The query and correction steps are repeated until the change in the three-dimensional coordinates calculated between two iterations is less than a preset convergence threshold. The converged three-dimensional coordinates are used as the final physical position of the pixel on the plate surface.

[0046] Specifically, the process involves iteratively approximating the two-dimensional pixel coordinates in the visual image to their three-dimensional physical locations on the metal plate surface. By scanning the metal plate surface, discrete height measurement points are obtained, exhibiting a non-uniform spatial distribution that reflects the actual height values ​​at each sampling location on the metal plate surface. Since these discrete measurement points cannot cover all locations on the metal plate surface, two-dimensional spatial interpolation is required to reconstruct the continuous height field distribution on the metal plate surface. During the iterative calculation, initialization is first performed, using perspective projection geometry to calculate the initial coordinates of the plate surface corresponding to each pixel in the visual image.

[0047] The core steps of the iterative process are as follows: Based on the current position coordinates, query the corresponding height value from the reconstructed height field; use the queried height value to correct the 3D coordinates of the pixel on the board surface; and then repeat the query and correction steps. After each iteration, compare the change in 3D coordinates between the two iterations. When this change is less than a preset convergence threshold, the iterative process is considered to have converged. Once the convergence condition is met, the converged 3D coordinates are taken as the final physical position of the pixel on the board surface.

[0048] In this embodiment, iterative 3D geometric reconstruction is used to achieve accurate mapping from visual pixels to the actual physical positions of the metal plate. First, a continuous height field is reconstructed based on discrete height points using 2D interpolation. Then, the projection relationship is initialized with a planar assumption to obtain the initial coordinates of the plate surface corresponding to the pixel. Subsequently, the actual height is queried based on the height field, and the 3D position of the pixel is continuously corrected until the coordinate change meets the convergence threshold, thereby obtaining a stable and accurate pixel-physical coordinate correspondence. This process effectively solves the projection error caused by the curvature and undulation of the plate surface, providing a reliable foundation for the accurate registration of multi-source data on a 3D curved surface.

[0049] Furthermore, the steps for correcting the relative positional offsets of multi-source data include:

[0050] Rolling stripes are extracted from the visual image data of the current scanning area, and the spatial phase distribution of the rolling stripes is calculated using Hilbert transform. The extracted stripe features are then matched with a pre-established multimodal feature fingerprint database to determine the sub-region identifier corresponding to the current area. The same rolling stripe extraction and phase calculation are performed on the laser topography data of the current scanning area. The phase difference between the stripe phases in the visual image data and the stripe phases in the laser topography data is calculated. Based on the known spatial period of the rolling stripes, the relative positional offset between the visual image data and the laser topography data is calculated using the linear relationship between the phase difference and spatial displacement. A translation transformation is performed on the spatial coordinates of the visual image data and the laser topography data to correct the relative positional offset. (See details...) Figure 3 .

[0051] The core of correcting relative position offset lies in utilizing the phase consistency of the common texture in visual and laser topography data to calculate the physical displacement through signal processing.

[0052] The specific steps are as follows: First, the rolling stripes are extracted from the visual image data of the current scanning area, and the extracted rolling stripe signals are processed using Hilbert transform to calculate the spatial phase distribution of the stripes in the visual image data. Next, the extracted stripe features are matched with a multimodal feature fingerprint database to find the corresponding sub-region identifier in the database, thereby determining the baseline parameters. Subsequently, the same operation is performed on the laser topography data of the same scanning area, namely, extracting the rolling stripes and calculating the spatial phase distribution of the stripes in the laser topography data using Hilbert transform. At this time, the system compares the phase values ​​of the visual image data and the laser topography data at the same position and calculates the phase difference between them. Based on the pre-known spatial period of the rolling stripes (i.e., the physical length corresponding to a complete stripe), the calculated phase difference value is converted into a specific physical distance value using the linear proportional relationship between the phase difference and the spatial displacement. This value is the relative position offset. Finally, the system performs a translation transformation operation on the spatial coordinate system of the visual image data or the laser topography data according to the direction and magnitude of the offset, that is, adding or subtracting the calculated offset from the coordinate values ​​of one of the data, thereby achieving precise alignment of the two data at the pixel level and in physical position.

[0053] A signal processing method based on stripe phase consistency is used to achieve high-precision positional registration between visual images and multi-source laser topography data. The core idea is to utilize the periodic texture of rolling stripes, which coexist in both types of data, to obtain their respective spatial phase distributions through Hilbert transform. Then, the visual and laser phases are compared point-by-point to calculate the phase difference, which is converted into a physical displacement based on the known spatial period of the stripes. Combining this with region matching from a multimodal feature fingerprint database ensures that the phase comparison is based on the correct sub-region features, improving registration reliability. Finally, a translation transformation of the coordinate system achieves pixel-level alignment of the two types of data, significantly reducing spatial offsets caused by sensor installation errors, scanning path differences, or surface geometric undulations. Overall, this method effectively improves the accuracy of multi-source data fusion, providing a more stable and reliable feature foundation for subsequent defect identification.

[0054] Furthermore, it also includes a defect early warning step based on the degradation of rolling stripes:

[0055] For the registered metal plate surface, the rolling stripe feature parameters of each sub-region are extracted, including the spatial period, orientation angle, and modulation amplitude of the stripes; the gradient of the rolling stripe feature parameters between adjacent sub-regions is calculated; regions where the gradient of the rolling stripe feature parameters is greater than a preset gradient threshold are identified as feature transition zones; continuity analysis is performed on the feature transition zones to determine whether the rolling stripes in the region exhibit regular degradation, including a gradual increase or decrease in stripe period, a gradual deviation of the stripe direction from the main rolling direction, and a gradual weakening of the stripe modulation amplitude; when the feature transition zone exhibits regular degradation and the degradation range exceeds a preset size threshold, the region is marked as a potential defect warning zone, and a warning signal is output; the warning zone is then subjected to focused scanning and defect verification.

[0056] By monitoring the spatial gradient characteristics of rolling stripes, early warning of defects based on texture degradation is achieved. After multi-source data registration, the stripe period, orientation angle, and modulation amplitude of each sub-region are extracted, and the cross-regional change gradient is calculated to identify characteristic transition zones with abnormal changes. By analyzing whether the stripe period, orientation, and amplitude show a continuous and regular degradation trend, potential structural anomalies can be identified. When the degradation range exceeds a set threshold, the system marks it as a defect warning zone and outputs a warning signal, enabling early location of potential defects and guiding focused scanning, thereby improving the sensitivity and reliability of overall detection.

[0057] Furthermore, the establishment of the defect correlation criterion includes:

[0058] Standard samples containing known defect types, including scratches and indentations, were collected. Multi-source data was acquired from the standard samples. Gray-level contrast, edge gradient, and shape features were extracted from visual image data; depth anomalies, width, edge slope, and roughness variations were extracted from laser morphology data; and impedance variation amplitude and phase shift were extracted from eddy current electromagnetic data. Statistical analysis was performed on the standard sample data for each defect type, calculating the mean, variance, and distribution range of each feature parameter. Response patterns for different defect types were analyzed to determine the correlation patterns between feature parameters. For scratches, the correlation pattern was that visual gray-level contrast, laser depth anomalies, and eddy current impedance variation amplitudes all showed anomalies. For indentations, the correlation pattern was that laser depth anomalies were significant, but the eddy current impedance variation was lower than that caused by scratches. A multi-dimensional feature space for defect types was established, defining reasonable value ranges for feature parameters and constraints between parameters. Anomalies were determined by obtaining the mean, variance, and distribution range of visual gray-level contrast, laser depth anomalies, and eddy current impedance variation amplitudes for the defect area; the 3σ principle was used to determine whether an anomaly was present.

[0059] The response pattern analysis and characteristic parameter correlation pattern determination for defect types are based on the logical construction of different physical sensors detecting different types of surface damage. For scratch defects, the response pattern is as follows: visually, there is a drastic change in grayscale due to light scattering; morphologically, there is material removal or abrupt changes in depth; electromagnetically, the eddy current path is obstructed due to the interruption of metal continuity. Therefore, the determined correlation pattern is: when the visual grayscale contrast is high, the laser depth anomaly value is significant, and the eddy current impedance change amplitude shows a significant anomaly, this combined pattern corresponds to a scratch defect.

[0060] For indentation defects, the response pattern is as follows: although there is obvious plastic deformation leading to depth changes in morphology, and shadows may be produced visually, the changes in conductivity and permeability are small because the metal material has not fractured or been missing. Therefore, the determined correlation mode is: when the laser depth anomaly is significant, but the eddy current impedance change is lower than the impedance change caused by scratches, this combination mode corresponds to indentation defects.

[0061] The parameters extracted from each sensor (including visual edge gradient, shape features, laser width, edge slope, roughness changes, eddy current phase shift, etc.) are logically combined according to physical mechanisms to form a correlation pattern that distinguishes the nature of different defects.

[0062] The establishment of the multidimensional feature space for defect types and the definition of parameter constraints involve defining the range of "true defects" in the data space through statistical analysis of a large number of known samples. Specifically, firstly, standard sample data containing known scratch and indentation defects are collected. For each standard sample, all feature parameters (such as visual grayscale contrast, laser depth anomalies, and eddy current impedance variation amplitude) are extracted. Subsequently, statistical analysis is performed on all sample data of the same defect type (e.g., scratches), calculating the mean, variance, and data distribution range of each feature parameter. Based on these statistics, a specific region is constructed in the multidimensional coordinate system; this region is the multidimensional feature space for that defect type.

[0063] Within this space, reasonable ranges for characteristic parameters are defined (e.g., depth outliers must fall within a certain statistical interval). Simultaneously, constraints between parameters are defined, describing the interrelationships between data from different sensors. For example, for scratches, the constraint is defined as: "visual grayscale contrast" and "eddy current impedance variation amplitude" must be positively correlated, and both must simultaneously exceed their respective baseline statistical limits. Only feature data combinations falling within this multidimensional feature space and satisfying the aforementioned constraints between parameters are considered to be genuine defects of that type.

[0064] In this embodiment, a defect correlation criterion based on physical mechanisms is constructed to achieve multi-source fusion discrimination of typical defects such as scratches and indentations. First, visual, laser, and eddy current multi-source data of standard samples are collected, and key features such as grayscale contrast, depth anomalies, and impedance changes are extracted. Statistical analysis is then performed on each type of defect sample to form the mean, variance, and value range of multi-dimensional features. Based on the material failure mechanisms of different defects, a correlation pattern between features is established: scratches exhibit strong anomalies in visual appearance, morphology, and electromagnetic properties, while indentations mainly show depth changes but have a weaker electromagnetic response. Furthermore, a multi-dimensional feature space for defects is constructed, and parameter constraints are set to describe the linkage between features from different sensors. Overall, this model achieves a structured expression of the essential characteristics of defects, enabling multi-source feature combinations to automatically match defect types based on physical logic, thus improving the accuracy and robustness of defect identification.

[0065] Furthermore, matching multi-source feature data with defect correlation criteria includes:

[0066] For the same registered board position, obtain multi-source feature data; calculate the similarity between the multi-source feature data and the feature space of each defect type in the defect correlation criterion; verify whether the multi-source feature data satisfies the parameter constraint relationship of a certain defect type; when the multi-source feature data conforms to the feature space of the defect type and satisfies the corresponding parameter constraint relationship, the position is determined as the real defect of the defect type; when the feature vector is greater than the first distance threshold with the feature space of all known defect types, and / or does not satisfy the parameter constraint relationship of any defect type (i.e., mismatch), the position is marked as a suspected defect and / or a non-defect feature.

[0067] The process of verifying whether multi-source feature data meets the parameter constraints of a certain defect type is as follows: perform cross-physical field logical verification on the multi-source feature data extracted at the current detection point; obtain the multi-source feature data of the current position to be determined, which includes gray-scale contrast, edge gradient, and shape features extracted from visual images, depth anomalies, width, edge slope, and roughness changes extracted from laser morphology, and impedance change amplitude and phase shift extracted from eddy current electromagnetic data.

[0068] The verification process involves comparing the real-time acquired feature values ​​with the constraint logic defined in the pre-established "defect correlation criterion" item by item; specifically, it checks whether the linkage between features from different sensors conforms to physical laws.

[0069] Verification of scratch type constraints: Analyze the feature vector at this location to determine whether the following conditions are met simultaneously: the grayscale contrast of the visual image data is in the high response range, the depth anomalies in the laser topography data indicate the presence of material loss or deformation, and the impedance change amplitude of the eddy current data shows a significant signal jump; this "strong interaction of all three" is the parameter constraint relationship of the scratch.

[0070] Verification of indentation type constraints: Analyze the eigenvectors at this location to determine if the following differential condition is met: the depth anomalies in the laser topography data indicate significant deformation, but the impedance variation in the eddy current data at the same location is in a low response range or shows no significant change. This specific combination of "strong deformation and weak electromagnetic interference" is the parameter constraint relationship of the indentation.

[0071] The logic for confirming a genuine defect:

[0072] When multi-source feature data is calculated and confirmed to fall within the multi-dimensional feature space pre-statistically derived for a certain type of defect (such as scratches or indentations) (i.e., all feature parameters are within the distribution range defined by the mean and variance), and simultaneously passes the logical verification of the aforementioned parameter constraint relationship, the match is deemed valid. At this point, since the data not only conforms to the sample distribution in terms of statistical values, but also conforms to the response law of multi-source sensors in terms of physical mechanisms, the system ultimately confirms the location as the real defect of the corresponding defect type.

[0073] The logic for marking features as suspected defects or non-defects:

[0074] When multi-source feature data exhibits one or both of the following conditions, an exclusion operation is performed:

[0075] Distance-based exclusion: Calculate the distance between the current multi-source feature data and the feature space center of all known defect types (scratches, indentations, etc.); if the calculated distance values ​​are all greater than the preset first distance threshold, it means that the combination pattern of the current feature data is outside the statistical distribution of known defects and does not belong to any known defect form.

[0076] Elimination based on constraint relationships: Verification revealed that the feature vector did not satisfy the parameter constraint relationship of any defect type; for example, the visual image showed obvious "black spots" (high grayscale contrast), but the laser topography data did not detect any depth anomalies, and the eddy current impedance did not change; in this case, although the visual features resembled defects, the lack of support from deformation and electromagnetic features violated the response pattern of real defects (which may only be surface oil or water stains).

[0077] In this embodiment, a matching mechanism based on multi-source features and defect correlation criteria is constructed to achieve accurate identification of real defects. The system aggregates multi-source features such as visual, laser, and eddy current features at the same location on the board surface. Similarity is calculated between these features and the multi-dimensional feature spaces of each defect type, and further verification is performed to confirm whether they satisfy the corresponding cross-physical field parameter constraints. If a feature vector falls within the feature space range and conforms to the physical linkage logic, it is determined to be a real defect; if it is significantly distant from the feature spaces of all defect types or does not satisfy any constraints, it is marked as a suspected defect or a non-defect. This mechanism integrates statistical distribution and physical response, significantly improving the accuracy and robustness of the determination.

[0078] Furthermore, it also includes an adaptive partitioning step based on the plate deformation field:

[0079] Based on the reconstructed height field distribution on the metal plate surface, the second-order partial derivative of the height field is calculated to obtain the surface curvature at the plate position. A curvature threshold is set, and regions with surface curvature greater than the curvature threshold are marked as high curvature regions, while regions with surface curvature less than the curvature threshold are marked as low curvature regions. The sub-region size of the head transition zone is set for the high curvature region, and the sub-region size of the central stabilization zone is set for the low curvature region.

[0080] In this embodiment, an adaptive partitioning strategy based on the deformation field is introduced to achieve intelligent adjustment of the size of sub-regions. The system calculates the surface curvature based on the reconstructed height field and distinguishes between high and low curvature regions using a curvature threshold, thereby identifying the transitional areas at the beginning and end of the plate surface with drastic deformation and the stable central region with a smooth morphology. Smaller sub-regions are used for high curvature regions to improve feature capture accuracy, while larger sub-regions are used for low curvature regions to improve processing efficiency, thus improving the overall adaptability and accuracy of the detection.

[0081] This invention achieves accurate detection of surface defects in metal plates by fusing multi-source sensor data, including visual, laser, and eddy current sensors. The method first divides the metal plate into functional regions based on the characteristics of the cold rolling process and the curvature of the plate surface, and adaptively sets the size of each sub-region. Then, it extracts features such as stripe direction angle, spatial period, and undulation amplitude through two-dimensional and one-dimensional Fourier analysis to construct a multi-modal feature fingerprint database. Subsequently, it establishes a mapping from pixels to the physical position on the plate surface based on three-dimensional height field reconstruction, and calculates the relative offset between visual and laser topography data by combining stripe phase consistency, achieving accurate registration of multi-source data on the three-dimensional curved surface. On this basis, it matches multi-source features with defect correlation criteria to achieve reliable identification of real defects such as scratches and indentations, and can provide early warning of defects by monitoring stripe gradient changes. Overall, this method achieves accurate modeling of plate surface texture, high-confidence data fusion, and highly reliable defect discrimination, significantly improving detection accuracy and robustness.

[0082] Example 2:

[0083] This invention provides a method for detecting surface defects in metal plates by fusing multi-source sensor data. Based on Embodiment 1, it further includes a multi-scan timing consistency verification step and a confidence-weighted step based on defect propagation risk. The technical solution is as follows:

[0084] Steps for verifying timing consistency across multiple scans:

[0085] When a location is initially identified as a suspected defect, its coordinates and multi-source feature data are recorded as the first measurement result. The sensor is then controlled to return to the location, and a second scan is performed after adjusting at least one detection parameter. These detection parameters include the illumination angle of the visual sensor, the incident angle of the laser sensor, or the operating frequency of the eddy current sensor. The multi-source feature data from the second scan is extracted, and the feature similarity between the two measurement results is calculated. If the feature similarity is greater than a preset similarity threshold, the defect feature at that location is determined to have temporal stability and is confirmed as a genuine defect. If the feature similarity is less than the preset similarity threshold, or if the defect feature disappears during the second scan, the location is determined to have temporary interference and is marked as a non-defect feature. For boundary cases that are difficult to determine, a third scan is performed, and a final determination is made based on the statistical consistency of the three measurement results.

[0086] Specifically, the process involves repeatedly scanning suspected defect locations under different detection conditions. By leveraging the temporal stability of real defect features and the randomness of temporary interference features, a reliable determination of suspected defects is achieved. When a location is initially identified as a suspected defect, its coordinate information and multi-source feature data are recorded as the first measurement result. The coordinate information includes the lateral and longitudinal positions of the location in the metal plate surface coordinate system. The multi-source feature data includes grayscale contrast, edge gradient, and shape features extracted from visual image data; depth anomalies, width, edge slope, and roughness changes extracted from laser topography data; and impedance change amplitude and phase shift extracted from eddy current electromagnetic data. The coordinate information and multi-source feature data are then associated and stored to establish the first measurement file for the suspected defect location. Further, the process of controlling the sensor system to return to the location for a second scan includes: the system controlling the metal plate transmission mechanism or sensor movement mechanism based on the recorded coordinate information to reposition the sensor array directly above the suspected defect location; before performing the second scan, at least one detection parameter is adjusted to change the detection conditions. The adjustment of the detection parameter includes at least one of the following three methods:

[0087] Adjusting the illumination angle of the vision sensor: By controlling the illumination direction of the ring light source or strip light source, the angle between the incident light and the surface of the metal plate is changed, thereby changing the surface reflection characteristics. The purpose of this adjustment is to distinguish between real defects and surface oil or water stains, because the geometry of real defects will still produce a stable grayscale response under different illumination angles, while the reflection characteristics of oil or water stains will change significantly with the change of illumination angle.

[0088] Adjusting the incident angle of the laser sensor: By changing the tilt angle of the scanning head of the laser displacement sensor, the laser beam is incident on the surface of the metal plate at different angles. The purpose of this adjustment is to verify the authenticity of the depth anomaly value, because the depth value measured by real morphological defects at different incident angles has geometric consistency, while the false depth response caused by the surface oxide layer or dust particles will have a significant deviation as the incident angle changes.

[0089] Adjusting the operating frequency of the eddy current sensor: By changing the operating frequency of the eddy current excitation signal, the skin depth and sensitivity characteristics of the eddy current are adjusted. The purpose of this adjustment is to distinguish between surface defects and subsurface defects, and at the same time to verify the stability of the impedance response. This is because real material damage will produce impedance changes that conform to physical laws at different operating frequencies, while false responses caused by electromagnetic interference or sensor noise do not have frequency-related physical consistency.

[0090] After adjusting the detection parameters, a second scan is performed to collect visual image data, laser topography data, and eddy current electromagnetic data at the location. Multi-source feature data from the second scan is extracted using the same method as the first scan, including grayscale features, depth features, and impedance features. Further, the process of calculating the feature similarity between the two measurements includes: constructing feature vectors from the multi-source feature data of the first and second measurements, each containing the values ​​of its respective feature parameters; normalizing the two feature vectors to eliminate the influence of differences in the dimensions of different feature parameters; calculating the similarity between the two normalized feature vectors using the cosine of the vector angle or a weighted Euclidean distance to characterize the closeness of the two measurement results in the multi-dimensional feature space; when the feature similarity is greater than a preset similarity threshold, it indicates that the location is close to the target location. Under different detection conditions, the system exhibits stable defect feature responses, consistent with the physical characteristics of real defects. The system determines that the defect feature at this location has temporal stability, confirms it as a real defect, and outputs the location and its defect type to the detection result database. When the feature similarity is less than the preset similarity threshold, it indicates that the feature response at this location changes significantly with the detection conditions, does not conform to the stability characteristics of real defects, and the system determines that there is temporary interference at this location, marks it as a non-defect feature, and removes the location from the suspected defect list. When the defect feature completely disappears in the second scan, that is, the multi-source feature data of the second measurement are all within the normal range, the system also determines that there is temporary interference at this location and marks it as a non-defect feature. For boundary cases that are difficult to determine, that is, when the feature similarity is close to the preset similarity threshold, or when the two measurement results have some features that are consistent and some features that are different, a third scan is performed.

[0091] During the third scan, the detection parameters are adjusted again to make the detection conditions different from the previous two scans. Multi-source feature data of the third measurement are collected. The process of making a final judgment based on the statistical consistency of the three measurement results includes: calculating the mean and standard deviation of the multi-source feature data of the three measurements, and analyzing the degree of fluctuation of each feature parameter in the three measurements; when the standard deviation of a certain feature parameter is less than the preset fluctuation threshold, the feature is determined to have statistical stability; when most feature parameters have statistical stability and the average feature vector of the three measurements falls within the feature space range of the defect correlation criterion, it is finally confirmed as a real defect; otherwise, it is finally marked as a non-defect feature.

[0092] In this embodiment, a multi-scan temporal consistency verification mechanism is introduced to effectively solve the problem of false responses that may occur during a single scan. This mechanism leverages the fundamental difference between the stability of the characteristic responses of genuine defects under different detection conditions and the randomness of the characteristic responses of temporary interference factors. By repeatedly verifying the response by changing detection parameters such as illumination angle, laser incident angle, or eddy current operating frequency, it ensures that only defect features with temporal stability can be identified as genuine defects. For boundary cases, statistical consistency analysis of three measurements is used to further improve the reliability of the judgment. This method significantly reduces the false judgment rate caused by environmental interference, surface contamination, or sensor noise, and improves the stability and reliability of the detection system in actual production environments.

[0093] This embodiment introduces a confidence-weighted mechanism based on defect propagation risk. It combines the geometric characteristics of defects with subsequent stamping process information to calculate the propagation risk coefficient for each defect, and adjusts the confidence level of defect assessment and processing strategy accordingly. The specific technical solution is as follows: Confidence-weighted steps based on defect propagation risk:

[0094] The system acquires subsequent processing information for the metal sheet to be inspected, including the forming direction of the stamped part, the expected strain distribution, and the location of high-strain regions. For each detected defect, its geometric features are extracted, including the long axis direction and the coordinates of the defect's location. The angle between the long axis direction of the defect and the direction of the principal stamping stress is calculated. It is determined whether the defect is located in a high-strain region. Based on the angle and location information, the system calculates the defect propagation risk coefficient. When the angle between the long axis direction of the defect and the principal stress direction is less than a preset angle threshold and the defect is located in a high-strain region, a high-risk coefficient is assigned. When the long axis direction of the defect is nearly perpendicular to the principal stress direction or the defect is located in a low-strain region, a low-risk coefficient is assigned. The propagation risk coefficient is used as a weighting factor to adjust the confidence level of defect judgment. The judgment threshold is lowered for high-risk defects to improve sensitivity, and the judgment threshold is raised for low-risk defects to reduce over-rejection. A test report containing risk classification is output to provide a basis for sorting decisions.

[0095] The process of obtaining subsequent processing information for the metal sheet to be inspected includes: reading the application information and corresponding stamping die parameters of the batch of metal sheets from the production management system or process database; determining the forming direction of the stamped part based on the geometry of the stamping die and the forming process, wherein the forming direction is the main direction of material flow during the stamping process; obtaining the expected strain distribution of the stamped part during the forming process based on stamping process simulation data or experience database, wherein the expected strain distribution characterizes the degree of deformation of each region of the metal sheet after stamping; identifying the location of high strain regions based on the expected strain distribution, wherein the high strain regions are regions where the expected strain value exceeds the set strain level, typically corresponding to the fillet transition area, deep drawing area, or flange area of ​​the stamped part.

[0096] For each detected defect, the process of extracting the geometric features of the defect includes: determining the major axis direction of the defect based on the boundary contour of the defect region using principal component analysis or ellipse fitting, wherein the major axis direction is the main direction of defect extension; and obtaining the coordinates of the location of the defect, wherein the coordinates are the position of the geometric center of the defect in the coordinate system of the metal plate surface.

[0097] The process of calculating the angle between the long axis direction of the defect and the direction of the principal stamping stress includes: converting the forming direction of the stamped part into the principal stamping stress direction vector in the coordinate system of the metal plate surface; representing the long axis direction of the defect as a unit vector; and calculating the angle between the long axis direction vector of the defect and the direction vector of the principal stamping stress, wherein the value of the angle ranges from zero degrees to ninety degrees.

[0098] The process of determining whether a defect is located in a high-strain region includes: matching the coordinates of the defect with the expected strain distribution map; querying the expected strain value corresponding to the coordinates; comparing the expected strain value with the threshold for determining a high-strain region; determining that the defect is located in a high-strain region when the expected strain value is greater than the threshold; and determining that the defect is located in a low-strain region when the expected strain value is less than or equal to the threshold.

[0099] The process of calculating the propagation risk coefficient of a defect based on angle and location information includes: establishing a propagation risk assessment model, using angle and location factors as input variables. When the angle between the defect's long axis and the main stamping stress direction is less than a preset angle threshold, it indicates that the defect's extension direction is nearly parallel to the stress direction. During stamping, the defect tip will bear a large opening stress and has a high tendency to propagate, thus assigning a high directional risk weight. When the defect's long axis is nearly perpendicular to the main stamping stress direction, the opening stress at the defect tip is small, and the tendency to propagate is low, assigning a low directional risk weight. When the defect is located in a high-strain region, this region will undergo large plastic deformation during stamping, and the driving force for defect propagation is strong, assigning a high location risk weight. When the defect is located in a low-strain region, the degree of plastic deformation is small, and the driving force for defect propagation is weak, assigning a low location risk weight. The directional risk weight and location risk weight are combined to obtain the defect propagation risk coefficient. Specifically, when the angle between the long axis of the defect and the principal stress direction is less than a preset angle threshold and the defect is located in a high strain region, the two risk factors are superimposed, and the system assigns a high risk coefficient to the defect, indicating that the defect has a high risk of propagation and expansion during the subsequent stamping process; when the long axis of the defect is nearly perpendicular to the principal stress direction, or the defect is located in a low strain region, at least one risk factor is at a low level, and the system assigns a low risk coefficient to the defect, indicating that the risk of propagation of the defect during the subsequent stamping process is controllable.

[0100] The process of adjusting the defect judgment confidence by using the propagation risk coefficient as a weighting factor includes: for defects assigned a high risk coefficient, the system lowers the judgment threshold for such defects, making the feature responses that were originally in a boundary state more likely to be judged as defects, thereby improving the detection sensitivity of high-risk defects and avoiding missed detection of high-risk defects due to excessively high detection thresholds; for defects assigned a low risk coefficient, the system raises the judgment threshold for such defects, so that slight feature responses are not judged as defects that need to be processed, thereby reducing the over-discarding of low-risk defects and reducing unnecessary material scrap costs.

[0101] The process of outputting an inspection report containing risk classification includes: summarizing all defect information detected on the current metal plate, including defect location, defect type, defect size, and propagation risk coefficient; classifying defects into high-risk, medium-risk, and low-risk levels based on the propagation risk coefficient; generating an inspection report, which includes a defect distribution map, detailed characteristic parameters of each defect, risk level labeling, and processing recommendations; and outputting the inspection report to the production management system to provide a basis for subsequent sorting decisions, allowing high-risk defective plates to be prioritized for manual re-inspection or downgrade processing, while low-risk defective plates can be selectively released according to actual quality requirements.

[0102] In this embodiment, a confidence-weighted mechanism based on defect propagation risk is introduced to achieve intelligent correlation between defect detection and subsequent processing technology. This mechanism comprehensively considers the angle between the defect's long axis and the main stamping stress direction, as well as the expected strain level at the defect's location, establishing a quantitative assessment model for defect propagation risk. By employing a differentiated strategy of lowering the judgment threshold for high-risk defects and raising the judgment threshold for low-risk defects, the excessive rejection rate is effectively reduced while ensuring that no critical defects are missed. This achieves a balance between detection sensitivity and economic benefits, and the output risk-level detection report provides a scientific basis for the production line's sorting decisions, improving the accuracy and intelligence of overall quality control.

[0103] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting surface defects of a metal sheet by fusing multi-source sensing data, characterized by, The method comprises the following steps: Collecting multi-source data of the surface of the metal sheet to be inspected; Extracting characteristic parameters of rolling stripes on the surface of the metal sheet; Dividing the surface of the metal sheet into sub-regions and establishing a multi-modal characteristic fingerprint library containing the characteristic parameters of the rolling stripes for each sub-region; The multi-source data comprises visual image data and laser topography data; the characteristic parameters comprise a stripe direction angle and a spatial period extracted from the visual image data, and a surface fluctuation period and amplitude extracted from the laser topography data; The step of dividing the surface of the metal sheet into sub-regions comprises: performing two-dimensional Fourier transform on the visual image data, identifying a dominant spatial frequency peak corresponding to the rolling stripes in the frequency domain, and obtaining the direction angle and the spatial period of the stripes through inverse transform and peak positioning; extracting a surface profile curve along a direction perpendicular to the rolling direction for the laser topography data, and extracting periodic components and amplitudes of surface fluctuations through one-dimensional Fourier analysis; verifying the deviation of the rolling stripe direction angle and the deviation of the spatial period extracted from the visual image data and the laser topography data, and determining a region meeting the condition as an effective registration region; dividing the surface of the metal sheet into a head transition zone, a tail transition zone, an edge region and a center stable zone according to the surface quality distribution law of the effective registration region in the cold rolling process; setting different sub-region sizes for different functional regions, extracting and storing the direction angle, the spatial period and the fluctuation amplitude of the rolling stripes for each sub-region, and establishing a multi-modal characteristic fingerprint of the sub-region; Reconstructing a height field distribution of the surface of the metal sheet according to real-time measured three-dimensional topography data of the metal sheet; projecting the multi-source data onto the reconstructed three-dimensional surface to establish a corresponding relationship between pixels and physical positions of the sheet surface; extracting rolling stripe characteristics of a current scanning region and matching them with the multi-modal characteristic fingerprint library respectively, and correcting the relative position offset of the multi-source data by calculating the relative difference of the stripe phase; Collecting eddy current electromagnetic data at the same position after registration, extracting impedance characteristics, and combining the gray scale characteristics and depth characteristics of the multi-source data in the defect region to obtain multi-source characteristic data; Establishing a defect correlation criterion to obtain a characteristic parameter response mode of the defect type; matching the multi-source characteristic data with the defect correlation criterion, and confirming a real defect when the combined mode of the characteristic data is consistent with the response mode of the defect type; marking as a suspected defect and / or a non-defect feature when there is no match. The establishment of the defect correlation criterion comprises: collecting standard samples containing known defect types, the defect types including scratch defects and indentation defects; performing multi-source data acquisition on the standard samples, extracting gray scale features from visual image data, including gray scale contrast, edge gradient and shape features, extracting depth features from laser topography data, including depth outliers, width, edge slope and roughness variation, and extracting impedance features from eddy current electromagnetic data, including impedance variation amplitude and phase shift; performing statistical analysis on the standard sample data of each defect type, calculating the mean, variance and distribution range of each feature parameter; analyzing the response modes of different defect types to determine the correlation mode between each feature parameter, wherein the correlation mode of scratch defects is that the visual gray scale contrast, laser depth outlier and eddy current impedance variation amplitude all show abnormalities, and the correlation mode of indentation defects is that the laser depth outlier is significant but the eddy current impedance variation is lower than that caused by scratches; establishing a multi-dimensional feature space of defect types, defining reasonable value ranges of feature parameters and constraint relationships between parameters; Matching the multi-source feature data with the defect correlation criterion comprises: For the same plate surface position after registration, multi-source feature data is obtained; the similarity of the multi-source feature data to each defect type feature space in the defect correlation criterion is calculated; it is verified whether the multi-source feature data meets the parameter constraint relationship of a certain defect type; when the multi-source feature data meets the feature space of the defect type and meets the corresponding parameter constraint relationship, the position is determined as a real defect of the defect type; when the feature vector distance from the feature space of the known defect type is greater than the first distance threshold, and / or does not meet the parameter constraint relationship of any defect type, the position is marked as a suspected defect and / or a non-defect feature.

2. The method of claim 1, wherein: The establishment of the correspondence between pixels and physical positions of the plate surface comprises: obtaining discrete distributed height measurement point data by scanning the surface of the metal plate; performing two-dimensional space interpolation on the height measurement point data to reconstruct the continuous height field distribution of the surface of the metal plate; initializing the iteration process, calculating the initial position coordinates of each pixel in the visual image corresponding to the plate surface by using perspective projection geometric relationship; querying the height value of the corresponding position from the reconstructed height field according to the initial position coordinates; correcting the three-dimensional coordinates of the plate surface corresponding to the pixel by using the queried height value; repeating the querying and correcting steps until the three-dimensional coordinates calculated by the previous two iterations change by less than a preset convergence threshold; taking the converged three-dimensional coordinates as the final corresponding physical positions of the plate surface.

3. The method of claim 1, wherein: The step of correcting the relative position offset of the multi-source data comprises: performing mill stripe extraction on the visual image data of the current scanning area, calculating the spatial phase distribution of the mill stripe by Hilbert transform; performing similarity matching of the extracted stripe features with a pre-established multi-modal feature fingerprint library to determine the sub-area identifier corresponding to the current area; performing the same mill stripe extraction and phase calculation on the laser topography data of the current scanning area; calculating the phase difference between the stripe phase in the visual image data and the stripe phase in the laser topography data; calculating the relative position offset between the visual image data and the laser topography data by using the linear relationship between the phase difference and the spatial displacement according to the known spatial period of the mill stripe; and performing translational transformation on the spatial coordinates of the visual image data and the laser topography data to correct the relative position offset.

4. The method of claim 1, wherein: The adaptive partitioning step based on the plate deformation field comprises: calculating the second-order partial derivative of the height field to obtain the surface curvature of the plate position according to the reconstructed metal plate surface height field distribution; setting a curvature threshold, marking the area with a surface curvature greater than the curvature threshold as a high-curvature area, and marking the area with a surface curvature less than the curvature threshold as a low-curvature area; setting the sub-area size of the head transition zone for the high-curvature area; and setting the sub-area size of the center stable zone for the low-curvature area.

Images