A complex curved surface polishing track automatic generation method based on point cloud segmentation

Abstract

The application discloses a complex curved surface polishing track automatic generation method based on point cloud segmentation, relates to the technical field of intelligent manufacturing and precision machining, and comprises the following steps: acquiring initial surface data of a curved workpiece, collecting point cloud information through sensor array scanning, performing preliminary filtering processing on a real-time change area, obtaining a preliminary point cloud dataset, identifying key features on the curved surface by using a deep learning algorithm, performing point cloud segmentation on the preliminary point cloud dataset, extracting curvature and roughness distribution features of each curved surface area after segmentation, and obtaining feature extraction results and point cloud segmentation results; the complex curved surface polishing track automatic generation method based on point cloud segmentation is helpful to improving the surface quality, contour accuracy and geometric consistency of complex curved surface machining, reducing machining defects and rework costs, and further improving the automation level, machining efficiency and final forming quality of complex curved surface polishing machining.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and precision machining technology, specifically to a method for automatically generating grinding trajectories for complex curved surfaces based on point cloud segmentation. Background Technology

[0002] Grinding is a crucial machining process in modern manufacturing, used to improve the surface quality, contour accuracy, and geometric consistency of workpieces. It plays a vital role, particularly in aerospace, automotive manufacturing, and precision mold making, where high surface performance and shape accuracy are paramount for components. For complex curved surface workpieces, the surface typically exhibits significant curvature variations, complex local morphology, and irregular machining areas. Therefore, accurately identifying complex curved surface regions and generating appropriate grinding trajectories has always been a key focus in this field.

[0003] Existing methods for grinding complex curved surfaces typically rely on preset trajectories, offline programming, or path planning based on empirical rules. In actual processing, due to manufacturing errors, clamping errors, and surface morphology changes caused by material removal during grinding, simply using a fixed path often fails to meet the actual processing requirements of complex curved surfaces. This can easily lead to insufficient or excessive grinding in certain areas, or poor processing uniformity, thus affecting the surface quality and geometric accuracy of the workpiece. With the development of sensing and digital processing technologies, using point cloud data to characterize the geometric morphology of workpiece surfaces has become an important technique in the processing of complex curved surfaces. By collecting point cloud information from the workpiece surface, the spatial morphological characteristics of the object being processed can be reflected to a certain extent. However, current technologies still have limited processing capabilities for point cloud data of complex curved surfaces. Especially when facing workpiece surfaces with complex curvature variations, unclear regional boundaries, or significant differences in local features, it is often difficult to effectively divide and accurately identify the point cloud data, resulting in insufficiently precise descriptions of the processing area, which in turn affects the rationality and targeting of subsequent grinding trajectory generation. Furthermore, grinding complex curved surfaces requires not only identifying the current surface state of the workpiece but also adjusting and optimizing the grinding path based on the geometric differences in different regions. Without effective extraction and classification of surface region features, relying solely on a single rule for trajectory planning makes it difficult to consider the processing needs of different regions, easily leading to a mismatch between the trajectory and the actual surface shape, reducing processing efficiency and quality. In addition, the workpiece surface continuously changes during actual grinding. If the post-processed surface state cannot be obtained and analyzed in a timely manner, it is difficult to determine whether the current grinding effect meets the expected requirements, and it is also difficult to provide a reliable basis for subsequent trajectory adjustments. Especially in high-precision machining scenarios, insufficient restoration of the geometric shape of the processing area, surface quality assessment, and feedback of processing results may lead to error accumulation, ultimately affecting the overall forming quality and consistency of complex curved surface workpieces. Summary of the Invention

[0004] The purpose of this invention is to provide an automatic generation method for grinding trajectories of complex curved surfaces based on point cloud segmentation, thereby solving the problems existing in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for automatically generating grinding trajectories for complex curved surfaces based on point cloud segmentation, comprising: S1. Obtain the initial surface data of the curved workpiece, collect point cloud information by scanning the sensor array, perform preliminary filtering on the real-time changing area, and obtain the preliminary point cloud dataset. S2. Use deep learning algorithms to identify key features on the surface, and perform point cloud segmentation on the preliminary point cloud dataset. Extract the curvature and roughness distribution features of each surface region after segmentation to obtain feature extraction results and point cloud segmentation results. S3. Based on the feature extraction results and point cloud segmentation results, generate a dynamic perception model, optimize the path for real-time changes, group the segmented similar curved surface regions, and generate an optimized polishing trajectory plan. S4. Obtain the current processing position data from the optimized grinding trajectory planning, combine it with the surface change information fed back by the sensor in real time, and determine if the change exceeds the preset threshold. Then adjust the grinding parameters to obtain the updated surface state. S5. Construct a 3D reconstruction model, fuse multiple frames of data through point cloud registration technology, determine the precise geometric shape, and obtain the restored geometric data; S6. Calculate the surface roughness and deviation index using statistical analysis methods. If the index meets the preset standard, the processed area is confirmed to be qualified and a quality assessment report is obtained. S7. The quality assessment report is fed back into the dynamic perception model to determine the final processing accuracy improvement and obtain the complete grinding process optimization results.

[0006] Preferably, S1 includes: The sensor array collects spatial coordinate data of the curved workpiece surface, and maps it to a unified three-dimensional coordinate system through a spatial coordinate transformation matrix to obtain the first point cloud dataset. The local curvature change rate is calculated based on the first point cloud dataset. If the local curvature change rate exceeds a preset threshold, the corresponding region is marked as a dynamically perceived region to obtain the second point cloud dataset. A bilateral filtering algorithm is used to remove outlier noise points from the second point cloud dataset to obtain the third point cloud dataset; A local surface normal vector field is constructed based on the third point cloud dataset, and redundant points are eliminated through normal vector consistency constraints to obtain the fourth point cloud dataset. The local surface equation is fitted based on the fourth point cloud dataset, and points whose Euclidean distance exceeds a preset threshold are removed to obtain the fifth point cloud dataset. Based on the cloud dataset in point 5, a dynamic perception model is constructed, and the initial surface data of complex curved workpieces is dynamically perceived and adjusted by updating the model parameters in real time.

[0007] Preferably, S2 includes: Multi-scale feature extraction is performed on the original point cloud using a convolutional neural network to obtain depth features on the curved surface; The point cloud topology is constructed using deep features, and the point cloud is divided into regions based on the density of the topology connections to obtain segmented point clouds with region labels. The local normal vectors of the segmentation points are calculated using cloud computing, thereby determining the discrete curvature of each surface region; Based on the fluctuation of discrete curvature in spatial distribution, the surface roughness of each surface region is extracted to form a set of geometric properties reflecting surface quality; If the values ​​in the geometric attribute set exceed the preset path constraints, the feed rate of the machining trajectory is adjusted to obtain the real-time analysis results of the machining path. By summarizing depth features, segmented point clouds, discrete curvature, and surface roughness, the system outputs feature extraction results and point cloud segmentation results, enabling real-time analysis of the processing path.

[0008] Preferably, S3 includes: The raw point cloud data is acquired and geometric features are extracted. Multiple segmented surfaces are obtained through point cloud segmentation. The segmented surface is grouped by similarity using a clustering algorithm to determine cluster regions with consistent physical properties; The real-time displacement data generated by the clustered regions during the processing is input into the dynamic sensing model to obtain the sensing state reflecting the deformation of the workpiece; The path is optimized by combining the perceived state with the preset processing accuracy requirements. If the perceived state deviates from the preset threshold, the optimized path that avoids obstacles and fits the surface is recalculated. The optimized path is transformed into a grinding trajectory plan through inverse kinematics, enabling real-time adjustments during the grinding operation and achieving automatic generation and optimization of the grinding trajectory under the guidance of a dynamic perception model.

[0009] Preferably, S4 includes: The trajectory planning coordinate data and sensor data are acquired and spatially mapped to obtain the surface roughness value of the current processing point; If the surface roughness value exceeds the preset surface morphology threshold, the correction amount for the current grinding pressure and feed speed is determined. The correction amount of the current grinding pressure and feed speed is applied to the grinding actuator to obtain the grinding trajectory and machining torque data; By performing a differential calculation between the grinding trajectory and processing torque data and the initial geometric model, a geometric deviation distribution map of the current processed surface is obtained. Based on the geometric deviation distribution map of the current processed surface, the final geometric shape restoration data is obtained through an iterative correction algorithm, thereby realizing geometric shape restoration.

[0010] Preferably, S5 includes: Surface state data is acquired and mapped to a unified global coordinate system to obtain the first spatial point cloud data; Based on the first spatial point cloud data, multi-frame feature point matching is performed to obtain the second spatial point cloud data; A three-dimensional reconstruction model is constructed by fitting a surface to the second spatial point cloud data, and then the third spatial point cloud data is obtained. If the deviation analysis result between the third spatial point cloud data and the preset geometric benchmark exceeds the preset threshold, then local point cloud resampling is performed on the third spatial point cloud data to obtain the fourth spatial point cloud data. Based on the consistency verification of the fourth spatial point cloud data and the determination of the precise geometric shape, the restored geometric data is obtained; Use the restored geometric data to prepare for the quality assessment.

[0011] Preferably, S6 includes: The original point cloud data of the processing area is acquired and the least squares method is used to fit and generate the restored geometric model. The spatial coordinate information of each sampling point on the surface of the restored geometric model is extracted. The normal distance of each sampling point relative to the theoretically designed curved surface is calculated based on the spatial coordinate information, and the deviation index distribution matrix reflecting the surface morphology of the processing area is obtained.

[0012] Preferably, S6 further includes: A sliding window algorithm is used to smooth the local area of ​​the deviation index distribution matrix, extract the surface contour features, and calculate the arithmetic mean roughness value corresponding to the surface contour features. If both the arithmetic mean roughness value and the maximum deviation value in the deviation index distribution matrix are less than the preset standard, the processed area is deemed qualified and a quality assessment report is generated to avoid processing defects.

[0013] Preferably, S7 includes: The quality assessment report is analyzed to extract feature vectors, which are then mapped to a dynamic perception model to obtain a feature distribution matrix. The residual value is calculated based on the feature distribution matrix. If the residual value exceeds the threshold, the cluster center coordinates and cluster radius are updated.

[0014] Preferably, S7 further includes: The updated cluster center coordinates and cluster radius are used to divide the grinding force feedback data, classify the grinding force feedback data into processing state clusters and determine the compensation vector; The compensation vector is loaded into the control system to correct the grinding trajectory. The correction of the grinding trajectory achieves dynamic compensation of the processing accuracy, resulting in a complete optimized grinding process.

[0015] As can be seen from the above technical solution, the present invention has the following beneficial effects: This method for automatically generating grinding trajectories for complex curved surfaces based on point cloud segmentation acquires initial surface point cloud data of complex curved workpieces and combines point cloud segmentation, key feature recognition, and curvature and roughness distribution extraction to effectively characterize the geometric features of different regions of the complex curved surface. This allows for the automatic generation and optimization of grinding trajectories based on the actual morphological differences of each region. Compared to existing methods relying on preset trajectories, offline programming, or empirical rules, this method better adapts to the dynamic changes in surface morphology during the processing of complex curved surfaces, improving the targeting and adaptability of trajectory planning. Furthermore, by introducing real-time feedback on surface changes, dynamic parameter adjustment, 3D reconstruction, point cloud registration and fusion, and surface roughness and deviation index evaluation during the grinding process, it not only promptly corrects processing deviations and reduces the probability of insufficient or excessive grinding in certain areas, but also helps improve the surface quality, contour accuracy, and geometric consistency of complex curved surface processing, reducing processing defects and rework costs. This further enhances the automation level, processing efficiency, and final forming quality of complex curved surface grinding. Attached Figure Description

[0016] Figure 1 This is a signal transmission diagram of the present invention; Figure 2 This is a structural block diagram of the local terminal of an exemplary electronic device of the present invention; Figure 3 This is a structural block diagram of the network terminal of an exemplary electronic device of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] As shown in the figure, this invention provides a technical solution: an automatic generation method for grinding trajectories of complex curved surfaces based on point cloud segmentation, comprising: S1. Acquire initial surface data of the curved workpiece, collect point cloud information by scanning with a sensor array, perform preliminary filtering on real-time changing areas, and obtain a preliminary point cloud dataset for subsequent dynamic sensing and adjustment. S2. Based on the preliminary point cloud dataset, a deep learning algorithm is used to identify key features on the surface, and the preliminary point cloud dataset is segmented into points. The curvature and roughness distribution features of each surface region after segmentation are extracted to obtain feature extraction results and point cloud segmentation results, which are used for real-time analysis of the processing path. S3. Based on the feature extraction results and point cloud segmentation results, generate a dynamic perception model, optimize the path for real-time changes, use a clustering algorithm to group similar curved surface regions after segmentation, and automatically generate an optimized polishing trajectory plan to guide adjustments during the polishing operation. S4. During the grinding operation, the current processing position data is obtained from the optimized grinding trajectory planning. Combined with the surface change information fed back by the sensor in real time, if the change exceeds the preset threshold, the grinding parameters are adjusted to obtain the updated surface state for geometric restoration. S5. Based on the updated surface state, construct a three-dimensional reconstruction model, fuse multiple frames of data through point cloud registration technology, determine the precise geometric shape, and obtain restored geometric data for quality assessment preparation. S6. For the restored geometric data, use statistical analysis methods to calculate the surface roughness and deviation index. If the index meets the preset standard, the processing area is confirmed to be qualified and a quality assessment report is obtained to avoid processing defects. S7. The quality assessment report is fed back into the dynamic perception model to update the clustering algorithm parameters to adapt to subsequent real-time changes, determine the final processing accuracy improvement, and obtain the complete grinding process optimization results.

[0019] In one specific implementation process, the surface condition is continuously acquired, regional differences are identified, and the grinding path is adjusted synchronously during the actual processing of the workpiece, thereby transforming the trajectory generation from a static preset method to a dynamic planning method oriented towards the real-time surface condition.

[0020] In this embodiment, the initial surface data of the curved workpiece is preferably acquired by a sensor array positioned around the grinding station. The sensor array may include a line laser profile sensor, a structured light sensor, an industrial vision camera, or other detection units capable of outputting spatial point cloud information. Its function is to scan the workpiece surface from multiple angles to form raw point cloud data covering the area to be processed. During the data processing stage in this embodiment, preliminary filtering can be performed on real-time changing areas and outliers caused by processing debris, reflective interference, and vibration disturbances. For example, statistical outlier removal, neighborhood radius filtering, and normal consistency screening can be used to obtain a preliminary point cloud dataset. Therefore, this step provides an effective input foundation for the subsequent dynamic perception model, enabling the system to respond instantly to changes in the workpiece's state.

[0021] In the above workflow, the initial point cloud dataset is further input into a deep learning algorithm module to identify key features on the surface and complete point cloud segmentation. As a preferred embodiment, a neural network model suitable for point cloud processing can be used to classify and identify surface boundaries, concave-convex transition zones, local abrupt change zones, and key grinding areas. In some embodiments, deep learning recognition results can be combined with traditional geometric feature analysis to balance recognition accuracy with real-time requirements in industrial scenarios. This method allows for the further extraction of local features such as curvature distribution, roughness distribution, and normal variation degree in each segmented surface region. Furthermore, these features not only describe the region's own geometric properties but also serve as a basis for subsequent decisions regarding trajectory density, feed direction, and tool attitude adjustment.

[0022] Based on the feature extraction and point cloud segmentation results, a dynamic perception model can be generated. This model is used to uniformly represent the correspondence between the current workpiece surface state, the category of the area to be processed, historical processing results, and the current trajectory execution state. Furthermore, a clustering algorithm can be used to group similar curved surface regions after segmentation, grouping regions with similar curvature, roughness levels, or consistent normal variation trends into the same processing category. This automatically increases the trajectory coverage density, and in flat areas, the spacing between adjacent trajectories can be appropriately increased to balance processing quality and cycle time efficiency. It should be noted that this grouping is not only for the immediate grinding of a single workpiece; its more practical value lies in establishing transferable process parameter mapping relationships for similar complex curved surface parts, thus demonstrating strong engineering promotion significance and industrial application value.

[0023] During the actual grinding stage, in this embodiment, the system obtains the current processing position data from the optimized grinding trajectory planning and continuously receives surface change information from the sensors to determine whether the current processing area deviates from the predetermined removal state. When the local material removal amount, surface height difference, roughness fluctuation value, or contour error exceeds a preset threshold, the control unit adjusts the grinding parameters, which may include grinding head rotation speed, feed speed, contact pressure, dwell time, and trajectory offset.

[0024] As can be seen from the above embodiments, this adjustment mechanism transforms the grinding process from a unidirectional execution of a preset trajectory into a closed-loop control process that integrates perception, judgment, and correction. This method can reduce the need for secondary rework caused by individual workpiece differences, clamping errors, or uneven material removal, which has direct practical significance for improving equipment utilization and reducing the intensity of manual intervention.

[0025] After adjusting local parameters, a 3D reconstruction model can be constructed based on the updated surface state. Multiple frames of data are then fused using point cloud registration technology to obtain more accurate restored geometric data. In another specific embodiment of this application, rigid registration can be used for workpieces with minimal posture changes during processing; while non-rigid registration or partitioned registration can be used for scenarios requiring slight deformation or local disturbance compensation. The restored geometric data after 3D reconstruction can be further used for quality assessment. Statistical analysis methods are used to calculate surface roughness and deviation indices to determine whether the current processing area meets preset standards. Therefore, quality judgment is based on multi-frame fusion and overall geometric restoration, which is more conducive to ensuring the consistency of processing quality across the entire complex curved surface.

[0026] Furthermore, it should be noted that the quality assessment report will also be fed back into the dynamic perception model to update clustering algorithm parameters, correct region classification boundaries, and optimize subsequent trajectory planning strategies. In other words, this application constructs an automatic trajectory generation mechanism that can continuously learn and correct itself during processing. Therefore, this embodiment, through the processing chain of "point cloud acquisition, feature recognition, region segmentation, dynamic modeling, trajectory optimization, real-time adjustment, 3D reconstruction, and quality feedback," not only improves the grinding accuracy of complex curved surfaces but also takes into account processing stability, cycle time control, and process reuse value, making it suitable for the automated processing and commercialization of high-value-added curved surface parts.

[0027] S1 includes: acquiring spatial coordinate data of the curved workpiece surface using a sensor array; mapping the data to a unified three-dimensional coordinate system using a spatial coordinate transformation matrix to obtain a first point cloud dataset; calculating the local curvature change rate based on the first point cloud dataset; marking the corresponding area as a dynamic sensing area to obtain a second point cloud dataset if the local curvature change rate exceeds a preset threshold; using a bilateral filtering algorithm to remove outlier noise points from the second point cloud dataset to obtain a third point cloud dataset; constructing a local surface normal vector field based on the third point cloud dataset; removing redundant points using normal vector consistency constraints to obtain a fourth point cloud dataset; fitting a local surface equation based on the fourth point cloud dataset; removing points whose Euclidean distance exceeds a preset threshold to obtain a fifth point cloud dataset; and constructing a dynamic sensing model based on the fifth point cloud dataset, achieving dynamic sensing adjustment of the initial surface data of the complex curved workpiece by updating model parameters in real time.

[0028] In this embodiment, the sensor array is installed at the circumferential and top positions of the workpiece in the grinding station. After the equipment is installed, each sensor undergoes internal parameter calibration, external parameter calibration, and workpiece coordinate system alignment calibration in sequence. After calibration, the spatial coordinate data output by each sensor includes corresponding installation pose information and acquisition timing information. Subsequently, the surface spatial coordinate data acquired by each sensor are uniformly converted to the workpiece's three-dimensional coordinate system and stitched together according to the time sequence within the same acquisition cycle to form the first point cloud dataset. After this processing, all point cloud data are under the same spatial reference, and subsequent calculations of curvature, normal vectors, and local surface morphology are all based on the same coordinate system, which can avoid boundary misalignment and region misjudgment caused by inconsistencies in the coordinate references of different sensors.

[0029] Furthermore, after the first point cloud dataset is formed, a fixed number of neighboring points are extracted for each sampling point as neighborhood points. The number of neighborhood points is set to 30; when there are sparse sampling areas on the workpiece surface, the number of neighborhood points is set to 24. Subsequently, local surface fitting is performed on each sampling point and its neighborhood points to obtain the local curvature value at the location of the sampling point; then, the local curvature value of the sampling point is sorted and compared with the local curvature values ​​corresponding to its neighboring points to calculate the local curvature change rate at the location of the sampling point. The threshold determination process is as follows: first, a standard sample from the same type of workpiece is selected, and three complete scans are performed to obtain the local curvature change rate data of all scan points; then, all local curvature change rate data are sorted according to the numerical size, and the first 10% of the values ​​are taken as the curvature change rate threshold; during formal processing, the current workpiece uses this threshold.

[0030] For the first prototype scenario, the initial scan results of the current workpiece are used as the sorting sample, and the starting value at the top 10% is still taken as the curvature change rate threshold. When the local curvature change rate of any point reaches or exceeds this threshold, the location of that point and the continuous area within a 5mm radius around it are collectively marked as the dynamic sensing area, and summarized to form the second point cloud dataset. With this setting, the areas entering the second point cloud dataset are all areas with clear surface shape changes and clear requirements for subsequent trajectory correction.

[0031] As a preferred embodiment, bilateral filtering is performed on the second point cloud dataset. During processing, neighboring points within a 3mm radius of each sampling point are extracted; simultaneously, the spatial distance difference and surface normal difference between the sampling point and its neighbors are calculated. The spatial parameter of the bilateral filter is set to three times the average point spacing of the batch of point clouds, and the attribute parameter is set to the residual values ​​of the top 15% after sorting the residuals of the benchmark sample scan. Neighboring points that meet the spatial distance constraint and whose attribute differences do not exceed the above attribute parameters participate in weighted smoothing; points that do not meet the conditions are identified as outliers and removed. Through this processing, the third point cloud dataset retains the boundary and transition shapes of the dynamically perceived area, while removing reflective interference points, debris occlusion points, and abrupt change points. The result of this step is that the data used for subsequent normal vector calculations are all valid points constrained by both spatial relationships and surface attributes.

[0032] In this embodiment, a local surface normal vector field is constructed based on the third point cloud dataset. The specific process is as follows: 24 neighboring points are re-extracted for each sampling point in the third point cloud dataset. These neighboring points form a local fitting plane, and the normal direction of the sampling point is determined. After processing all sampling points, a normal vector field covering the entire dynamic sensing area is obtained. Subsequently, a consistency check is performed on spatially adjacent sampling points. When the distance between two sampling points does not exceed half the average distance between points in this batch of point clouds, and their normal deflection angle does not exceed 8 degrees, these two sampling points are identified as duplicate sampling points expressing the same local surface trend. Among the duplicate sampling points, the points with later sampling times are retained, and the remaining points are deleted, thus completing the redundant point removal and obtaining the fourth point cloud dataset.

[0033] Furthermore, the octave threshold here is obtained by scanning the standard sample five times consecutively: the normal deflection angles of adjacent points are sorted, and the largest value outside the top 5% is taken as the upper limit of normal fluctuation of the normal surface, which is finally fixed at 8 degrees in this embodiment. After adopting this processing method, the points in the fourth point cloud dataset not only maintain the local surface orientation, but also eliminate the data accumulation caused by repeated sampling.

[0034] In this embodiment, after the fourth point cloud dataset is formed, a local surface equation is fitted to each dynamically perceived region. For regions with an area not exceeding 400 square millimeters, quadratic surface fitting is used; for regions with an area exceeding 400 square millimeters, the region is first divided into multiple 400 square millimeter sub-regions, and then quadratic surface fitting is performed on each sub-region. After fitting, the Euclidean distance from each sampling point to the corresponding local surface is calculated point by point. The process for determining the Euclidean distance threshold is as follows: First, the standard sample is scanned and fitted three times, and the distance data from all sampling points to the fitted surface is statistically analyzed; then, the distance data is sorted from smallest to largest, and the top 95% of the values ​​are taken as the Euclidean distance threshold. During formal processing, any sampling point whose Euclidean distance to the fitted surface reaches or exceeds this threshold is identified as an abnormal deviation point and is removed. The remaining points are summarized to form the fifth point cloud dataset. For the finishing stage, the Euclidean distance threshold is directly set to 1 / 2 of the product profile tolerance; for the roughing stage, the Euclidean distance threshold is directly set to the product profile tolerance. After this processing, the fifth cloud dataset can accurately reflect the true surface shape of the workpiece and maintain consistency with the accuracy required for subsequent trajectory planning and quality assessment.

[0035] As a preferred embodiment, when constructing the dynamic perception model based on the fifth point cloud dataset, each dynamic perception region is first divided into several computational units. Each computational unit records the curvature value, roughness estimate, normal distribution value, region area value, and boundary position value. Subsequently, new point cloud data is continuously received according to the acquisition cycle, and the new data is input into the dynamic perception model in the same processing order as described above: coordinate unification, curvature filtering, bilateral filtering, normal constraint, and surface fitting.

[0036] Furthermore, the model parameter update rules are set as follows: when the local curvature change rate of the same calculation unit reaches the curvature change rate threshold for three consecutive acquisition cycles, the calculation unit is marked as a key correction unit; when the Euclidean distance of the same calculation unit reaches the Euclidean distance threshold for two consecutive acquisition cycles, the calculation unit is marked as a shape deviation unit; when the normal deflection angle of the same calculation unit does not exceed 8 degrees for four consecutive acquisition cycles, the calculation unit is marked as a stable unit. Key correction units, shape deviation units, and stable units are respectively written into the dynamic sensing model for subsequent region priority allocation, path spacing setting, and grinding parameter adjustment during grinding trajectory generation. Thus, the dynamic sensing adjustment of the initial surface data of complex curved workpieces no longer relies on a single scan result, but is transformed into a region state description result that is continuously updated during the processing.

[0037] S2 includes: extracting multi-scale features from the original point cloud using a convolutional neural network to obtain depth features on the surface; constructing topological connections of the point cloud using the depth features, and dividing the point cloud into regions based on the tightness of the topological connections to obtain segmented point clouds with region labels; calculating local normal vectors for the segmented point clouds to determine the discrete curvature of each surface region; extracting the surface roughness of each surface region based on the fluctuation of the discrete curvature in spatial distribution to form a set of geometric attributes reflecting surface quality; if the values ​​in the set of geometric attributes exceed the preset path constraints, adjusting the feed speed of the processing trajectory to obtain the real-time analysis results of the processing path; and outputting the feature extraction results and point cloud segmentation results by summarizing the depth features, segmented point clouds, discrete curvature, and surface roughness to complete the real-time analysis of the processing path.

[0038] In this embodiment, firstly, the original point cloud output by S1 is divided into multiple local point cloud blocks according to spatial continuity, and each local point cloud block contains 2048 sampling points; when the number of sampling points of a certain local point cloud block is less than 2048, the nearest neighbor copying method is used to supplement the sampling points to 2048; when the number of sampling points of a certain local point cloud block exceeds 2048, 2048 sampling points are retained according to the uniform sampling method.

[0039] Furthermore, after unifying the number of points, coordinate normalization is performed on each local point cloud block to ensure that the geometric center of the local point cloud block falls into a unified reference position and the coordinate range is limited to a preset input range. Then, the 3D coordinates, corresponding normal vector values, and the dynamic perception region label values ​​obtained in S1 of each sampling point are used as network input. The convolutional neural network adopts a 3-layer feature extraction structure. The first layer searches for 16 neighboring points for each sampling point and outputs 64-dimensional features; the second layer searches for 32 neighboring points for each sampling point and outputs 128-dimensional features; and the third layer searches for 64 neighboring points for each sampling point and outputs 256-dimensional features. The outputs of the three layers are concatenated by channel to form the depth feature data corresponding to the sampling point. In other words, this step outputs depth features that characterize the local boundary relationships, regional continuity relationships, and surface segment relationships of each sampling point.

[0040] Furthermore, for each sampling point, its 20 nearest neighbor sampling points are searched. Three connectivity criteria are calculated for the current sampling point and its 20 nearest neighbor sampling points: spatial distance, depth feature difference, and normal deflection angle. A topological connection edge is established between two sampling points only when all three connectivity criteria simultaneously meet preset conditions. These conditions are: spatial distance not exceeding twice the average spacing between points in the current local point cloud; depth feature difference within the top 85% of all neighboring point differences after sorting; and normal deflection angle not exceeding 10°. After determining the connectivity of all sampling points, each sampling point corresponds to several valid topological connection edges. Subsequently, the number of valid topological connection edges corresponding to each sampling point and their continuous spatial distribution are used as the evaluation result of the topological connectivity tightness of that sampling point.

[0041] In this embodiment, during region division, the process begins with sampling points having at least 12 effective topological connection edges, serving as the starting point for region growth. Then, the process expands outwards along these established effective topological connection edges, grouping all sampling points directly or indirectly connected to the starting point into the same region. When no more sampling points are added to a region, the division of that region is complete, and a unique region label is assigned to it. This process is repeated until all sampling points are labeled. The resulting segmented point cloud contains each sampling point simultaneously with its original spatial coordinates, depth features, and the label of its region. Subsequent curvature calculations and roughness extractions are based on this labeled segmented point cloud.

[0042] As a preferred embodiment, after obtaining the segmented point cloud with region labels, 24 nearest neighbor sampling points are extracted for each sampling point within the same region label. These 24 neighboring points are then used to fit a local plane, and the normal direction of the current sampling point is determined by the vertical direction of this local plane. Considering that the normal directions of adjacent sampling points within the same region need to be consistent, a unified correction of the normal directions of adjacent sampling points within the same region is also required. The correction rule is set as follows: when the angle between the normals of two adjacent sampling points is greater than 90°, the normal direction of the post-processed sampling point is reversed.

[0043] Furthermore, after unifying the normal vectors, a local quadratic surface is fitted based on each sampling point and its 24 neighboring points. The curvature of this local quadratic surface at the current sampling point is used as the discrete curvature value of that sampling point. Then, the discrete curvature values ​​of all sampling points within the same region label are recorded in spatial order to form the discrete curvature distribution result corresponding to that region. Here, "discrete curvature" provides a specific curvature value for each sampling point within the region, allowing subsequent steps to directly determine whether the region needs to reduce the feed rate, narrow the trajectory spacing, or increase the number of coverages based on the curvature distribution.

[0044] Furthermore, within each region label, local statistical units are divided according to a 5 mm × 5 mm grid. When the number of sampling points in a statistical unit is less than 30, the statistical unit is merged with its adjacent statistical units until the merged statistical unit contains no less than 30 sampling points. Subsequently, four curvature fluctuation data are statistically analyzed within each statistical unit, including the maximum discrete curvature, the minimum discrete curvature, the average discrete curvature, and the average curvature difference between adjacent sampling points, to reflect the spatial distribution variation of curvature within the statistical unit.

[0045] Furthermore, based on this, a local reference plane is fitted using all sampling points within the statistical unit, and the average absolute value of the distances from all sampling points to this local reference plane is calculated. This average value is used as the surface roughness value of the statistical unit. After completing the calculation for all statistical units, all statistical units within the same region label are summarized to obtain the region's average discrete curvature value, maximum discrete curvature value, average roughness value, boundary position value, and area value. These data are then combined to form the geometric attribute set corresponding to the surface region. In this way, the geometric attribute set not only includes curvature-related data but also roughness, boundary, and area data, simultaneously reflecting the geometric shape and surface quality state of the region.

[0046] Following this, in this embodiment, discrete curvature constraint values ​​and surface roughness constraint values ​​are first set respectively. The process for determining the discrete curvature constraint value is as follows: select 6 standard workpieces of the same model, complete scanning, region division, and discrete curvature statistical processing, sort all region average discrete curvature values ​​from smallest to largest, and take the top 90% of the sorted values ​​as discrete curvature constraint values. The process for determining the surface roughness constraint value is as follows: select 6 standard workpieces of the same model, perform roughness detection on their qualified machined surfaces, and match the detection results with the region roughness values ​​extracted from the point cloud. Then, sort all the matched region roughness values ​​from smallest to largest, and take the top 90% of the sorted values ​​as surface roughness constraint values. During formal processing, the geometric attribute set of each curved surface region is compared one by one. If the region average discrete curvature value of the region reaches or exceeds the discrete curvature constraint value, or the region average roughness value of the region reaches or exceeds the surface roughness constraint value, it is determined that the region has a path adjustment requirement.

[0047] Furthermore, the corresponding feed rate adjustment rules are as follows: when only the discrete curvature reaches or exceeds the constraint value, the feed rate of the current area is adjusted to 70% of the reference feed rate; when only the surface roughness reaches or exceeds the constraint value, the feed rate of the current area is adjusted to 80% of the reference feed rate; when both the discrete curvature and surface roughness reach or exceed the constraint values, the feed rate of the current area is adjusted to 60% of the reference feed rate. The reference feed rate is preset in the process document before machining and remains consistent throughout the machining process of the same batch of workpieces. After processing according to these rules, a clear speed adjustment result can be output for each surface area, which constitutes one of the real-time analysis results of the machining path.

[0048] Finally, after completing depth feature extraction, region segmentation, discrete curvature calculation, and surface roughness extraction, all intermediate results are archived uniformly, organized by region label. The archived content includes at least: the depth feature data corresponding to each sampling point, the region label to which each sampling point belongs, the discrete curvature value of each sampling point, the roughness value of each statistical unit, the boundary position value, area value, and feed rate adjustment value of each surface region. Then, two types of output data are generated. The first type of output data is the feature extraction result, which includes at least the depth feature description of each surface region, the average discrete curvature of the region, the average roughness of the region, and the region path adjustment status. The second type of output data is the point cloud segmentation result, which includes at least the region labels of all sampling points, the region boundary positions, and the connectivity relationships between regions.

[0049] Furthermore, after receiving the above two types of output data, the control system can map different curved surface regions to different trajectory control units and send each trajectory control unit to S3. In other words, the final output of S2 is real-time path analysis data that already includes region semantics, geometric quantization results, and feed rate analysis results.

[0050] S3 includes acquiring raw point cloud data and extracting geometric features, and obtaining multiple segmented surfaces through point cloud segmentation. The segmented surfaces are grouped by similarity using a clustering algorithm to determine clustered regions with consistent physical properties. Real-time displacement data generated by the clustered regions during processing is input into the dynamic perception model to obtain a perception state reflecting the workpiece deformation. The perception state is combined with preset processing accuracy requirements for path optimization. If the perception state deviates from the preset threshold, an optimized path that avoids obstacles and conforms to the surface is recalculated. The optimized path is transformed into a grinding trajectory planning through inverse kinematics, completing real-time adjustments during the grinding operation and realizing automatic generation and optimization of the grinding trajectory under the guidance of the dynamic perception model.

[0051] In this embodiment, S3 receives the feature extraction results and point cloud segmentation results output by S2. Its purpose is to unify the obtained regional information, geometric attribute information and real-time processing status information into a grinding trajectory that the equipment can execute.

[0052] In this embodiment, the point cloud with region labels output by S2 is directly used as the original point cloud data, and all sampled points are grouped according to the region labels for the first time. Subsequently, six geometric features are extracted from each point cloud group: the average discrete curvature of the region, the maximum discrete curvature of the region, the average surface roughness of the region, the area of ​​the region, the length of the region boundary, and the mean value of the region normal direction. After extraction, each geometric feature is uniformly normalized so that the six features participate in subsequent segmentation processing within the same numerical range. Next, a second segmentation is performed on the original point cloud according to the region connectivity: when the area of ​​a region is greater than 400 square millimeters, it is segmented along the boundary of the discrete curvature abrupt change within the region, and the area of ​​the single surface after segmentation is no greater than 400 square millimeters; when the area of ​​a region is less than 40 square millimeters, the region is merged into the adjacent region that shares the longest boundary and whose normal deflection angle does not exceed 8°. After the above processing, multiple segmented surfaces with clear boundaries, well-defined areas, and clear geometric features are obtained.

[0053] Furthermore, this embodiment employs a clustering method with 4 categories. Each segmented surface reads the aforementioned 6 geometric features, and simultaneously reads 4 corresponding process attributes from the process database. These 4 process attributes are the target removal amount, the reference feed rate, the contact pressure, and the grinding wheel speed. The term "consistent physical attributes" in this embodiment refers to all segmented surfaces within the same cluster region sharing the same set of target removal amount, reference feed rate, contact pressure, and grinding wheel speed.

[0054] Specifically, during clustering, all segmented surfaces are first divided into four categories based on the average discrete curvature of the regions. Then, category correction is performed based on the average surface roughness of the regions and the target removal amount. Finally, the final category is determined by boundary continuity as a constraint. The classification results are four types: Category 1 is the refinement region, Category 2 is the uniform removal region, Category 3 is the transition region, and Category 4 is the boundary trimming region. Segmented surfaces of the same category and sharing a common boundary are uniformly merged into the same cluster. After this processing, each cluster corresponds to specific processing parameters, eliminating the need for point-by-point judgment during subsequent path calculations. This method can directly transform the geometric differences of complex surfaces into differences in processing regions.

[0055] After entering the processing stage, in this embodiment, the equipment collects real-time displacement data once every 20 milliseconds. The real-time displacement data consists of three parts: the first part is the surface point offset data obtained by the rescanning of the sensor array, the second part is the current position data fed back by the robot end encoder, and the third part is the clamping position change data fed back by the clamp displacement monitoring unit.

[0056] Specifically, these three data sets are first mapped to the workpiece coordinate system, and then assigned to clustered regions for matching. For each clustered region, four region state variables are calculated: the region center point offset, the region average normal deflection angle, the region boundary point average offset, and the cumulative material removal amount. These four region state variables are then written into the region state table corresponding to the dynamic sensing model.

[0057] As a preferred embodiment, the dynamic perception model stores eight state data items for each clustered region: region number, theoretical center position, actual center position, theoretical normal direction, actual normal direction, current cumulative removal amount, target cumulative removal amount, and state update time. Therefore, the perception state in this embodiment is a comprehensive result that simultaneously reflects position changes, posture changes, and processing progress.

[0058] In this embodiment, after the dynamic perception model outputs the current perception state, the process document pre-sets four machining accuracy requirements: contour error no greater than 0.03 mm, surface roughness no greater than 0.8 μm, adjacent trajectory coverage width of 35% of the effective tool width, and minimum safe distance of 2 mm between the tool center and the fixture outer contour. The perception state of the current clustered region is compared with each of the above four machining accuracy requirements, and three deviation judgment conditions are used to trigger path recalculation: 1. The offset of the region center point reaches or exceeds 0.20 mm; 2. The average normal deflection angle of the region reaches or exceeds 6°; 3. The average offset of the region boundary points reaches or exceeds 0.10 mm. Path optimization is immediately triggered when any one of the three conditions is met.

[0059] Furthermore, during path optimization, the local surface is first refitted with the actual point cloud of the current cluster region, and then all trajectory points on the original path are reprojected onto the new local surface to ensure the trajectory fits the current actual surface. Next, obstacle collision checks are performed on each of the updated trajectory points. The obstacle model is fixed and includes the outer contour of the fixture, the sensor mounting bracket, adjacent unprocessed protrusions, and the grinding head envelope. Any trajectory point whose distance to any obstacle is less than 2 mm is identified as a collision risk point. For collision risk points, they are first translated 1 mm in the opposite direction to the current surface normal, and then reprojected onto the local surface. If the distance is still less than 2 mm after reprojection, the corresponding trajectory segment is laterally offset by 0.5 times the effective tool width until all trajectory points meet the 2 mm safety distance. Finally, cubic spline smoothing is performed on the updated trajectory segment to eliminate trajectory inflection points and attitude change points, resulting in an optimized path that avoids obstacles and fits the surface.

[0060] Finally, in this embodiment, the grinding equipment uses a 6-axis robot, and each optimized path point needs to be converted into the target position value of the robot's 6 joints. Inverse kinematics is performed point-by-point on all trajectory points in the optimized path, while simultaneously satisfying three trajectory execution constraints: First, the angle between the grinding head axis and the current surface normal is fixed at 15°; second, the end effector displacement between two adjacent trajectory points is no greater than 1.5 mm; third, all joint positions are within the robot's allowable travel range. When there are more than two executable inverse kinematics results for the same trajectory point, the result with the smallest difference in total joint displacement from the previous trajectory point is selected to reduce sudden changes in robot posture. After the inverse kinematics is completed, each optimized path is converted into a time-ordered sequence of control instructions. The control instruction sequence includes at least the trajectory point coordinates, tool posture angle, feed rate, contact pressure, and dwell time. The controller continuously receives the latest control instruction sequence at a 20-millisecond update cycle and, upon receiving a new sensing state, directly replaces the subsequent trajectory segments that have not yet been executed, thereby achieving real-time adjustments during the grinding process. This method enables the dynamic perception model, path optimization results, and robot execution trajectory to maintain a synchronous update relationship, ultimately achieving automatic generation and online optimization of complex surface grinding trajectories.

[0061] S4 includes acquiring trajectory planning coordinate data and sensor data and performing spatial mapping to obtain the surface roughness value of the current processing point; if the surface roughness value exceeds the preset surface morphology threshold, then determining the correction amount of the current grinding pressure and feed speed; applying the correction amount of the current grinding pressure and feed speed to the grinding actuator to obtain grinding trajectory and processing torque data; performing a difference operation between the grinding trajectory and processing torque data and the initial geometric morphology model to obtain the geometric deviation distribution map of the current processed surface; and obtaining the final geometric morphology restoration data through an iterative correction algorithm based on the geometric deviation distribution map of the current processed surface to achieve geometric morphology restoration.

[0062] In this embodiment, S4 corresponds to the online correction and geometric restoration process during the grinding of complex curved surfaces. First, the controller reads the trajectory planning coordinate data once every 20 ms and simultaneously reads the sensor data once. The trajectory planning coordinate data includes at least the coordinates of the current trajectory point, the tool posture angle, the current target feed speed, the current target grinding pressure, and the current trajectory segment number. The sensor data includes at least the surface point cloud data, local contour height data, and reflection intensity data of the current processing area.

[0063] Subsequently, the trajectory planning coordinate data and sensor data are uniformly converted into the workpiece coordinate system, and a local surface sampling area with a radius of 2mm is extracted with the current trajectory point as the center. When the number of effective sampling points in this local surface sampling area is not less than 50, a local reference plane is directly fitted to all sampling points in the area, and the absolute value of the distance from each sampling point to the local reference plane is calculated. The average value of all absolute distances is recorded as the surface roughness value of the current machining point. When the number of effective sampling points is less than 50, the sampling radius is expanded to 3mm and the local surface sampling area is re-extracted. If the number of effective sampling points is still less than 50 after expansion, the surface roughness value corresponding to the previous sampling cycle is used as the surface roughness value of the current machining point.

[0064] In this embodiment, after obtaining the surface roughness value of the current processing point, in order to have a clear basis for the preset surface morphology threshold, six qualified workpieces of the same model are selected, and three final inspection scans are performed on each workpiece to obtain a total of 18 sets of surface roughness reference data. Then, the 18 sets of surface roughness reference data are sorted according to the size of the values, and the end value in the first 90% position is taken as the surface morphology threshold.

[0065] In this embodiment, the surface morphology threshold is fixed at 0.80 μm. During actual processing, if the surface roughness value of the current processing point is less than 0.80 μm, the current grinding pressure and feed rate remain unchanged; if the surface roughness value of the current processing point reaches or exceeds 0.80 μm but is less than 1.00 μm, the current grinding pressure is increased by 8 N and the current feed rate is decreased by 12%; if the surface roughness value of the current processing point reaches or exceeds 1.00 μm but is less than 1.20 μm, the current grinding pressure is increased by 12 N and the current feed rate is decreased by 18%; if the surface roughness value of the current processing point reaches or exceeds 1.20 μm, the current grinding pressure is increased by 16 N and the current feed rate is decreased by 25%.

[0066] Additionally, it should be noted that when the shortest distance between the current processing point and the boundary of the current cluster region is no more than 3 mm, to prevent over-grinding of the boundary, the grinding pressure is further reduced by 4 N and the feed rate is reduced by 6% based on the above correction results. This method can establish a one-to-one correspondence between the surface roughness state and the pressure correction amount and the speed correction amount.

[0067] In this embodiment, after completing the correction calculation, the controller immediately sends the corrected grinding pressure command and the corrected feed speed command to the grinding execution mechanism. The grinding execution mechanism includes a robot end effector unit, a pressure control unit, a spindle drive unit, and a trajectory following unit. The pressure control unit adjusts the end-effector contact pressure at 5ms intervals, the trajectory following unit adjusts the trajectory running speed and attitude following results at 20ms intervals, and the spindle drive unit maintains stable output according to the spindle speed setting. After the correction action is executed, two types of execution data are collected synchronously. The first type is grinding trajectory data, which includes at least the actual trajectory point coordinates, actual tool attitude angle, actual running speed, actual contact path length, and actual running timestamp. The second type is machining torque data, which includes at least the spindle output torque, normal contact torque, and tangential resistance torque. Both types of execution data have a unified timestamp and are written to the execution buffer in the order of trajectory segments. Therefore, "obtaining the grinding trajectory and machining torque data" refers to the set of actual execution data synchronously output by the position feedback unit and the force feedback unit after the correction is applied.

[0068] As a preferred embodiment, after the aforementioned execution data has been formed, the initial geometric model established in S1 is first invoked, and all actual grinding trajectory points of the current trajectory segment are mapped to the theoretical surface positions corresponding to the initial geometric model. Then, taking each actual grinding trajectory point as the center, a theoretical local region with a radius of 1.5mm and the corresponding actual local region are extracted, and the theoretical surface height value and the actual surface height value are obtained respectively. At the same time, the spindle output torque, normal contact torque, and tangential resistance torque at the corresponding moment of the trajectory point are read, and the processing torque data at that moment is converted into the actual material removal depth value according to the pre-established material removal mapping table. Next, the difference between the theoretical surface height value and the actual surface height value is calculated, and the difference is corrected in conjunction with the actual material removal depth value to obtain the geometric deviation value of the trajectory point.

[0069] Furthermore, after calculating the geometric deviation values ​​of all trajectory points, the workpiece surface is divided into grids of 2mm × 2mm. The average of all geometric deviation values ​​falling within the same grid is then used as the geometric deviation result for that grid. In this way, all grids on the workpiece surface obtain corresponding deviation values, ultimately generating a geometric deviation distribution map of the currently processed surface. In this embodiment, grids with an absolute deviation value less than 0.02mm are marked as qualified grids, grids with an absolute deviation value reaching or exceeding 0.02mm but less than 0.05mm are marked as correction grids, and grids with an absolute deviation value reaching or exceeding 0.05mm are marked as key correction grids.

[0070] In this embodiment, the iterative correction algorithm uses grid cells as the basic processing object and executes them in the order of "prioritizing the correction of key grid cells, followed by correction of other grid cells, and keeping qualified grid cells unchanged." In the first iteration, for the area where the key correction grid cells are located, the trajectory spacing of the corresponding trajectory segments is reduced to 70% of the original trajectory spacing, and the number of grinding coverage times for this area is increased by 1; for the area where the correction grid cells are located, the trajectory spacing of the corresponding trajectory segments is reduced to 85% of the original trajectory spacing, and the original number of coverage times remains unchanged; for the area where qualified grid cells are located, the trajectory spacing and the number of coverage times remain unchanged.

[0071] Furthermore, after completing the first iteration, local grinding is re-executed, and updated trajectory data, machining torque data, and surface point cloud data are re-acquired to generate a new geometric deviation distribution map. In the second and subsequent iterations, the system compares the deviation changes of the same mesh between the two iterations. When the deviation change of a mesh is less than 0.01 mm, the mesh is considered to have entered a stable state and no further adjustments are made. When the deviation change of a mesh reaches or exceeds 0.01 mm, the iteration correction privilege for that mesh is retained. The entire iteration correction process is executed for a maximum of four rounds. When all meshes have entered a stable state, or after the fourth iteration is completed, the correction is immediately stopped, and the final geometric shape restoration data is output.

[0072] Specifically, the final geometric shape restoration data includes at least the restored surface point cloud, the restored grid deviation values, the restored region surface roughness values, and the restored trajectory coverage record. Based on the aforementioned embodiments, it can be seen that after the above multiple rounds of deviation identification, local correction, and re-detection processing, the current surface shape of the workpiece will continuously converge towards the target geometric shape, thereby achieving geometric shape restoration.

[0073] In another specific embodiment of this application, when the workpiece material is a metal material with a more stable removal response, the number of iterations is fixed at 3 rounds; when the workpiece material is a resin-based composite material, the number of iterations is fixed at 2 rounds to prevent surface damage caused by repeated corrections in local areas. As a preferred embodiment, after all iterations are completed, a full-surface verification scan can be performed on the final geometric restoration data. If the surface roughness and mesh deviation obtained from the full-surface verification scan meet the process requirements, the restoration data is directly evaluated for quality.

[0074] S5 includes acquiring surface state data and mapping it to a unified global coordinate system to obtain first spatial point cloud data; performing multi-frame feature point matching based on the first spatial point cloud data to obtain second spatial point cloud data; constructing a three-dimensional reconstruction model by fitting the surface using the second spatial point cloud data to obtain third spatial point cloud data; if the deviation analysis result between the third spatial point cloud data and the preset geometric benchmark exceeds a preset threshold, performing local point cloud resampling on the third spatial point cloud data to obtain fourth spatial point cloud data; verifying consistency and determining the precise geometric shape based on the fourth spatial point cloud data to obtain restored geometric data; and using the restored geometric data to complete the quality assessment preparation.

[0075] In this embodiment, S5 corresponds to the process of restoring and reconstructing the surface state after polishing. Based on the foregoing embodiments, this method can further transform the surface state data collected at the end of the polishing process into restored geometric data that can be used for quality assessment.

[0076] First, in this embodiment, the surface state data is sourced from three parts. The first part is the surface point cloud data acquired by the sensor array after S4; the second part is the current position and attitude data fed back by the robot end effector or multi-axis actuator; and the third part is the workpiece clamping position data output by the fixture positioning unit. A unified timestamp is appended to the above three parts of data with a 20ms acquisition cycle, and then time alignment processing is performed to ensure a one-to-one correspondence between the surface point cloud data, current position data, and clamping position data within the same acquisition cycle. Subsequently, a unified global coordinate system is established using the workpiece clamping reference surface, the center of the workpiece positioning hole, and the fixture mounting reference edge, and the coordinates of all surface points within the same acquisition cycle are transformed into this unified global coordinate system.

[0077] Furthermore, after the conversion, points located inside the fixture entity envelope are deleted, invalid points with reflection intensity below a set lower limit are deleted, and isolated points that have not appeared repeatedly in three adjacent sampling periods are deleted. The remaining points are considered valid surface points. The valid surface points from six consecutive sampling periods are then summarized in chronological order to obtain the first spatial point cloud data. Therefore, the first spatial point cloud data is standardized spatial point cloud data that has undergone time synchronization, coordinate unification, and invalid point removal.

[0078] In this embodiment, after obtaining the first spatial point cloud data, six consecutive frames of the first spatial point cloud data are taken as a registration unit, and the first frame is used as the reference frame. The second to sixth frames are aligned with the reference frame sequentially. To ensure the stability of the matching result, three types of feature points are extracted in each frame of point cloud: boundary inflection points, curvature abrupt change points, and normal abrupt change points. Boundary inflection points are obtained by determining missing edge neighborhoods, curvature abrupt change points are obtained by selecting the top 10% of points after sorting local curvature, and normal abrupt change points are obtained by points where the angle between the normals of adjacent points reaches or exceeds 8°. Subsequently, the feature points of the current frame are compared point by point with the feature points of the previous frame. Only point pairs that simultaneously meet the following three conditions are determined as valid matching point pairs: spatial position difference is not greater than 0.30mm, normal deflection angle is not greater than 8°, and local curvature difference is not greater than 15% of the average curvature value of the current frame. After forming valid matching point pairs between every two frames, these valid matching point pairs are used as registration constraints to perform frame-by-frame rigid alignment processing. If the number of valid matching point pairs between two frames is less than 30, the alignment process is not initiated directly. Instead, local surface data is sampled again for the current frame, feature points are extracted again, and the matching is repeated. If the number of valid matching point pairs after rematching is still less than 30, the current frame is discarded. After aligning all available frames frame by frame, the aligned multi-frame point clouds are merged to obtain the second spatial point cloud data.

[0079] In this embodiment, the second spatial point cloud data is first divided into local reconstruction units according to the surface unfolding range, with each local reconstruction unit having a fixed coverage area of ​​1 mm × 10 mm. If the number of sampling points in a local reconstruction unit is less than 80, the unit is merged with the adjacent unit sharing the longest boundary until the number of sampling points after merging is no less than 80. Then, quadratic surface fitting is performed on all sampling points in each local reconstruction unit to obtain the local surface patch corresponding to the unit. After the fitting of each local surface patch is completed, the common boundary of adjacent local surface patches is continuously stitched together. During stitching, the boundary position difference, boundary normal difference, and boundary height difference are checked respectively. Only when the boundary position difference is no greater than 0.02 mm, the boundary normal deflection angle is no greater than 5°, and the boundary height difference is no greater than 0.02 mm, is the current stitching result directly retained; if any of the above conditions are not met, the neighborhood weighted fitting is re-performed on the common boundary region of two adjacent local surface patches until the stitching conditions are met or the upper limit of two refits is reached.

[0080] Furthermore, after all the local surface patches are stitched together, a three-dimensional reconstruction model of the workpiece's current surface is constructed. Then, all sampling points in the second spatial point cloud data are reprojected onto the surface of this three-dimensional reconstruction model, and the projected point cloud result is recorded as the third spatial point cloud data. As can be seen above, the third spatial point cloud data is consistent with the three-dimensional reconstruction model, and subsequent deviation analysis is carried out based on the reconstructed point cloud constrained by the model.

[0081] Furthermore, in this embodiment, the preset geometric datum is fixed as the workpiece design surface model, and has been converted to a coordinate datum consistent with the unified global coordinate system during the process preparation stage. During deviation analysis, each sampling point in the third spatial point cloud data is taken as the object, and the shortest distance from that sampling point to the corresponding position of the preset geometric datum is calculated. This shortest distance is recorded as the local deviation value of that sampling point. Then, the entire surface is divided into a 2mm × 2mm grid, and the average value of all local deviation values ​​in each grid is calculated as the deviation analysis result for that grid.

[0082] Furthermore, the process for determining the preset threshold is as follows: select 6 qualified workpieces of the same model, perform full-surface scanning, alignment, reconstruction and deviation calculation for each workpiece to obtain all mesh deviation results; then sort all mesh deviation results from smallest to largest, and take the first 95% of the final values ​​after sorting as the deviation threshold.

[0083] In this embodiment, the deviation threshold is set to 0.05 mm. During formal processing, if the deviation analysis result of any grid reaches or exceeds 0.05 mm, it is determined that the reconstruction accuracy of the corresponding area of ​​that grid is insufficient, and local point cloud resampling needs to be performed. During local point cloud resampling, the surface area within a radius of 4 mm is extracted with the center of the out-of-tolerance grid as the center, and the sampling resolution of this area is adjusted to twice the original sampling resolution. The surface point cloud is then collected again. After resampling, the original point cloud data of the corresponding area in the third spatial point cloud data is replaced with the newly collected high-density point cloud, while the remaining non-out-of-tolerance areas remain unchanged. The point cloud result after replacement is recorded as the fourth spatial point cloud data. After this processing, the out-of-tolerance areas obtain denser surface sampling information, which can reduce local reconstruction errors.

[0084] In this embodiment, after the fourth spatial point cloud data is formed, to ensure that the final reconstruction result has full-surface credibility, the system fixedly executes three consistency checks. The first item is position consistency check, and the check method is to compare the height difference of corresponding points on both sides of the common boundary of adjacent reconstruction units. If the height difference on both sides of the common boundary is no more than 0.02mm, the position consistency is determined to be qualified. The second item is normal consistency check, and the check method is to compare the normal deflection angles on both sides of the common boundary of adjacent reconstruction units. If the normal deflection angle is no more than 5°, the normal consistency is determined to be qualified. The third item is sampling density consistency check, and the check method is to count the difference in the number of sampling points per square millimeter of adjacent reconstruction units. If the difference in the number of sampling points is no more than 15, the density consistency is determined to be qualified. Only the areas that simultaneously meet the above three consistency conditions are retained as effective reconstruction areas. For the areas that fail the consistency check, one boundary smoothing process is performed first, then one neighborhood reprojection process is performed, and then the consistency is rechecked. If the conditions are still not met after the recheck, local resampling is performed again and the above process is repeated, but the total number of processing times does not exceed 2. After all areas pass the consistency check, the full-surface effective reconstruction areas are spliced into a complete surface, and then the full-surface boundary position, full-surface normal distribution and full-surface height distribution are uniformly calculated to obtain restored geometric data.

[0085] Further, the restored geometric data at least includes the restored full-surface point cloud, the restored surface mesh model, the restored boundary position data and the restored normal distribution data.

[0086] Finally, the restored geometric data is used to complete the preparation for quality assessment. In this embodiment, it is organized into standard input data that can be directly called by the subsequent quality assessment module. During specific processing, the restored geometric data is first reorganized according to the data format requirements of the quality assessment module to form an assessment input file including full-surface point cloud coordinates, surface mesh topological relationship, region boundary number, local normal information and deviation analysis index. Then, according to the division rule of the functional areas of the workpiece, the restored geometric data is split into a plurality of to-be-assessed area data packets, and each to-be-assessed area data packet corresponds to one subsequent quality detection area. For each to-be-assessed area data packet, three types of reference information are additionally attached, which are the corresponding design surface reference data, the corresponding process tolerance data and the corresponding surface quality detection parameters. After all the to-be-assessed area data packets are generated, they are written into the quality assessment buffer, and an assessment preparation completion identifier of this reconstruction result is synchronously generated.

[0087] S6 includes acquiring the original point cloud data of the processing area and using the least squares method to fit and generate a restored geometric model, extracting the spatial coordinate information of each sampling point on the surface of the restored geometric model; calculating the normal distance of each sampling point relative to the theoretically designed curved surface based on the spatial coordinate information, and obtaining a deviation index distribution matrix reflecting the surface morphology of the processing area; using a sliding window algorithm to smooth the local area of ​​the deviation index distribution matrix, extracting surface contour features, and calculating the arithmetic mean roughness value corresponding to the surface contour features; if the arithmetic mean roughness value and the maximum deviation value in the deviation index distribution matrix are both less than the preset standard, the processing area is determined to be qualified and a quality assessment report is generated to avoid processing defects.

[0088] In this embodiment, S6 corresponds to the final quality assessment of the processed area. This embodiment integrates the original point cloud data of the processed area, the theoretically designed surface, the deviation distribution results, and the roughness results into the same evaluation chain, and finally outputs a quality assessment report.

[0089] First, in this embodiment, the original point cloud data of the current processing area is retrieved from the data package of the area to be evaluated output by S5. This original point cloud data includes at least the three-dimensional coordinates of all surface sampling points within the current processing area, the area boundary data, and the corresponding normal data. After retrieval, data purification processing is performed. The specific processing order is as follows: delete transition points within a 1mm radius of the boundary, delete isolated points with an average distance of 0.20mm or more between adjacent points, and delete abnormal points that do not appear repeatedly in two consecutive scans. After purification, the current processing area is divided into multiple local fitting units of 5mm × 5mm. Each local fitting unit retains no fewer than 60 sampling points. When the number of sampling points in a local fitting unit is less than 60, the fitting unit is merged with the adjacent fitting unit sharing the longest boundary until the number of sampling points reaches or exceeds 60.

[0090] Subsequently, least-squares surface fitting is performed on all sampling points within each local fitting unit, with the fitted surface consistently using quadratic surface patches. After all local surface patches are fitted, continuity stitching is performed on the common boundaries of adjacent local surface patches. The stitching verification conditions are fixed at three conditions: the height difference between the two sides of the common boundary is no greater than 0.02 mm, the normal deflection angle between the two sides of the common boundary is no greater than 5°, and the position difference between the boundary points on the two sides of the common boundary is no greater than 0.02 mm. If any condition is not met, a neighborhood weighted fitting is performed again for the corresponding common boundary region; if the conditions are still not met after refitting, a second boundary fitting is performed, and this process is not repeated. Once all local surface patches are stitched together, the restored geometric model is formed.

[0091] Then, uniform resampling is performed on the surface of the restored geometric model at a sampling interval of 0.5 mm to obtain uniformly distributed model sampling points, and the spatial coordinate information of each model sampling point is extracted.

[0092] As a preferred embodiment, after obtaining the spatial coordinate information of each sampling point on the surface of the restored geometric model, the theoretical design surface model corresponding to the current processing area is first read, and the theoretical design surface model is transformed into a global coordinate system consistent with the restored geometric model. Subsequently, for each sampling point on the surface of the restored geometric model, the model normal direction of the sampling point is read, and a corresponding projection search is performed on the theoretical design surface along the normal direction. The search range is fixed at 1 mm before and after the normal direction; when a corresponding intersection point of the theoretical design surface is found within this range, the normal distance between the current sampling point and the intersection point is directly calculated; when no corresponding intersection point is found within this range, the nearest theoretical surface patch within 1 mm of the current sampling point is used as a substitute corresponding surface, and the normal distance is calculated again.

[0093] Furthermore, if the current sampling point is located outside the theoretically designed surface, the normal distance is recorded as a positive value; if the current sampling point is located inside the theoretically designed surface, the normal distance is recorded as a negative value. After the normal distances of all sampling points are calculated, they are stored in a matrix according to their two-dimensional arrangement in the processing area. The row direction of the matrix is ​​consistent with the length direction of the workpiece, and the column direction of the matrix is ​​consistent with the width direction of the workpiece. Each matrix element corresponds to one sampling grid position, and the value of each matrix element is the normal distance value corresponding to that position. To ensure matrix integrity, when a sampling grid position lacks a valid normal distance value, the average of the known normal distances of its four adjacent grid positions (up, down, left, and right) is used to fill the gap; when there are fewer than two valid values ​​among the four adjacent grid positions, the normal distance value of the corresponding position in the previous evaluation cycle is directly called.

[0094] In this embodiment, the deviation index distribution matrix is ​​first smoothed locally. The sliding window size is fixed at 5×5, and the window movement step size is fixed at 1 matrix unit. Each time the window moves, 25 matrix elements within the window are read, and the deviation values ​​are sorted according to their numerical values. The deviation value at the middle position is taken as the smoothed deviation value at the center position of the window, and this smoothed deviation value is written into a new smoothing matrix. After the window has moved completely, the smoothed deviation matrix corresponding to the current processing area is obtained. Subsequently, surface contour features are extracted from the smoothed deviation matrix. Specifically, one longitudinal contour line is extracted every 2 mm along the length of the workpiece, and one transverse contour line is extracted every 2 mm along the width of the workpiece. For each contour line, the position of the contour peak point, the position of the contour valley point, the peak-valley height difference, and the position of the contour average line are determined sequentially, and the above results are used as the surface contour features of the contour line. Then, using the contour average line of each contour line as a baseline, the absolute deviation value of all sampling points of the contour line relative to the contour average line is calculated point by point, and the average of all absolute deviation values ​​is then used as the arithmetic mean roughness value corresponding to the contour line. After the arithmetic mean roughness values ​​of all longitudinal and transverse contour lines are calculated, the average value of the arithmetic mean roughness values ​​of all contour lines is taken as the final arithmetic mean roughness value of the current machining area.

[0095] In this embodiment, if the number of valid sampling points in a certain contour line is less than 20, then that contour line is not included in the roughness statistics; if the number of valid contour lines is less than 6, then contour lines are re-sampled at 1mm intervals and the above calculation is repeated. Therefore, the arithmetic mean roughness value in this embodiment comes from the comprehensive evaluation of multiple contour lines in the entire processing area.

[0096] Furthermore, in this embodiment, the preset standards are divided into two categories: roughness standards and deviation standards. The process for determining the roughness standard is as follows: Six qualified workpieces of the same model are selected, and the surface roughness of the corresponding processed area of ​​each workpiece is tested. The test results are then matched one-to-one with the arithmetic mean roughness values ​​of the area obtained according to this embodiment. Then, all the matched arithmetic mean roughness values ​​are sorted from smallest to largest, and the top 90% of the sorted values ​​are taken as the roughness standard. In this embodiment, the roughness standard is fixed at 0.80 μm. The process for determining the deviation standard is as follows: Six qualified workpieces of the same model are selected, and the reconstruction geometric model and deviation index distribution matrix are constructed for each workpiece. The matrix element with the largest absolute value is extracted from the deviation index distribution matrix of each workpiece as the maximum deviation value of that workpiece. Then, all maximum deviation values ​​are sorted from smallest to largest, and the top 95% of the sorted values ​​are taken as the deviation standard. In this embodiment, the deviation standard is fixed at 0.05 mm. During the formal evaluation, the system first reads the final arithmetic mean roughness value of the current processing area, and then extracts the matrix element with the largest absolute value from the deviation index distribution matrix as the maximum deviation value of the current processing area. When the final arithmetic mean roughness value is less than 0.80μm and the maximum deviation value is less than 0.05mm, the processing area is deemed qualified. When either of the two values ​​reaches or exceeds the corresponding standard, the processing area is deemed unqualified.

[0097] Furthermore, after the judgment is completed, the system automatically generates a quality assessment report. The quality assessment report includes at least the processing area number, assessment time, final arithmetic mean roughness value, maximum deviation value, deviation distribution map, contour feature extraction results, pass / fail judgment result, and defect location index. If the judgment result is unqualified, the coordinate range of the out-of-tolerance area, the corresponding grid number, and the suggested rework area number are further marked in the quality assessment report for direct reference during subsequent rework.

[0098] S7 includes parsing the quality assessment report to extract feature vectors, mapping the feature vectors to the dynamic perception model to obtain the feature distribution matrix; calculating the residual value based on the feature distribution matrix, and updating the cluster center coordinates and cluster radius if the residual value exceeds the threshold; using the updated cluster center coordinates and cluster radius to divide the grinding force feedback data, classifying the grinding force feedback data into processing state clusters and determining the compensation vector; loading the compensation vector into the control system to correct the grinding trajectory, and achieving dynamic compensation of processing accuracy through the correction of the grinding trajectory to obtain the complete grinding process optimization result.

[0099] In this embodiment, S7 corresponds to the closed-loop feedback correction process of the entire polishing process. This embodiment rewrites the regional quality results from the quality assessment report into the dynamic perception model, and updates the clustering parameters, redistributes the force feedback state, and generates compensation vectors accordingly, ultimately transforming the backend quality conclusions into the basis for frontend trajectory correction.

[0100] In this embodiment, immediately after step S6, the quality assessment report corresponding to the current workpiece is read and parsed zone by zone according to the processing area number. Nine characteristic data items are extracted from each processing area, in the following order: processing area number, arithmetic mean roughness value, maximum deviation value, average absolute deviation value, defect area, transverse coordinate value of the defect area center point, longitudinal coordinate value of the defect area center point, area acceptance mark, and rework level. The area acceptance mark uses "1" to indicate acceptance and "0" to indicate rejection; the rework level is fixed at three levels, where "1" indicates no rework required, "2" indicates partial rework, and "3" indicates key rework. After reading, the above nine characteristic data items are processed using unified dimensions: the arithmetic mean roughness value, maximum deviation value, average absolute deviation value, and defect area are converted proportionally according to the pre-stored range upper limit in the process database; the transverse and longitudinal coordinate values ​​of the defect area center point remain unchanged in millimeters under the workpiece's global coordinate system; the area acceptance mark and rework level are directly written as integer values. After dimension unification, each processing area forms a feature vector of fixed length.

[0101] Furthermore, the feature vector is then written into the region status table corresponding to the dynamic sensing model. The region status table in the dynamic sensing model has 11 fixed fields: region number, historical average roughness, historical average maximum deviation, current roughness value, current maximum deviation value, average absolute deviation, defect area, defect center lateral coordinate, defect center longitudinal coordinate, region pass / fail flag, and rework level. After writing all regions, the entire region status table is expanded into a two-dimensional matrix according to the actual arrangement of the workpiece surface processing areas. The matrix rows are aligned with the workpiece length direction, and the matrix columns are aligned with the workpiece width direction. Each matrix unit corresponds to the status result of one processing area, thus obtaining the feature distribution matrix. Therefore, the feature distribution matrix in this embodiment transforms the quality assessment results into a spatial distribution result that can be directly called by the dynamic sensing model.

[0102] In this embodiment, after the feature distribution matrix is ​​constructed, the residual value is the comprehensive deviation between the prediction result of the dynamic sensing model and the quality assessment result of the current round. Specifically, the process first reads the three predicted values ​​saved by the dynamic sensing model after the previous round of trajectory optimization: predicted roughness value, predicted maximum deviation value, and predicted defect area value. Then, it reads the three actual values ​​corresponding to the current processing area from the feature distribution matrix: the current round roughness value, the current round maximum deviation value, and the current round defect area value. Next, it calculates the absolute values ​​of the differences between the predicted roughness value and the current round roughness value, the absolute values ​​of the differences between the predicted maximum deviation value and the current round maximum deviation value, and the absolute values ​​of the differences between the predicted defect area value and the current round defect area value, and performs cumulative processing with fixed weights. Specifically, the weights for the roughness difference, maximum deviation, and defect area are fixed at 35%, 35%, and 30%, respectively. After cumulative processing, the residual value for the current processing area is obtained. After calculating the residual values ​​for all processing areas, the residual values ​​are written back to the corresponding feature distribution matrix units. The threshold determination process is fixed as follows: Select 8 workpieces of the same model, complete the full grinding process and quality assessment respectively, and record the residual value of each processing area; then sort all the residual values ​​according to the numerical size, and take the end value in the first 90% as the residual threshold.

[0103] In this embodiment, the residual threshold is set to 0.12. During actual operation, when the residual value of any processing area reaches or exceeds 0.12, it is determined that there is a significant deviation between the prediction result of the dynamic perception model for that area and the actual quality result, and the clustering parameters need to be updated. The clustering parameter update is divided into two parts: the first part is the update of the cluster center coordinates, and the second part is the update of the cluster radius. For the update of the cluster center coordinates, the system reads the original horizontal and vertical coordinate values ​​of the center of the current cluster region, and reads the horizontal and vertical coordinate values ​​of the center point of the current defect region; when the distance between the original cluster center and the center of the defect region reaches or exceeds 1.0 mm but is less than 2.0 mm, the cluster center moves 0.5 mm along the direction of the defect region center; when the distance reaches or exceeds 2.0 mm, the cluster center moves 1.0 mm along the direction of the defect region center; when the distance is less than 1.0 mm, the cluster center remains unchanged. For cluster radius updates, the processing is based on the defect area: when the defect area reaches or exceeds 25 mm², the cluster radius increases by 1.0 mm; when the defect area reaches or exceeds 9 mm² but is less than 25 mm², the cluster radius increases by 0.5 mm; when the defect area is less than 9 mm², the cluster radius remains unchanged.

[0104] Next, the grinding force feedback data corresponding to the current workpiece is read from the execution data cache of S4. This data includes five items: spindle output torque, normal contact force, tangential resistance, actual trajectory running speed, and removal amount per unit path. The sampling period is fixed at 20ms. After reading, all grinding force feedback data is mapped to the updated clustering region based on the trajectory position coordinates corresponding to each force feedback sampling point. Sampling points whose trajectory position coordinates fall within the same cluster center and the defined cluster radius are grouped into the same data group to be classified.

[0105] Then, for each data group to be classified, four state characteristics are calculated: average normal contact force, normal contact force fluctuation value, average tangential resistance, and average unit path removal. All data groups are then categorized into three machining state clusters according to preset rules. The first cluster is the stable machining state cluster, defined as follows: average normal contact force fluctuation value not exceeding 6N, average tangential resistance fluctuation ratio not exceeding 5%, and average unit path removal deviation not exceeding 0.01mm. The second cluster is the over-cutting state cluster, defined as follows: average unit path removal reaches or exceeds 120% of the target removal amount, or average spindle output torque reaches or exceeds 85% of the spindle's rated torque. The third cluster is the under-cutting state cluster, defined as follows: average unit path removal is less than 85% of the target removal amount, or average normal contact force is less than 80% of the target contact force. After the state clusters are classified, compensation vectors are generated according to the cluster category.

[0106] Specifically, the compensation vector is fixed and consists of three compensation amounts: trajectory normal compensation, feed rate compensation, and contact pressure compensation. For stable machining states, the trajectory normal compensation is set to 0 mm, the feed rate compensation to 0%, and the contact pressure compensation to 0 N. For overcutting states, the trajectory normal compensation is set to offset outward by 0.10 mm along the current surface normal, the feed rate compensation is increased by 8%, and the contact pressure compensation is decreased by 6 N. For undercutting states, the trajectory normal compensation is set to offset inward by 0.08 mm along the current surface normal, the feed rate compensation is decreased by 10%, and the contact pressure compensation is increased by 8 N.

[0107] Finally, in this embodiment, before each trajectory segment begins execution, the control system first reads the processing state cluster category and compensation vector content corresponding to that trajectory segment. The compensation vector loading sequence is fixed at three steps. Step 1: The trajectory normal compensation is superimposed on all trajectory points of the current trajectory segment, and the spatial position of each trajectory point is recalculated along its surface normal direction. Step 2: The feed rate compensation is written into the trajectory parameter table, and the target feed rate of all trajectory points in the trajectory segment is uniformly corrected. Step 3: The contact pressure compensation is written into the force control parameter table, and the target contact pressure of the trajectory segment is synchronously corrected. After the trajectory correction is completed, the inverse kinematics process is performed again on the corrected trajectory points to generate a new equipment control command sequence, and the subsequent trajectory segments that have not yet been executed in the original control command sequence are replaced.

[0108] Furthermore, to prevent trajectory jumps after compensation loading, this embodiment also performs continuity checks on adjacent trajectory points. The continuity check includes two conditions: the positional difference between two adjacent trajectory points does not exceed 1.5mm, and the attitude angle change between two adjacent trajectory points does not exceed 4°. If either condition is not met, spline smoothing is performed on the current corrected trajectory segment, and then control commands are reissued. After all trajectory segments have completed compensation loading, the system writes the updated cluster center coordinates, cluster radius, processing state cluster category, compensation vector content, corrected trajectory parameters, and quality improvement results of this round into the process optimization record table, forming a complete grinding process optimization result. This complete grinding process optimization result includes at least the region optimization number, updated cluster parameters, state cluster classification results, compensation vector content, corrected trajectory parameters, roughness value before correction, roughness value after correction, maximum deviation value before correction, and maximum deviation value after correction. Thus, S7 outputs the complete set of process optimization results after quality feedback, parameter updates, force feedback reclassification, and trajectory compensation.

[0109] The present invention also discloses a machine-executable program that can be automatically executed by a machine to realize the method for automatically generating grinding trajectories of complex curved surfaces based on point cloud segmentation as described above.

[0110] The machine (electronic device) mentioned above is, for example, a microcontroller, a single-board computer, a desktop computer, a laptop computer, a server, a programmable controller, or a field-programmable gate array.

[0111] Figure 2 This is a structural block diagram of the local terminal of an exemplary electronic device (machine) of the present invention; as shown... Figure 2 As shown, the electronic device of the present invention includes a processor 11, a memory 12 and a storage space 13 for storing a machine-executable program 14, the machine-executable program 14 being used to execute the above-described control logic.

[0112] Figure 3 This is a structural block diagram of the network end of an exemplary electronic device of the present invention; as shown below. Figure 3 As shown, the present invention also provides an electronic device (machine), which may include at least one processor 210, at least one memory 230 communicatively connected to the processor, and a communication bus 240 connecting different system components (including memory 230 and processor 210); wherein, memory 230 stores a machine executable program that can be executed by the processor, and processor 210 can execute the above-mentioned control logic by calling the machine executable program.

[0113] Communication bus 240 represents one or more of several bus architectures, including a memory bus or memory controller, peripheral bus, graphics acceleration port, processor, or local bus using any of the various bus architectures. For example, these architectures include, but are not limited to, Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MAC) bus, Enhanced ISA bus, Video Electronics Standards Association (VESA) local bus, and Peripheral Component Interconnection (PCI) bus.

[0114] Electronic devices typically include a variety of computer system readable media, which can be any available media that can be accessed by the electronic device, including volatile and non-volatile media, and removable and non-removable media.

[0115] Memory 230 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. The electronic device may further include other removable / non-removable, volatile / non-volatile computer system storage media. Memory 230 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the control logic described above.

[0116] A program / utility having a set (at least one) of program modules can be stored in memory 230. Such program modules include, but are not limited to, an operating system, one or more applications, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment.

[0117] Machine-executable programs for performing this invention can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages ​​such as Java, C++, and Python, and may also include specialized engineering languages ​​such as R. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0118] The present invention also discloses a storage medium on which a machine-executable program as described above is stored.

[0119] The aforementioned storage medium may be any combination of one or more computer-readable media. Computer-readable media may be, for example, computer-readable signal media or computer-readable storage media. Computer-readable storage media include, but are not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), or flash memory, optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable storage medium may be, for example, any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0120] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0121] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0122] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for automatically generating grinding trajectories for complex curved surfaces based on point cloud segmentation, characterized in that, include: S1. Obtain the initial surface data of the curved workpiece, collect point cloud information by scanning the sensor array, perform preliminary filtering on the real-time changing area, and obtain the preliminary point cloud dataset. S2. Use deep learning algorithms to identify key features on the surface, and perform point cloud segmentation on the preliminary point cloud dataset. Extract the curvature and roughness distribution features of each surface region after segmentation to obtain feature extraction results and point cloud segmentation results. S3. Based on the feature extraction results and point cloud segmentation results, generate a dynamic perception model, optimize the path for real-time changes, group the segmented similar curved surface regions, and generate an optimized polishing trajectory plan. S4. Obtain the current processing position data from the optimized grinding trajectory planning, combine it with the surface change information fed back by the sensor in real time, and determine if the change exceeds the preset threshold. Then adjust the grinding parameters to obtain the updated surface state. S5. Construct a 3D reconstruction model, fuse multiple frames of data through point cloud registration technology, determine the precise geometric shape, and obtain the restored geometric data; S6. Calculate the surface roughness and deviation index using statistical analysis methods to obtain a quality assessment report; S7. Read the quality assessment report corresponding to the current workpiece, extract multiple feature data, and the feature data index includes at least the arithmetic mean roughness value, the maximum deviation value, and the area qualification mark. Perform unified dimension processing on the multiple feature data to form a fixed-length feature vector. Parse the quality assessment report to extract the feature vector and map the feature vector to the dynamic perception model to obtain the feature distribution matrix. The residual value is calculated based on the feature distribution matrix. If the residual value exceeds the threshold, the cluster center coordinates and cluster radius are updated. The updated cluster center coordinates and cluster radius are used to divide the grinding force feedback data, classify the grinding force feedback data into processing state clusters, and determine the compensation vector. The compensation vector is loaded into the control system to correct the grinding trajectory. The correction of the grinding trajectory achieves dynamic compensation of the processing accuracy, and the complete grinding process optimization result is obtained. S3 includes: The raw point cloud data is acquired and geometric features are extracted. Multiple segmented surfaces are obtained through point cloud segmentation. The segmented surface is grouped by similarity using a clustering algorithm to determine cluster regions with consistent physical properties; The real-time displacement data generated by the clustered regions during the processing is input into the dynamic sensing model to obtain the sensing state reflecting the deformation of the workpiece; The path is optimized by combining the perceived state with the preset processing accuracy requirements. If the perceived state deviates from the preset threshold, the optimized path that avoids obstacles and fits the surface is recalculated. The optimized path is transformed into a grinding trajectory plan through inverse kinematics, enabling real-time adjustments during the grinding operation and achieving automatic generation and optimization of the grinding trajectory under the guidance of a dynamic perception model.

2. The method for automatically generating grinding trajectories of complex curved surfaces based on point cloud segmentation according to claim 1, characterized in that: S1 includes: The sensor array collects spatial coordinate data of the curved workpiece surface, and maps it to a unified three-dimensional coordinate system through a spatial coordinate transformation matrix to obtain the first point cloud dataset. The local curvature change rate is calculated based on the first point cloud dataset. If the local curvature change rate exceeds a preset threshold, the corresponding region is marked as a dynamically perceived region to obtain the second point cloud dataset. A bilateral filtering algorithm is used to remove outlier noise points from the second point cloud dataset to obtain the third point cloud dataset; A local surface normal vector field is constructed based on the third point cloud dataset, and redundant points are eliminated through normal vector consistency constraints to obtain the fourth point cloud dataset. The local surface equation is fitted based on the fourth point cloud dataset, and points whose Euclidean distance exceeds a preset threshold are removed to obtain the fifth point cloud dataset. Based on the cloud dataset in point 5, a dynamic perception model is constructed, and the initial surface data of complex curved workpieces is dynamically perceived and adjusted by updating the model parameters in real time.

3. The method for automatically generating grinding trajectories for complex curved surfaces based on point cloud segmentation according to claim 1, characterized in that: S2 includes: Multi-scale feature extraction is performed on the original point cloud using a convolutional neural network to obtain depth features on the curved surface; The point cloud topology is constructed using deep features, and the point cloud is divided into regions based on the density of the topology connections to obtain segmented point clouds with region labels. The local normal vectors of the segmentation points are calculated using cloud computing, thereby determining the discrete curvature of each surface region; Based on the fluctuation of discrete curvature in spatial distribution, the surface roughness of each surface region is extracted to form a set of geometric properties reflecting surface quality; If the values ​​in the geometric attribute set exceed the preset path constraints, the feed rate of the machining trajectory is adjusted to obtain the real-time analysis results of the machining path. By summarizing depth features, segmented point clouds, discrete curvature, and surface roughness, the system outputs feature extraction results and point cloud segmentation results, enabling real-time analysis of the processing path.

4. The method for automatically generating grinding trajectories for complex curved surfaces based on point cloud segmentation according to claim 1, characterized in that: S4 includes: The trajectory planning coordinate data and sensor data are acquired and spatially mapped to obtain the surface roughness value of the current processing point; If the surface roughness value exceeds the preset surface morphology threshold, the correction amount for the current grinding pressure and feed speed is determined. The correction amount of the current grinding pressure and feed speed is applied to the grinding actuator to obtain the grinding trajectory and machining torque data; By performing a differential calculation between the grinding trajectory and processing torque data and the initial geometric model, a geometric deviation distribution map of the current processed surface is obtained. Based on the geometric deviation distribution map of the current processed surface, the final geometric shape restoration data is obtained through an iterative correction algorithm, thereby realizing geometric shape restoration.

5. The method for automatically generating grinding trajectories for complex curved surfaces based on point cloud segmentation according to claim 1, characterized in that: S5 includes: Surface state data is acquired and mapped to a unified global coordinate system to obtain the first spatial point cloud data; Based on the first spatial point cloud data, multi-frame feature point matching is performed to obtain the second spatial point cloud data; A three-dimensional reconstruction model is constructed by fitting a surface to the second spatial point cloud data, and then the third spatial point cloud data is obtained. If the deviation analysis result between the third spatial point cloud data and the preset geometric benchmark exceeds the preset threshold, then local point cloud resampling is performed on the third spatial point cloud data to obtain the fourth spatial point cloud data. Based on the consistency verification of the fourth spatial point cloud data and the determination of the precise geometric shape, the restored geometric data is obtained; Use the restored geometric data to prepare for the quality assessment.

6. The method for automatically generating grinding trajectories for complex curved surfaces based on point cloud segmentation according to claim 1, characterized in that: S6 includes: The original point cloud data of the processing area is acquired and the least squares method is used to fit and generate the restored geometric model. The spatial coordinate information of each sampling point on the surface of the restored geometric model is extracted. The normal distance of each sampling point relative to the theoretically designed curved surface is calculated based on the spatial coordinate information, and the deviation index distribution matrix reflecting the surface morphology of the processing area is obtained.

7. The method for automatically generating grinding trajectories for complex curved surfaces based on point cloud segmentation according to claim 6, characterized in that: S6 further includes: A sliding window algorithm is used to smooth the local area of ​​the deviation index distribution matrix, extract the surface contour features, and calculate the arithmetic mean roughness value corresponding to the surface contour features. If both the arithmetic mean roughness value and the maximum deviation value in the deviation index distribution matrix are less than the preset standard, the processed area is deemed qualified.

Images