Tool management method and system for large gantry machine tool based on visual recognition

Abstract

The application provides a tool management method and device for large gantry machine tools based on visual recognition, and relates to the field of industrial automation. The technical problem that the prior art cannot realize high-robustness wear quantification and residual life prediction due to unreliable tool model recognition and inaccurate wear geometric feature extraction caused by poor three-dimensional reconstruction quality in complex industrial environments such as high light reflection, oil stains and local shielding of metal tools is solved. The method comprises the following steps: projecting a coded light pattern on a tool to be evaluated, and collecting a plurality of non-coplanar perspective structured light images; based on the structured light images and known coded light pattern information, generating a three-dimensional geometric model of the tool to be evaluated, extracting its structured geometric features, determining its model and recognition confidence; based on the recognition confidence; constructing a nonlinear state space model, obtaining the current wear level and residual life prediction value, and evaluating the health status of the tool to be evaluated. The application is used in the tool management process.

Description

Technical Field

[0001] This application relates to the field of industrial automation, and in particular to a tool management method and system for a large gantry milling machine based on vision recognition. Background Technology

[0002] In the high-precision and high-efficiency machining process of large gantry milling machines, the tool, as a key component that directly participates in cutting, has its wear condition directly affecting the surface quality, dimensional accuracy, and production safety of the workpiece.

[0003] Traditional tool management relies heavily on manual experience or threshold alarms based on indirect signals such as spindle current and vibration, making it difficult to accurately reflect the true geometric wear morphology. While existing visual inspection methods can acquire three-dimensional morphology, in complex industrial environments such as strong reflections from metal tools, oil contamination, and partial occlusion, the reconstructed model is prone to holes, artifacts, or geometric distortions. This leads to unreliable extraction of key features such as the width of the flank wear band and the radius of the cutting edge passivation, resulting in misidentification of tool type and deviations in wear assessment. Consequently, it cannot support highly robust online life prediction and intelligent tool change decisions.

[0004] Therefore, there is an urgent need for a closed-loop tool health management system that integrates anti-interference 3D reconstruction, geometric feature analysis and physical mechanism modeling, in order to achieve accurate perception and reliable early warning of tool status under complex working conditions. Summary of the Invention

[0005] This application provides a tool management method and system for large gantry milling machines based on vision recognition. It solves the technical problems of existing technologies in complex industrial environments such as high reflectivity of metal tools, oil stains, and partial obstruction, where poor 3D reconstruction quality leads to unreliable tool model identification and inaccurate extraction of wear geometric features, thus making it impossible to achieve highly robust wear quantification and remaining life prediction.

[0006] To achieve the above objectives, this application adopts the following technical solution: Firstly, a tool management method for a large gantry milling machine based on vision recognition is provided, including: A coded light pattern is projected onto the tool to be evaluated using a structured light projection unit, and structured light images from multiple non-coplanar viewpoints are acquired. The structured light images include co-polarized images and orthogonal polarized images. The co-polarized images are those in which the camera polarization direction is parallel to the projection light polarization direction, and the orthogonal polarized images are those in which the camera polarization direction is orthogonal to the projection light polarization direction. Based on the structured light image and the known coded light pattern information, phase calculation and triangulation are performed to generate a three-dimensional geometric model of the tool to be evaluated. The structured geometric features of the tool to be evaluated are extracted from the three-dimensional geometric model. The structured geometric features include the tool holder diameter, the helix angle of the cutting edge, the effective cutting length, and the wear geometry parameters. The structured geometric features are matched with a pre-stored standard tool template library to determine the tool model to be evaluated and the identification confidence level. When the identification confidence level is higher than the preset threshold, the wear geometry parameters and cumulative usage time of the tool to be evaluated are used as state variables to construct a nonlinear state-space model, and the extended Kalman filter is used to recursively estimate the model to obtain the current wear level and remaining life prediction value of the tool to be evaluated. Based on the wear level and remaining life prediction, the health status of the tool to be evaluated is assessed, and corresponding operation and maintenance control signals are generated.

[0007] Based on the above technical solutions, in the tool management method for large gantry milling machines based on vision recognition provided in this application, in the high-precision and high-efficiency machining scenarios of large gantry milling machines, the tool wear state directly affects workpiece quality and production safety. However, existing methods are difficult to achieve reliable recognition and accurate wear assessment in complex industrial environments such as strong reflection, oil stains, and obstructions. Traditional vision solutions suffer from 3D reconstruction distortion due to interference from metal surfaces, leading to model misjudgment; while life prediction based on a single signal or static threshold lacks physical basis and has poor robustness. To address this, this solution constructs a closed-loop "perception-recognition-modeling-decision" chain: it suppresses specular reflection through multi-view imaging with polarized structured light and generates a high-fidelity 3D model by fusing phase information; it extracts standard-compatible structured geometric features from the model and combines weighted similarity, matching discrimination, and topological constraints to generate quantifiable recognition confidence, ensuring reliable model judgment; only under high confidence conditions is a mechanism-driven nonlinear state-space model constructed based on wear parameters and usage time, and extended Kalman filtering is used for recursive estimation to achieve physically consistent wear level and remaining life prediction. This technical solution deeply integrates optical anti-interference, geometric analysis, and dynamic state estimation, significantly improving the accuracy, interpretability, and engineering applicability of tool health management, effectively supporting intelligent operation and maintenance decisions, and avoiding unplanned downtime and machining accidents.

[0008] In conjunction with the first aspect above, in one possible implementation, the method for reconstructing the three-dimensional geometric model of the tool to be evaluated includes: Based on the orthogonal polarization image, the absolute phase map of each viewpoint is calculated, and in the region where the phase position confidence is lower than the preset confidence threshold, the phase information of the same polarization image is fused in a weighted manner to generate a fused phase map of each viewpoint. The initial 3D point cloud is calculated based on the fused phase map from each viewpoint and the camera calibration parameters. The initial 3D point cloud is registered from multiple perspectives, and point cloud weights are constructed based on local polarization difference and surface gradient stability indices. A weighted Poisson surface reconstruction algorithm is used to generate a three-dimensional geometric model of the tool to be evaluated.

[0009] In conjunction with the first aspect above, in one possible implementation, the method for constructing the point cloud weights includes: Based on the pixel intensity in orthogonal polarization images and co-polarization images, calculate the local polarization difference of each 3D point cloud point. The gradient magnitude of the orthogonal polarization image at this point is calculated and marked as a surface gradient stability index; Based on the local polarization difference degree and surface gradient stability index, a first weighting factor and a second weighting factor are generated by monotonically decreasing and monotonically increasing mappings, respectively, and the two are multiplied to obtain the point cloud weight of the point.

[0010] In conjunction with the first aspect above, in one possible implementation, the extraction of the structured geometric features of the tool to be evaluated includes: Curvature analysis is performed on the reconstructed three-dimensional geometric model of the tool to be evaluated; regions with curvature less than a preset curvature threshold are marked as candidate tool holder regions; a random sampling consensus algorithm is used to perform cylindrical fitting on the candidate tool holder regions to obtain the spindle direction and tool holder diameter of the tool to be evaluated; A local coordinate system for the tool to be evaluated is established with the spindle direction as the reference, and the point cloud is segmented and projected along the axis of the coordinate system; the periodicity of the gradient signal caused by the spiral groove in the projected image is analyzed, the starting axial coordinate and ending axial coordinate of the spiral groove are identified, the distance between the two coordinates is calculated, and the effective cutting length is obtained. Multiple equally spaced cross-sectional profiles are extracted within the cutting edge region, and the helix angle of the cutting edge is calculated by fitting a helical model. The wear geometry parameters include the tip curvature radius and the width of the flank wear band; the tip curvature radius is calculated based on the principal curvature of the surface; the flank wear band width is obtained by calculating the distance deviation from the actual flank area to the ideal cutting surface reference plane and projecting it along the tool spindle direction. The tool holder diameter, cutting edge helix angle, effective cutting length, tool tip curvature radius, and wear geometry parameters are combined into a structured geometric feature vector.

[0011] In conjunction with the first aspect above, in one possible implementation, the method for obtaining the ideal cutting edge reference plane includes: Based on the model of the tool to be evaluated, the theoretical wedge angle and helix angle corresponding to the tool to be evaluated are obtained from the standard tool database; Extract the cutting edge ridge line from the three-dimensional geometric model and determine a reference point on the cutting edge; A local right-handed coordinate system is constructed with the reference point as the origin, the cutting edge tangential direction as the first axis, and the tool spindle direction as the third axis. In the local coordinate system, the normal vector of the ideal back face is calculated based on the theoretical wedge angle and helix angle; An ideal cutting edge reference plane is constructed using the reference point and the normal vector.

[0012] In conjunction with the first aspect above, in one possible implementation, the method for obtaining the blade tip curvature radius includes: The cutting edge ridge line is determined from the three-dimensional geometric model of the tool to be evaluated, and the foremost point of the cutting edge ridge line is selected as the center point of the cutting tip. With the center point of the blade tip as the center, a local point cloud neighborhood is extracted within a preset physical radius; a tangent plane is fitted within the local point cloud neighborhood, and a local orthogonal coordinate system is established based on the tangent plane, and the point cloud coordinates are transformed to the coordinate system; A weighted least squares method is used to fit a quadratic surface function to obtain the coefficients of the quadratic term characterizing the local surface shape. A Hessian matrix is ​​constructed based on these coefficients, and the two principal curvatures are calculated through eigenvalue decomposition. , Through formula The radius of curvature of the tool tip was calculated. .

[0013] In conjunction with the first aspect above, in one possible implementation, the method for constructing the nonlinear state-space model includes: After the confidence level of tool model identification is higher than a preset threshold, a state vector is defined. The state vector includes the width of the flank wear band and the cumulative effective cutting time calculated by the three-dimensional geometric model. Based on Arcard's wear law, and combined with real-time acquired normal cutting force, spindle speed, and chip temperature estimated from spindle power, a state transition equation containing an exponential thermal softening term is constructed. The width of the wear band on the back face measured in real time by the three-dimensional vision system is used as the observation. An identity mapping relationship is established between the observed value and the corresponding element in the state vector to form the observation equation. The state transition equation is combined with the observation equation to form a nonlinear state-space model.

[0014] In conjunction with the first aspect above, in one possible implementation, the method for obtaining the confidence level includes: The structured geometric features are compared with the features of each candidate template in the standard tool template library to calculate the weighted similarity and obtain the distance metric for each candidate model. Match discrimination is evaluated based on the ratio of the minimum distance to the second minimum distance; Simultaneously verify whether the three-dimensional geometric model satisfies the topological constraints of the candidate model; By combining the distance metric, matching discrimination, and topology consistency, a recognition confidence value between 0 and 1 is generated.

[0015] In conjunction with the first aspect above, in one possible implementation, the assessment of the health status of the tool to be assessed includes: The tool model, wear level, and predicted remaining life value are sent to the machine tool controller. The machine tool controller determines whether the tool is in a healthy state based on the allowable wear threshold and minimum remaining life corresponding to the tool model. If yes, no action is taken; otherwise, a tool change command or maintenance warning signal is generated.

[0016] Secondly, this application provides a tool management system for a large gantry milling machine based on vision recognition, comprising: an acquisition module, an identification module, and an evaluation module; wherein, the acquisition module is used to project an coded light pattern onto the tool to be evaluated through a structured light projection unit and acquire structured light images from multiple non-coplanar perspectives; the identification module is used to perform phase calculation and triangulation based on the structured light images and known coded light pattern information to generate a three-dimensional geometric model of the tool to be evaluated; extract the structured geometric features of the tool to be evaluated from the three-dimensional geometric model; match the structured geometric features with a pre-stored standard tool template library to determine the model and identification confidence level of the tool to be evaluated; the evaluation module is used to construct a nonlinear state-space model using the wear geometric parameters and cumulative usage time of the tool to be evaluated as state variables when the identification confidence level is higher than a preset threshold, and recursively estimate the model using an extended Kalman filter to obtain the current wear level and remaining life prediction value of the tool to be evaluated; evaluate the health status of the tool to be evaluated based on the wear level and remaining life prediction value, and generate corresponding operation and maintenance control signals.

[0017] This application provides a tool management method and system for large gantry milling machines based on vision recognition. It effectively solves core problems in existing technologies caused by high reflectivity of metal tools, oil contamination, and partial occlusion in complex industrial environments, such as distorted 3D reconstruction, unreliable model identification, and inaccurate wear assessment. By deploying a multi-view, non-coplanar polarized structured light sensor array in the tool changing area, it simultaneously acquires co-directional and orthogonal polarized images, suppressing specular reflection and enhancing the fringe signal-to-noise ratio. Combining phase fusion and weighted Poisson reconstruction, a high-fidelity 3D geometric model is generated. Based on this, standard-compatible structured geometric features (such as tool holder diameter, helix angle, effective cutting length, and wear parameters) are extracted. A multi-dimensional criterion combining weighted similarity, matching discrimination, and topological constraints is used to output a recognition confidence level of 0 to 1, ensuring reliable model identification. Only when the confidence level is higher than a threshold is a nonlinear state-space model constructed using wear parameters and cumulative usage time as state variables. Extended Kalman filtering is then used to recursively estimate physically consistent wear levels and remaining life. This solution achieves a closed-loop process across the entire chain, from "anti-interference perception to high-reliability identification to mechanism-driven prediction to intelligent operation and maintenance decision-making." It significantly improves the accuracy, robustness, and interpretability of tool health management, effectively supports predictive maintenance, reduces the risk of unplanned downtime, and ensures machining quality and production safety.

[0018] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0019] Figure 1 A system architecture diagram of a tool management system for a large gantry milling machine based on vision recognition, provided for embodiments of this application; Figure 2 A flowchart illustrating a tool management method for a large gantry milling machine based on vision recognition, provided in an embodiment of this application; Figure 3 This is a flowchart illustrating a method for generating a three-dimensional geometric model of a cutting tool, as provided in an embodiment of this application. Detailed Implementation

[0020] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The tool management method for large gantry milling machines based on vision recognition provided in this application embodiment can be applied to the tool management system of large gantry milling machines based on vision recognition, such as... Figure 1 As shown, the system includes: a data acquisition module, an identification module, and an evaluation module; The acquisition module is used to project a coded light pattern onto the tool to be evaluated through the structured light projection unit and acquire structured light images from multiple non-coplanar viewpoints. The recognition module is used to perform phase calculation and triangulation based on structured light images and known coded light pattern information to generate a three-dimensional geometric model of the tool to be evaluated. The structured geometric features of the tool to be evaluated are extracted from the three-dimensional geometric model; the structured geometric features are matched with a pre-stored standard tool template library to determine the tool model and identification confidence level. The evaluation module is used to construct a nonlinear state-space model by using the wear geometry parameters and cumulative usage time of the tool to be evaluated as state variables when the identification confidence level is higher than a preset threshold. The extended Kalman filter is then used to recursively estimate the model to obtain the current wear level and remaining life prediction value of the tool to be evaluated. Based on the wear level and remaining life prediction value, the health status of the tool to be evaluated is assessed, and corresponding operation and maintenance control signals are generated.

[0022] To address the technical problems of unreliable tool model identification and inaccurate extraction of wear geometry features due to poor 3D reconstruction quality in complex industrial environments such as high reflectivity, oil contamination, and partial obstruction of metal cutting tools, which ultimately prevent the achievement of robust wear quantification and remaining life prediction, this application provides a tool management method for large gantry milling machines based on vision recognition. This method includes: projecting an coded light pattern onto the tool to be evaluated through a structured light projection unit and acquiring structured light images from multiple non-coplanar viewpoints. These structured light images include co-polarized images and orthogonally polarized images. The co-polarized image is an image where the camera polarization direction is parallel to the projection light polarization direction, and the orthogonally polarized image is an image where the camera polarization direction is orthogonal to the projection light polarization direction. Based on the structured light images and known coded light pattern information, phase calculation and triangulation are performed to generate a 3D geometric model of the tool to be evaluated. The structured geometric features of the tool to be evaluated are extracted from the 3D geometric model, including the tool holder diameter, the helix angle of the cutting edge, etc. The system analyzes the cutting length and wear geometry parameters. It matches structured geometric features with a pre-stored standard tool template library to determine the tool model and its identification confidence level. When the identification confidence level exceeds a preset threshold, it constructs a nonlinear state-space model using the wear geometry parameters and cumulative usage time of the tool as state variables. An extended Kalman filter is then used to recursively estimate this model, obtaining the current wear level and predicted remaining life of the tool. Based on the wear level and predicted remaining life, the system assesses the health status of the tool and generates corresponding maintenance control signals. Using polarized structured light multi-view imaging, it effectively suppresses surface reflection and oil contamination of metal tools, achieving high-precision 3D reconstruction. A multi-dimensional matching mechanism between geometric features and the standard template library ensures reliable tool model identification and outputs a quantified confidence level. Only under high confidence conditions is a mechanism-driven nonlinear state-space model constructed by fusing wear parameters and usage time, and an extended Kalman filter is used to dynamically estimate the wear level and remaining life. This method achieves closed-loop management from perception and identification to prediction, significantly improving the accuracy, robustness, and engineering practicality of tool health assessment, providing a scientific basis for intelligent operation and maintenance, and effectively avoiding machining accidents and unplanned downtime.

[0023] like Figure 2 As shown in the embodiments of this application, the tool management method for a large gantry milling machine based on vision recognition includes: S201. Project an coded light pattern onto the tool to be evaluated through a structured light projection unit, and acquire structured light images from multiple non-coplanar viewpoints.

[0024] In this system, multiple non-coplanar industrial cameras and structured light projection units are arranged in the tool changing area and near the spindle of the gantry milling machine. The structured light images include co-polarized images and orthogonally polarized images; co-polarized images are those where the camera polarization direction is parallel to the projection light polarization direction, and orthogonally polarized images are those where the camera polarization direction is orthogonal to the projection light polarization direction. In some implementations, the structured light projection unit collaborates with multiple industrial cameras to form a multi-view active vision sensing array. Specifically, the structured light projection unit projects an coded light pattern (such as a combination of Gray code and phase-shifted sinusoidal fringes) onto the tool to be evaluated, located in the tool changing area or near the spindle of a gantry milling machine. Simultaneously, non-coplanar industrial cameras positioned at different locations acquire images of the modulated light fringes. Through multi-view geometric constraints and phase unwrapping algorithms, the problems of local overexposure, phase jumps, and data loss caused by the high reflectivity of the tool's metal surface can be effectively overcome, thus providing complete and high-precision raw observation data for subsequent 3D reconstruction.

[0025] It should be noted that the placement of the industrial cameras and structured light projection units is not arbitrary, but optimized based on the mechanical structure, tool change path, and typical tool dimensions of the large gantry milling machine. They are typically distributed in a ring or hemispherical shape around the tool changer arm to ensure that at least three non-coplanar viewing angles simultaneously cover the tool cutting edge and tool holder area. Simultaneously, a sufficiently large baseline angle (generally not less than 15°) is maintained between the optical axis of each camera and the projection direction to improve triangulation accuracy. Furthermore, all imaging devices integrate hardware trigger interfaces, with a central controller issuing microsecond-level synchronization signals to ensure strict consistency in the acquisition time of multi-view images, avoiding motion blur or phase mismatch caused by tool micro-vibration or coolant flow. This sensor layout scheme balances the spatial constraints of the industrial environment with the requirements of measurement robustness, and is a key hardware foundation for achieving highly reliable online tool identification.

[0026] S202. Based on the structured light image and the known coded light pattern information, perform phase calculation and triangulation to generate a three-dimensional geometric model of the tool to be evaluated.

[0027] S203. Extract the structured geometric features of the tool to be evaluated from the three-dimensional geometric model.

[0028] The process involves extracting the structured geometric features of the tool to be evaluated from the three-dimensional geometric model. These features include the shank diameter, helix angle, effective cutting length, and wear geometry parameters. The wear geometry parameters include the width of the wear band on the flank face and the radius of curvature of the tool tip.

[0029] It should be noted that steps S202 and S203 constitute a tightly coupled geometric perception closed loop: the quality of 3D point cloud reconstruction directly determines the extraction accuracy of structured geometric features, while the robustness of feature extraction, in turn, verifies the reliability of the reconstruction results. To address interference from factors such as metal reflection, oil contamination, and local occlusion on the tool surface in large gantry milling machine environments, this solution introduces a polarization imaging and weighted Poisson fusion strategy in the reconstruction stage, and employs an analytical method based on differential geometry and prior tool knowledge in the feature extraction stage, avoiding reliance on data-driven models. Furthermore, "structured geometric features" are not arbitrary geometric quantities, but specifically refer to a set of key parameters with clear physical meaning and compatibility with tool model standards (such as ISO 13399), including tool holder diameter, helix angle, effective cutting length, tool tip curvature radius, and flank face geometry, ensuring interpretability and manufacturing consistency in subsequent matching and recognition.

[0030] S204. Match the structured geometric features with the pre-stored standard tool template library to determine the tool model to be evaluated and the confidence level of identification.

[0031] S205. When the identification confidence level is higher than the preset threshold, the wear geometry parameters and cumulative usage time of the tool to be evaluated are used as state variables to construct a nonlinear state-space model, and the extended Kalman filter is used to recursively estimate the model to obtain the current wear level and remaining life prediction value of the tool to be evaluated.

[0032] S206. Based on the wear level and remaining life prediction, assess the health status of the tool to be evaluated and generate corresponding operation and maintenance control signals.

[0033] Specifically, the tool model, wear level, and predicted remaining life of the tool to be evaluated are sent to the machine tool controller. Based on the allowable wear threshold and minimum remaining life corresponding to the tool model, the machine tool controller determines whether the current tool is in a healthy state. If yes, no action is taken; otherwise, a tool change command or maintenance warning signal is generated.

[0034] Based on the above technical solutions, the tool management method for large gantry milling machines based on vision recognition provided in this application addresses the challenges posed by the complex machining environment of large gantry milling machines, where metal tool surfaces often suffer from strong reflections, oil stains, and partial obstructions. This makes it difficult for traditional methods to accurately identify tool models and reliably assess their wear status. Existing technologies either rely on easily interfered single sensor signals or employ static threshold judgments lacking physical basis, which are insufficient to meet the demands of high-precision intelligent manufacturing. To address this issue, this solution proposes a complete closed-loop technical approach: It utilizes polarized structured light to acquire co-polarized and orthogonally polarized images from multiple perspectives, effectively suppressing specular reflection, and reconstructs a high-fidelity 3D geometric model through phase fusion. Based on this, it extracts standard-compliant structured geometric features (such as tool holder diameter and helix angle), and combines weighted matching, discrimination analysis, and topological consistency verification to generate a recognition confidence level of 0 to 1, ensuring reliable model identification. Only when the confidence level meets the standard is a nonlinear state-space model based on the Archard wear mechanism constructed using wear parameters and cumulative usage time as state variables. Dynamic recursive estimation is then achieved through extended Kalman filtering, thereby obtaining physically meaningful wear levels and remaining life predictions. This solution deeply integrates anti-interference perception, geometric analysis, and mechanism modeling, significantly improving the accuracy, robustness, and interpretability of tool health assessment, providing a scientific basis for intelligent operation and maintenance, and effectively avoiding unplanned downtime and machining quality accidents.

[0035] In one possible implementation of the embodiments of this application, such as Figure 3 As shown, the above S202 can be specifically implemented through the following S301-S303, which are explained in detail below: S301. Based on the orthogonal polarization image, calculate the absolute phase map under each viewpoint, and in the region where the phase position confidence is lower than the preset confidence threshold, fuse the phase information of the same polarization image in a weighted manner to generate the fused phase map of each viewpoint.

[0036] In some implementations, absolute phase map calculation uses algorithms (such as Fourier transform profilometry, phase shifting method, etc.) to extract the absolute phase distribution map from orthogonal polarization images at each viewpoint.

[0037] Confidence assessment and processing: For the phase information of each pixel, the confidence level is determined based on its phase stability or noise level. If the confidence level of a certain region is lower than a set threshold, the data reliability of that region is considered insufficient.

[0038] Weighted fusion strategy: Introduce phase information from co-polarized images into low-confidence regions and use weighted averaging or other optimization algorithms for data fusion to improve the accuracy and robustness of the final phase map.

[0039] S302. Calculate the initial 3D point cloud based on the fused phase map from each viewpoint and the camera calibration parameters.

[0040] In some implementations, camera calibration requires precise calibration of all cameras involved in the shooting process, including internal parameters (such as focal length, principal point coordinates, distortion coefficients, etc.) and external parameters (such as the relative positions and attitudes between cameras). This step can be accomplished using Zhang's calibration method or other advanced calibration techniques.

[0041] Stereo matching: Using the fused phase maps obtained from various viewpoints, a stereo matching algorithm (such as region-based correlation matching or feature point matching) is executed to find corresponding points and build a disparity map.

[0042] 3D Reconstruction: Combining camera calibration parameters and parallax information, the spatial coordinates of each point are calculated using the principle of triangulation, thereby constructing an initial 3D point cloud model of the target object.

[0043] It should be noted that some post-processing operations may be required on the generated 3D point cloud, such as denoising, surface smoothing, and hole filling, in order to improve the quality and usability of the model.

[0044] S303. Perform multi-view registration on the initial 3D point cloud and construct point cloud weights based on local polarization difference and surface gradient stability indices; use a weighted Poisson surface reconstruction algorithm to generate a 3D geometric model of the tool to be evaluated.

[0045] The method for constructing point cloud weights is as follows: Based on the pixel intensity in orthogonal polarization images and co-polarization images, calculate the local polarization difference of each 3D point cloud point. The gradient magnitude of the orthogonal polarization image at this point is calculated as an index of surface gradient stability. Based on the local polarization difference degree and surface gradient stability index, the first weight factor and the second weight factor are generated by monotonically decreasing and monotonically increasing mapping respectively, and the two are multiplied to obtain the point cloud weight of the point. Among them, the smaller the local polarization difference and the higher the surface gradient stability, the greater the point cloud weight.

[0046] It should be noted that the weighted Poisson surface reconstruction algorithm significantly improves the 3D reconstruction quality of metal tools under conditions of high reflectivity, oil contamination, or partial occlusion by introducing local confidence weights from point cloud data. Specifically, each 3D point cloud point is assigned a weight between 0 and 1, which comprehensively reflects its reliability in polarization imaging: on the one hand, the local polarization difference is calculated based on orthogonal and co-polarized images; the smaller the difference (indicating weaker specular reflection), the higher the weight. On the other hand, the surface fringes sharpness is evaluated by combining the gradient magnitude at that point in the orthogonal polarized image; the larger the gradient (the more stable the phase solution), the higher the weight. These two indicators are fused into the final point-level weight after nonlinear mapping and used to modulate the contribution intensity of the normal vector field in the Poisson equation—high-weight regions dominate isosurface generation, while low-weight regions (such as overexposed or blurred areas) are automatically suppressed. The resulting 3D geometric model not only fully preserves key structures such as the cutting edge, helical groove, and flank face, but also effectively avoids the "artifacts," "holes," or "oversmoothing" problems commonly found on metal surfaces in traditional uniform reconstruction methods, providing a high-fidelity geometric basis for subsequent tool model identification and wear quantification.

[0047] Based on the above technical solutions, in the complex industrial environment of large gantry milling machines, metal tool surfaces commonly suffer from strong reflections, oil contamination, and localized occlusion, leading to phase calculation distortion or failure in traditional structured light 3D reconstruction, severely affecting point cloud integrity and geometric accuracy. Therefore, it is urgent to solve the technical challenge of acquiring high-fidelity 3D models under interference conditions. This solution introduces orthogonal polarization images to suppress specular reflections and combines them with co-polarization images to supplement information in weakly textured areas, achieving phase map fusion and significantly improving data reliability. Furthermore, point cloud weights are constructed based on local polarization differences and surface gradient stability, enabling the Poisson reconstruction algorithm to automatically suppress the influence of low-quality areas. This technical solution cleverly utilizes polarization characteristics and geometric priors to achieve a closed-loop processing of "anti-interference acquisition—robust fusion—intelligent weighting," effectively overcoming the interference of high reflectivity and oil contamination on reconstruction, and improving the geometric integrity of key areas such as the cutting edge and flank face. This provides a high-precision foundation for subsequent wear parameter extraction and life prediction, demonstrating good industrial applicability and promotional value.

[0048] In one possible implementation of this application embodiment, the above-mentioned S203 can be specifically described as follows: Curvature analysis was performed on the reconstructed three-dimensional geometric model of the tool to be evaluated.

[0049] Among them, the region with curvature less than the preset curvature threshold is marked as the candidate tool holder region; the random sampling consensus algorithm is used to perform cylindrical fitting on the candidate tool holder region to obtain the spindle direction and tool holder diameter of the tool to be evaluated.

[0050] A local coordinate system for the tool to be evaluated is established with the spindle direction as the reference, and the point cloud is segmented and projected along the axis of the coordinate system. The periodicity of the gradient signal caused by the spiral groove in the projected image is analyzed, the starting axial coordinate and ending axial coordinate of the spiral groove are identified, the distance between the two coordinates is calculated, and the effective cutting length is obtained.

[0051] In some implementations, the periodicity of the gradient signal is achieved by: projecting the point cloud along the principal axis onto a vertical plane to obtain a grayscale projection image I(z) (where z is the axial coordinate); calculating the first derivative of I(z) along the z-direction to obtain the gradient signal G(z); performing a fast Fourier transform on G(z) to identify the dominant frequency f0; and then determining the wavelength corresponding to f0. Determine the spiral groove period; take the position where the gradient amplitude first increases significantly as the starting point and the position where it last decreases significantly as the ending point.

[0052] Effective cutting length: Orthogonally project the point cloud along the tool spindle direction to generate a grayscale projection image; average this projection image circumferentially to obtain a one-dimensional axial intensity signal; differentiate this signal to obtain a gradient signal; detect the first and last local maxima positions in the gradient signal where the amplitude exceeds a threshold T, and use these as the starting positions of the cutting edge. and termination position Then the effective cutting length .

[0053] Multiple equally spaced cross-sectional profiles are extracted within the cutting edge region, and the helix angle of the cutting edge is calculated by fitting a helical model.

[0054] In some implementations, the calculation of the helix angle of the cutting edge is achieved through the following steps: Within the identified cutting edge region, multiple cross-sections perpendicular to the spindle are cut along the tool spindle direction at fixed intervals (e.g., 0.5 mm or adaptively set according to the tool diameter), and the set of cutting edge points on each cross-section is extracted as the cross-sectional profile; each profile is subjected to circle fitting or Gauss-Newton optimization to determine its center coordinates and radius, thereby obtaining a series of discrete cutting edge spatial points; these points are projected onto a cylindrical coordinate system (with the tool spindle as the axis) to obtain the corresponding circumferential angle. Axial position The sequence; a linear relationship is fitted using the least squares method. And based on the slope Calculate the helix angle ;in, The average cutting edge radius, The initial circumferential angle is given. This method fully utilizes the geometric periodicity of the helical groove, and even under conditions of local wear or partial obstruction, it can still robustly estimate the helical angle through redundant information from multiple cross sections, avoiding the problem of single-section fitting being susceptible to noise interference, and significantly improving the accuracy of tool type identification.

[0055] A quadratic surface is fitted in a local neighborhood of the cutting edge tip, and the radius of curvature of the cutting edge tip is calculated based on the principal curvature of the surface.

[0056] In some implementations, the calculation process for the blade tip curvature radius is as follows: The cutting edge ridge is located in the 3D geometric model, and its foremost point is selected as the blade tip center. A local spherical neighborhood is constructed with this center point as the sphere center and a physical radius (e.g., 0.2–0.5 mm) is set. All point cloud data within this neighborhood are extracted. A tangent plane is fitted within this neighborhood, and a local orthogonal coordinate system is established. The point cloud is then transformed to a local coordinate system centered on the blade tip, allowing the surface to be approximated as a height function. Based on this, a weighted least squares method is used to fit a quadratic surface model. The weights decrease Gaussianly with the distance from the point to the blade tip to enhance the dominance of the blade tip region. Furthermore, a Hessian matrix is ​​constructed from the fitting coefficients, and two principal curvatures are obtained through eigenvalue decomposition. and Based on the principles of differential geometry, the radius of curvature of the blade tip is defined as the harmonic average of the radii of curvature corresponding to the two principal curvatures, i.e. This method can effectively quantify the degree of edge dulling caused by wear, and has strong robustness to point cloud noise and local defects, making it suitable for on-machine vision inspection of tools in large gantry milling machines.

[0057] Based on the surface normal vector distribution and cutting edge position of the three-dimensional geometric model, the local area located on the non-rake face side of the cutting edge and whose normal vector direction forms a preset wedge angle range with the tool spindle is the flank face region; the distance deviation from each point in the flank face region to the ideal cutting edge reference plane is calculated to obtain the wear geometric parameters.

[0058] The method for obtaining the ideal cutting edge reference plane is as follows: Based on the model of the tool to be evaluated, the theoretical wedge angle and helix angle corresponding to the tool to be evaluated are obtained from the standard tool database; Extract the cutting edge ridge line from the three-dimensional geometric model and determine a reference point on the cutting edge; A local right-handed coordinate system is constructed with the reference point as the origin, the cutting edge tangential direction as the first axis, and the tool spindle direction as the third axis. In the local coordinate system, the normal vector of the ideal back face is calculated based on the theoretical wedge angle and helix angle; The ideal cutting edge reference plane is constructed using the reference point and the normal vector.

[0059] The tool holder diameter, cutting edge helix angle, effective cutting length, tool tip curvature radius, and wear geometry parameters are combined into a structured geometric feature vector.

[0060] In some implementations, the identification of the flank face region and the extraction of wear geometry parameters specifically include the following steps: calculating the surface normal vector of each point from the reconstructed 3D geometric model, and determining the cutting edge position based on high curvature ridge detection; within the cutting edge neighborhood, excluding the set of points located on the rake face side (usually facing the chip discharge direction), retaining only the candidate region on the non-rake face side; within this candidate region, further filtering points whose normal vector direction and the angle between the tool spindle and the point are within a preset wedge angle range (e.g., 6°-15°, according to ISO standards or tool model parameters), and identifying them as the flank face region; retrieving the theoretical wedge angle according to the identified tool model, and constructing an ideal cutting edge reference plane by combining the cutting edge reference point and the helix angle; calculating the signed vertical distance from each point in the flank face region to this ideal plane, and statistically analyzing points whose absolute distance value exceeds a micrometer threshold, defining the maximum projection span along the spindle direction as the width of the flank face wear band, or integrating all deviations to characterize the overall wear volume trend, and this quantification result is the wear geometry parameter used for tool condition assessment. This method does not rely on unworn samples and can achieve highly robust wear quantification through geometric priors and physical constraints, making it suitable for online monitoring scenarios in industrial settings.

[0061] Based on the above technical solutions, in the management of tools for large gantry milling machines, traditional methods rely on manual experience or simple geometric feature identification, which is insufficient to cope with the problems of diverse tool models and variable wear states under complex working conditions. Therefore, it is urgent to solve the technical challenge of how to automatically and accurately extract structured geometric parameters from a 3D model to achieve reliable identification and wear assessment. This solution locates candidate tool holder regions through curvature analysis and uses the Random Sample Consensus (RANSAC) algorithm to fit the cylindrical surface, accurately obtaining the spindle direction and tool holder diameter. A local coordinate system is established based on the spindle, and a point cloud is projected along the axial direction. The periodic gradient signal caused by the helical groove is used to identify the starting and ending axial coordinates, thereby calculating the effective cutting length. A helical model is fitted with an equally spaced cross-sectional profile to accurately calculate the helix angle of the cutting edge. The radius of curvature of the cutting edge is calculated by combining the principal curvature of the surface, and the wear band width is statistically calculated based on the distance deviation from the flank face to the ideal reference plane. This technical solution fully utilizes the physical consistency and topological regularity of three-dimensional geometric information to achieve high-precision and automated extraction of key structural parameters. It avoids the dependence on image quality in traditional methods, significantly improves the robustness of tool model recognition and the accuracy of wear assessment, and provides a reliable geometric representation basis for intelligent tool management systems.

[0062] In one possible implementation of this application embodiment, the confidence level acquisition method in S204 above can be specifically implemented through the following S401-S403, which are described in detail below: S401. Perform weighted similarity calculation between the structured geometric features and the features of each candidate template in the standard tool template library to obtain the distance metric for each candidate model. S402. Evaluate matching discrimination based on the ratio of the minimum distance to the second smallest distance; S403. Simultaneously verify whether the three-dimensional geometric model satisfies the topological constraints of the candidate model; By combining distance metric, matching discrimination, and topology consistency, a recognition confidence value between 0 and 1 is generated.

[0063] In some implementations, the generation process of recognition confidence integrates three criteria: statistical similarity, matching uniqueness, and geometric rationality. The extracted structured geometric features (such as tool holder diameter, helix angle, and effective cutting length) are weighted and similar to each candidate model in the standard tool template library. The weights are dynamically allocated based on the sensitivity of each feature to model discrimination (e.g., tool holder diameter has a higher weight than wear-related features), and Mahalanobis distance, which considers manufacturing tolerance covariance, is used as the distance metric. The ratio between the first candidate model corresponding to the minimum distance and the second candidate model corresponding to the second minimum distance is calculated. If this ratio is close to 1, it indicates that the matching result is ambiguous and has low discriminative power. Simultaneously, the system automatically verifies whether the 3D geometric model satisfies the topological constraints of the first candidate model, including prior rules such as the number of cutting edges, number of grooves, helix direction, and symmetry. If a conflict exists (e.g., a 4-cutting edge is detected but the candidate has 2 cutting edges), the confidence is forcibly reduced or the candidate is excluded. The information from the above three dimensions is fused into a recognition confidence value between 0 and 1 using a nonlinear mapping function (such as Sigmoid or piecewise linear combination), which is then used for subsequent decision threshold judgment.

[0064] It's important to note that the "topological constraint" doesn't rely on image semantic segmentation or deep learning classification, but rather on rigid rules based on the inherent geometry of the cutting tool. For example, the number of cutting edges on an end mill must be an integer and correspond one-to-one with the helical grooves, and the insert mounting positions on a face mill have periodic symmetry. Furthermore, this confidence mechanism effectively avoids mismatches caused by localized wear, occlusion, or reconstruction noise. Even if a feature (such as the tool tip curvature radius) shifts due to wear, high confidence can be maintained as long as the core geometric parameters (such as diameter, helix angle, and topology) remain consistent. Conversely, if key parameters are ambiguous or multiple candidates compete fiercely, the confidence drops significantly, triggering manual review or rejection. This design significantly improves the reliability and safety of the automatic tool identification system in complex industrial environments.

[0065] In one possible implementation of this application embodiment, the above-mentioned S205 can be specifically implemented by the following S501-S504, which will be described in detail below: S501. After the confidence level of tool model identification is higher than the preset threshold, define a state vector. The state vector includes the width of the flank wear band and the cumulative effective cutting time calculated by the three-dimensional geometric model.

[0066] S502. Based on Arcard's wear law, and combined with the real-time acquired normal cutting force, spindle speed, and chip temperature estimated from spindle power, a state transition equation containing an exponential thermal softening term is constructed to describe the nonlinear evolution of the wear band width on the flank face with the machining process.

[0067] In some implementations, the specific construction process of the state transition equation is as follows: Based on Arcard's wear law, the increment of the wear band width on the flank face is modeled as a function related to the normal cutting force, spindle speed, and material hardness; considering that high temperature will significantly reduce the hardness of the tool material, the chip temperature estimated in real time by the spindle power is introduced, and the thermal softening effect is characterized by an Arrhenius-type exponential term, resulting in the discrete-time state transition equation: It should be noted that the chip temperature is not directly measured by an infrared sensor, but is estimated online based on the difference between the spindle power and the theoretical no-load power, combined with the empirical thermal efficiency coefficient. This avoids the engineering challenge of installing high-temperature sensors in the confined tool-changing area. Furthermore, as the cutting temperature increases, the tool material hardness decreases, and the wear rate accelerates non-linearly. This mechanism allows the model to accurately reflect the different wear dynamics characteristics under roughing (high temperature, high speed) and finishing (low temperature, low speed). This mechanism-driven modeling approach is significantly superior to purely data-driven black-box models, possessing strong generalization ability and interpretability.

[0068] S503. Using the width of the wear band on the back face measured in real time by the three-dimensional vision system as the observation, establish the identity mapping relationship between the observed value and the corresponding element in the state vector to form the observation equation.

[0069] The identity mapping relationship between corresponding elements in the state vector is mapped to form the observation equation; the corresponding elements are analyzed to the width of the surface wear zone, and the uniqueness of the material and the empirical thermal efficiency coefficient are evaluated online; the low temperature constant is set, and the material hardness decreases; In some implementations, the construction of the observation equation specifically includes: extracting the point cloud of the flank face region from the three-dimensional geometric model, and calculating the vertical distance from each point to the plane based on the ideal cutting surface reference plane corresponding to the confirmed tool model; projecting the distance value along the tool spindle direction and calculating the maximum effective width as the flank face wear band width at the current moment, and establishing an identity mapping relationship with the wear state variable in the state vector to obtain the observation equation.

[0070] It should be noted that "identity mapping" does not assume error-free observation, but rather that the observations directly correspond to the state variables themselves mathematically. The uncertainty is characterized by observation noise and dynamically adjusted through the covariance matrix in the extended Kalman filter. Furthermore, to improve robustness, the system only outputs valid observations when the visual reconstruction quality meets the standards (e.g., point cloud density and normal consistency meet thresholds); otherwise, it marks them as invalid frames and skips updates, thus avoiding contamination of state estimation by low-quality geometric data. This direct observation mechanism based on high-fidelity visual geometry is a key prerequisite for achieving high-precision online wear state estimation.

[0071] S504. Combine the state transition equation with the observation equation to form a nonlinear state-space model, and configure an extended Kalman filter to perform recursive state estimation based on this model, outputting the current wear level and remaining life prediction value.

[0072] Based on the above technical solutions, in the scenario of automatic tool identification for large gantry milling machines, due to the large variety of tool types, similar geometric parameters, and the fact that the actual 3D model is often affected by wear, occlusion, or reconstruction noise, relying solely on single feature matching can easily lead to misclassification of tool types, thus causing subsequent wear assessments to lose their benchmark. Therefore, it is urgent to solve the technical problem of how to achieve highly reliable tool type identification and quantify its credibility under the presence of interference and uncertainty. This solution proposes a multi-dimensional confidence assessment mechanism that integrates statistical similarity, matching uniqueness, and geometric rationality: weighted similarity calculation highlights the discriminative role of key features (such as tool holder diameter); the ratio of the minimum and second minimum distances is introduced to determine whether the matching has significant distinguishability, avoiding fuzzy matching; at the same time, topological constraints (such as the number of cutting edges and symmetry) are embedded for hard verification to eliminate physically unreasonable candidates. The three factors work together to generate an identification confidence level of 0 to 1, providing a clear threshold basis for subsequent decision-making. This technical solution effectively overcomes the shortcomings of traditional nearest neighbor matching, which is susceptible to local distortion interference. It significantly improves identification robustness and system security, ensuring that model-based wear analysis is only enabled under high confidence conditions, thereby guaranteeing the reliability and engineering practicality of the entire tool health management process.

[0073] Some of the data in the above formula are calculated by removing dimensions and taking their numerical values. The formula is the closest to the real situation obtained by software simulation of a large amount of collected data. The preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

Claims

1. A tool management method for large gantry milling machines based on vision recognition, characterized in that, include: A coded light pattern is projected onto the tool to be evaluated using a structured light projection unit, and structured light images from multiple non-coplanar viewpoints are acquired. The structured light images include co-polarized images and orthogonal polarized images. The co-polarized images are those in which the camera polarization direction is parallel to the projection light polarization direction, and the orthogonal polarized images are those in which the camera polarization direction is orthogonal to the projection light polarization direction. Based on the structured light image and the known coded light pattern information, phase calculation and triangulation are performed to generate a three-dimensional geometric model of the tool to be evaluated. The structured geometric features of the tool to be evaluated are extracted from the three-dimensional geometric model. The structured geometric features include the tool holder diameter, the helix angle of the cutting edge, the effective cutting length, and the wear geometry parameters. The structured geometric features are matched with a pre-stored standard tool template library to determine the tool model to be evaluated and the identification confidence level. When the identification confidence level is higher than the preset threshold, the wear geometry parameters and cumulative usage time of the tool to be evaluated are used as state variables to construct a nonlinear state-space model, and the extended Kalman filter is used to recursively estimate the model to obtain the current wear level and remaining life prediction value of the tool to be evaluated. Based on the wear level and remaining life prediction, the health status of the tool to be evaluated is assessed, and corresponding operation and maintenance control signals are generated.

2. The tool management method for a large gantry milling machine based on vision recognition according to claim 1, characterized in that, The method for reconstructing the three-dimensional geometric model of the tool to be evaluated includes: Based on the orthogonal polarization image, the absolute phase map of each viewpoint is calculated, and in the region where the phase position confidence is lower than the preset confidence threshold, the phase information of the same polarization image is fused in a weighted manner to generate a fused phase map of each viewpoint. The initial 3D point cloud is calculated based on the fused phase map from each viewpoint and the camera calibration parameters. The initial 3D point cloud is registered from multiple perspectives, and point cloud weights are constructed based on local polarization difference and surface gradient stability indices. A weighted Poisson surface reconstruction algorithm is used to generate a three-dimensional geometric model of the tool to be evaluated.

3. The tool management method for a large gantry milling machine based on vision recognition according to claim 2, characterized in that, The method for constructing the point cloud weights includes: Based on the pixel intensity in orthogonal polarization images and co-polarization images, calculate the local polarization difference of each 3D point cloud point. The gradient magnitude of the orthogonal polarization image at this point is calculated and marked as a surface gradient stability index; Based on the local polarization difference degree and surface gradient stability index, a first weighting factor and a second weighting factor are generated by monotonically decreasing and monotonically increasing mappings, respectively, and the two are multiplied to obtain the point cloud weight of the point.

4. The tool management method for a large gantry milling machine based on vision recognition according to claim 1, characterized in that, The extraction of the structured geometric features of the tool to be evaluated includes: Curvature analysis is performed on the reconstructed three-dimensional geometric model of the tool to be evaluated; regions with curvature less than a preset curvature threshold are marked as candidate tool holder regions; a random sampling consensus algorithm is used to perform cylindrical fitting on the candidate tool holder regions to obtain the spindle direction and tool holder diameter of the tool to be evaluated; A local coordinate system for the tool to be evaluated is established with the spindle direction as the reference, and the point cloud is segmented and projected along the axis of the coordinate system; the periodicity of the gradient signal caused by the spiral groove in the projected image is analyzed, the starting axial coordinate and ending axial coordinate of the spiral groove are identified, the distance between the two coordinates is calculated, and the effective cutting length is obtained. Multiple equally spaced cross-sectional profiles are extracted within the cutting edge region, and the helix angle of the cutting edge is calculated by fitting a helical model. The wear geometry parameters include the tip curvature radius and the width of the flank wear band; the tip curvature radius is calculated based on the principal curvature of the surface; the flank wear band width is obtained by calculating the distance deviation from the actual flank area to the ideal cutting surface reference plane and projecting it along the tool spindle direction. The tool holder diameter, cutting edge helix angle, effective cutting length, tool tip curvature radius, and wear geometry parameters are combined into a structured geometric feature vector.

5. The tool management method for a large gantry milling machine based on vision recognition according to claim 4, characterized in that, The method for obtaining the ideal cutting edge reference plane includes: Based on the model of the tool to be evaluated, the theoretical wedge angle and helix angle corresponding to the tool to be evaluated are obtained from the standard tool database; Extract the cutting edge ridge line from the three-dimensional geometric model and determine a reference point on the cutting edge; A local right-handed coordinate system is constructed with the reference point as the origin, the cutting edge tangential direction as the first axis, and the tool spindle direction as the third axis. In the local coordinate system, the normal vector of the ideal back face is calculated based on the theoretical wedge angle and helix angle; An ideal cutting edge reference plane is constructed using the reference point and the normal vector.

6. The tool management method for a large gantry milling machine based on vision recognition according to claim 4, characterized in that, The method for obtaining the radius of curvature of the blade tip includes: The cutting edge ridge line is determined from the three-dimensional geometric model of the tool to be evaluated, and the foremost point of the cutting edge ridge line is selected as the center point of the cutting tip. With the center point of the blade tip as the center, a local point cloud neighborhood is extracted within a preset physical radius; a tangent plane is fitted within the local point cloud neighborhood, and a local orthogonal coordinate system is established based on the tangent plane, and the point cloud coordinates are transformed to the coordinate system; A weighted least squares method is used to fit a quadratic surface function to obtain the coefficients of the quadratic term characterizing the local surface shape. A Hessian matrix is ​​constructed based on these coefficients, and the two principal curvatures are calculated through eigenvalue decomposition. , Through formula The radius of curvature of the tool tip was calculated. .

7. The tool management method for a large gantry milling machine based on vision recognition according to claim 1, characterized in that, The method for constructing the nonlinear state-space model includes: After the confidence level of tool model identification is higher than a preset threshold, a state vector is defined. The state vector includes the width of the flank wear band and the cumulative effective cutting time calculated by the three-dimensional geometric model. Based on Arcard's wear law, and combined with real-time acquired normal cutting force, spindle speed, and chip temperature estimated from spindle power, a state transition equation containing an exponential thermal softening term is constructed. The width of the wear band on the back face measured in real time by the three-dimensional vision system is used as the observation. An identity mapping relationship is established between the observed value and the corresponding element in the state vector to form the observation equation. The state transition equation is combined with the observation equation to form a nonlinear state-space model.

8. The tool management method for a large gantry milling machine based on vision recognition according to claim 1, characterized in that, The method for obtaining the confidence level includes: The structured geometric features are compared with the features of each candidate template in the standard tool template library to calculate the weighted similarity and obtain the distance metric for each candidate model. Match discrimination is evaluated based on the ratio of the minimum distance to the second minimum distance; Simultaneously verify whether the three-dimensional geometric model satisfies the topological constraints of the candidate model; By combining the distance metric, matching discrimination, and topological consistency, a recognition confidence value between 0 and 1 is generated.

9. The tool management method for a large gantry milling machine based on vision recognition according to claim 1, characterized in that, The assessment of the health status of the tool to be assessed includes: The tool model, wear level, and predicted remaining life value are sent to the machine tool controller. The machine tool controller determines whether the tool is in a healthy state based on the allowable wear threshold and minimum remaining life corresponding to the tool model. If yes, no action is taken; otherwise, a tool change command or maintenance warning signal is generated.

10. A tool management system for a large gantry milling machine based on vision recognition, operating based on the tool management method for a large gantry milling machine based on vision recognition as described in any one of claims 1-9, characterized in that, It includes a data acquisition module, an identification module, and an evaluation module; The acquisition module is used to project an coded light pattern onto the tool to be evaluated through a structured light projection unit and acquire structured light images from multiple non-coplanar viewpoints. The recognition module is used to perform phase calculation and triangulation based on the structured light image and known coded light pattern information to generate a three-dimensional geometric model of the tool to be evaluated. Extract the structured geometric features of the tool to be evaluated from the three-dimensional geometric model; The structured geometric features are matched with a pre-stored standard tool template library to determine the tool model to be evaluated and the identification confidence level. The evaluation module is used to construct a nonlinear state-space model by taking the wear geometry parameters and cumulative usage time of the tool to be evaluated as state variables when the identification confidence level is higher than a preset threshold, and to recursively estimate the model by using an extended Kalman filter to obtain the current wear level and remaining life prediction value of the tool to be evaluated. Based on the wear level and remaining life prediction, the health status of the tool to be evaluated is assessed, and corresponding operation and maintenance control signals are generated.

Images