A visual deviation correction tool verification method and device

Abstract

This invention relates to a visual correction method and apparatus for tooling verification, primarily addressing the problem of inefficient identification and real-time adaptive compensation of tooling positioning deviations under heterogeneous working conditions. The core solution includes: pre-setting geometrically robust physical markers on the surfaces of various tooling components; acquiring reference images under multiple lighting conditions, materials, and orientations; and extracting the topological relationships of the markers to generate rotation- and scale-invariant semantic fingerprints. Based on these fingerprints, a deviation pattern map indexed by inherent deviation features is established, enabling incremental learning and dynamic expansion for new working conditions. In the online stage, high-precision, low-latency real-time deviation correction is achieved through sub-pixel-level positioning, rapid fingerprint matching, and piecewise linear function compensation. This method improves the adaptability of the detection system to complex environmental changes and enhances position calibration accuracy, effectively supporting the intelligent and self-evolving upgrades of industrial production lines.

Description

Technical Field

[0001] This invention relates to the field of machine vision tooling verification and dynamic self-calibration technology, and in particular to a tooling verification method and apparatus for visual correction. Background Technology

[0002] Currently, visual alignment and tooling verification systems are widely used in automated assembly, precision manufacturing, and other fields. Mainstream visual alignment technologies in the industry typically rely on deep learning feature extraction, template matching, external environment perception sensors, environmental simulation data augmentation, or traditional feature point tracking to achieve workpiece / tooling positioning and deviation correction. In actual production line deployments, considering adaptability, accuracy, and consistency, commonly used technical approaches include: Multi-condition feature extraction based on neural networks improves robustness to complex lighting, material and pose changes through end-to-end model training. However, the data required for training and upgrading is extremely complex, resulting in heavy burdens in terms of memory resources, inference time and model switching.

[0003] Additional environmental information can be collected using external sensors (such as temperature, humidity, ambient light intensity, reflectivity, etc.) to assist in the dynamic adjustment of the parameters of the decision vision module, or a multi-model switching / fusion scheme can be adopted to switch vision models between scenes based on the transmission sensor signals. However, this approach involves high hardware investment, complex system integration, and is susceptible to environmental interference.

[0004] This method expands the model's applicability by simulating various working environments using a large-scale simulation image library, diverse data augmentation, and noise injection. However, this approach is highly dependent on computational resources, lacks real-world adaptability to specific tooling and production lines, and has limited generalization effectiveness.

[0005] Alignment methods based on process feature points / standard structural parts templates are simple to implement, but they lack feature expression power when dealing with non-standard parts, new models, reflective materials, or variable working conditions, making it difficult to meet the requirements of high consistency and high precision.

[0006] Applicable to general tooling verification and assembly processes, this type of technology can often achieve good positioning and correction in typical single environments, such as the alignment of automotive parts, the assembly and adjustment of intelligent equipment, and the precision inspection and processing of electronic components.

[0007] However, existing technologies generally suffer from the following major technical defects under various operating conditions or actual production line environments: The model has insufficient generalization ability. When the light intensity, tooling material, posture angle or background environment change, the existing deviation analysis module is difficult to accurately express the actual geometric characteristics of the tooling body under complex working conditions. The model needs to be frequently switched or incrementally trained for different working conditions, which makes it difficult to guarantee consistency and robustness across working conditions.

[0008] The system relies on additional hardware or redundant sensing paths to improve multi-condition adaptability, which increases system integration and maintenance costs, while also bringing potential stability risks and affecting the deployment density and scalability of large-scale flexible production lines.

[0009] Traditional methods rely heavily on pixel-level or texture information and are easily affected by surface changes such as reflection, wear, and contamination, leading to unstable feature point detection. In complex working conditions or when expanding to new models, model retraining and parameter adjustment are extremely costly and difficult to adapt quickly.

[0010] While existing simulation and noise injection solutions can alleviate some of the problem of insufficient data coverage, they rely heavily on complex modeling of preset working conditions, human experience, and the process itself. They lack autonomous and interpretable expression of the geometric features of the tooling itself, making it difficult to support the needs of self-evolving and self-growing production lines.

[0011] The actual production line is highly flexible, with numerous assembly conditions and tooling models. Adding a new tooling usually requires several hours to several days for model acquisition, training, and verification, which greatly affects the efficiency of production line switching and expansion.

[0012] Current technology has not yet developed a generalized deviation analysis and online self-calibration system that can be built based on the tooling itself, without additional hardware investment, by fully utilizing the coupling relationship between its own geometric features and historical deviation behavior, and possessing high robustness, high speed, and low storage requirements.

[0013] With the increasing demand for intelligent manufacturing, flexible high-end equipment production lines, and multi-condition assembly, the market demand for visual correction systems that offer "rapid generalization, zero hardware upgrade costs, seamless production line expansion, and long-term online self-evolution" is growing stronger. There is an urgent need for an innovative self-calibration method that can eliminate systematic errors in deviation analysis modules under multiple conditions, break the strong dependence of traditional technologies on environment, materials, or models, and significantly improve the model's cross-condition generalization ability and assembly consistency. To this end, this invention proposes a lightweight online self-calibration scheme based on tooling semantic fingerprints and deviation pattern maps. Through feature extraction dominated by tooling geometric attributes and a systematic deviation compensation strategy, it achieves high adaptability, rapid response, and dynamic iterative optimization, solving the technical bottlenecks of existing solutions such as weak generalization ability, insufficient system stability, and high production line maintenance costs. Summary of the Invention

[0014] This application provides a tooling verification method and apparatus for visual correction, which aims to solve one of the problems or problems of the prior art mentioned in the background art.

[0015] This application provides a tooling verification method for visual correction, specifically including: S1: For all tooling models involved in the target production line, preset geometrically robust physical marker points on the surface of each tooling and collect reference images to obtain a multi-condition reference image dataset. S2: Based on the multi-condition benchmark image dataset, extract the relative topological relationship between each physical marker point and encode it into a geometric invariant vector that does not include the original pixel information, thereby constructing the tooling semantic fingerprint; S3: Input the multi-condition reference image dataset into the deviation analysis module to obtain the position deviation result and compare it with the measured true value. Calculate the inherent deviation features including deviation direction preference, amplitude saturation range and response delay characteristics, and then construct a deviation pattern map with tooling semantic fingerprint as the index key. S4: During the online operation phase, acquire the real-time image of the current tooling and perform sub-pixel level marker point localization and topological relationship encoding. Match the corresponding tooling semantic fingerprint in the deviation pattern map by hash comparison to lock the current working condition label. S5: Load the corresponding low-order piecewise linear mapping function cluster from the deviation pattern map according to the successfully matched current working condition label, and input the original position deviation output by the deviation analysis module into the mapping function cluster for real-time interpolation correction to obtain the calibrated position deviation; S6: Determine whether the tooling semantic fingerprint under the current working condition exists in the deviation pattern map. If it does not exist, package the tooling semantic fingerprint and the original position deviation of multiple consecutive frames into a new working condition data packet. S7: Upload the new working condition data packet to the edge training node and automatically expand the deviation pattern map during the offline period, only updating the local mapping function to complete the dynamic adaptive update of the deviation pattern map; S8: Based on the updated deviation pattern map, semantic fingerprint matching and deviation compensation mapping are re-executed on the subsequently acquired real-time images to complete the verification of the tooling.

[0016] This application also provides a tooling calibration device for visual correction, specifically including a device employing a tooling calibration method for visual correction as described above.

[0017] The tooling verification method for visual correction provided in this application has the following beneficial effects: (1) To address the problems of weak deviation modeling ability, poor environmental adaptability, and reliance on high-computing-power deep learning models in existing industrial vision positioning systems under complex working conditions, this solution constructs a "deviance pattern map" that couples the "semantic fingerprint" based on the geometric invariance of the tooling ontology with historical deviation behavior, thereby achieving accurate characterization and rapid compensation of position deviation features. Traditional methods usually rely on texture information or end-to-end neural networks for attitude estimation, which are easily affected by factors such as changes in illumination, surface wear, and background interference, leading to frequent positioning drift or recalibration. However, this solution abandons the reliance on texture and reflective properties and instead utilizes the spatial configuration invariance of geometrically robust markers (such as micro-engraved cross grooves, coating color difference rings, etc.) pre-set on the tooling to extract a compact vector containing only geometrically invariant variables such as angle, scale, and connectivity order as a unique identifier, ensuring stable recognition and matching under multi-transformation imaging conditions. This semantic fingerprint does not carry the original pixel data and has scale, rotation and illumination invariance, which significantly improves the reliability and generalization ability of feature representation. It effectively overcomes the risk of mismatch caused by fluctuations in imaging parameters (such as exposure, gain and white balance shifts) and provides a high-confidence index basis for subsequent deviation compensation.

[0018] (2) Furthermore, this scheme innovatively models the inherent deviation characteristics of each tooling under different working conditions—including deviation direction preference, amplitude saturation range, and response delay—into a “deviation pattern map” with semantic fingerprints as keys and working condition labels as indexes, and stores a lightweight piecewise linear compensation function cluster, thereby achieving real-time dynamic correction without model retraining. Compared with traditional compensation mechanisms that rely on online deep inference or external sensor fusion, the lightweight semantic fingerprint extraction unit in this scheme only needs to perform sub-pixel-level localization and topology coding, and the hash comparison time is less than 0.8ms, which greatly reduces the computational overhead and avoids introducing additional inference delay; after a successful match, the corresponding mapping function can be called immediately to complete the interpolation correction and output the calibrated position deviation result, ensuring the real-time performance and deterministic response of the system. More importantly, when encountering new working conditions not covered by the graph, the system automatically triggers an incremental learning mechanism, packaging and uploading the semantic fingerprints, original biases, and manually verified ground truth values ​​of five consecutive frames to the edge training node. This completes local graph expansion and function updates during offline periods, ensuring online service continuity and achieving closed-loop adaptive evolution capabilities. This significantly improves the system's flexibility in responding to scenarios such as frequent model changes and sudden disturbances in flexible production lines.

[0019] The aforementioned technologies collectively construct a lightweight, interpretable, and self-evolving visual compensation system inherent in the analysis of tooling characteristics and historical behavior. This system requires no additional sensors, does not alter existing image acquisition links, does not increase the load on deep networks, and does not require building digital twins or conducting lighting simulations. All enhancements stem from the coupled modeling of the topological relationships and deviation patterns of physical markers. Consequently, the system not only possesses the high-density deployment capability of supporting over 2000 tooling maps on a single edge device, but also supports the completion of full-condition map construction for newly added tooling within 2 hours, significantly shortening the deployment cycle. This effectively meets the comprehensive requirements of modern intelligent manufacturing for high precision, fast response, strong robustness, and low maintenance costs, demonstrating broad applicability in high-end manufacturing fields such as automotive assembly, aerospace component docking, and precision electronic assembly. Attached Figure Description

[0020] Figure 1 This is the main flowchart of a tooling verification method for visual correction.

[0021] Figure 2 This is a sub-flowchart of a tooling verification method for visual correction.

[0022] Figure 3 This is another sub-flowchart of a tooling verification method for visual correction. Detailed Implementation

[0023] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0024] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0025] like Figure 1 As shown, this application provides a tooling verification method for visual correction, specifically including: S1: For all tooling models involved in the target production line, preset geometrically robust physical marker points on the surface of each tooling and collect reference images to obtain a multi-condition reference image dataset. S2: Based on the multi-condition benchmark image dataset, extract the relative topological relationship between each physical marker point and encode it into a geometric invariant vector that does not include the original pixel information, thereby constructing the tooling semantic fingerprint; S3: Input the multi-condition reference image dataset into the deviation analysis module to obtain the position deviation result and compare it with the measured true value. Calculate the inherent deviation features including deviation direction preference, amplitude saturation range and response delay characteristics, and then construct a deviation pattern map with tooling semantic fingerprint as the index key. S4: During the online operation phase, acquire the real-time image of the current tooling and perform sub-pixel level marker point localization and topological relationship encoding. Match the corresponding tooling semantic fingerprint in the deviation pattern map by hash comparison to lock the current working condition label. S5: Load the corresponding low-order piecewise linear mapping function cluster from the deviation pattern map according to the successfully matched current working condition label, and input the original position deviation output by the deviation analysis module into the mapping function cluster for real-time interpolation correction to obtain the calibrated position deviation; S6: Determine whether the tooling semantic fingerprint under the current working condition exists in the deviation pattern map. If it does not exist, package the tooling semantic fingerprint and the original position deviation of multiple consecutive frames into a new working condition data packet. S7: Upload the new working condition data packet to the edge training node and automatically expand the deviation pattern map during the offline period, only updating the local mapping function to complete the dynamic adaptive update of the deviation pattern map; S8: Based on the updated deviation pattern map, semantic fingerprint matching and deviation compensation mapping are re-executed on the subsequently acquired real-time images to complete the verification of the tooling.

[0026] Step S1: For all tooling models involved in the target production line, pre-set geometrically robust physical marker points on the surface of each tooling, and acquire reference images to obtain a multi-condition reference image dataset. Specifically, this includes: S1.1: For all tooling models involved in the target production line, physical marker points are preset on the surface of each tooling based on the principle of geometric invariance to obtain a physical marker point array with spatial configuration invariance. This step takes the tooling body as the input object and uses geometric feature definition technology such as micro-engraved cross grooves or precision hole array edges to output a physical marker point array with anti-lighting interference and anti-texture loss characteristics, which serves as a reference for subsequent image acquisition.

[0027] For all tooling models involved in the target production line, the 3D geometric structure data of the tooling body is used as input, and a high-precision surface scanning unit is called to obtain a complete surface point cloud matrix. A surface coordinate system is established based on this surface point cloud matrix, and geometric feature saliency analysis is performed in this coordinate system to screen curvature extrema, normal vector abrupt change points, and local plane intersection points as candidate marker locations. For the candidate marker location set, a cross-shaped groove is etched at each selected location using a micro-engraving method, with the etching depth, width, and groove arm length meeting preset geometric robustness standards. The groove arm length is set according to the following proportional relationship: Where L is the groove arm length and d is the etching depth, to ensure that the edge response of the groove shape remains constant when the incident light angle changes. For areas with high reflectivity on the material surface, a micro-ring coating with high saturation and low reflectivity is deposited at selected locations using a coating color difference ring technique. The formula is optimized by combining the ring width w and the color difference ΔC: Where R is the surface reflectance coefficient and ΔC is the color difference value, ensuring that the color ring boundary can be stably identified under multiple illumination intensity gradients. For the marked positions at the edges of the precision aperture array, aperture edge enhancement processing is performed, which improves the edge gradient by etching micro-chamfers. The chamfer radius r satisfies: Here, 'a' represents the aperture, ensuring that the topological connectivity remains consistent regardless of changes in imaging scale. All processed marker positions are indexed according to a surface coordinate system, forming a physical marker array with an invariant spatial configuration. After array formation, a geometric consistency check is performed to verify whether the relative distances and angles of each marker point in the 3D coordinate system meet preset error tolerances. A qualified array is output upon successful geometric consistency check. Through physical processing and surface feature enhancement, the geometry of the tooling body from the previous step is transformed into a physical marker array with resistance to illumination interference and texture loss, achieving a stable and repeatable reference object construction effect.

[0028] For example, on a smart manufacturing production line, the surface of tooling with model A103 is marked. The surface scanning resolution is 0.01mm, and 20 extreme curvature points are selected as candidate marking positions. The depth d of the etched cross groove is set to 0.5mm, and the groove arm length L is calculated to be 1.5mm according to the formula. A color difference ring with a width w=0.8mm is deposited in the polished area with a reflectivity R of 0.8. ΔC=35 is obtained through the color difference optimization formula to ensure stable recognition under high reflectivity conditions. A chamfer with a radius r=1mm is applied to the edge of the precision hole array with an aperture a of 5mm to maintain the topology. After processing, a surface coordinate system index is established and geometric consistency verification is performed. The relative distance error is controlled within 0.02mm and the included angle error is controlled within 0.1°. The final output physical marker array maintains stable recognition under laboratory multi-light intensity tests (300lx-1200lx), different material surfaces (aluminum, steel, composite materials), and multiple installation postures, becoming a reliable reference for subsequent acquisition of benchmark images.

[0029] S1.2: Based on the physical marker array, a standardized image acquisition operation is performed under a multi-dimensional working condition combination covering five types of illumination intensity gradients, four main material surfaces, three typical installation posture angles, and two types of background interference to obtain the original working condition image sequence. This step uses the physical marker array and the multi-dimensional working condition combination as input conditions, and adopts a synchronous trigger control method to coordinate the imaging device and environmental variables, and outputs an original working condition image sequence that includes rich environmental change features to ensure the completeness of data coverage.

[0030] The physical marker array generated in step S1.1, combined with preset multi-dimensional working conditions, is used as input to load working condition variables and configure synchronous triggering on the environmental interaction interface of the imaging device. This ensures that illumination, material, installation posture, and background interference are controllable and adjustable during the acquisition process. For five types of illumination intensity gradients, two sub-states—constant illuminance and dynamic illuminance—are set. A PWM dimming driver is used to perform graded and periodic modulation of the light source output power to achieve stable imaging testing of the marker array under different brightness conditions. Uniform coatings or polishing are applied to the surfaces of four main materials, and reflectivity values ​​are obtained in real time using a surface reflectivity meter. These values ​​are used as reference inputs for the gain setting of the imaging device, ensuring that the influence of material characteristics on image features is fully captured. Precision adjustable platforms are constructed for three typical installation posture angles. Discrete adjustment of posture angle displacement is achieved through stepper motor drive, and the imaging device is triggered to capture the marker array image under the current posture after each adjustment. For two types of background interference, regular textures or random noise are applied using a background projection screen or diffuser. The interference intensity is monitored by a background illumination sensor, and this intensity is fed back to the imaging trigger control unit for optimization of synchronous imaging delay. By comprehensively collecting multi-dimensional operating condition variables, a sequence of original operating condition images covering all expected environmental changes is generated, enabling comprehensive image recording of a single physical marker array under multiple operating conditions. Through a synchronous trigger control method, the timestamps of each operating condition acquisition result are bound to the current operating condition number, forming a raw data index that allows S1.3 to read imaging parameters, thus ensuring both data coverage and retrieval capability.

[0031] Through the above-mentioned multi-dimensional working condition loading and synchronous triggering processing method, the physical marker array of the previous step is transformed into a sequence of original working condition images covering five types of light intensity gradients, four types of material surfaces, three types of installation posture angles, and two types of background interference, so as to realize the full acquisition of original working condition image data and the complete preservation of environmental change characteristics.

[0032] S1.3: For the original working condition image sequence, synchronously capture and record imaging parameters such as exposure time, gain value, white balance offset and lens distortion coefficient corresponding to each group of images to obtain structured imaging parameter metadata; this step uses the original working condition image sequence as the execution object, uses the camera driver interface protocol to perform parameter reading and timestamp alignment processing, and outputs structured imaging parameter metadata that is strictly bound to the image frame, providing a quantitative basis for the accurate mapping of subsequent working condition labels.

[0033] S1.4: Based on the original working condition image sequence and structured imaging parameter metadata, and combined with preset working condition classification rules, perform automated labeling processing of working condition labels to obtain a set of working condition labels with multi-dimensional attribute identifiers; this step takes the original working condition image sequence and structured imaging parameter metadata as input information, applies the rule engine matching method to discretize and encode environmental variables, and outputs a set of working condition labels that uniquely correspond to a specific environmental combination, thereby realizing the conversion of unstructured image data into semi-structured data.

[0034] The original working condition image sequence and structured imaging parameter metadata are input into the rule engine matching method to perform multi-dimensional feature extraction on environmental variables. The parameters are discretized and encoded by comparing the exposure time, gain value, white balance offset and lens distortion coefficient in the structured metadata with the numerical range of the preset working condition classification rule table.

[0035] The discretized coding results are jointly mapped with the illumination intensity gradient, material surface type, installation posture angle, and background interference mode of the original working condition image, and a preliminary environmental combination identifier sequence is generated through multi-dimensional combination indexing.

[0036] Conflict detection and uniqueness verification are performed on the initial environmental combination identifier sequence. The hash collision resolution method is used to ensure that each combination identifier uniquely corresponds to a specific label in the working condition label set.

[0037] The verified combined identifier sequence is converted into symbolic working condition labels by a label generator, and multi-dimensional attribute identifiers are attached, including illumination level, material category, attitude code and background mode code, forming a set of working condition labels with full parameter coverage.

[0038] The rule engine's index binding mechanism establishes a bidirectional association between the generated set of working condition labels and the corresponding original working condition image frames and structured imaging parameter metadata, thereby realizing the conversion of unstructured image data into semi-structured data and its searchable storage.

[0039] By using automated labeling processing, the results of the previous step are transformed into working condition label data with multi-dimensional attribute identifiers, enabling a fast, unified, and quantifiable description of the working environment.

[0040] For example, in the tooling verification scenario of a smart manufacturing production line, the original working condition image sequence includes five types of illumination intensity gradients (120Lux, 300Lux, 500Lux, 800Lux, and 1200Lux), four main material surfaces (aluminum alloy, carbon steel, stainless steel, and polymer composite materials), three installation posture angles (0°, 45°, and 90°), and two types of background interference (dynamic personnel flow and static equipment images). In the structured imaging parameter metadata, the exposure time setting ranges from 1ms to 20ms, the gain value is between 1 and 8 times, the white balance offset ranges from -100 to +100, and the lens distortion coefficient ranges from 0.95 to 1.05. The rule engine matching method discretizes and encodes the above numerical parameters; for example, an exposure time of 1ms to 5ms is mapped to illumination level code A1, and a gain value of 1 to 3 times is mapped to material environment code M2, etc. After joint mapping, a combined identifier sequence is formed, such as illumination level code A3, material environment code M1, attitude code P2, and background mode code B1, corresponding to a specific working condition combination. After performing conflict detection, the combined identifier sequence that passes uniqueness verification is converted into a working condition label "TAG_A3_M1_P2_B1", and an attribute identifier is attached to the label set for easy retrieval. This label establishes a bidirectional index relationship with the original image frame and imaging metadata, which can be instantly retrieved during retrieval. Verification results show that the label generation accuracy is significantly improved, and the annotation processing time remains at the 1ms level.

[0041] S1.5: For the original working condition image sequence, structured imaging parameter metadata, and working condition label set, perform data encapsulation and index construction operations to obtain a multi-working condition benchmark image dataset; this step integrates the above three as objects, uses database serialization technology to establish an association index between image content and metadata, and outputs a multi-working condition benchmark image dataset that supports fast retrieval and batch calling, completing the final transformation from scattered data collection to systematic training resources.

[0042] Step S2: Based on the multi-condition benchmark image dataset, extract the relative topological relationships between each physical marker point and encode them into geometrically invariant vectors that do not include original pixel information, thereby constructing a tooling semantic fingerprint. Specifically, this includes: S2.1: Gaussian pyramid layering and adaptive thresholding are performed on the original working condition image sequences in the multi-working condition benchmark image dataset to suppress background noise interference under different illumination intensity gradients and highlight the edge contours of physical marker point arrays such as micro-engraved cross grooves, thereby generating a binarized marker point mask image with a high bit signal-to-noise ratio.

[0043] Using the original working condition image sequence in the multi-working condition reference image dataset as the input execution object, and combining it with the multi-working condition coverage dataset output from the previous step S1.5, it is determined that the sequence includes a comprehensive working condition combination of five types of illumination intensity gradients, four main material surfaces, three installation attitude angles, and two types of background interference.

[0044] Gaussian pyramid layering is applied to each frame of the image to downsample the original image layer by layer to generate a multi-scale image set. The spatial configuration details of the physical markers are preserved in each scale image. The optimal level is determined based on the empirical variance index for edge feature enhancement.

[0045] The adaptive threshold segmentation results are mapped to a binary image, and edge connectivity analysis is performed to remove isolated noise points with an area smaller than a set lower limit, while retaining connected components that conform to geometric marker features.

[0046] Morphological opening operations are performed on the preserved connected regions to eliminate high-frequency noise protrusions while maintaining the edge integrity of the micro-etched cross grooves and precision hole arrays.

[0047] Through the above-mentioned layered and threshold chain processing method, the original image sequence under multiple working conditions is transformed into a structured binary marker mask image with high signal-to-noise ratio, thereby achieving stable visualization of physical marker features under different imaging conditions.

[0048] For example, in a smart manufacturing production line, this step is performed on a batch of tooling images with anodized aluminum surfaces, moderate illumination intensity, and low-contrast textured background interference. The Gaussian pyramid layering parameter is set to 3 layers, the optimal processing layer is the second layer, the adaptive threshold segmentation neighborhood window is 15×15 pixels, and the threshold adjustment coefficient k is set to 0.4. After edge connectivity analysis, noise points with an area less than 20 pixels are removed from the binarized results of each frame. After processing with a circular structuring element with a morphological opening radius of 2 pixels, the edges of the marker points are clear and the background noise is significantly eliminated. In the subsequent sub-pixel level coordinate extraction in S2.2, the marker point positioning accuracy of this binarized mask image remains at the 0.1 pixel level. The verification results show that the signal-to-noise ratio of the mask image is significantly improved compared to the original image under this condition, ensuring the stability and robustness of tooling semantic fingerprint extraction.

[0049] S2.2: Based on the binarized marker mask image, perform the gray-scale centroid method combined with the least squares ellipse fitting method to perform sub-pixel-level iterative optimization calculations on the center coordinates of the physical markers in each connected domain, so as to overcome the discretization error of the digital image and output a set of sub-pixel marker coordinates with an accuracy of one pixel.

[0050] Using the pixel distribution matrix of the binarized marker mask image as input, gray-level centroid calculation is performed for each connected component. A centroid vector is constructed using the gray values ​​and coordinate positions of all pixels within the connected component to obtain a preliminary estimate of the marker center position. The preliminary center position estimate is then processed using a least-squares elliptic fitting method to establish an elliptical model with the centroid as the reference. The symmetry and radial consistency of the marker geometry are optimized by fitting the ellipse parameters, eliminating local offset errors introduced by irregular shapes at the mask edges. Iterative optimization is performed in the parameter space of the ellipse fitting, using a convergence criterion that the fitting residual is less than a preset threshold. Combined with a gradient descent-based dynamic step size adjustment strategy, the center position coordinates are continuously corrected and the error magnitude is reduced during the iteration process. Each coordinate update during the iterative optimization process...

[0051] S2.3: Construct a complete graph topology using the sub-pixel marker point coordinate set, calculate the normalized angle ratio and Euclidean distance ratio sequence between any two points, and generate an initial geometric invariant vector that does not include the original pixel information and has rotation and scale invariance through orthogonal transformation encoding.

[0052] S2.4: Perform principal component analysis for dimensionality reduction and quantization compression on the initial geometric invariant vector, remove redundant dimensions and retain the feature components with the highest variance contribution rate, to obtain a compact geometric feature descriptor with fixed dimensions and extremely low storage overhead.

[0053] S2.5: Based on the compact geometric feature descriptor, perform hash mapping and check code appending operations to convert the continuous numerical sequence into a unique binary string identifier, thereby generating a tooling semantic fingerprint for constant-time complexity retrieval in the deviation pattern map.

[0054] like Figure 2 As shown, step S3 involves inputting the multi-condition reference image dataset into the deviation analysis module to obtain the position deviation results and comparing them with the measured true values. This process calculates inherent deviation features, including deviation direction preference, amplitude saturation range, and response delay characteristics, thereby constructing a deviation pattern map indexed by the tooling semantic fingerprint. Specifically, this includes: S3.1: Perform batch inference processing on the multi-condition reference image dataset, use the preset deviation analysis module to extract the original position deviation vector of each frame image, and perform differential operation between the original position deviation vector and the actual position measurement value collected by the high-precision laser tracker to obtain the initial deviation residual sequence including systematic error information.

[0055] S3.2: Based on the initial deviation residual sequence, perform multidimensional statistical feature mining, use the direction histogram clustering method to identify deviation direction preference, use the amplitude saturation detection method to locate the amplitude saturation interval, and calculate the response delay characteristics through time-series lag cross-correlation, thereby synthesizing an inherent deviation feature triplet that characterizes the inherent error behavior of the system under the current operating condition.

[0056] The initial deviation residual sequence generated in step S3.1 of the multi-condition reference image dataset is used as input. The statistical window, threshold, and delayed quantization reference value required for direction analysis, amplitude determination, and timing calculation are set. For deviation direction preference identification, a direction histogram clustering method is used to discretize the deviation residual vector into a fixed number of direction intervals according to the polar angle distribution. The frequency of occurrence within each direction interval is statistically analyzed, and a direction probability distribution matrix is ​​constructed. The deviation direction preference component is extracted through the direction vector with the highest frequency. For amplitude saturation interval location, an amplitude saturation detection method is executed. The absolute value sequence of the deviation residuals is compared with a preset amplitude response threshold. Intervals that continuously exceed the threshold are marked, and their start and end points are calculated to form an amplitude saturation interval index.

[0057] S3.3: Perform low-order piecewise linear fitting processing on the inherent deviation feature triplet, apply the least squares regression method to construct a low-order piecewise linear mapping function cluster that can describe the mapping relationship from input deviation to output correction, and compress and encode the low-order piecewise linear mapping function cluster into a deviation compensation mapping function package with fewer than 200 bytes of parameters.

[0058] The inherent bias feature triplet from the previous step is used as the input data set. The inherent bias feature triplet includes three components: bias direction preference coefficient, amplitude saturation interval threshold, and response delay compensation factor. Each component is a quantitative parameter extracted by statistical methods under the current operating conditions.

[0059] The relationship between input deviation and output correction is established by combining the deviation direction preference coefficient with the amplitude saturation interval threshold. The input variable is the original position deviation component, and the output variable is the corresponding correction component. Different amplitude intervals are used as segmentation boundary conditions.

[0060] The least squares regression method is used to construct a low-order linear mapping function within each segmented interval. The mapping coefficients and residual sums of squares are obtained by minimizing the sum of squared residuals for each interval sample. The calculation formula is as follows: Where y is the measured correction amount. is the fitted predicted value for the i-th sample.

[0061] The timing phase of each piecewise mapping function is adjusted by incorporating a response delay compensation factor. Lag effect correction is achieved by adding a delay compensation term to the coefficient matrix. The delay compensation calculation formula is as follows: Where A is the original mapping coefficient matrix, λ is the delay compensation factor, and D is the timestamp difference matrix.

[0062] After completing the fitting and compensation of all segments, the low-order linear mapping functions and their coefficient parameter sets of each segment are uniformly encoded, and the total number of parameters is controlled within 200 bytes by quantization compression. High compression ratio storage is achieved by pruning the parameter value range and converting floating-point numbers to fixed-point numbers.

[0063] Through the above processing method, the inherent bias feature triples are transformed into a low-order piecewise linear mapping function family, forming a bias compensation mapping function package with controlled capacity that can be loaded in sub-millisecond time, realizing the core computing model required for online retrieval and real-time correction.

[0064] S3.4: Using the tooling semantic fingerprint generated in the previous steps as the primary key index, and combining it with the recorded working condition tags as the secondary index, a hash table construction operation is performed, and the deviation compensation mapping function package is mounted to the corresponding index node, ultimately forming a deviation pattern graph data structure that supports sub-millisecond retrieval response.

[0065] The binary string identifier of the tooling semantic fingerprint output from the previous step is used as the primary key input, and combined with the stored set of working condition tags as the secondary index, a hash index initialization parameter is established, and the number of hash slots and load factor are defined to meet the sub-millisecond retrieval latency requirement.

[0066] Perform a fixed-length block segmentation and bitwise XOR operation on the primary key string to generate a hash value, and then map it to a preset hash slot number through a modulo operation to determine the index address.

[0067] Within a defined hash slot, a collision handling strategy is implemented. An open-chain method is used to establish an index node linked list. The working condition label and its corresponding low-order piecewise linear mapping function package are used as the node data field, and a checksum is recorded to ensure data consistency.

[0068] Memory-resident optimization is performed on the mapped function package. The compressed parameter structure is loaded into the cache area, and a memory pointer is attached to the inode to enable direct addressing access to the function package.

[0069] The search performance of the constructed hash table is verified by calling the combination of random tool semantic fingerprint and working condition label for retrieval, measuring the access path length and function package loading time, and ensuring that constant time complexity is maintained under the target load conditions.

[0070] By constructing a hash index and using memory mapping, the deviation compensation mapping function package from the previous step is transformed into a deviation pattern graph data structure that supports sub-millisecond retrieval response, achieving the expected technical effect of online fast matching and invocation.

[0071] like Figure 3 As shown, step S4: During the online operation phase, the current tooling real-time image is acquired and sub-pixel level marker point localization and topological relationship encoding are performed. The corresponding tooling semantic fingerprint is matched in the deviation pattern map via hash comparison to lock the current working condition label. Specifically, this includes: S4.1: Gaussian filtering and contrast enhancement preprocessing are performed on the real-time image of the current fixture transmitted by the image acquisition module to eliminate noise interference caused by ambient light fluctuations and highlight the edge features of physical marker points, thereby obtaining a preprocessed fixture image with a significantly improved signal-to-noise ratio.

[0072] S4.2: Based on the preprocessed tooling image, a sub-pixel level marker point positioning method combining grayscale centroid method and least squares method fitting is performed to iteratively optimize the preset physical marker point center coordinates in order to overcome pixel discretization error and obtain a sub-pixel marker point coordinate set with an accuracy of 0.1 pixel level.

[0073] Based on the real-time image data of the tooling after Gaussian filtering for noise reduction and contrast enhancement preprocessing, a preset array of physical marker points is used as the localization input. The gray-scale centroid method is called to perform center position estimation processing on the gray-scale distribution of each candidate region of the marker point, forming an initial pixel-level coordinate set. On the basis of the above coordinate set, an elliptical boundary fitting model is constructed, and the boundary parameters of each connected region are analytically calculated using the least squares method to generate the curve equation closest to the actual geometric shape and output the fitted center position. The fitted center position and the gray-scale centroid method result are subjected to difference analysis to extract the offset vector. This offset vector is used to drive the iterative optimization process, updating the center position estimate in each iteration until the Euclidean length of the offset vector is less than a threshold. For each marker point center position estimate, the residual deviation under the illumination gradient is corrected a second time as needed. The positioning robustness under complex working conditions is improved by recalculating the gray-scale moment intercept parameter. The center position coordinates after iterative optimization and secondary correction are converted into sub-pixel-level coordinate output to ensure that the positioning accuracy reaches the target of 0.1 pixels. By using a combined iterative optimization method of gray-scale centroid and least squares fitting, the preprocessed image results from the previous step are transformed into high-precision sub-pixel level marker coordinate data, enabling reliable identification and accurate positioning of the physical marker array.

[0074] For example, in a high-speed assembly scenario on a smart manufacturing production line, a 1920×1080 resolution industrial camera is used to acquire pre-processed tooling images. The diameter of the physical marker points is set to 1.2 mm, and the radius of the grayscale centroid search window is configured to 5 pixels. The initial coordinate set is used to estimate the center position through grayscale moment calculation. The ellipse fitting model uses the least squares method, with a fixed number of 32 boundary sampling points. The allowable value for the initial fitting center error is set to 0.3 pixels. An offset vector is generated through fitting difference to drive iterative updates. The upper limit of the number of iterations is set to 10, and the termination condition is that the Euclidean length of the offset vector is less than 0.05 pixels. After iteration, a secondary correction is performed on each center position by recalculating the grayscale moment intercept to compensate for residual deviations under highlight or shadow conditions. In the test, under conditions of rapidly changing light intensity (maximum brightness fluctuation of up to 2 times) and background noise interference, the accuracy of the output sub-pixel coordinate set is stable within 0.1 pixels, which greatly improves the robustness and accuracy of marker point positioning, ensuring that subsequent topology construction and semantic fingerprint encoding processes are based on high-quality geometric coordinate data.

[0075] S4.3: Construct a complete graph topology using the sub-pixel marker coordinate set, calculate the normalized angle ratio and distance ratio sequence between the lines connecting each physical marker, and generate a tooling semantic fingerprint vector that does not include the original pixel information and has rotation and scale invariance through orthogonal transformation encoding.

[0076] Based on the sub-pixel marker coordinate set obtained in the previous steps, this coordinate set is used as input for constructing geometric topological relationships.

[0077] For each pair of physical markers in the coordinate set, an edge set of the complete graph is established. The direction vectors of each connection are obtained through vector difference operation, and normalization is performed to eliminate scale dependence.

[0078] The arctangent function is applied to the normalized direction vector to obtain the angle value of the connecting line. This angle value is then divided by the reference angle base to form a normalized angle ratio sequence, ensuring the consistency of angle parameters under different attitude conditions.

[0079] The normalized angle ratio sequence and the normalized distance ratio sequence are combined and encoded to form a geometrically invariant matrix that does not contain the original pixel information but has rotation and scale invariance.

[0080] Perform an orthogonal transformation operation on the geometrically invariant matrix to map the matrix to a rotationally symmetric space, thereby achieving stable representation of features under different imaging conditions.

[0081] The above orthogonal transformation outputs a tooling semantic fingerprint vector, which establishes unique and low-dimensional geometric coding data for subsequent hash mapping, enabling efficient retrieval required for real-time matching.

[0082] By using methods such as complete graph construction, normalized angle and distance calculation, and orthogonal transformation coding, the high-precision marker coordinate data from the previous step is transformed into a tooling semantic fingerprint vector that is invariant to rotation and scale, thus establishing a robust working condition recognition foundation under multiple working conditions.

[0083] For example, on a smart manufacturing production line, for a tooling piece equipped with a cross-groove and circular hole array, the coordinate set obtained by sub-pixel level marker positioning includes 12 points distributed near four fixed-position components. After constructing the complete graph, the number of edges is 66. For each edge, a direction vector is calculated and normalized, with the reference angle benchmark set to 45°, resulting in normalized angle ratios ranging from 0.11 to 2.45. The Euclidean distance calculated to be between 12.3 and 55.8 mm is normalized using 40 mm as the reference distance benchmark to generate a distance ratio sequence. After combination encoding, a 66-row, 2-column geometric invariant matrix is ​​formed. After orthogonal transformation and mapping to rotationally symmetric space, the vector length is compressed to 20 dimensions. When this semantic fingerprint vector is retrieved and verified in the deviation pattern map, the matching time is 0.5 ms, and the matching success rate is still significantly improved under the condition of ±30% change in illumination intensity. The output results can be directly used for subsequent deviation compensation mapping calls to complete real-time dynamic condition recognition.

[0084] S4.4: Perform local sensitive hash mapping operation based on the tooling semantic fingerprint vector to compress the high-dimensional geometric invariant into a fixed-length binary hash code, so as to establish a lightweight tooling semantic fingerprint hash key for fast index retrieval.

[0085] S4.5: Based on the semantic fingerprint hash key of the lightweight chemical equipment, perform a hash table lookup operation with constant time complexity in the deviation pattern map, match the stored index key value and extract the associated metadata information, thereby locking the unique current working condition label that represents the current environmental conditions.

[0086] A constant-time retrieval mechanism based on a lightweight chemical equipment semantic fingerprint hash key is constructed. Using a hash matching module as the execution object, a fixed-length binary hash code generated in the preceding sub-steps is used as the unique index input signal. The direct addressing characteristic of the hash function is used to locate candidate index slots in the primary key space of the deviation pattern map. Bit-by-bit comparison operations are performed on the index key values ​​stored in the candidate slots. XOR operations combined with bitmasks are used to accumulate the matching results, generating a matching score vector. Based on the comparison relationship between the score vector and a preset threshold, it is determined whether the index key value is completely consistent with the input hash key, triggering consistency confirmation logic to ensure the uniqueness and conflict-free nature of the matching results. When consistency confirmation is successful, the associated metadata information is retrieved and dereferenced from the index node structure, and the working condition label identifier field included in the metadata is output through pointer relocation. This label identifier field is used as a unique code for the environmental conditions, completing the locking process of the current working condition label and providing an accurate working condition index foundation for the subsequent loading of low-order piecewise linear mapping function clusters. By using constant-time hash lookup and conflict-free index matching, the geometric invariant encoding result of the previous step is transformed into a unique label for environmental conditions, enabling real-time identification under multiple operating conditions.

[0087] For example, during the online operation of tooling deviation analysis in a smart manufacturing production line, the real-time generated tooling semantic fingerprint hash key is a 512-bit binary code. A multiplicative hybrid hash function is used to generate the index slot number, with a total of 4096 slots configured. When performing bit-by-bit XOR matching, the bitmask is set to all 1s, and a cumulative XOR result of 0 indicates a complete match. The matching threshold is set to be consistent when the cumulative number of matching differences is less than 1. In a retrieval task, the input hash key is directly addressed in slot 153, and a candidate index key value appears. The cumulative XOR result is 0, satisfying the consistency confirmation condition. The dereferenced condition tag in the associated metadata structure is "L3-M2-A1-B0," representing illumination level 3, material level 2, installation attitude angle level 1, and background interference level 0. This tag is sent to the secondary index of the deviation pattern map for use in step S5 to call the corresponding low-order piecewise linear compensation function cluster. Under the above operating conditions, the time for a single index match is 0.04 milliseconds, and the retrieval accuracy is stable within the coverage of the entire production line, achieving unique locking of environmental conditions under multiple working conditions on the high-speed production line.

[0088] Step S5: Based on the successfully matched current operating condition label, load the corresponding low-order piecewise linear mapping function cluster from the deviation pattern map, and input the original position deviation output by the deviation analysis module into the mapping function cluster for real-time interpolation correction to obtain the calibrated position deviation. Specifically, this includes: S5.1: Based on the current operating condition label locked in the previous steps, retrieve and load the corresponding low-order piecewise linear mapping function cluster parameter set from the index structure of the deviation pattern map to obtain a multi-dimensional correction parameter vector including deviation direction preference coefficient, amplitude saturation interval threshold and response delay compensation factor, so as to provide a dynamically adapted mathematical model basis for subsequent deviation correction calculation.

[0089] S5.2: Normalize the raw position deviation data output in real time by the deviation analysis module, use the amplitude saturation interval threshold in the multidimensional correction parameter vector to determine the boundary, cut off the raw position deviation that exceeds the linear response region to the effective calculation domain, and generate a standardized raw position deviation sequence to eliminate the risk of numerical divergence caused by sensor noise under extreme working conditions.

[0090] S5.3: Based on the standardized original position deviation sequence and the deviation direction preference coefficient in the multidimensional correction parameter vector, perform piecewise linear interpolation to construct a local error compensation surface and calculate the instantaneous compensation vector corresponding to the current input deviation value, so as to accurately quantify the nonlinear system error components caused by changes in illumination or material differences.

[0091] The standardized original position deviation sequence output from the preceding steps and the deviation direction preference coefficient in the multidimensional correction parameter vector are used as parallel input conditions.

[0092] The standardized original position deviation sequence is multiplied by the deviation direction preference coefficient to generate a direction-weighted deviation sequence, so as to highlight the directional error components related to changes in illumination or material differences under the current working conditions.

[0093] Linear interpolation is performed on the direction-weighted deviation sequence within the domain of each segment. The interpolation weights are determined using the mapping coefficients of the endpoints of adjacent segments, resulting in a normalized compensation coefficient matrix within each segment.

[0094] The normalized compensation coefficient matrix and the direction-weighted deviation sequence are multiplied to obtain the initial value sequence of piecewise compensation, which serves as the basic data for constructing the local error surface.

[0095] Two-dimensional interpolation fitting is performed on the initial value sequence of piecewise compensation. On the interpolation plane, with the deviation value and direction coefficient as coordinate axes, a local error compensation surface is generated by approximation through a surface function, which is used to describe the relationship between nonlinear system error and input deviation.

[0096] The instantaneous compensation vector corresponding to the current input deviation is calculated using the function value of the local error compensation surface.

[0097] in, For instantaneous compensation vector, To standardize the original position deviation value, The bias direction preference coefficient matrix, This is a surface function for local error compensation.

[0098] By using piecewise interpolation and surface fitting, the standardized deviation sequence from the previous step is transformed into an instantaneous compensation vector that can be directly used to cancel errors, thus achieving accurate quantification of nonlinear error components.

[0099] S5.4: The response delay compensation factor in the multidimensional correction parameter vector is used to perform time-phase correction on the instantaneous compensation vector. Combined with the timestamp information of the current image acquisition frame, a dynamic hysteresis correction term is generated, and the time-synchronization compensation vector is output to solve the problem of asynchronous feedback control caused by the delay of the image processing pipeline in high-speed production lines.

[0100] Using the instantaneous compensation vector as input, the response delay compensation factor in the multidimensional correction parameter vector is read and a phase correction coefficient matrix is ​​established to numerically adjust the timing characteristics of the compensation vector.

[0101] The timestamp information attached to the current image acquisition frame is precisely aligned with the generation time corresponding to the compensation vector. The processing delay difference between the two is calculated, and this difference is input into the phase correction coefficient matrix to form a delay correction factor sequence for the current frame.

[0102] The compensation vector after time synchronization is normalized in vector space to ensure that all components are within a unified dimensional standard, which facilitates superposition with the subsequent deviation sequence.

[0103] By using vector index mapping, the processed timing synchronization compensation vector is passed to the error cancellation stage of the next sub-step, thereby suppressing the feedback control asynchrony problem caused by image processing pipeline delay in high-speed production line environments.

[0104] By introducing a phase correction and amplitude normalization processing method based on timestamp difference, the instantaneous compensation vector of the previous step is transformed into compensation data with timing synchronization characteristics, thereby improving the time consistency of the feedback control link under high-speed conditions.

[0105] S5.5: Perform algebraic superposition operation on the timing synchronization compensation vector and the standardized original position deviation sequence, perform the final reverse error cancellation processing, generate calibrated position deviation data, and send the calibrated position deviation data to the motion control interface to drive the actuator to complete high-precision tooling attitude adjustment.

[0106] Step S6: Determine whether the tooling semantic fingerprint under the current working condition exists in the deviation pattern map. If not, package the tooling semantic fingerprints of multiple consecutive frames and the original position deviation into a new working condition data packet. Specifically, this includes: S6.1: Perform hash comparison processing on the index key in the pre-stored deviation pattern map of the current tooling semantic fingerprint generated in real time to calculate the fingerprint matching degree and generate the working condition coverage judgment result, thereby identifying whether the current working condition belongs to a new working condition not covered by the map.

[0107] S6.2: Based on the new working condition signal indicated in the working condition coverage determination result, perform sub-pixel level marker point localization and topological relationship encoding processing on the five consecutive real-time images output by the image acquisition module to obtain a set of consecutive multi-frame semantic fingerprints with time sequence consistency.

[0108] Upon receiving a new working condition signal from the working condition coverage determination module as a trigger condition, five consecutive frames of real-time tooling images output by the image acquisition module are used as input. Sub-pixel level marker point localization is used to iteratively optimize the center coordinates of preset physical marker points in each frame. In each iteration, the initial value of the gray-scale centroid method and the least squares fitting result are used for error feedback to ensure positioning accuracy reaches the zero-pixel level. Based on the set of marker point coordinates after localization, a complete graph topology structure corresponding to each frame is constructed. The normalized angle ratio, Euclidean distance ratio sequence, and connectivity information between any two points are calculated. Each feature value is then orthogonally transformed and encoded into a geometric invariant vector excluding the original pixel data. For the geometric invariant vector sequence generated for consecutive frames, a timestamp synchronization and sequence consistency constraint detection mechanism is used to filter out distorted frames caused by instantaneous imaging anomalies, ensuring the stability of semantic features in the time dimension. The geometric invariant vector sequence that passes the consistency detection is subjected to hash mapping processing to generate a fixed-length binary hash code set, which serves as the final encoding result of the semantic fingerprint set for consecutive frames. Through the above processing method, the working condition coverage determination result of the previous step is transformed into structurally stable and searchable continuous multi-frame semantic fingerprint data, so as to achieve temporal consistency and spatial uniqueness of subsequent deviation data synchronous extraction.

[0109] For example, in a smart manufacturing production line, when the working condition coverage determination module identifies that the lighting conditions and material combinations in the current assembly process do not match the deviation pattern map, and outputs a new working condition signal, the image acquisition module acquires five consecutive frames of tooling images at a sampling rate of 50fps, with a resolution of 1920×1080 pixels and an exposure time of 2ms. In the sub-pixel level marker point localization stage, the initial value of the grayscale centroid method is combined with least-squares ellipse fitting and iterated three times, converging the average localization deviation to 0.08 pixels. In the complete graph topology construction stage, each frame contains 8 marker points, and the normalized angle ratio and Euclidean distance ratio are orthogonally transformed and encoded to generate a 64-dimensional geometric invariant vector. In the time consistency detection stage, the threshold for the change in the geometric invariant Euclidean distance between each frame and the previous frame is set to 0.005, eliminating one frame of distorted data caused by mechanical vibration. In the hash mapping stage, the retained 5-frame geometric invariant vector sequence is compressed into a 128-bit fixed-length binary code, forming a continuous multi-frame semantic fingerprint set. When using this set to drive the deviation analysis module to extract the original position deviation synchronously, strict alignment of different frame data in terms of time and space features is achieved, ensuring the accuracy of subsequent manual verification of true value comparison and generation of new working condition data packets.

[0110] S6.3: Using the set of semantic fingerprints from multiple consecutive frames as a query index, the deviation analysis module is driven to synchronously extract the original position deviation data at the corresponding time, and combined with the manual verification truth value fed back by the high-precision laser tracker for spatiotemporal alignment processing to obtain a triplet data stream including input features, system output and true benchmark.

[0111] For the set of consecutive multi-frame semantic fingerprints, a mapping table between time index and frame number is established to ensure accurate location of the image frame corresponding to each semantic fingerprint when the deviation analysis module is invoked. Based on this mapping table, the deviation analysis module executes a synchronous data extraction program, sequentially reading and caching the original position deviation vector of each frame in a time series buffer, maintaining consistency between frame intervals and acquisition timing. Using the interface protocol of the high-precision laser tracker, a ground truth acquisition process is initiated to obtain the spatial position ground truth data corresponding to the frame number and store it in the ground truth sequence buffer. Spatiotemporal alignment is performed using frame timestamps, and the sampling points of the deviation vector and ground truth sequence are adjusted through interpolation and resampling methods to ensure precise matching of their correspondence on the time axis. Through the above processing, the result of the consecutive multi-frame semantic fingerprint set from the previous step is transformed into structured triplet data that can describe the working condition characteristics and deviation analysis performance, achieving accurate capture of new working condition data and availability for subsequent incremental training.

[0112] S6.4: Perform outlier removal and noise smoothing filtering on the triplet data stream to eliminate measurement jitter caused by transient interference, thereby obtaining a high-confidence cleaned working condition feature sequence.

[0113] S6.5: Encapsulate the high-confidence cleaned working condition feature sequence into a standardized data structure, add timestamp tags and production line identification metadata, and obtain a new working condition data package for offline expansion of the map entries for edge training nodes.

[0114] The cleaned working condition feature sequence with high confidence is used as the encapsulated source data. The semantic fingerprint vector, original position deviation value and manual verification truth value of each frame in the sequence are referenced. Data structure standardization transformation is performed to map the input multidimensional feature set into a unified field type and fixed byte alignment format to ensure cross-platform parsing consistency.

[0115] During the aforementioned standardization conversion process, a timestamp label is attached to each record according to the system's preset working condition data encapsulation protocol. This timestamp uses UTC format and is accurate to the millisecond level to ensure that the data stream can be reconstructed in time series when the deviation pattern map is expanded offline at the edge training node.

[0116] The production line identifier metadata is extracted from the interface call results of the production scheduling system, and its uniqueness is confirmed by hash verification. The production line identifier is used as the production source mark of the data packet so as to realize cross-production line data isolation and differentiated modeling during the map expansion process.

[0117] The encapsulated structured record set is stored as a serialized file format that supports streaming transmission. Combined with metadata indexing operations, the total number of records, field descriptors, and checksums are defined in the file header, thereby providing the necessary mechanisms for subsequent fast loading and integrity verification.

[0118] By using encapsulation conversion, timestamp appending, production line identifier binding, and serialization index construction, the cleaned working condition feature sequence from the previous step is transformed into a new working condition data packet that can be directly consumed by edge training nodes, thus achieving efficient transmission and secure storage of new working condition information during the offline training phase.

[0119] Step S7: Upload the new operating condition data packet to the edge training node and automatically expand the deviation pattern map during offline periods, updating only the local mapping function to complete the dynamic adaptive update of the deviation pattern map. Specifically, this includes: S7.1: Spatiotemporal alignment and outlier removal are performed on the semantic fingerprints of consecutive multiple frames, original positional deviations, and manually verified ground truth values ​​in the new working condition data packets to obtain a high-confidence incremental training sample set, providing a clean data input foundation for subsequent model updates.

[0120] S7.2: Calculate the inherent deviation feature vector under the new working condition based on the incremental training sample set, and use the least squares method to fit and generate a candidate low-order piecewise linear mapping function cluster to construct a local mapping function prototype that can accurately describe the deviation compensation relationship under the current new working condition.

[0121] Based on a high-confidence incremental training sample set, the semantic fingerprints of multiple consecutive frames, along with the corresponding original positional deviations and manually verified ground truth values, are used as a set of input vectors to form a multi-dimensional data matrix for deviation feature analysis.

[0122] The multidimensional data matrix is ​​subjected to differential operation in frame order to extract the error vector between the original position deviation and the true value of each frame. The direction component, amplitude component and response time component are calculated on the error vector to form a ternary structure with inherent deviation characteristics.

[0123] Normalization is performed on the ternary structure with inherent bias characteristics to fit different components at a uniform scale. The input of the mapping model is established with the directional component as the segmentation condition, the amplitude component as the height of the mapping surface, and the response time component as the temporal weight.

[0124] The relationship between the normalized features and the manually verified true values ​​is fitted by the least squares regression method. A multi-segment low-order linear function that minimizes the total residual is constructed. Each segment of the function corresponds to a specific directional preference interval and the function switches at the boundary of the amplitude saturation interval.

[0125] A local mapping function prototype is constructed using a function cluster generated by piecewise fitting. This prototype fully describes the mapping relationship between the input deviation vector and the required compensation amount under the new operating conditions, and provides a mathematical model basis for subsequent quantization compression and boundary smoothing.

[0126] Through the above processing method, the incremental training sample set of the previous step is transformed into a prototype of a local mapping function that accurately represents the deviation compensation relationship of the new working condition, thereby realizing the dynamic correction capability to adapt to the current working condition.

[0127] S7.3: Perform parameter quantization compression and boundary smoothing optimization on the prototype of the local mapping function to obtain a target local mapping function that meets storage constraints and has numerical stability, ensuring that the total number of parameters of the updated function cluster is controlled within a preset byte threshold.

[0128] S7.4: Using the newly generated tooling semantic fingerprint as an index key, the target local mapping function is written into the uncovered area of ​​the deviation pattern map to complete the entry expansion operation of the deviation pattern map and realize the dynamic absorption of new working condition features by the map knowledge structure.

[0129] Upon receiving the target local mapping function generated after parameter quantization compression and boundary smoothing optimization, and combining it with the new tooling semantic fingerprint generated from the incremental training sample set as the unique index key, the input conditions for the write operation are established. For the uncovered areas of the deviation pattern map, index structure localization is performed, using a hash table primary key matching mechanism to lock the blank nodes corresponding to the tooling semantic fingerprint, ensuring consistency between the write location and the data structure. The complete parameter set of the target local mapping function is serialized and encoded according to the storage format of a low-order piecewise linear mapping function cluster, and necessary version identifiers and working condition metadata are attached, achieving readable and parsable structured storage of parameters in the graph. Through the metadata mounting mechanism of the index node, the serialized local mapping function is associated with the index key, forming a bidirectional mapping relationship between the primary key and the function cluster, ensuring that the compensation strategy can be directly extracted in constant time complexity during subsequent retrieval. Consistency verification is performed on the written nodes, including hash code verification, parameter byte count verification, and boundary value simulation calculations, ensuring that the new entry does not compromise the numerical stability and retrieval performance of the existing graph. Through the above processing method, the target local mapping function of the previous step is transformed into a searchable new entry in the deviation pattern map, realizing the dynamic absorption of new working condition characteristics by the map knowledge structure, and providing readily available strategy resources for subsequent online deviation compensation.

[0130] S7.5: Based on the updated deviation pattern map, atomic version switching and memory hot loading operations are performed to activate the latest target local mapping function and release old resources, thereby completing seamless iterative upgrades of the deviation analysis model while maintaining online service continuity.

[0131] Using the updated deviation pattern map instance as input, the loading trigger signal is offline expanded by the edge training nodes and written into the status flag control of the new working condition local mapping function. A dual-version mapping region is established in memory for the updated deviation pattern map. A version control index table is used to map old and new version entries one-to-one, ensuring the index table remains consistent during switching. Atomic version switching operations are applied to the version control index table. The version pointer jump from the old to the new version is completed in an indivisible transaction through unified batch updates of index status flags, avoiding half-updated states during online retrieval. The atomically switched version control index table is bound to the running deviation analysis module through a memory hot-loading mechanism. Segment mapping operations during hot loading directly replace the old version entry page with the new version entry page, ensuring the real-time performance of module retrieval calls. The resource release command for the old version entry is written to a delayed release queue. Reference counting monitoring ensures that the release operation is executed after all calls are completed, avoiding null pointer exceptions caused by premature release. The released memory resources are added to a reusable memory pool, providing reusable storage space for subsequent dynamic updates. By employing the atomic version switching and memory hot-loading methods described above, the target local mapping function from the previous step is transformed into a real-time computing resource available to the deviation analysis module while maintaining the continuity of online services. This enables seamless iterative upgrades of the deviation analysis model and enhances its multi-condition adaptive capabilities.

[0132] For example, in a smart manufacturing production line, the updated deviation pattern map includes 2000 tooling semantic fingerprint entries, with each mapping function package being 150 bytes in size. The length and number of entries in the old and new version index tables are exactly the same. When the version control index table is constructed, each entry's status flag is set to 1 to indicate that it points to the old version address. During the atomic switch transaction, all status flags are updated to 2 in batches, indicating that they point to the new version address. The entire batch update takes 0.15ms in the transaction log. During memory hot loading, the memory page number of the new version entry is swapped with the old version's page number in the page table mapping through segment mapping replacement operations. The total number of page swaps is 2000, taking 0.35ms. The old version resource release command waits in the delayed release queue for the reference count to reach zero. The zeroing detection frequency is once every 10ms. After release, the memory reclamation amount is 300KB, which is added to the memory pool for use in new working condition updates. Tests showed that the online deviation retrieval response time remained stable within 0.8ms during the switching process, and the deviation compensation calculation delay did not change significantly. After the update, the system could immediately call the new working condition mapping function to complete the position deviation correction. After calibration, the deviation value fluctuation was significantly reduced, and the assembly consistency was significantly improved.

[0133] Step S8: Based on the updated deviation pattern map, semantic fingerprint matching and deviation compensation mapping are re-executed on the subsequently acquired real-time images to complete the verification of the tooling. Specifically, this includes: S8.1: Obtain the latest deviation pattern map data after offline expansion and updating of the local mapping function by the edge training nodes. Perform integrity verification and index reconstruction on the latest deviation pattern map data to obtain an online available deviation pattern map instance with real-time loading capability, which serves as the benchmark data source for subsequent real-time matching and compensation calculations.

[0134] S8.2: Based on the online available deviation pattern map instance, receive a new round of real-time tooling images transmitted by the image acquisition module, perform sub-pixel level physical marker point localization and topological relationship encoding operations on the new round of real-time tooling images, and obtain the real-time tooling semantic fingerprint vector corresponding to the current frame, which serves as the unique input identifier for working condition identification.

[0135] S8.3: Utilize the real-time tooling semantic fingerprint vector to perform hash key-value comparison and nearest neighbor search operations in the online available deviation pattern map instance to lock the successfully matched target tooling label and its associated low-order piecewise linear mapping function cluster, thereby determining a dedicated deviation compensation strategy suitable for the current imaging environment.

[0136] S8.4: Receive the original position deviation data output by the deviation analysis module according to the dedicated deviation compensation strategy, input the original position deviation data into the low-order piecewise linear mapping function cluster for real-time interpolation correction calculation, obtain the calibrated position deviation result, and eliminate the systematic measurement error caused by operating condition drift.

[0137] S8.5: Based on the position deviation result after calibration, the feedback calibration module is driven to perform tooling attitude adjustment. At the same time, it monitors whether the newly generated working condition data stream triggers the incremental learning condition, so as to form a complete closed-loop control process including perception acquisition, feature extraction, deviation correction and model self-updating links, and realize continuous iterative optimization of system performance.

[0138] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0139] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0140] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A tooling verification method for visual correction, specifically including: S1: For all tooling models involved in the target production line, preset geometrically robust physical marker points on the surface of each tooling and collect reference images to obtain a multi-condition reference image dataset. S2: Based on the multi-condition benchmark image dataset, extract the relative topological relationship between each physical marker point and encode it into a geometric invariant vector that does not include the original pixel information, thereby constructing the tooling semantic fingerprint; S3: Input the multi-condition reference image dataset into the deviation analysis module to obtain the position deviation result and compare it with the measured true value. Calculate the inherent deviation features including deviation direction preference, amplitude saturation range and response delay characteristics, and then construct a deviation pattern map with tooling semantic fingerprint as the index key. S4: During the online operation phase, acquire the real-time image of the current tooling and perform sub-pixel level marker point localization and topological relationship encoding. Match the corresponding tooling semantic fingerprint in the deviation pattern map by hash comparison to lock the current working condition label. S5: Based on the successfully matched current working condition label, load the corresponding low-order piecewise linear mapping function cluster from the deviation pattern map, and input the original position deviation output by the deviation analysis module into the mapping function cluster for real-time interpolation correction to obtain the calibrated position deviation.

2. The tooling verification method for visual correction according to claim 1, characterized in that, The step S5 is followed by: S6: Determine whether the tooling semantic fingerprint under the current working condition exists in the deviation pattern map. If it does not exist, package the tooling semantic fingerprint and the original position deviation of multiple consecutive frames into a new working condition data packet. S7: Upload the new working condition data packet to the edge training node and automatically expand the deviation pattern map during the offline period, only updating the local mapping function to complete the dynamic adaptive update of the deviation pattern map; S8: Based on the updated deviation pattern map, semantic fingerprint matching and deviation compensation mapping are re-executed on the subsequently acquired real-time images to complete the verification of the tooling.

3. The tooling verification method for visual correction according to claim 1, characterized in that, The acquisition of reference images includes acquiring reference images under various combinations of working conditions covering multiple light intensities, material surfaces, and installation postures.

4. The tooling verification method for visual correction according to claim 1, characterized in that, The multi-condition benchmark image dataset includes recorded imaging parameters and condition labels.

5. The tooling verification method for visual correction according to claim 1, characterized in that, Step S3 specifically includes: Batch inference processing is performed on the multi-condition benchmark image dataset. The original position deviation vector of each frame image is extracted using the preset deviation analysis module. The original position deviation vector is then differentially analyzed with the measured true value to obtain the initial deviation residual sequence including systematic error information. Based on the initial deviation residual sequence, multidimensional statistical feature mining is performed to synthesize an intrinsic deviation feature triplet that characterizes the inherent error behavior of the system under the current operating condition; Low-order piecewise linear fitting is performed on the triplet of inherent bias characteristics. The least squares regression method is used to construct a low-order piecewise linear mapping function cluster that can describe the mapping relationship from input bias to output correction. The low-order piecewise linear mapping function cluster is compressed and encoded into a bias compensation mapping function package with fewer than 200 bytes of parameters. Using the tooling semantic fingerprint generated in the previous steps as the primary key index, and combining it with the recorded working condition tags as the secondary index, a hash table construction operation is performed, and the deviation compensation mapping function package is mounted to the corresponding index node, ultimately forming a deviation pattern map that supports sub-millisecond retrieval response.

6. The tooling verification method for visual correction according to claim 5, characterized in that, The process of performing multidimensional statistical feature mining based on the initial deviation residual sequence to synthesize an intrinsic deviation feature triplet characterizing the inherent error behavior of the system under the current operating condition includes: Based on the initial deviation residual sequence, multidimensional statistical feature mining is performed. The deviation direction preference is identified by the direction histogram clustering method, the amplitude saturation interval is located by the amplitude saturation detection method, and the response delay characteristics are calculated by the time lag cross-correlation, thereby synthesizing an inherent deviation feature triplet that characterizes the inherent error behavior of the system under the current operating condition.

7. The tooling verification method for visual correction according to claim 1, characterized in that, Step S4 specifically includes: Gaussian filtering and contrast enhancement preprocessing are performed on the current real-time image of the tooling to eliminate noise interference caused by ambient light fluctuations and highlight the edge features of physical markers, thereby obtaining a preprocessed tooling image with a significantly improved signal-to-noise ratio. Based on the preprocessed tooling image, a sub-pixel level marker point positioning method combining grayscale centroid method and least squares fitting is performed to iteratively optimize the preset physical marker point center coordinates and obtain a sub-pixel marker point coordinate set with an accuracy of 0.1 pixel level. A complete graph topology is constructed using the sub-pixel marker coordinate set. The normalized angle ratio and distance ratio sequence between the lines connecting each physical marker are calculated. A tooling semantic fingerprint vector with rotation and scale invariance is generated by orthogonal transformation encoding, which does not include the original pixel information. Locality-sensitive hash mapping operation is performed based on the tooling semantic fingerprint vector to compress high-dimensional geometric invariants into fixed-length binary hash codes, so as to establish a lightweight tooling semantic fingerprint hash key for fast index retrieval. Based on the semantic fingerprint hash key of the lightweight chemical equipment, a hash table lookup operation with constant time complexity is performed in the deviation pattern map to match the stored index key value and extract the associated metadata information, thereby locking the unique current working condition label that represents the current environmental conditions.

8. The tooling verification method for visual correction according to claim 3, characterized in that, The acquisition of reference images includes acquiring reference images under various working conditions covering multiple light intensities, material surfaces, and installation postures. Specifically, it involves acquiring reference images covering five types of light intensities, four types of material surfaces, and three installation postures. The acquired reference images are structurally bound to the working condition labels.

9. The tooling verification method for visual correction according to claim 1, characterized in that, The process of extracting the relative topological relationships between each physical marker point and encoding them into a geometrically invariant vector that does not include the original pixel information includes multi-layer downsampling and layering of each frame image, adaptive thresholding to form a binary mask, combining the gray-level centroid of the marker point region with least squares ellipse fitting to optimize the sub-pixel coordinates, and normalizing and orthogonally transforming and compressing the connection direction and distance to finally obtain the geometrically invariant vector.

10. A tooling calibration device for visual correction, characterized in that, An apparatus employing a tooling verification method for visual correction as described in claims 1-9.

Images