AI vision-based intelligent recognition and cutting control method and system for cutter sharpener

Abstract

The present application relates to the field of industrial automation and intelligent manufacturing, and particularly relates to a tool cutter intelligent identification and cutting control method and system based on AI vision, which comprises: acquiring workpiece images from multiple perspectives and establishing coordinate correspondence; edge detection to generate candidate boundaries; multi-frame dynamic scoring and cross-perspective verification to form a reference boundary; generating a cutting trajectory and process parameters; issuing a controller to drive cutting; real-time calculation of geometric deviation and optimization of trajectory and tool movement; accurate identification of workpiece boundaries based on multiple continuous images and exclusion of fixture interference; real-time calculation of geometric deviation and optimization of tool posture, speed and trajectory; automatic triggering of deviation correction or safety control; and realization of high-precision, safe and intelligent cutting processing.

Description

Technical Field

[0001] This invention relates to the field of industrial automation and intelligent manufacturing, and more specifically, to an AI vision-based intelligent identification and cutting control method and system for a cutting machine. Background Technology

[0002] Existing vision-based cutting control methods for cutting machines have made some progress in multi-view image acquisition and boundary recognition, enabling dynamic drift compensation using continuous frame images to maintain cutting accuracy when the workpiece shifts due to minor displacement or machine tool vibration. However, in actual machining processes, potential recognition errors still exist, especially when workpiece boundaries and structures such as fixtures and guards appear simultaneously in the field of view. Existing methods may misidentify the edges of fixed structures as workpiece contours. This misidentification may be "correctly compensated" by a high-confidence model during dynamic drift compensation, causing the tool to shift along the wrong object—a typical problem known as "correct drift compensation, but compensating for the wrong object."

[0003] The severity of this problem lies in the fact that even if the cutting path and compensation algorithm themselves are logically accurate, the system may still stably perform compensation and cutting along the fixed fixture or protective cover, causing machining errors to accumulate and resulting in irreversible damage. This phenomenon reflects the potential catastrophic risk of existing technologies in high-confidence scenarios; that is, when the model has excessive confidence in the boundary, it lacks an effective means to distinguish between the workpiece and the fixed structure, thus failing to guarantee cutting safety and machining accuracy. Although existing methods can mitigate this through manual intervention or the addition of visual markers, there is a lack of an automated, continuous, and dynamically verifiable and corrective technique for the target workpiece boundary.

[0004] Therefore, existing technologies still struggle to achieve highly reliable and accurate cutting boundary recognition and dynamic compensation in complex environments with multiple structures. In particular, when the fixture and workpiece boundaries are close, the shape is complex, or there are obstruction conditions, high-confidence catastrophic deviations are more likely to occur. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of this invention is to provide an intelligent recognition and cutting control method and system for a cutting machine based on AI vision.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The AI ​​vision-based intelligent recognition and cutting control method for knife-cutting machines includes the following steps:

[0008] Step 1: Acquire image information of the workpiece and its surrounding environment from different perspectives using at least two vision sensors, and establish the correspondence between the image coordinate system of each vision sensor and the machine tool coordinate system of the cutting machine.

[0009] Step 2: Perform edge detection and geometric analysis on the workpiece image to generate a candidate boundary set that includes the workpiece contour boundary and the fixed structure boundary.

[0010] Step 3: Based on continuous multi-frame image information, perform dynamic benchmark affinity scoring and cross-view consistency verification on each candidate boundary in the candidate boundary set to form the benchmark boundary of the target workpiece and exclude fixed structure boundaries.

[0011] Step 4: Generate the cutting trajectory in the machine tool coordinate system based on the target workpiece reference boundary, and generate a set of cutting process parameters;

[0012] Step 5: Send the cutting trajectory and cutting process parameters to the motion controller to drive the cutting machine to perform cutting;

[0013] Step 6: Calculate the geometric deviation based on the images and trajectory information obtained during the cutting process, and optimize the cutting trajectory, tool posture, local speed or local acceleration in real time.

[0014] Furthermore, at least two vision sensors include a primary view sensor and an auxiliary view sensor. The primary view sensor is set at a fixed position on the cutting machine, facing the main cutting surface of the workpiece. The auxiliary view sensor is mounted on the moving arm of the cutting machine and can adjust the viewing angle according to the actual cutting requirements to obtain image information of the workpiece from different angles.

[0015] Furthermore, the image information is processed through image registration technology to ensure the image stitching accuracy between the two viewpoints, providing accurate image data for boundary detection, cutting trajectory planning, and accuracy correction.

[0016] Furthermore, using a calibration plate or a workpiece with known geometric features as a reference object, at least two vision sensors are used to acquire multi-angle images of the workpiece. By matching the relationship between feature points in the images and known geometric features, and combining with a 3D reconstruction algorithm, a geometric correspondence between the image coordinate system of each vision sensor and the machine tool coordinate system of the cutting machine is established. The accuracy of the coordinate system transformation is ensured by optimizing the calculation using the least squares method, which is then used for subsequent image data processing and cutting trajectory generation.

[0017] Furthermore, the candidate boundary set generation includes performing Gaussian filtering on the workpiece image to suppress noise, extracting edge information, performing contour tracking to form a preliminary boundary set, and using geometric feature filtering to distinguish between workpiece contour boundaries and fixed structure boundaries, thereby generating the final candidate boundary set.

[0018] Furthermore, a dynamic benchmark affinity score is performed on each candidate boundary in the candidate boundary set. This includes comprehensively calculating the boundary position and contour changes in consecutive multi-frame images, weighting each boundary by simulating the potential energy distribution and local stability index in the physical field, and ranking the importance of the boundaries by combining multi-frame spatial consistency and historical trajectory smoothness. The cross-view consistency of the images is used to further ensure the accuracy of the score. The boundary score is updated through iterative optimization, so that the final selected target workpiece benchmark boundary has the best stability and spatial consistency in the multi-frame time series.

[0019] Furthermore, a set of cutting process parameters is generated based on the workpiece's material properties, the curvature of the reference boundary, and the cutting quality requirements.

[0020] Furthermore, based on the geometric deviation calculated in step six, a high-order orthogonal decomposition is performed in the controllable coordinate space of the motion controller to obtain the true deviation of the workpiece, and the subsequent cutting process is further optimized based on this deviation.

[0021] Furthermore, when the geometric deviation exceeds a preset safety threshold, it triggers the replanning of the remaining cutting trajectory or safety degradation control.

[0022] Furthermore, the AI ​​vision-based intelligent recognition and cutting control system for the cutting machine includes: an image acquisition module for acquiring image information of the workpiece and its surrounding environment;

[0023] The boundary generation module is used to perform edge detection and geometric analysis on the workpiece image, and generate a set of candidate boundaries that includes the workpiece contour boundary and the fixed structure boundary.

[0024] The datum determination module is used to form the datum boundary of the target workpiece and exclude fixed structural boundaries;

[0025] The trajectory generation module is used to generate a cutting trajectory in the machine tool coordinate system based on the target workpiece reference boundary, and to generate a set of cutting process parameters;

[0026] The execution control module is used to send the cutting trajectory and cutting process parameters to the motion controller to drive the cutting machine to perform cutting.

[0027] The deviation optimization module is used to calculate geometric deviations based on the images and trajectory information acquired during the cutting process, and to optimize the cutting trajectory, tool posture, local speed or local acceleration in real time. When the geometric deviation exceeds the preset safety threshold, it triggers the replanning of the remaining cutting trajectory or safety degradation control.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] This invention can acquire continuous images of the workpiece and its surrounding environment from different perspectives using at least two vision sensors, and establish a precise correspondence between the visual coordinate system and the machine tool coordinate system, thereby achieving accurate identification of the target workpiece boundary. During the identification process, this invention combines multi-frame dynamic benchmark affinity scoring with cross-view consistency verification to effectively eliminate interference from fixed fixtures, protective covers, and non-machined objects. This avoids the high-confidence disaster problem in existing technologies where "drift compensation is correct but the wrong object is being compensated," thus significantly improving the safety and reliability of the cutting process and ensuring that the tool always moves along the true workpiece contour.

[0030] By integrating the cutting trajectory and cutting process parameters and sending them to the machine tool motion controller, and by acquiring images and trajectory information in real time during the cutting process, the geometric deviation of the workpiece is calculated, enabling dynamic optimization of tool posture, local speed, and local acceleration. This method can automatically adjust motion parameters for complex curved surfaces, high-curvature edges, or thin-walled areas, ensuring balanced cutting force and excellent surface quality. It also improves cutting efficiency while enhancing machining accuracy, making it particularly suitable for high-precision machining of aerospace parts, titanium alloys, or complex multi-curved workpieces.

[0031] This invention utilizes a high-order orthogonal decomposition method to map real-time calculated geometric deviations to a controllable coordinate space, quantifying the actual deviations of the workpiece in three-dimensional space, and optimizing subsequent cutting trajectories and processing parameters accordingly. When the geometric deviation exceeds a preset safety threshold, this invention can automatically trigger replanning of the remaining trajectory or safety degradation control, achieving dynamic correction and protection, thereby reducing the risk of workpiece damage, minimizing manual intervention, and significantly enhancing the intelligence, adaptability, and controllability of the entire cutting process, providing reliable technical support for high-precision automated machining. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the overall process of an AI vision-based intelligent recognition and cutting control method for a cutting machine.

[0033] Figure 2 A schematic diagram illustrating the process of generating the candidate boundary set and determining the target workpiece reference boundary;

[0034] Figure 3 This is a schematic diagram of the intelligent recognition and cutting control system for a blade cutter based on AI vision. Detailed Implementation

[0035] Example 1: Refer to Figures 1 to 3 The AI ​​vision-based intelligent recognition and cutting control method for a knife-cutting machine includes the following steps:

[0036] Step one involves acquiring image information of the workpiece and its surrounding environment from different perspectives using at least two vision sensors, and establishing the correspondence between the image coordinate systems of each vision sensor and the machine tool coordinate system. This step is used to acquire the spatial geometric information of the workpiece and environmental reference information, providing basic data for subsequent boundary detection and cutting trajectory planning. By establishing the correspondence between the image coordinate system and the machine tool coordinate system, a precise correlation between visual information and mechanical actions can be achieved, enabling the cutting machine to perform precise positioning and motion control based on visual feedback. This step ensures that subsequent boundary detection, reference boundary determination, and cutting trajectory generation are performed under a unified coordinate system, thereby improving cutting accuracy and system stability.

[0037] In one specific embodiment, to achieve high-precision cutting of the workpiece, at least two vision sensors acquire image information of the workpiece and its surrounding environment from different perspectives. The at least two vision sensors include a primary view sensor and an auxiliary view sensor. The primary view sensor is fixedly mounted on one side of the cutting machine, facing the main cutting surface of the workpiece, and is used to acquire a complete image of the workpiece's front side. The auxiliary view sensor is mounted on the moving arm of the cutting machine, and its viewing angle can be flexibly adjusted according to different cutting process requirements to acquire image information of the workpiece's side and edge areas, thereby covering the workpiece's three-dimensional contour features. In actual operation, for example, when cutting an aluminum workpiece with a complex contour, the primary view sensor can capture the outer contour and key holes on the workpiece's front side, while the auxiliary view sensor can capture the unevenness and cutting blind spots on the workpiece's side, effectively compensating for the visual blind spots under a single view.

[0038] The acquired multi-view images are processed using image registration technology to match and spatially correct feature points, ensuring accurate stitching of images from different perspectives within a unified coordinate system. The registered images not only allow for complete reconstruction of the workpiece contour but also precise mapping of the workpiece's position relative to the machine tool coordinate system, providing a reliable data foundation for subsequent boundary detection, cutting trajectory planning, and accuracy correction. For example, when processing workpieces with irregular curved surfaces, the registered images can identify minute deviations in surface variations, allowing for pre-compensation during cutting trajectory generation and ensuring the accuracy and stability of the tool movement along the surface.

[0039] Based on the registered image information and combined with image data from multiple consecutive frames, edge extraction and geometric analysis are performed on the workpiece contour to obtain a complete set of candidate boundaries. Interfering boundaries from fixed devices or non-cutting areas are then eliminated through cross-view spatial consistency verification. During this process, dynamic weights can be assigned to different contour features, prioritizing boundaries with high stability and consistency with the actual geometry of the workpiece as target boundaries. For example, when cutting a steel plate with tilted holes and grooves, the system can automatically identify these feature boundaries, ensuring that the generated cutting trajectory accurately covers all processing areas. This achieves high-precision, high-reliability intelligent cutting operations, significantly improving processing efficiency and workpiece quality.

[0040] In one specific implementation, at least two vision sensors are deployed at different locations on the cutting machine, using a calibration plate or a workpiece with known geometric features as a reference object, to acquire image data of the workpiece from multiple perspectives. The calibration plate or the workpiece with known geometric features, as a standard reference object, provides accurate geometric features to ensure that the vision sensors can accurately capture the three-dimensional shape and spatial relationships of the workpiece. For example, when machining a workpiece with complex holes, pre-marked geometric features (such as rectangular or circular marks) on the calibration plate serve as key points, assisting the sensors in acquiring the workpiece's contour information from different angles.

[0041] Based on this multi-angle image information, by matching the relationship between feature points in the images and known geometric features on the calibration object, and combining this with a 3D reconstruction algorithm, the geometric correspondence between the image coordinate systems of each vision sensor and the machine tool coordinate system is further established. In this process, by using image information acquired from different perspectives, computer vision technology is used to extract feature points on the workpiece and calibration plate, and their positions in 3D space are calculated through geometric projection, thereby obtaining an accurate mapping between each vision sensor and the machine tool coordinate system. For example, the projection of a hole on the workpiece will differ under different perspectives; the 3D reconstruction algorithm can accurately restore the position of this hole in the machine tool coordinate system.

[0042] The least squares method is used to optimize the correspondence between all image coordinate systems and the machine tool coordinate system, ensuring the accuracy of coordinate system transformation. As a commonly used optimization method, the least squares method minimizes the errors at multiple acquisition points, precisely adjusting the deviation between the vision sensor coordinate system and the machine tool coordinate system. This process significantly improves the accuracy of image data processing and ensures the synchronization and consistency between the target boundary captured by the vision sensor and the actual machine tool operation when the cutting trajectory is generated. For example, if the workpiece has slight tilt or deformation, the optimized coordinate system transformation ensures that the cutting trajectory of the cutting machine can accurately adapt to these changes, thereby achieving high-precision machining results.

[0043] Step two involves edge detection and geometric analysis of the workpiece image to generate a candidate boundary set containing both the workpiece contour boundary and the fixed structure boundary. This step extracts the shape features of the workpiece and distinguishes the boundary information between the workpiece and the fixed structure, providing a candidate set for forming the target workpiece reference boundary. Through edge detection and geometric analysis, the system can obtain accurate contour information of the workpiece, providing a complete and reliable data source for reference boundary selection, while also eliminating interference from fixed devices or non-cutting areas in complex environments.

[0044] In one specific implementation, to obtain the accurate contour of the workpiece and eliminate interference from fixed structures, Gaussian filtering is applied to workpiece images acquired from at least two different perspectives to suppress noise and minor interference in the images while maintaining the integrity of edge information. After Gaussian filtering, the grayscale changes in the image are smoother, which helps subsequent edge detection algorithms accurately identify the workpiece contour. In practical applications, such as when machining an aluminum alloy workpiece with grooves and holes, Gaussian filtering can remove random noise caused by light reflection and workpiece surface texture, thereby preventing erroneous edges from being identified as valid contours.

[0045] Based on the filtered image, edge extraction technology is applied to detect the workpiece contour, and a preliminary boundary set is generated by combining it with a contour tracking algorithm. Contour tracking forms closed or continuous contour paths by sequentially traversing edge pixels to accurately reflect the geometry of the workpiece. For workpieces with complex shapes, such as metal plates with beveled edges or irregular surfaces, contour tracking can retain the geometric information of all key features of the workpiece, providing data support for precise cutting.

[0046] A preliminary set of boundaries is filtered using geometric features to distinguish between workpiece contour boundaries and fixed structure boundaries, thereby generating a final set of candidate boundaries. Geometric feature filtering includes analyzing parameters such as boundary length, curvature, closure, and spatial position, excluding boundaries irrelevant to the actual cutting area of ​​the workpiece, such as the edges of the machine tool fixture or the contours of surrounding support structures. This method generates a candidate boundary set that completely covers the workpiece's cutting area while eliminating interfering boundaries, achieving high-precision contour extraction for complex workpieces and providing a reliable basis for subsequent baseline boundary determination and cutting trajectory planning. In practical operation, for example, when cutting steel plates with multiple openings and curved edges, this method can effectively identify each valid cutting edge while ignoring the contours of the fixed fixture, ensuring that the tool movement accurately covers the required machining area of ​​the workpiece.

[0047] Step three involves performing dynamic benchmark affinity scoring and cross-view consistency verification on each candidate boundary in the candidate boundary set based on continuous multi-frame image information to form the target workpiece benchmark boundary and exclude fixed structure boundaries. This step is used to select stable and reliable target workpiece benchmark boundaries from the candidate boundary set. Through continuous multi-frame image analysis and dynamic scoring, the system can smooth boundary changes and eliminate the influence of instantaneous noise. Simultaneously, cross-view consistency verification ensures that the selected benchmark boundaries have spatial consistency under different viewpoints, thereby providing an accurate reference for cutting trajectory generation and improving cutting quality and positioning reliability.

[0048] In one specific implementation, to ensure accurate identification of the target workpiece's reference boundary from the candidate boundary set, multiple consecutive frames of images are simultaneously acquired to capture minute displacements or deformations that may occur in the workpiece before and during cutting. By calibrating and tracking the position of the candidate boundaries in each frame, spatial variation data of each boundary over time is obtained, forming a complete boundary motion trajectory. When machining an aluminum workpiece with complex arcs and beveled edges, this step can capture the edge offset caused by clamping or minute vibrations, thus providing accurate data support for subsequent boundary stability assessment.

[0049] A dynamic baseline affinity scoring model is constructed by comprehensively calculating the boundary position and contour changes in consecutive multi-frame images. This scoring model simulates the potential energy distribution and local stability indices in the physical field, and assigns a weighted score to each candidate boundary to quantify its stability over a multi-frame time series. For example, for the edge of a hole or thin-walled region on a workpiece, slight vibrations may occur due to minute deformations during processing. The potential energy distribution simulation can identify the physical reliability of the boundary, thereby distinguishing between stable boundaries and temporary noise boundaries.

[0050] The importance of boundaries is ranked by multi-frame spatial consistency and historical trajectory smoothness, and the score is further optimized by combining cross-view consistency verification. In this process, boundary data from different viewpoints are registered, and the geometric consistency of boundaries under each viewpoint is analyzed to eliminate false boundaries caused by viewpoint deviations or lighting variations. For example, when cutting a steel plate with bevels and grooves, the main viewpoint sensor may not be able to fully capture the bottom edge of the groove, while the side information provided by the auxiliary viewpoint sensor can be used to verify the authenticity of the boundary, ensuring that the final score accurately reflects the actual geometric features of the workpiece.

[0051] The boundary scores are updated through iterative optimization. By combining dynamic scoring results and cross-view verification results, a stable and spatially consistent target workpiece reference boundary is formed. During the iterative optimization process, boundaries with low scores and non-geometric consistency are excluded, while the weights of high-scoring boundaries are adjusted to ensure their consistency across consecutive time frames. In actual machining, such as machining an aerospace part with complex openings and curved contours, this method can automatically select reliable reference boundaries while eliminating boundary interference from fixed fixtures or support structures. This allows the cutting trajectory planning to accurately cover all machining areas of the workpiece, thereby significantly improving cutting accuracy, machining efficiency, and workpiece surface quality, achieving highly reliable and stable intelligent machining operations.

[0052] Example of implementing dynamic baseline affinity scoring: In one implementation, for each candidate boundary... In continuous Extracting the boundary point set from the frame image And calculate the following stability and consistency characteristics:

[0053] Position stability: The boundary centroids of each frame are... After mapping to the machine tool coordinate system, calculate the centroid variance: ;in for The average centroid.

[0054] Shape stability: Curvature sequence of boundaries for each frame Calculate inter-frame changes: ;

[0055] Historical trajectory smoothness: based on the baseline boundary determined in the previous time step. Predict the current boundary and calculate the prediction residual. (For example, using a constant velocity model or first-order filtering prediction). Example 1: Based on "boundary centroid + constant velocity prediction": Predicting candidate boundaries in consecutive frames. Extracting the centroid , Then, the velocity is estimated using the first two frames. Predict the centroid of the current frame Define the prediction residual. The calculation formula is: If the boundary motion is stable, the prediction error is small; if the boundary undergoes abrupt changes (such as noise, occlusion, or false boundaries), the prediction error is large, thus reducing the score. Example 2: Based on "Boundary Control Point Sequence + First-Order Filter (Exponential Smoothing)": In this implementation, the boundary... Parameterized into control point vectors sampled with equal arc length Prediction formula using first-order filtering Perform boundary prediction. Calculate the residuals as root mean square error. The formula is as follows: This method not only constrains the positional changes of the boundary, but also constrains the smooth evolution of the shape over time, making it suitable for complex workpiece contours.

[0056] Spatial consistency penalty: When the overlap ratio between the candidate boundary and the preset "non-processing area" (such as the mask of the area where the fixture is located) is higher than a threshold, a penalty term is introduced. Example 1: Based on "jig / non-processing area mask + overlap ratio": Obtain the mask for the jig or non-processing area. , representing the geometric extent in the machine tool coordinate system. Calculate the candidate boundary pixel set. Overlap ratio with mask : Set a threshold The penalty term is defined as follows: When the overlap ratio exceeds the threshold, a penalty term is applied. Activated to prevent fixtures or non-machined areas from being misidentified as workpiece boundaries. Example 2: Based on "cross-viewpoint fixation discrimination": Candidate boundaries are mapped to the machine tool coordinate system from multiple viewpoints to obtain the boundary pose at each viewpoint. If the boundary hardly changes with workpiece movement in the machine tool coordinate system, it is determined to be a fixed structure, and its fixation is calculated. : The penalty term is defined as follows: ;in It is the penalty coefficient. When the value is large, the penalty term is also large, thus ensuring that only the workpiece boundary is preserved.

[0057] Based on the above characteristics, a dynamic baseline affinity score is constructed: ;in The weights are non-negative. Weights can be obtained through at least one of the following methods: collecting sample sequences of "workpiece boundary / fixed structure boundary" during equipment calibration or trial cutting, and determining them using grid search or minimizing misclassification loss; or setting them empirically and adjusting them during trial operation according to the boundary misclassification rate. During score iteration updates, the stopping condition is either "the score ranking remains unchanged for two consecutive rounds" or "the score gain is less than the preset convergence threshold," thus ensuring the algorithm's repeatability. An example of allocation is as follows: Positional stability (centroid variance) has the highest weight because the overall positional stability of the boundary directly affects the accuracy of the workpiece boundary. Shape stability (curvature change), medium weight, to ensure that the boundary shape does not fluctuate due to minor noise. Historical trajectory smoothness, with a higher weight, is used to predict boundary trends using continuous frame information and improve boundary continuity. Spatial consistency penalty (overlapping with non-processed areas) has the lowest weight, but it is still necessary to prevent the wrong selection of fixed fixture boundaries. The reason for this allocation is that the spatial position error of the boundary has the greatest impact on the cutting deviation, so positional stability has the highest weight; the smoothness of the historical trajectory helps to reduce the wrong judgment caused by short-term noise; shape stability and spatial consistency provide auxiliary constraints to ensure the overall stability and reliability of the scoring model, while preventing wrong rejection due to an excessively large single penalty term.

[0058] In one implementation, cross-view consistency verification includes at least one of the following methods: Method 1 (reprojection error): The set of boundary points of the main viewpoint... The coordinates are transformed to the machine tool coordinate system using the calibrated extrinsic parameters, and then reprojected onto the auxiliary viewpoint image plane using the extrinsic and intrinsic parameters, resulting in a set of reprojected points. The set of boundary points actually detected from an auxiliary perspective. For reference, calculate the average reprojection error. ;in, Represents the reprojection point to the set The distance (e.g., the Euclidean distance to the nearest neighbor). When When determining candidate boundaries Consistency verification passed.

[0059] Method 2 (3D Reconstruction Consistency): Triangulation of corresponding boundary point pairs from two viewpoints yields a 3D boundary point cloud. And calculate the two-view reprojection residuals or point cloud registration residuals. ;when When determining candidate boundaries Consistency verification passed. The threshold... The threshold can be determined by comprehensively considering the upper bound of the calibration residual, pixel resolution, and allowable trajectory error; for example, the quantiles of the reprojection residuals can be used as a source of the threshold during the calibration phase. If a candidate boundary fails the consistency verification, a penalty is imposed on its score (e.g., setting the threshold to...). (or reduce its weight), and prioritize eliminating the candidate boundary in the next iteration.

[0060] Step four involves generating a cutting trajectory in the machine tool coordinate system based on the target workpiece's reference boundary, and also generating a set of cutting process parameters. This step transforms the target workpiece's reference boundary into a specific cutting trajectory and operating parameters, providing instructions and a plan for the actual cutting by the cutting machine. By generating the trajectory in the machine tool coordinate system, the system can directly control the machine tool to execute the cutting action. Simultaneously, it generates a set of cutting process parameters based on the workpiece material, boundary curvature, and quality requirements, ensuring that the cutting process conforms to process specifications and guarantees cutting accuracy, surface quality, and processing efficiency.

[0061] In one specific implementation, to achieve precise cutting of the target workpiece, the previously determined reference boundary of the target workpiece is mapped to the machine tool coordinate system of the cutting machine, establishing the precise positional relationship of the boundary in space. By sequentially sorting and curvature analyzing the boundary points in the machine tool coordinate system, information on the continuity and local geometric changes of the cutting path is obtained, providing basic data for subsequent trajectory generation. For example, when cutting an aluminum alloy workpiece with arc grooves and sharp corner contours, this step can accurately calibrate the spatial coordinates of each arc segment and sharp corner, ensuring that the tool precisely fits the workpiece surface when moving along the path, avoiding offset or collision.

[0062] A set of cutting process parameters is generated based on the workpiece material properties, the boundary curvature of the reference boundary, and the cutting quality requirements. Workpiece material properties include hardness, toughness, and thermal conductivity; boundary curvature reflects the degree of local bending along the cutting path; and cutting quality requirements involve surface finish, kerf accuracy, and processing efficiency. In practical operation, for example, when cutting high-hardness steel plates, for boundary segments with large curvature, the tool movement speed can be appropriately reduced and the sampling point density of the feed path increased to ensure the stability of the tool movement along complex contours and processing accuracy. Conversely, for straight segments with small curvature, the speed can be appropriately increased to improve processing efficiency.

[0063] The cutting trajectory is integrated with the generated set of cutting process parameters to form a complete instruction sequence that can be directly used for machine tool driving. During trajectory generation, smooth interpolation and local speed optimization methods are employed to ensure a smooth transition between curved boundaries and straight segments, avoiding cutting errors or surface scratches caused by sudden acceleration changes. In this way, the system can automatically generate optimal cutting trajectories for workpieces of different shapes and materials, and dynamically adjust tool motion parameters according to boundary features and processing requirements, thereby achieving high-precision and high-efficiency intelligent cutting. For example, when cutting aerospace parts with multiple curves and concave holes, this method ensures that the cutting trajectory accurately covers all contours while optimizing cutting speed and tool posture to achieve high-quality part forming.

[0064] Step five involves sending the cutting trajectory and cutting process parameters to the motion controller to drive the cutting machine to perform the cutting. This step enables the transmission and execution of the trajectory and process parameters to the machine tool control system, converting the visual calculation results into actual mechanical actions. After receiving the trajectory and parameters, the motion controller can control the tool to move along the planned path to complete the cutting of the workpiece. This step ensures seamless integration of visual information and machine tool operation, achieving the unification of intelligent control and automated processing.

[0065] In one specific implementation, to achieve precise workpiece cutting, the previously generated cutting trajectory and the set of cutting process parameters are integrated to form a complete operating instruction that can be directly used for machine tool driving. The cutting trajectory includes the spatial path points and movement direction information of the tool in the machine tool coordinate system, while the set of cutting process parameters includes tool speed, feed rate, depth of cut, cutting angle, and acceleration curve, etc. In actual operation, for example, when cutting a titanium alloy workpiece with complex arcs and grooves, by integrating the trajectory and process parameters, the movement of the tool along the curved surface and groove edges can be precisely controlled, ensuring that the tool always maintains a stable cutting posture, while avoiding vibration or surface scratches caused by speed or feed mismatch, thereby improving cutting accuracy and surface quality.

[0066] The integrated cutting commands are sent to the motion control system of the cutting machine to drive the tool to perform the cutting operation. During execution, the motion control system dynamically adjusts the motion of each axis according to the coordinates of the trajectory points and process parameters, achieving continuous and smooth tool movement along complex contours. This process can be optimized in real time for local geometric features of the workpiece. For example, when the tool passes through high-curvature edges or concave holes, the cutting speed is automatically reduced and the density of trajectory points is increased to ensure accurate coverage of the cutting path and local cutting quality. In actual machining, such as when cutting an aerospace part with multiple curves and deep holes, this method can ensure that the tool moves steadily along complex contours while maintaining a balanced distribution of cutting force and machining heat, thereby significantly improving machining efficiency, workpiece surface flatness, and dimensional accuracy, achieving highly reliable and high-precision automated intelligent cutting.

[0067] Step six involves calculating geometric deviations based on the images and trajectory information acquired during the cutting process, and optimizing the cutting trajectory, tool posture, local speed, or local acceleration in real time. This step dynamically monitors cutting accuracy, ensuring that deviations are corrected promptly during the cutting process. By calculating and analyzing geometric deviations, the system can adjust the cutting trajectory and tool motion parameters in real time, thereby compensating for deviations caused by workpiece position errors, clamping errors, or cutting forces. This step improves machining accuracy, reduces the risk of cutting defects, and ensures the stability and safety of the machining process.

[0068] In one specific implementation, to achieve high-precision control and adaptive optimization during the cutting process, image information of the workpiece is continuously acquired and the tool movement trajectory is recorded synchronously. Real-time image processing methods are used to extract and update the workpiece surface contour, thereby calculating the geometric deviation between the actual tool path and the target reference boundary. The calculation of geometric deviation includes not only the spatial positional deviation between the tool and the workpiece edge, but also the difference in cutting depth caused by changes in edge curvature, as well as local errors caused by material elastic deformation or machine tool vibration. For example, when machining aerospace parts with complex concave holes and multiple curved surfaces, real-time images can capture minute offsets generated during tool cutting, thus providing accurate deviation data.

[0069] The calculated geometric deviations are mapped to the controllable coordinate space of the motion controller. High-order orthogonal decomposition is then used to decompose and quantify the deviations, revealing the actual deviation components of the workpiece in three-dimensional space. This method decomposes geometric deviations into independent components along each axis and analyzes their corresponding local rotation and displacement characteristics, providing precise correction information for motion control. For example, when a tool cuts along a curved groove, orthogonal decomposition can distinguish between errors caused by tool posture deviation and deviations caused by material elastic rebound, thus enabling more precise adjustments.

[0070] Based on the above deviation analysis results, the cutting trajectory, tool posture, and local speed or acceleration are optimized in real time to dynamically adapt the tool movement to the actual shape of the workpiece and cutting conditions. During the optimization process, for high curvature or thin-walled areas, the system can automatically reduce the cutting speed and adjust the tool posture to reduce cutting force concentration and workpiece surface scratches; while in straight sections or low curvature areas, the feed rate can be appropriately increased to improve machining efficiency. For example, when cutting aluminum alloy workpieces with large thickness variations, this optimization strategy can ensure that the tool moves smoothly along the curved surface while maintaining a balanced cutting force, ensuring surface finish and dimensional accuracy.

[0071] When the calculated geometric deviation exceeds a preset safety threshold, the system triggers a replanning of the remaining cutting trajectory or safety degradation control to prevent workpiece damage or machine tool malfunction. Replanning involves generating a new cutting path and motion parameters based on the current workpiece shape, while safety degradation control keeps the cutting process within a controllable range by reducing tool speed, decreasing depth of cut, or adjusting feed strategy. Taking the machining of titanium alloy parts with deep holes and curved grooves as an example, when the geometric deviation of a certain local edge exceeds the safety threshold, the tool automatically adjusts its motion trajectory or reduces its speed for gentle cutting, thereby ensuring cutting accuracy and workpiece safety. Simultaneously, it provides a reliable basis for correction in subsequent cutting, achieving high-precision, high-reliability, and intelligent cutting processing.

[0072] In one embodiment, geometric deviation Use at least one of the following definitions: Definition 1 (Normal closest distance): Represent the target reference boundary as a curve / surface in the machine tool coordinate system. For the tool center trajectory point Calculated to The nearest point and take the normal distance As for the geometric deviation, the formula is as follows: ;in, Represents the dot product of vectors. It is a curve / surface At point The normal vector at that location.

[0073] Definition 2 (Point Set Registration Residual): At time t: Obtain the visual boundary point set With tool trajectory sampling point set Registration was performed and the root mean square error was used. Characterization bias, the formula is as follows: ;in, It is the visual boundary point set The Middle One point, It is the tool trajectory sampling point set The Middle To ensure temporal consistency between the image and the trajectory, both image acquisition and controller trajectory sampling are time-stamped. When the sampling periods of the two are inconsistent, the trajectory is linearly interpolated to align to the image timestamp. The deviation sequence can be filtered using median filtering or first-order low-pass filtering to suppress noise, and obvious outliers (such as those exceeding several times the standard deviation of the mean or exceeding the physical upper limit) are removed to avoid unnecessary trajectory oscillations caused by instantaneous false detections.

[0074] Preset safety threshold Determination: In one embodiment, a preset safety threshold is used. The determination of this includes at least one of the following criteria:

[0075] Based on equipment calibration and measurement errors: The upper bound / statistical distribution of the reprojection residual obtained during the calibration stage is determined by superimposing it with the allowable error of the trajectory.

[0076] Based on process safety distance: combining the minimum safe clearance between the fixture and the workpiece, the tool geometry and the allowable depth of cut, determine the upper limit of the deviation that must not be exceeded.

[0077] Adaptive based on historical data: The quantiles (e.g., high quantiles) of the deviation during statistically stable machining are used as the source of the threshold, and the data is recalibrated and updated after changing materials / tools.

[0078] To prevent threshold-triggered jitter, a "continuous" approach is adopted. The strategy of "triggering only when the sampling limit is exceeded" is implemented, and a hysteresis zone is set: when the deviation falls back to... The over-limit state will only be lifted under the following circumstances. After triggering, replanning will include at least: regenerating the remaining trajectory based on the currently updated baseline boundary; safety degradation control will include at least: reducing tool speed, decreasing depth of cut, or limiting the rate of change of acceleration to ensure that the cutting process is within a controllable range.

[0079] Example 2: Intelligent recognition and cutting control system for a cutting machine based on AI vision, including: an image acquisition module for acquiring image information of the workpiece and its surrounding environment;

[0080] The boundary generation module is used to perform edge detection and geometric analysis on the workpiece image, and generate a set of candidate boundaries that includes the workpiece contour boundary and the fixed structure boundary.

[0081] The datum determination module is used to form the datum boundary of the target workpiece and exclude fixed structural boundaries;

[0082] The trajectory generation module is used to generate a cutting trajectory in the machine tool coordinate system based on the target workpiece reference boundary, and to generate a set of cutting process parameters;

[0083] The execution control module is used to send the cutting trajectory and cutting process parameters to the motion controller to drive the cutting machine to perform cutting.

[0084] The deviation optimization module is used to calculate geometric deviations based on the images and trajectory information acquired during the cutting process, and to optimize the cutting trajectory, tool posture, local speed or local acceleration in real time. When the geometric deviation exceeds the preset safety threshold, it triggers the replanning of the remaining cutting trajectory or safety degradation control.

[0085] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included 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. An AI vision-based intelligent recognition and cutting control method for a knife-cutting machine, characterized in that, Includes the following steps: Step 1: Acquire image information of the workpiece and its surrounding environment from different perspectives using at least two vision sensors, and establish the correspondence between the image coordinate system of each vision sensor and the machine tool coordinate system of the cutting machine. Step 2: Perform edge detection and geometric analysis on the workpiece image to generate a candidate boundary set that includes the workpiece contour boundary and the fixed structure boundary. Step 3: Based on continuous multi-frame image information, perform dynamic benchmark affinity scoring and cross-view consistency verification on each candidate boundary in the candidate boundary set to form the benchmark boundary of the target workpiece and exclude fixed structure boundaries. Step 4: Generate the cutting trajectory in the machine tool coordinate system based on the target workpiece reference boundary, and generate a set of cutting process parameters; Step 5: Send the cutting trajectory and cutting process parameters to the motion controller to drive the cutting machine to perform cutting; Step 6: Calculate the geometric deviation based on the images and trajectory information obtained during the cutting process, and optimize the cutting trajectory, tool posture, local speed or local acceleration in real time.

2. The AI ​​vision-based intelligent identification and cutting control method for a cutting machine according to claim 1, characterized in that, At least two vision sensors include a primary view sensor and an auxiliary view sensor. The primary view sensor is set at a fixed position on the cutting machine, facing the main cutting surface of the workpiece. The auxiliary view sensor is mounted on the moving arm of the cutting machine and can adjust the viewing angle according to the actual cutting requirements to obtain image information of the workpiece from different angles.

3. The AI ​​vision-based intelligent identification and cutting control method for a knife-cutting machine according to claim 2, characterized in that, Image information is processed through image registration technology to ensure the image stitching accuracy between two viewpoints, providing accurate image data for boundary detection, cutting trajectory planning, and accuracy correction.

4. The AI ​​vision-based intelligent identification and cutting control method for a knife-cutting machine according to claim 1, characterized in that, Using a calibration plate or a workpiece with known geometric features as a reference object, multi-angle images of the workpiece are acquired using at least two vision sensors. By matching the relationship between feature points in the images and known geometric features, and combining with a 3D reconstruction algorithm, a geometric correspondence is established between the image coordinate system of each vision sensor and the coordinate system of the cutting machine. The accuracy of the coordinate system transformation is ensured by optimizing the calculation using the least squares method, which is then used for subsequent image data processing and cutting trajectory generation.

5. The AI ​​vision-based intelligent identification and cutting control method for a knife-cutting machine according to claim 1, characterized in that, The candidate boundary set generation process includes performing Gaussian filtering on the workpiece image to suppress noise, extracting edge information, performing contour tracking to form a preliminary boundary set, and using geometric feature filtering to distinguish between workpiece contour boundaries and fixed structure boundaries, thereby generating the final candidate boundary set.

6. The AI ​​vision-based intelligent identification and cutting control method for a cutting machine according to claim 1, characterized in that, For each candidate boundary in the candidate boundary set, a dynamic benchmark affinity score is performed. This includes comprehensively calculating the boundary position and contour changes in consecutive multi-frame images, weighting each boundary by simulating the potential energy distribution and local stability index in the physical field, and ranking the importance of the boundaries by combining multi-frame spatial consistency and historical trajectory smoothness. The cross-view consistency of the images is used to further ensure the accuracy of the score. The boundary score is updated through iterative optimization, so that the final selected target workpiece benchmark boundary has the best stability and spatial consistency in the multi-frame time series.

7. The AI ​​vision-based intelligent identification and cutting control method for a cutting machine according to claim 1, characterized in that, A set of cutting process parameters is generated based on the workpiece's material properties, the curvature of the reference boundary, and the cutting quality requirements.

8. The AI ​​vision-based intelligent identification and cutting control method for a cutting machine according to claim 1, characterized in that, Based on the geometric deviation calculated in step six, a high-order orthogonal decomposition is performed in the controllable coordinate space of the motion controller to obtain the true deviation of the workpiece, and the subsequent cutting process is further optimized based on this deviation.

9. The AI ​​vision-based intelligent identification and cutting control method for a cutting machine according to claim 1, characterized in that, When the geometric deviation exceeds the preset safety threshold, it triggers the replanning of the remaining cutting trajectory or safety degradation control.

10. An AI vision-based intelligent recognition and cutting control system for a cutting machine, applied to the AI ​​vision-based intelligent recognition and cutting control method for a cutting machine as described in any one of claims 1-9, characterized in that, include: The image acquisition module is used to acquire image information of the workpiece and its surrounding environment; The boundary generation module is used to perform edge detection and geometric analysis on the workpiece image, and generate a set of candidate boundaries that includes the workpiece contour boundary and the fixed structure boundary. The datum determination module is used to form the datum boundary of the target workpiece and exclude fixed structural boundaries; The trajectory generation module is used to generate a cutting trajectory in the machine tool coordinate system based on the target workpiece reference boundary, and to generate a set of cutting process parameters; The execution control module is used to send the cutting trajectory and cutting process parameters to the motion controller to drive the cutting machine to perform cutting. The deviation optimization module is used to calculate geometric deviations based on the images and trajectory information acquired during the cutting process, and to optimize the cutting trajectory, tool posture, local speed or local acceleration in real time. When the geometric deviation exceeds the preset safety threshold, it triggers the replanning of the remaining cutting trajectory or safety degradation control.

Images