Polishing parameter determination method, electronic equipment and medium

Abstract

The invention provides a polishing parameter determination method, electronic equipment and a medium. The method comprises the steps that point cloud data of a to-be-polished surface of a workpiece and image data of the to-be-polished surface are collected; the point cloud data are input into a classification model, the classification model outputs an initial classification result, and the initial classification result is used for representing the probability that the point cloud coordinate of each piece of point cloud data is convex and / or concave; obtaining an initial convex region and an initial concave region based on the initial classification result; correcting the initial convex area and the initial concave area based on the image data to obtain a final convex area and a final concave area of the to-be-polished surface; according to a preset algorithm, polishing parameter range values are obtained according to the point cloud coordinates corresponding to the final convex area and the final concave area; and based on a genetic algorithm model, determining an optimal polishing parameter in the range values of the polishing parameters. The polishing effect is improved.

Description

Technical Field

[0001] The present invention relates to the field of polishing technology, and in particular to a polishing parameter determination method, electronic equipment and medium. Background Art

[0002] In today's manufacturing industry, surface grinding is an essential step in the production of numerous products. However, traditional manual polishing methods have numerous drawbacks. Traditional manual methods require manual identification of bumps and depressions on the polished surface, making it difficult to identify even small irregularities and prone to significant machining errors. Summary of the Invention

[0003] In view of this, one of the objectives of the embodiments of the present application is to provide a method for determining polishing parameters, which can improve the problem of large machining errors.

[0004] To achieve the above technical objectives, the technical solutions adopted in this application are as follows:

[0005] In a first aspect, an embodiment of the present application provides a method for determining polishing parameters, the method comprising:

[0006] Collecting point cloud data of the surface to be polished of the workpiece and image data of the surface to be polished;

[0007] Inputting the point cloud data into a classification model, the classification model outputting an initial classification result, wherein the initial classification result is used to characterize the probability that the point cloud coordinates of each point cloud data are convex and / or concave;

[0008] Based on the initial classification result, an initial convex area and an initial concave area are obtained;

[0009] Correcting the initial raised area and the initial recessed area based on the image data to obtain a final raised area and a final recessed area of the surface to be polished;

[0010] Obtaining polishing parameter range values based on the point cloud coordinates corresponding to the final raised area and the final recessed area according to a preset algorithm;

[0011] Based on the genetic algorithm model, the optimal polishing parameters are determined within the range of values of the polishing parameters.

[0012] Furthermore, the final concave region includes a plurality of concavities, and the final convex region includes a plurality of convexities;

[0013] The point cloud coordinates corresponding to the final raised area and the final recessed area are used to obtain a polishing parameter range value according to a preset algorithm, including:

[0014] Extracting the maximum depth of the concave, the coverage of the final concave area, and the edge sharpness of the final concave area, as well as the maximum height of the convex, the main curvature radius of the final convex area, and the distribution density of the convex in the final convex area;

[0015] The polishing parameter range value is determined according to a preset algorithm based on the maximum depth of the depression, the coverage of the final depression area, the edge sharpness of the final depression area, the maximum height of the protrusion, the main curvature radius of the final protrusion area and the distribution density of the protrusions in the final protrusion area.

[0016] Further, the polishing parameters include polishing force, feed speed and dwell time, and the polishing parameter range values include polishing force parameter range values, feed speed range values and dwell time range values;

[0017] Determining the polishing parameter range value according to a preset algorithm includes:

[0018] The polishing force parameter range, feed speed range, and dwell time range are determined according to a preset algorithm, wherein the preset algorithm is:

[0019] F1=k1·h 1.2 +F base

[0020]

[0021] F min =min{F1, F2}

[0022] F max =0.8σ y A1 0.5

[0023] The polishing force parameter range is expressed as F∈(F min , F max );

[0024] k1 and k2 represent the material hardness coefficient of the workpiece and the wear compensation coefficient of the polishing tool respectively;

[0025] F base Indicates the base pressure;

[0026] F min Indicates the minimum polishing force;

[0027] F max Indicates the maximum polishing force;

[0028] F represents polishing force;

[0029] σ yrepresents the yield strength of the workpiece;

[0030] A1 represents the effective area of the grinding wheel of the polishing machine;

[0031] H represents the maximum height of the bulge, which is the Z-direction value of the bulge’s point cloud coordinates;

[0032] R represents the main curvature radius of the final convex area;

[0033] h represents the maximum depth of the depression;

[0034] v 1max =20+5e -0.3A2

[0035] v 1min =20

[0036] v 2max =15·tanh(0.1R)

[0037] v 2min =1.2·(R·ρ1) 0.3

[0038] v min =min{v 1min ,v 2min}

[0039] v max =max{v 1max ,v 2max}

[0040] The feed speed range value is expressed as: v∈(v max ,v min );

[0041] Wherein, v represents the feed speed;

[0042] ρ1 represents the distribution density of the protrusions in the final protrusion area;

[0043] A2 represents the coverage of the final concave area;

[0044] t min =0

[0045]

[0046] m=max{H,h

[0047] The dwell time range value can be expressed as: t∈(t min ,t max )

[0048] Among them, c m represents the material coefficient of the workpiece;

[0049] ρ2 represents the density of the workpiece.

[0050] Furthermore, the correcting the initial raised area and the initial recessed area based on the image data to obtain the final raised area and the final recessed area of the surface to be polished includes:

[0051] grayscale processing of the image data;

[0052] Screening all first pixel points whose brightness values are greater than a first threshold value and all second pixel points whose brightness values are less than a second threshold value in the image data after grayscale processing;

[0053] Based on all the first pixel points and all the second pixel points, the initial convex area and the initial concave area are corrected to obtain the final convex area and the final concave area.

[0054] Furthermore, the correcting the initial convex area and the initial concave area based on all the first pixel points and all the second pixel points to obtain an optimized classification result includes:

[0055] Based on all the first pixel points and all the second pixel points, a plurality of first camera areas and a plurality of second camera areas are obtained respectively;

[0056] Mapping the first pixel point of the first camera area and the second pixel point of the second camera area to the coordinate system corresponding to the point cloud data, to obtain a first point cloud area corresponding to the first camera area and a second point cloud area corresponding to the second camera area, respectively;

[0057] The first point cloud region and the initial convex region are merged to obtain a final convex region, and the second point cloud region and the initial concave region are merged to obtain a final concave region.

[0058] Furthermore, the genetic algorithm model is used to determine the optimal polishing parameters within the range of polishing parameters, taking the minimum surface roughness of the surface to be polished as the surface roughness function, including:

[0059] Based on the polishing force parameter range value, the feed speed range value and the dwell time range value, a plurality of polishing parameter discrete points are obtained;

[0060] Inputting the point cloud coordinates corresponding to the polishing parameter discrete points, the final raised area, and the final recessed area into a preset deep learning model to obtain the surface roughness corresponding to the polishing parameter discrete points;

[0061] Taking the polishing parameter discrete points corresponding to the surface roughness being less than or equal to the surface roughness threshold as parent discrete points, and iterating until the optimal polishing parameters are obtained;

[0062] The iterations include:

[0063] Based on the parent discrete points, crossover and / or mutation are performed to obtain offspring discrete points;

[0064] Inputting the offspring discrete points and optimized classification results into the deep learning model;

[0065] Screening the child discrete points whose surface roughness is less than or equal to the surface roughness threshold as the parent discrete points;

[0066] Repeat the crossover and / or mutation based on the parent discrete points to obtain child discrete points; input the child discrete points and the optimized classification results into the deep learning model, and screen the child discrete points whose surface roughness is less than or equal to the surface roughness threshold as the parent discrete points.

[0067] Inputting the polishing parameter discrete points into a preset deep learning model to obtain the surface roughness corresponding to the polishing parameter discrete points includes:

[0068] Collecting historical data, the historical data including historical optimization classification results, historical polishing forces, historical feed speeds, historical dwell times, and historical surface roughness of the workpiece in a one-to-one correspondence;

[0069] Based on the historical data, a preset deep learning model is trained;

[0070] The polishing parameter discrete points are input into a preset deep learning model to obtain the surface roughness corresponding to the polishing parameter discrete points.

[0071] Furthermore, the classification model includes an input layer, a feature extraction layer, a global feature fusion layer, a feature propagation layer and a classification output layer;

[0072] The input layer is configured to receive each point cloud coordinate in the point cloud data;

[0073] The feature extraction layer is configured to select M point cloud coordinates from the point cloud data as the center point cloud coordinates based on the farthest point sampling algorithm, and divide the point cloud data into M local areas with the center point cloud coordinates as the center and n as the radius, and extract local features of each local area based on the MLP algorithm to form M local feature matrices;

[0074] The global feature fusion layer is configured to perform global pooling on the local feature matrix to obtain a global feature matrix;

[0075] The global feature fusion layer is further configured to concatenate the global feature matrix with the local features of each local region to obtain a feature enhancement matrix;

[0076] The feature propagation layer is configured to obtain a feature vector for each point cloud coordinate based on a feature enhancement matrix according to a distance weighted algorithm;

[0077] The classification output layer is configured to output a probability of the point cloud coordinates being convex or concave based on the feature vector of the point cloud coordinates.

[0078] Furthermore, after determining the optimal polishing parameter within the range of polishing parameters based on the genetic algorithm model, the method further includes:

[0079] Based on the optimal polishing parameters, controlling a polishing machine to polish the surface to be polished;

[0080] collecting the actual surface roughness of the surface to be polished, and when the difference between the actual surface roughness and the preset surface roughness is greater than a preset difference, adjusting the optimal polishing parameters so that the difference between the actual surface roughness and the preset surface roughness is no greater than the preset difference;

[0081] The adjusting of the optimal polishing parameters comprises:

[0082] Based on the optimal polishing parameters and the preset adjustment amount, obtaining the adjusted optimal polishing parameters;

[0083] Inputting the adjusted optimal polishing parameters into a surface roughness calculation model to obtain an estimated surface roughness, and if a difference between the estimated surface roughness and the preset surface roughness is greater than a preset difference, repeating the following steps until the difference between the estimated surface roughness and the preset surface roughness is greater than the preset difference;

[0084] Obtaining an estimated surface roughness based on the adjusted optimal polishing parameters, the preset adjustment amount, and the surface roughness calculation model;

[0085] The surface roughness calculation model is:

[0086]

[0087] Among them, Ra represents the surface roughness;

[0088] K represents the comprehensive correction factor;

[0089] F represents polishing force;

[0090] α represents the nonlinear effect of polishing force on surface roughness, α<0;

[0091] v represents the feed speed;

[0092] β represents the nonlinear effect of feed rate on surface roughness, β<0;

[0093] t represents the dwell time;

[0094] γ represents the nonlinear effect of residence time on surface roughness, γ<0;

[0095] C represents the basic surface roughness constant;

[0096] P i Indicates the process parameters that affect surface roughness, δ i represents an independent index corresponding to the process parameters, wherein the process parameters include abrasive particle size, abrasive concentration, material hardness of the workpiece, temperature of the polishing area, and vibration frequency of the polishing tool.

[0097] In a second aspect, an embodiment of the present application provides an electronic device, which includes a processor and a memory coupled to each other, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the electronic device executes the above method.

[0098] In a third aspect, an embodiment of the present application provides a computer-readable storage medium, characterized in that a computer program is stored in the computer-readable storage medium, and when the computer program is run on a computer, the computer executes the above method.

[0099] The invention adopting the above technical solution has the following advantages:

[0100] In the technical solution provided in this application, the three-dimensional geometric features of the surface to be polished are captured based on point cloud data, which improves the accuracy of identifying protrusions and depressions;

[0101] In the technical solution provided by this application, the probability of each point cloud coordinate being concave or convex is determined by a classification model, and then the concave area and convex area are obtained, thereby improving the accuracy of identifying the concave area and convex area;

[0102] In the technical solution provided in this application, the initial concave area and the initial convex area are further corrected based on the image data, thereby further improving the accuracy of identifying the concave area and the convex area;

[0103] In the technical solution provided in the present application, polishing parameters are optimized through a genetic algorithm, thereby further improving the accuracy of identifying concave areas and convex areas, thereby improving the effect after polishing. BRIEF DESCRIPTION OF THE DRAWINGS

[0104] The present application may be further illustrated by the non-limiting embodiments provided in the accompanying drawings. It should be understood that the following drawings illustrate only certain embodiments of the present application and are therefore not to be construed as limiting the scope of the present application. It is understood that a person skilled in the art can derive other relevant drawings from these drawings without inventive effort.

[0105] Figure 1 A flowchart provided for an embodiment of the present application.

[0106] Figure 2 This is a sub-flowchart of S140 provided in an embodiment of the present application.

[0107] Figure 3 This is a sub-flowchart of S141 provided in an embodiment of the present application.

[0108] Figure 4 This is a sub-flowchart of S160 provided in an embodiment of the present application.

[0109] Figure 5 This is a sub-flowchart of S170 provided in an embodiment of the present application. DETAILED DESCRIPTION

[0110] The present application will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that similar or identical parts in the drawings or descriptions are numbered the same. Implementations not shown or described in the drawings are known to those of ordinary skill in the art. In the description of this application, the terms "first," "second," etc. are used solely to distinguish descriptions and are not to be construed as indicating or implying relative importance.

[0111] An embodiment of the present application provides an electronic device including a processing module and a storage module. The storage module stores a computer program, which, when executed by the processing module, enables the electronic device to perform corresponding steps in the polishing parameter determination method described below.

[0112] Please refer to Figure 1 The present application also provides a method for determining polishing parameters. The voice message method may include the following steps:

[0113] S110, collecting point cloud data of the surface to be polished of the workpiece, and image data of the surface to be polished;

[0114] S120, inputting the point cloud data into a classification model, the classification model outputting an initial classification result, the initial classification result being used to characterize the probability that each point cloud coordinate of the point cloud data is a convexity and / or a concaveity;

[0115] S130, obtaining an initial convex region and an initial concave region based on the initial classification result;

[0116] S140: Correcting the initial classification result based on the image data to obtain an optimized classification result, wherein the optimized classification result includes a final raised area and a final recessed area of the surface to be polished;

[0117] S150, obtaining a polishing parameter range value based on the point cloud coordinates corresponding to the optimization classification result according to a preset algorithm;

[0118] S160 , determining an optimal polishing parameter within the range of polishing parameters based on a genetic algorithm model.

[0119] The following are the steps of the polishing parameter determination method:

[0120] In step 110, point cloud data refers to a set of vectors representing an object or scene in a three-dimensional coordinate system. Each point contains three-dimensional coordinates (X, Y, Z) and may carry additional information such as color, reflection intensity, and time. It is acquired by discretely sampling the surface of an object using a device such as a 3D scanner or lidar, and can accurately express the spatial distribution and surface characteristics of the target.

[0121] In this embodiment, a point cloud acquisition module is composed of a line laser scanner (accuracy ±5μm) and an industrial camera. The scanning frequency is 100Hz, generating a high-density point cloud containing X, Y, and Z coordinates and normal vectors. This scan is used to obtain initial point cloud data from the surface to be polished. The initial point cloud data is then processed using a Gaussian filter algorithm to smooth out noise and generate the point cloud data in step 110 for subsequent calculations.

[0122] The image data of the surface to be polished is two-dimensional plane data. When acquiring the image data of the surface to be polished of the workpiece in step 110, high-quality acquisition can be achieved by the following methods to ensure effective complementarity with the point cloud data:

[0123] 1. Image acquisition device selection

[0124] Industrial cameras: Prioritize high-resolution (e.g., 20 megapixels or above) and low-distortion industrial cameras that support RAW format output to preserve original details.

[0125] Lens configuration: Select a fixed-focus or zoom lens according to the size of the workpiece. The focal length needs to cover the area to be polished, and the aperture should be adjusted to a moderate value (such as F5.6-F8) to balance the depth of field and resolution.

[0126] Light source system:

[0127] Ring Light: Evenly illuminates curved surfaces or complex geometry, reducing shadows.

[0128] Coaxial light: For highly reflective materials (such as metal), reducing mirror reflection interference.

[0129] Multi-spectral light source: Optional UV or infrared light source can enhance the contrast of specific defects (such as scratches and oxide layers).

[0130] 2. Image Acquisition Strategy

[0131] Multi-angle shooting:

[0132] For large workpieces, a turntable + multi-angle shooting (such as collecting images every 30°) is used to cover the entire surface to be polished.

[0133] For small precision parts, a microscope lens is used in conjunction with an automatic translation stage to perform area-by-area scanning.

[0134] Resolution and overlap:

[0135] The resolution of a single image must be below 0.05mm / pixel, and the overlap rate of adjacent images must be ≥30% to ensure stitching accuracy.

[0136] Trigger synchronization:

[0137] If working synchronously with a laser scanner, a hardware trigger signal is required to achieve millisecond-level synchronous acquisition to avoid motion blur.

[0138] 3. Image Preprocessing

[0139] Denoising: Use median filtering or the BM3D algorithm to remove sensor noise.

[0140] Enhancements:

[0141] Histogram equalization improves dark details.

[0142] Anisotropic diffusion filtering smoothes surface textures while preserving edges.

[0143] Dimension alignment:

[0144] Affix coded markers (such as ArUco codes) to the surface of the workpiece for subsequent spatial alignment with point cloud data.

[0145] 4. Data Fusion Preparation

[0146] Coordinate system alignment:

[0147] The intrinsic parameter matrix is obtained through camera calibration (Zhang Zhengyou calibration method).

[0148] Use marker points or common feature points (such as edges and holes) to realize the conversion between the image pixel coordinate system and the point cloud three-dimensional coordinate system.

[0149] Unified data format:

[0150] Convert the image to grayscale or HSV color space to facilitate correlation analysis with point cloud intensity information.

[0151] 5. Special scene processing

[0152] Transparent / Translucent Materials:

[0153] Capture after spraying developer (such as matte white paint), or use a backlight + polarizer combination.

[0154] Dynamically deforming surfaces:

[0155] Use a high-speed camera (>1000fps) with a stroboscopic light source to freeze the motion moment.

[0156] In step 120, the classification model can be established in the following manner:

[0157] 1. Data Collection

[0158] Collect historical point cloud data from multiple workpiece surfaces using LiDAR or 3D reconstruction technology. Use specialized tools (such as LabelMe) to segment the point cloud data into regions, labeling them as smooth, convex, or concave. Color and texture features in the image can help correct for annotation errors (for example, color differences in oxidized areas may be mistakenly labeled as concave).

[0159] Data preprocessing: Denoising - using statistical filtering (eliminating outliers whose distance standard deviation exceeds the threshold) or radius filtering (removing isolated points); Downsampling - reducing the amount of data through voxel grid method (Voxel Downsampling), such as dividing the point cloud into 1mm 3 Voxels, each voxel retains the center point. Feature enhancement - calculates geometric features such as normal vectors and curvature, and fuses image RGB values into additional attributes of the point cloud

[0160] 2. Model Construction and Training

[0161] In view of the characteristics of point cloud, PointNet++ or 3D-CNN architecture is preferred. When PointNet++ is used, the disordered point cloud data is processed through layered sampling and local feature aggregation. When 3D-CNN is used, the point cloud is converted into a voxel grid (such as dividing the space into 0.5mm 3 grid), extract features through 3D convolution

[0162] 3. Training process:

[0163] Forward propagation: input point cloud data, extract local geometric features (such as edges, curvature changes) through the convolution layer, reduce the dimension through the pooling layer, and output the classification probability through the fully connected layer.

[0164] Loss calculation: Use weighted cross entropy loss function to adjust weights for class imbalance problems (such as fewer samples in concave areas)

[0165] Backpropagation: Optimize network parameters through gradient descent and dynamically adjust the learning rate using the Adam optimizer (the initial value is often set to 0.001)

[0166] 4. Training Optimization Strategy

[0167] Data enhancement: Randomly rotate (±15°), translate (±5mm), and scale (0.9-1.1 times) the point cloud to simulate scanning results from different perspectives

[0168] Regularization: Add a Dropout layer (dropout rate 0.3-0.5) or L2 regularization to prevent the model from overfitting.

[0169] Multi-scale training: Mixing point clouds of different resolutions (such as original point clouds and downsampled point clouds) in the input stage to improve model robustness.

[0170] 5. Model Validation and Tuning

[0171] Classification accuracy: Calculate the confusion matrix and calculate the accuracy and recall rate (for example, the recall rate of the concave area must be >90%).

[0172] Edge consistency: Use conditional random fields (CRF) to force the classification results of adjacent points to be smooth and reduce isolated error points

[0173] Hyperparameter search: Use the grid search method to optimize the convolution kernel size (commonly used 3×3×3 or 5×5×5) and the learning rate decay strategy (such as 50% decay every 10 rounds).

[0174] Multi-model integration: Train multiple networks (such as 3D-CNN and PointNet++ in parallel), fuse classification results through a voting mechanism, and improve generalization performance.

[0175] Therefore, in S120, the point cloud data obtained in S110 is input into the classification model. The classification model outputs the probability that the point cloud coordinates of each point cloud data are the point cloud coordinates of a smooth area, the point cloud coordinates of a concave area, and the point cloud coordinates of a convex area. Then, based on the "probabilities of the point cloud coordinates of the smooth area, the point cloud coordinates of the concave area, and the point cloud coordinates of the convex area", the initial smooth area, the initial convex area, and the initial concave area are obtained. The initial smooth area, initial convex area, and initial concave area are all obtained based on the point cloud data.

[0176] The classification model in this embodiment includes an input layer, a feature extraction layer, a global feature fusion layer, a feature propagation layer, and a classification output layer;

[0177] The input layer is configured to receive each point cloud coordinate in the point cloud data;

[0178] The feature extraction layer is configured to select M point cloud coordinates from the point cloud data as the center point cloud coordinates based on the farthest point sampling algorithm, and divide the point cloud data into M local areas with the center point cloud coordinates as the center and n as the radius, and extract local features of each local area based on the MLP algorithm to form M local feature matrices;

[0179] The global feature fusion layer is configured to perform global pooling on the local feature matrix to obtain a global feature matrix;

[0180] The global feature fusion layer is further configured to concatenate the global feature matrix with the local features of each local region to obtain a feature enhancement matrix;

[0181] The feature propagation layer is configured to obtain a feature vector for each point cloud coordinate based on a feature enhancement matrix according to a distance weighted algorithm;

[0182] The classification output layer is configured to output a probability of the point cloud coordinates being convex or concave based on the feature vector of the point cloud coordinates.

[0183] Exemplarily: the classification model includes:

[0184] 1. Input Layer

[0185] Input data: original point cloud data, with a shape of N×3 (N points, each point contains x, y, z coordinates).

[0186] Preprocessing:

[0187] Coordinate normalization: Scale the coordinates to the range of [-1, 1] to accelerate model convergence.

[0188] Normal calculation: Calculate the normal vector (3D) of each point, expand the input to N×6, and enhance the local geometric representation.

[0189] 2. Hierarchical Feature Extraction

[0190] A multi-level Set Abstraction (SA) module is used to downsample and extract local features layer by layer:

[0191] SA Module 1:

[0192] Sampling: The farthest point sampling (FPS) selects 512 center points.

[0193] Grouping: With each center point as the center, the ball query radius r1 = 0.1 is used to aggregate the neighboring points.

[0194] Feature extraction: MLP (64→64→128), extracting 128-dimensional features for each local point set.

[0195] Output: 512×128 feature matrix.

[0196] SA Module 2:

[0197] Sampling: FPS selects 256 center points.

[0198] Grouping: radius r2 = 0.2, expand the receptive field.

[0199] Feature extraction: MLP (128 → 128 → 256), output 256 × 256.

[0200] SA Module 3:

[0201] Sampling: FPS selects 128 center points.

[0202] Grouping: radius r3 = 0.4, capturing global context.

[0203] Feature extraction: MLP (256 → 256 → 512), output 128 × 512.

[0204] 3. Global feature fusion layer

[0205] Global pooling: Perform maximum pooling on the output of SA module 3 to obtain 1×512 global features.

[0206] Feature replication and splicing: The global features are copied and spliced to each point feature of SA module 3 to form an enhanced feature matrix of 128×1024.

[0207] 4. Feature Propagation with Skip Connections

[0208] Recover detailed information through interpolation and skip connections:

[0209] FP module 3→2:

[0210] Interpolation: Based on distance weighting, the 128-point feature is upsampled to 256 points.

[0211] Skip connection: concatenate the intermediate features (256 dimensions) of SA module 2 to form 256×(256+512).

[0212] MLP processing: MLP(256→256→256), output 256×256.

[0213] FP module 2 → 1:

[0214] Interpolation: Upsample 256 points to 512 points.

[0215] Skip connection: concatenate the features of SA module 1 (128 dimensions) to form 512×(128+256).

[0216] MLP processing: MLP (128 → 128 → 128), output 512 × 128.

[0217] FP module 1 → original number of points (N):

[0218] Interpolation: Upsample 512 points to the original number of points N.

[0219] Jump connection: concatenate the normalized coordinates (or normals) of the input layer to form N×(3+128).

[0220] MLP processing: MLP (64 → 32 → 16), output N × 16.

[0221] 5. Classification Output Layer (Probability Prediction)

[0222] Fully connected layer: maps 16-dimensional features to 2 dimensions (convex / concave).

[0223] Activation function: The Softmax function outputs the probability that each point belongs to two categories, with a shape of N×2.

[0224] In S130, the initial convex area and the initial concave area are generated based on the classification probability of each coordinate point in the point cloud data. The output of the classification model and the spatial clustering algorithm need to be combined. The specific implementation process is as follows:

[0225] 1. Probability threshold determination and preliminary classification model output:

[0226] The classification model (such as improved PointNet++ or 3D-CNN) outputs the probability values of smooth, convex, and concave categories for each point (for example: smooth probability 0.8, convex 0.1, concave 0.1)

[0227] Category assignment rule: Maximum probability method: the category with the highest probability is used as the initial label of the point (for example, the point with the highest smoothing probability is marked as the smoothing area point).

[0228] Threshold filtering method: If the probability of a certain category exceeds the threshold (such as smoothed probability ≥ 0.7), it is marked as the corresponding category; otherwise, it is regarded as an uncertain point and subsequently corrected through neighborhood relationships.

[0229] 2. Spatial clustering to form continuous areas

[0230] Euclidean clustering algorithm: divides points of the same category into independent regions based on spatial proximity.

[0231] Input: All points marked with the same category (e.g., all “concave” points). Parameter settings: Set the neighborhood search radius (e.g., 0.5 mm) and the minimum number of points (e.g., 50 points) to remove isolated noise points.

[0232] Output: Multiple spatially continuous clusters (e.g., clustering scattered depression points into depression areas). Example: If a depression area consists of three independent pits, three depression area clusters will be generated after clustering.

[0233] 3. Geometric feature verification and regional correction Local curvature analysis: Smooth area: curvature value close to 0 (no obvious concave-convex change). Convex area: curvature value significantly positive (local surface convex). Concave area: curvature value significantly negative (local surface concave)

[0234] Verification logic: If the classification label of a point does not match the curvature feature (for example, it is marked as concave but the curvature is positive), the label is reassigned.

[0235] Normal vector consistency: The normal vector directions of points in the same area should be consistent. If there is a sudden change, it may be a classification error and the boundary needs to be adjusted.

[0236] 3. Post-processing optimization of the region boundary conditional random field (CRF): enforce the consistency of adjacent point categories and eliminate isolated misclassified points (such as sporadic convex points in smooth areas).

[0237] Region merging and splitting:

[0238] Merge: If the distance between two adjacent concave areas is less than a threshold (such as 1mm), they are merged into the same area.

[0239] Segmentation: If a convex region has concave features inside, it is segmented into independent sub-regions 36.

[0240] Manual verification: Use visualization tools to check the classification results and manually correct obvious errors (for example, adjust a decorative pattern that was mistakenly identified as a concave area to a smooth area).

[0241] 4. Output initial regional parameter regional attribute statistics:

[0242] Smooth area: has the largest area share (e.g., 70% of the surface) and the smallest curvature standard deviation.

[0243] Raised / depressed areas: Statistical height extremes (such as depression depth 0.2mm to 1.5mm), distribution density, etc.

[0244] Data structure: The region division results are stored in the form of a point cloud index list or a three-dimensional mask matrix.

[0245] In S140, if Figure 2 As shown, the following steps are included:

[0246] S141: grayscale processing of the image data;

[0247] S142: Filtering the image data for all first pixel points whose brightness values are greater than a first threshold, and all second pixel points whose brightness values are less than a second threshold;

[0248] S143: Based on all the first pixel points and all the second pixel points, correct the initial convex area and the initial concave area to obtain the final convex area and the final concave area.

[0249] In this embodiment, the first pixel point has a higher pixel value, which is the convex position of the convex area, and the second pixel point has a lower pixel value, which is the concave position of the concave area. The reason is:

[0250] 1. Causes of high brightness values in raised areas

[0251] ① When the light source shines at a specific angle (such as oblique or low-angle ring light), the local geometric height changes of the raised area will cause the light reflection direction to change. For example, the top of the raised area may directly reflect more light to the camera sensor, which appears as a higher brightness value.

[0252] ② In the three-dimensional grayscale matrix model, the protrusion corresponds to the local grayscale peak, and its height is positively correlated with the grayscale value.

[0253] 2. Causes of low brightness values in concave areas

[0254] ① The intensity of light entering the camera is weakened due to light blocking or scattering in the concave area: A shadow may form at the bottom of the concave area, resulting in a significant decrease in brightness.

[0255] ② In frequency domain processing, the concave area may correspond to the grayscale attenuation after the low-frequency component is suppressed.

[0256] In this embodiment, the first threshold and the second threshold can be obtained by acquiring multiple grayscale image data of other polished surfaces, as follows:

[0257] 1. Data preparation and annotation

[0258] Image acquisition: Acquire multiple grayscale images of the polished surface, covering different lighting conditions, surface textures, and defect morphologies (such as bumps and depressions).

[0259] Pixel-level annotation: Manually or semi-automatically annotate the type of each pixel (convex, concave, or normal area) to form a training dataset.

[0260] 2. Feature extraction and mathematical modeling

[0261] Grayscale features: Extract the grayscale value of the pixel and the statistics of its neighborhood (such as mean, variance, gradient direction, etc.) as basic features.

[0262] Morphological features: Combine the physical characteristics of convexity / concavity to extract gradient direction vectors, centripetal ratios, or aggregated discrete values (such as the features obtained through neighborhood path analysis in Webpage 2).

[0263] Regression model construction:

[0264] Logistic regression: Take the pixel type (convex / concave) as the dependent variable and the grayscale value, gradient direction, etc. as the independent variables to establish a probability prediction model. The threshold can be determined by the probability cutoff point.

[0265] The feature space is divided by a hyperplane, and the optimal threshold for distinguishing two types of pixels is automatically learned.

[0266] 3. Optimization of machine learning methods

[0267] Supervised learning: Use labeled data to train a neural network (such as a fully convolutional network FCN). The model outputs the convexity / concavity probability of each pixel. The first threshold and the second threshold can be set to the high confidence interval of the probability distribution (for example, a probability > 0.8 is convex and < 0.2 is concave).

[0268] Reinforcement learning: Using the Q-learning method, the threshold is dynamically adjusted to maximize the objective function (such as segmentation accuracy) by interacting with the environment (image data).

[0269] In S143, Figure 3 As shown, S132 may specifically include the following steps:

[0270] S1431: Based on all the first pixel points and all the second pixel points, obtain a plurality of first camera areas and a plurality of second camera areas respectively;

[0271] S1432: Mapping the first pixel point of the first camera area and the second pixel point of the second camera area to the coordinate system corresponding to the point cloud data, to obtain a first point cloud area corresponding to the first camera area and a second point cloud area corresponding to the second camera area, respectively;

[0272] S1433: Merging the first point cloud area and the initial convex area to obtain a final convex area, and merging the second point cloud area and the initial concave area to obtain a final concave area.

[0273] In S1431, all the first pixel points and the second pixel points are clustered to obtain a plurality of first camera areas and a plurality of second camera areas.

[0274] In S1432, the coordinate system corresponding to the point cloud data in this embodiment is the geodetic coordinate system.

[0275] The merging described in S1433 is represented by the union of the first point cloud area and the initial convex area, which is the final convex area; and the union of the second point cloud area and the initial concave area, which is the final concave area.

[0276] In S150, Figure 4 As shown, the following steps may be specifically included, wherein the final concave area includes several concavities, and the final convex area includes several convexities.

[0277] S151: Extracting the maximum depth of the concave, the coverage of the final concave area, and the edge sharpness of the final concave area, as well as the maximum height of the convex, the main curvature radius of the final convex area, and the distribution density of the convex in the final convex area;

[0278] S152: Based on the maximum depth of the depression, the coverage of the final depression area, the edge sharpness of the final depression area, the maximum height of the protrusion, the main curvature radius of the final protrusion area and the distribution density of the protrusions in the final protrusion area, the polishing parameter range value is determined according to a preset algorithm.

[0279] The polishing parameters include polishing force, feed speed and dwell time, and the polishing parameter range values include polishing force parameter range values, feed speed range values and dwell time range values.

[0280] The default algorithm is:

[0281] F1=k1·h 1.2 +F base

[0282]

[0283] F min =min{F1, F2}

[0284] F max =0.8σ y A1 0.5

[0285] The polishing force parameter range is expressed as F∈(F min , F max );

[0286] k1 and k2 represent the material hardness coefficient of the workpiece and the wear compensation coefficient of the polishing tool respectively;

[0287] F base Indicates the base pressure, which is determined by the material properties of the workpiece and can represent the basic bearing capacity of the material used to manufacture the workpiece;

[0288] F min Indicates the minimum polishing force;

[0289] F max Indicates the maximum polishing force;

[0290] F represents polishing force;

[0291] σ y represents the yield strength of the workpiece;

[0292] A1 represents the effective area of the grinding wheel of the polishing machine;

[0293] H represents the maximum height of the protrusion;

[0294] R represents the principal curvature radius of the final raised area. The shape of the final raised area is approximated by fitting the quadratic surface of the final raised area, and the principal curvature radius is calculated accordingly. k max Indicates the maximum principal curvature of the quadratic surface. The maximum principal curvature is calculated by calculating the principal curvature corresponding to each data point, with the data points corresponding to the point cloud data one by one, and then selecting the largest value among all the principal curvatures.

[0295] h represents the maximum depth of the depression, which is the value of the depression in the Z direction in the geodetic coordinate system;

[0296] v 1max =20+5e -0.3A2

[0297] v 1min =20

[0298] v 2max =15·tanh(0.1R)

[0299] v 2min =1.2·(R·ρ1) 0.3

[0300] v min =min{v 1min ,v 2min}

[0301] v max =max{v 1max ,v 2max}

[0302] The feed speed range value is expressed as: v∈(v max ,v min );

[0303] Wherein, v represents the feed speed;

[0304] ρ1 represents the distribution density of the protrusions in the final protrusion area, that is, the number of protrusions per unit area in the final protrusion area;

[0305] A2 represents the coverage of the final depression area, which is expressed as the projection area of the final depression area on the XY plane of the geodetic coordinate system;

[0306] t min =0

[0307]

[0308] m=max{H,h

[0309] The dwell time range value can be expressed as: t∈(t min ,t max )

[0310] Among them, c m represents the material coefficient of the workpiece;

[0311] ρ2 represents the density of the workpiece.

[0312] According to the above preset algorithm, the polishing force parameter range value, the feed speed range value and the dwell time range value are determined.

[0313] In S160, if Figure 4 As shown, the specific steps include:

[0314] S161: Based on the polishing force parameter range value, the feed speed range value, and the dwell time range value, a plurality of polishing parameter discrete points are obtained;

[0315] S162: Inputting the polishing parameter discrete points, the final raised area, and the final recessed area into a preset deep learning model to obtain the surface roughness corresponding to the polishing parameter discrete points;

[0316] S163: taking the polishing parameter discrete points corresponding to the surface roughness being less than or equal to the surface roughness threshold as parent discrete points, and performing iteration until the optimal polishing parameters are obtained;

[0317] The iteration includes:

[0318] Based on the parent discrete points, crossover and / or mutation are performed to obtain offspring discrete points;

[0319] Inputting the offspring discrete points and optimized classification results into the deep learning model;

[0320] Screening the descendant discrete points whose surface roughness is less than or equal to the surface roughness threshold as the parent discrete points;

[0321] Repeat the crossover and / or mutation based on the parent discrete points to obtain child discrete points; input the child discrete points and the optimized classification results into the deep learning model, and screen the child discrete points whose surface roughness is less than or equal to the surface roughness threshold as the parent discrete points.

[0322] The genetic algorithm process mentioned in S160 includes:

[0323] Initialization: Set the evolutionary generation counter t = 0, set the maximum evolutionary generation T, and randomly generate M individuals as the initial population P(0). Selection operation: Apply the selection operator to the population. The purpose of selection is to directly pass on the optimized individuals to the next generation or to generate new individuals through pairing and crossover and then pass them on to the next generation. The selection operation is based on the fitness evaluation of individuals in the population. Crossover operation: Apply the crossover operator to the population. Crossover refers to the operation of replacing and recombining part of the structure of two parent individuals to generate new individuals. It plays a core role in the biological evolution process in nature and also in genetic algorithms. Mutation operation: Apply the mutation operator to the population. That is, the gene values at certain loci of the individual strings in the population are changed. The basic contents of the mutation operator include real-valued mutation and binary mutation. Termination condition judgment: If t = T, the individual with the maximum fitness obtained in the evolution process is output as the optimal solution and the calculation is terminated. The termination condition also includes when the fitness of the optimal individual reaches a given threshold or when the fitness of the optimal individual and the fitness of the population no longer increase.

[0324] The polishing parameter range can be the range of polishing force, dwell time and feed speed, and the range cannot be too large, for example, F∈[0,15], V∈[0,5], T∈[0,10]. Then, discrete points are randomly selected within these ranges, and these discrete points are integrated to form an initialized group. The initialized group includes several polishing parameter discrete points.

[0325] In this embodiment, the deep learning model is used to predict the surface roughness based on the polishing parameter discrete points and the optimized point cloud data, where the optimized point cloud data includes the final raised area and the final recessed area.

[0326] The deep learning model can be a CNN model, etc. The training method of the deep learning model is:

[0327] Collect historical data, wherein the historical data includes a one-to-one correspondence of historical optimization classification results, historical polishing force, historical feed speed, historical dwell time, and historical surface roughness of the workpiece; train a preset deep learning model based on the historical data; input the discrete points of the polishing parameters into the preset deep learning model to obtain the surface roughness corresponding to the discrete points of the polishing parameters.

[0328] In this embodiment, after S160, S170 is further included. Figure 5 As shown, S170 includes the following steps:

[0329] S171: Based on the optimal polishing parameters, controlling a polishing machine to polish the surface to be polished;

[0330] S172: Collecting the actual surface roughness of the surface to be polished. When the difference between the actual surface roughness and the preset surface roughness is greater than the preset difference, adjusting the optimal polishing parameters so that the difference between the actual surface roughness and the preset surface roughness is not greater than the preset difference.

[0331] Adjusting the optimal polishing parameters includes:

[0332] Based on the optimal polishing parameters and the preset adjustment amount, obtaining the adjusted optimal polishing parameters;

[0333] Inputting the adjusted optimal polishing parameters into a surface roughness calculation model to obtain an estimated surface roughness, and if a difference between the estimated surface roughness and the preset surface roughness is greater than a preset difference, repeating the following steps until the difference between the estimated surface roughness and the preset surface roughness is greater than the preset difference;

[0334] Obtaining an estimated surface roughness based on the adjusted optimal polishing parameters, the preset adjustment amount, and the surface roughness calculation model;

[0335] The surface roughness calculation model is:

[0336]

[0337] Among them, Ra represents the surface roughness;

[0338] K represents the comprehensive correction factor;

[0339] F represents polishing force;

[0340] α represents the nonlinear effect of polishing force on surface roughness, α<0;

[0341] v represents the feed speed;

[0342] β represents the nonlinear effect of feed rate on surface roughness, β<0;

[0343] t represents the dwell time;

[0344] γ represents the nonlinear effect of residence time on surface roughness, γ<0;

[0345] C represents the basic surface roughness constant;

[0346] P i Indicates the process parameters that affect surface roughness, δ i represents an independent index corresponding to the process parameters, wherein the process parameters include abrasive particle size, abrasive concentration, material hardness of the workpiece, temperature of the polishing area, and vibration frequency of the polishing tool.

[0347] In this embodiment, P i The values are shown in the following table:

[0348]

[0349] In this embodiment, the processing module can be an integrated circuit chip with signal processing capabilities. The above-mentioned processing module can be a general-purpose processor. For example, the processor can be a central processing unit (CPU), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, which can implement or execute the disclosed methods, steps and logic block diagrams in the embodiments of the present application.

[0350] The memory module may be, but is not limited to, a random access memory, a read-only memory, a programmable read-only memory, an erasable programmable read-only memory, an electrically erasable programmable read-only memory, and the like.

[0351] It should be noted that those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working process of the electronic device described above can refer to the corresponding processes of each step in the aforementioned method, and will not be elaborated here.

[0352] The present application also provides a computer-readable storage medium that stores a computer program, which, when executed on a computer, causes the computer to execute the polishing parameter determination method described in the above embodiment.

[0353] Through the description of the above implementation methods, those skilled in the art can clearly understand that the present application can be implemented through hardware or by means of software plus a necessary general hardware platform. Based on this understanding, the technical solution of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.), including a number of instructions for enabling a computer device (which can be a personal computer, an electronic device, or a network device, etc.) to execute the methods described in each implementation scenario of the present application.

[0354] In the embodiments provided in the present application, it should be understood that the disclosed method can also be implemented in other ways. The method embodiments described above are merely schematic. For example, the flow charts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the methods and computer program products according to the multiple embodiments of the present application. In this regard, each box in the flow chart or block diagram can represent a module, a program segment or a part of code, and a part of the module, program segment or code includes one or more executable instructions for implementing the specified logical function. It should also be noted that each box in the block diagram and / or flow chart, and the combination of the boxes in the block diagram and / or flow chart can be implemented by a dedicated hardware-based system that performs the specified function or action, or can be implemented by a combination of dedicated hardware and computer instructions. In addition, the functional modules in the various embodiments of the present application can be integrated together to form an independent part, or each module can exist alone, or two or more modules can be integrated to form an independent part.

[0355] The above description is merely an embodiment of the present application and is not intended to limit the scope of protection of the present application. For those skilled in the art, various modifications and variations of the present application are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A method for determining polishing parameters, characterized in that: The method comprises: Collecting point cloud data of the surface to be polished of the workpiece and image data of the surface to be polished; Inputting the point cloud data into a classification model, the classification model outputting an initial classification result, wherein the initial classification result is used to characterize the probability that the point cloud coordinates of each point cloud data are convex and / or concave; Based on the initial classification result, an initial convex area and an initial concave area are obtained; Correcting the initial raised area and the initial recessed area based on the image data to obtain a final raised area and a final recessed area of the surface to be polished; Obtaining polishing parameter range values based on the point cloud coordinates corresponding to the final raised area and the final recessed area according to a preset algorithm; Based on the genetic algorithm model, the optimal polishing parameters are determined within the range of values of the polishing parameters.

2. The method according to claim 1, characterized in that The final concave region includes a plurality of concavities, and the final convex region includes a plurality of convexities; The point cloud coordinates corresponding to the final raised area and the final recessed area are used to obtain polishing parameter range values according to a preset algorithm, including: Extracting the maximum depth of the concave, the coverage of the final concave area, and the edge sharpness of the final concave area, as well as the maximum height of the convex, the main curvature radius of the final convex area, and the distribution density of the convex in the final convex area; The polishing parameter range value is determined according to a preset algorithm based on the maximum depth of the depression, the coverage of the final depression area, the edge sharpness of the final depression area, the maximum height of the protrusion, the main curvature radius of the final protrusion area and the distribution density of the protrusions in the final protrusion area.

3. The method according to claim 2, characterized in that The polishing parameters include polishing force, feed speed and dwell time, and the polishing parameter range values include polishing force parameter range values, feed speed range values and dwell time range values; Determining the polishing parameter range value according to a preset algorithm includes: The polishing force parameter range, feed speed range, and dwell time range are determined according to a preset algorithm, wherein the preset algorithm is: F1=k1·h 1.2 +F base <h2 style=";text-align:left;direction:ltr">F<h2 style=";text-align:left;direction:ltr"> min <h2 style=";text-align:left;direction:ltr"> = min{F1, F2} F max =0.8σ y ·A1 0.5 The polishing force parameter range is expressed as F∈(F min , F max ); k1 and k2 represent the material hardness coefficient of the workpiece and the wear compensation coefficient of the polishing tool respectively; F base Indicates the base pressure; F min Indicates the minimum polishing force; F max Indicates the maximum polishing force; F represents polishing force; σ y represents the yield strength of the workpiece; A1 represents the effective area of the grinding wheel of the polishing machine; H represents the maximum height of the bulge, which is the Z-direction value of the bulge’s point cloud coordinates; R represents the main curvature radius of the final convex area; h represents the maximum depth of the depression; v 1max =20+5e -0.3A2 v 1min ＝20 v 2max =15·tanh(0.1R) v 2min =1.2·(R·ρ1) 0.3 v min =min{v 1min ,v 2min } v max =max{v 1max ,v 2max } The feed speed range value is expressed as: v∈(v max ,v min ); Wherein, v represents the feed speed; ρ1 represents the distribution density of the protrusions in the final protrusion area; A2 represents the coverage of the final concave area; t min ＝0 m = max{H, h} The dwell time range value can be expressed as: t∈(t min ,t max ) Among them, c m represents the material coefficient of the workpiece; ρ2 represents the density of the workpiece.

4. The method according to claim 1, wherein The step of correcting the initial raised area and the initial recessed area based on the image data to obtain the final raised area and the final recessed area of the surface to be polished comprises: grayscale processing of the image data; Screening all first pixel points whose brightness values are greater than a first threshold value and all second pixel points whose brightness values are less than a second threshold value in the image data after grayscale processing; Based on all the first pixel points and all the second pixel points, the initial convex area and the initial concave area are corrected to obtain the final convex area and the final concave area.

5. The method according to claim 4, characterized in that The correcting the initial convex area and the initial concave area based on all the first pixel points and all the second pixel points to obtain an optimized classification result includes: Based on all the first pixel points and all the second pixel points, a plurality of first camera areas and a plurality of second camera areas are obtained respectively; Mapping the first pixel point of the first camera area and the second pixel point of the second camera area to the coordinate system corresponding to the point cloud data, to obtain a first point cloud area corresponding to the first camera area and a second point cloud area corresponding to the second camera area, respectively; The first point cloud region and the initial convex region are merged to obtain a final convex region, and the second point cloud region and the initial concave region are merged to obtain a final concave region.

6. The method according to claim 1, characterized in that The method of determining the optimal polishing parameters based on the genetic algorithm model within the range of polishing parameters and taking the minimum surface roughness of the surface to be polished as the surface roughness function includes: Based on the polishing force parameter range value, the feed speed range value and the dwell time range value, a plurality of polishing parameter discrete points are obtained; Inputting the point cloud coordinates corresponding to the polishing parameter discrete points, the final raised area, and the final recessed area into a preset deep learning model to obtain the surface roughness corresponding to the polishing parameter discrete points; Taking the polishing parameter discrete points corresponding to the surface roughness being less than or equal to the surface roughness threshold as parent discrete points, and iterating until the optimal polishing parameters are obtained; The iterations include: Based on the parent discrete points, crossover and / or mutation are performed to obtain offspring discrete points; Inputting the offspring discrete points and optimized classification results into the deep learning model; Screening the child discrete points whose surface roughness is less than or equal to the surface roughness threshold as the parent discrete points; Repeating the crossover and / or mutation based on the parent discrete points to obtain child discrete points; inputting the child discrete points and the optimized classification results into the deep learning model, and screening the child discrete points whose surface roughness is less than or equal to the surface roughness threshold as the parent discrete points; Inputting the polishing parameter discrete points into a preset deep learning model to obtain the surface roughness corresponding to the polishing parameter discrete points includes: Collecting historical data, the historical data including historical optimization classification results, historical polishing forces, historical feed speeds, historical dwell times, and historical surface roughness of the workpiece in a one-to-one correspondence; Based on the historical data, a preset deep learning model is trained; The polishing parameter discrete points are input into a preset deep learning model to obtain the surface roughness corresponding to the polishing parameter discrete points.

7. The method according to claim 1, characterized in that The classification model includes an input layer, a feature extraction layer, a global feature fusion layer, a feature propagation layer and a classification output layer; The input layer is configured to receive each point cloud coordinate in the point cloud data; The feature extraction layer is configured to select M point cloud coordinates from the point cloud data as the center point cloud coordinates based on the farthest point sampling algorithm, and divide the point cloud data into M local areas with the center point cloud coordinates as the center and n as the radius, and extract local features of each local area based on the MLP algorithm to form M local feature matrices; The global feature fusion layer is configured to perform global pooling on the local feature matrix to obtain a global feature matrix; The global feature fusion layer is further configured to concatenate the global feature matrix with the local features of each local region to obtain a feature enhancement matrix; The feature propagation layer is configured to obtain a feature vector for each point cloud coordinate based on a feature enhancement matrix according to a distance weighted algorithm; The classification output layer is configured to output a probability of the point cloud coordinates being convex or concave based on the feature vector of the point cloud coordinates.

8. The method according to claim 1, characterized in that After determining the optimal polishing parameter within the range of polishing parameters based on the genetic algorithm model, the method further includes: Based on the optimal polishing parameters, controlling a polishing machine to polish the surface to be polished; collecting the actual surface roughness of the surface to be polished, and when the difference between the actual surface roughness and the preset surface roughness is greater than a preset difference, adjusting the optimal polishing parameters so that the difference between the actual surface roughness and the preset surface roughness is no greater than the preset difference; The adjusting of the optimal polishing parameters comprises: Based on the optimal polishing parameters and the preset adjustment amount, obtaining the adjusted optimal polishing parameters; Inputting the adjusted optimal polishing parameters into a surface roughness calculation model to obtain an estimated surface roughness, and if a difference between the estimated surface roughness and the preset surface roughness is greater than a preset difference, repeating the following steps until the difference between the estimated surface roughness and the preset surface roughness is greater than the preset difference; Obtaining an estimated surface roughness based on the adjusted optimal polishing parameters, the preset adjustment amount, and the surface roughness calculation model; The surface roughness calculation model is: Among them, Ra represents the surface roughness; K represents the comprehensive correction factor; F represents polishing force; α represents the nonlinear effect of polishing force on surface roughness, α<0; v represents the feed speed; β represents the nonlinear effect of feed rate on surface roughness, β<0; t represents the dwell time; γ represents the nonlinear effect of residence time on surface roughness, γ<0; C represents the basic surface roughness constant; P i Indicates the process parameters that affect surface roughness, δ i represents an independent index corresponding to the process parameters, wherein the process parameters include abrasive particle size, abrasive concentration, material hardness of the workpiece, temperature of the polishing area, and vibration frequency of the polishing tool.

9. An electronic device, characterized in that: The electronic device includes a processor and a memory coupled to each other, wherein a computer program is stored in the memory. When the computer program is executed by the processor, the electronic device executes the method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is run on a computer, the computer is caused to execute the method according to any one of claims 1 to 8.

Images