Workpiece machining quality inspection method and system based on artificial intelligence

Abstract

Provided in some embodiments of the present application is a workpiece machining quality inspection method based on artificial intelligence. In the embodiments of the present application, the method comprises: acquiring workpiece images of the current workpiece from at least two angles, and determining a camera pose and image depth information; establishing a three-dimensional voxel model of the current workpiece; acquiring attribute information of the current workpiece, assigning the attribute information to each voxel of the three-dimensional voxel model, and performing Gaussian splatting on each voxel; determining a mask of each workpiece image in a workpiece image set, and on the basis of the masks, classifying the workpiece images into workpiece structural images and background images; and determining a workpiece point cloud of the current workpiece, calculating the similarity between the workpiece point cloud and a preset standard workpiece point cloud, and if the similarity exceeds a preset similarity threshold, determining the current workpiece to be a qualified workpiece.

Description

A workpiece machining quality inspection method and system based on artificial intelligence

[0001] Cross-referencing

[0002] This application incorporates Chinese Patent Application No. 202411418200.9, filed on October 11, 2024, entitled "A workpiece processing quality inspection method and system based on artificial intelligence", which is incorporated herein by reference in its entirety. Technical Field

[0003] This application relates to the field of artificial intelligence technology, and in particular to a workpiece processing quality inspection method and system based on artificial intelligence. Background Technology

[0004] The quality of workpiece processing directly affects the final performance and lifespan of the product. Inspection ensures that parameters such as the workpiece's dimensions, shape, and positional accuracy meet design requirements and technical standards, thus guaranteeing the overall quality of the product. Currently, workers manually conduct comprehensive inspections of the workpieces, from appearance and dimensions to functional performance, strictly adhering to standards to ensure that every detail meets quality requirements, thereby guaranteeing the overall quality of the product.

[0005] The inventors have discovered at least the following technical problems in the related technology:

[0006] Manual inspection requires operators to examine each workpiece individually, and complex or subtle defects may require even longer observation and judgment, resulting in relatively low inspection efficiency. Furthermore, manual inspection often lacks quantifiable quality standards, making accurate comparison and analysis of inspection results difficult. Summary of the Invention

[0007] The purpose of some embodiments of this application is to provide a workpiece processing quality inspection method based on artificial intelligence.

[0008] The method includes: acquiring workpiece images from at least two angles of the current workpiece; generating a workpiece image set based on the workpiece images; determining camera pose and image depth information based on the workpiece image set; establishing a three-dimensional voxel model of the current workpiece based on the workpiece image set, camera pose, and image depth information; acquiring attribute information of the current workpiece; allocating the attribute information to each voxel of the three-dimensional voxel model; performing Gaussian sputtering on each voxel based on the attribute information of each voxel, camera pose, and a preset Gaussian function; wherein the attribute information includes density, color, and transparency; determining the mask for each workpiece image in the workpiece image set based on a preset mask extraction model; dividing the workpiece images into workpiece structure images and background images based on the masks; determining the workpiece point cloud of the current workpiece based on the Gaussian sputtered three-dimensional voxel model and the divided workpiece images; calculating the similarity between the workpiece point cloud and a preset workpiece standard point cloud; if the similarity exceeds a preset similarity threshold, determining the current workpiece as a qualified workpiece.

[0009] The purpose of some embodiments of this application is also to provide a workpiece processing quality inspection system based on artificial intelligence.

[0010] The system includes: an image acquisition module for acquiring workpiece images from at least two angles, generating a workpiece image set based on the workpiece images, and determining camera pose and image depth information based on the workpiece image set; a model building module for building a three-dimensional voxel model of the current workpiece based on the workpiece image set, camera pose, and image depth information; a Gaussian sputtering module for acquiring attribute information of the current workpiece, distributing the attribute information to each voxel of the three-dimensional voxel model, and performing Gaussian sputtering on each voxel based on the attribute information of each voxel, camera pose, and a preset Gaussian function; wherein the attribute information includes density, color, and transparency; a mask determination module for determining the mask of each workpiece image in the workpiece image set based on a preset mask extraction model, and dividing the workpiece image into a workpiece structure image and a background image based on the mask; and a workpiece detection module for determining the workpiece point cloud of the current workpiece based on the Gaussian sputtered three-dimensional voxel model and the divided workpiece images, calculating the similarity between the workpiece point cloud and a preset workpiece standard point cloud, and determining the current workpiece as a qualified workpiece if the similarity exceeds a preset similarity threshold. Attached Figure Description

[0011] One or more embodiments are illustrated by way of example with reference to the accompanying drawings, and these illustrative descriptions do not constitute a limitation on the embodiments.

[0012] Figure 1 is an exemplary flowchart of a workpiece machining quality inspection method based on artificial intelligence according to the first embodiment of this application;

[0013] Figure 2 is an exemplary flowchart of a workpiece processing quality inspection method based on artificial intelligence according to the second embodiment of this application;

[0014] Figure 3 is an exemplary flowchart of a workpiece processing quality inspection method based on artificial intelligence according to the third embodiment of this application;

[0015] Figure 4 is an exemplary structural diagram of an artificial intelligence-based workpiece processing quality inspection system according to the fourth embodiment of this application. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, some embodiments of this application will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0017] The first embodiment of this application relates to a workpiece machining quality inspection method based on artificial intelligence. As shown in Figure 1, the method may include the following steps:

[0018] S101, acquire workpiece images from at least two angles of the current workpiece, generate a workpiece image set based on the workpiece images, and determine the camera pose and image depth information based on the workpiece image set.

[0019] First, this solution can be used to acquire workpiece images and build a 3D voxel model of the workpiece, then determine the workpiece point cloud, and finally determine whether the workpiece is qualified by comparing the workpiece point cloud with a standard workpiece point cloud.

[0020] Based on the above usage scenarios, it is understandable that the implementing entity of this application can be an artificial intelligence-based workpiece processing quality inspection system, without further limitations here.

[0021] In this solution, the current workpiece can refer to an object or part that is being inspected and processed. It refers to a specific workpiece on the production line or in the inspection system that is in the inspection, analysis, or processing stage.

[0022] Workpiece images can refer to two-dimensional pictures of a current workpiece taken from different angles. Each image captures a specific viewpoint of the workpiece and shows some or all of its surface features.

[0023] A workpiece image set can be a collection of multiple workpiece images. These images can be taken from different angles to ensure comprehensive capture of the workpiece's three-dimensional structural information.

[0024] Camera pose refers to the specific position and orientation of the camera when capturing each image of a workpiece. It includes the camera's spatial position (i.e., three-dimensional coordinates) and orientation (i.e., rotation angle).

[0025] Image depth information refers to the distance from each pixel in an image to the camera. Depth information transforms an image from two-dimensional to three-dimensional data with spatial distances.

[0026] Once the workpiece enters the detection area and triggers the grating, the rotating platform begins to rotate at a constant speed. During this rotation, the workpiece is captured by a high-resolution camera from multiple angles. The camera employs advanced optical technology to ensure clear images even during high-speed rotation; these images include at least two different views of the workpiece. The multi-angle images captured by the camera are saved, forming an image set containing multiple views of the workpiece. This image set includes workpiece images taken from different angles, ensuring comprehensive coverage of all aspects of the workpiece. The generated workpiece image set can then be imported into COLMAP software. COLMAP is a powerful computer vision tool that can acquire 3D reconstruction information through the processing of multi-view images. COLMAP first extracts features from the images in the image set and matches feature points in different images. Using these matched feature points, COLMAP calculates the camera's pose information (including position and orientation) when capturing each image. COLMAP further analyzes the geometric relationships of each image and, combined with the camera pose, generates corresponding image depth information. The depth information represents the distance from each pixel in the image to the camera, thus helping to construct a 3D model of the workpiece. After COLMAP completes the calculation, it outputs the camera pose information (including camera extrinsic parameters such as position and orientation) and the corresponding image depth information for each image.

[0027] S102, establish a three-dimensional voxel model of the current workpiece based on the workpiece image set, camera pose, and image depth information.

[0028] A 3D voxel model can refer to representing the three-dimensional structure of an object by dividing its space into uniform small cubes (called voxels). Each voxel represents a volume unit in space and has specific properties such as density, color, and transparency. A 3D voxel model is a discrete representation of an object in three-dimensional space, similar to pixels in a 2D image, but it has depth information in three-dimensional space.

[0029] For each image's depth map, using camera intrinsics (focal length, optical center, etc.) and camera pose, each pixel (and its depth value) in the depth map can be projected into 3D space, forming a point in the point cloud. Each point is represented by its X, Y, Z coordinates in 3D space. The point clouds generated from all images are then merged into a single global coordinate system. This involves transforming each point cloud from its respective camera coordinate system to the global coordinate system based on the camera pose. Then, statistical filters (such as radius filters and voxel grid filters) are used to remove noise and outliers from the point cloud. Radius filters remove noise by deleting points with too few points within a certain radius of each point. Voxel grid filters reduce the size and noise of the point cloud by dividing the point cloud into small 3D voxels (voxel grids) and approximating all points within that voxel using the centroid of all points within that voxel. Finally, surface reconstruction algorithms (such as Poisson surface reconstruction, moving least squares, sphere rotation, etc.) are used to generate a 3D surface model of the workpiece from the processed point cloud. These algorithms estimate the geometry of the workpiece surface based on the points in the point cloud and their neighborhood relationships. Then, select a suitable voxel size (i.e., the size of the 3D space represented by each voxel) and define a voxel mesh containing the workpiece based on its dimensions. Traverse each point (or face) on the 3D surface model and check if they intersect with voxels in the voxel mesh. If they intersect, mark the voxel as occupied (or perform more complex interpolation as needed). Finally, optimize the voxel model, such as removing isolated voxels and merging adjacent occupied voxels, to improve the model's accuracy.

[0030] S103: Obtain the attribute information of the current workpiece, assign the attribute information to each voxel of the 3D voxel model, and perform Gaussian sputtering on each voxel according to the attribute information of each voxel, the camera pose, and the preset Gaussian function; wherein, the attribute information includes density, color, and transparency.

[0031] Attribute information can be the physical and visual characteristics of a workpiece, such as density, color, and transparency.

[0032] Density can describe the compactness of material within a voxel and is typically related to the material and structure of the workpiece. It can be obtained from known properties of the material or inferred from scan data.

[0033] Color can be a color attribute representing a voxel, typically extracted from the color information of an image.

[0034] Transparency can describe the ability of light to penetrate a voxel. Transparency can be based on material properties or obtained through measuring tools such as optical sensors.

[0035] The preset Gaussian function can be a probability density function, typically used to represent the scattering of light in space. The preset Gaussian function is used to simulate the interaction between light rays and voxels.

[0036] Attributes can be read from the workpiece's CAD model or design file. Since the voxel model is generated from point cloud data, it represents the workpiece's position and structure in 3D space. Each voxel can be associated with a corresponding attribute value (such as density, color, and transparency). Specifically, color information can be extracted from images taken from multiple angles and mapped to voxels based on camera pose. Density can be assigned to the corresponding voxels based on the workpiece's material or density information obtained through techniques such as CT scans. Transparency information can be assigned to each voxel based on material properties or optical measurement results. For example, if a part of the workpiece is red, the corresponding voxel will be assigned the red attribute. Density values ​​are assigned to voxels based on the material's density distribution. Then, the Gaussian function value is calculated based on each voxel's attribute information (such as density, color, and transparency). When light passes through the voxel space, it interacts with the voxel attributes. The propagation of light is affected by the voxel's density and transparency. Based on the Gaussian function value, the effect of light in the camera's viewpoint is splashed onto the surrounding area of ​​the voxel. The range and intensity of the splashing are determined by the voxel's attribute information and the parameters of the Gaussian function. The Gaussian distributions of different voxels accumulate during the rendering process, eventually forming a three-dimensional voxel model that simulates the effects of light propagation and scattering.

[0037] Based on the above technical solution, in an exemplary embodiment, the preset Gaussian function is:

[0038] Where G(x, D, C, T) is the voxel contribution value; x is the distance variable, representing the distance between the voxel center point and the camera; D is the voxel density, i.e., the material concentration of the voxel in three-dimensional space; C is the voxel color, which can represent color intensity or color information; T is the voxel transparency, i.e., the transparency of the voxel; σ is the standard deviation of the Gaussian function, which determines the degree of sputtering diffusion; W D Voxel density weights reflect the influence of voxel density on the Gaussian function; W C Voxel color weights reflect the influence of voxel colors on the Gaussian function; W T The voxel transparency weights reflect the influence of voxel transparency on the Gaussian function; The effect of voxels on light was calculated; is the normalization factor of the Gaussian function, ensuring that the integral of the entire function is 1.

[0039] In this scheme, for each voxel, the straight-line distance *x* from its center point to the camera (or viewpoint) can be calculated. This can be done using the distance formula in a 3D coordinate system. Substituting the distance *x* into... The influence of voxels on light is calculated based on the distance between the voxel center and the camera. Then, combining the voxel's density, color, and transparency weights, as well as a normalization factor, the voxel contribution value is calculated and applied to the light propagation path. This means that during rendering, the voxel's influence simulates light scattering in voxel space according to a Gaussian distribution. The voxel contributions are then weighted by a Gaussian function and added to the light propagation path to form the final rendered 3D voxel model.

[0040] Based on the above technical solution, in an exemplary embodiment, before performing Gaussian sputtering on each voxel according to the attribute information of each voxel, the camera pose, and a preset Gaussian function, the method further includes:

[0041] Obtain the preset weight allocation rules, the initial Gaussian function standard deviation, and the preset standard deviation update formula;

[0042] Accordingly, Gaussian sputtering is performed on each voxel based on its attribute information, camera pose, and a preset Gaussian function, including:

[0043] Based on the preset weight allocation rules and the attribute information of each voxel, determine the voxel density weight value, voxel color weight value, and voxel transparency weight value of each voxel in the preset Gaussian function.

[0044] The standard deviation of the Gaussian function is updated based on the voxel density weight value, the initial standard deviation of the Gaussian function, and the preset standard deviation update formula.

[0045] Obtain the center point position of each voxel, and determine the distance variable value of each voxel based on the center point position of each voxel and the camera pose;

[0046] Gaussian sputtering is performed on each voxel based on voxel density weight value, voxel color weight value, voxel transparency weight value, updated Gaussian function standard deviation, distance variable value, and preset Gaussian function.

[0047] In this scheme, preset weighting rules can be used to assign weights to voxel attributes (such as density, color, and transparency). These rules are typically determined through experimentation or standardization processes to ensure that the effects of different attributes on Gaussian sputtering are as expected. Specifically, this may include: a method for calculating voxel density weights, such as linear scaling of voxel density values, where higher density results in a greater weight, which can be calculated using the following formula: W D =Density × Scale D ;

[0048] Density is the voxel density value, which can be determined from the attribute information; Scale DThis is an adjustment factor for voxel density, used to control the degree of influence of density on weights.

[0049] The calculation method for voxel color weights: To map color values ​​to weight ranges, the following formula can be used:

[0050] R, G, and B represent the intensity values ​​of the red, green, and blue channels, respectively, and can be obtained from the attribute information; Scale C This is a factor for adjusting the color of voxels;

[0051] The calculation method for voxel transparency weights, such as the normalization of transparency values, can be calculated using the following formula:

[0052] Here, Transparency is the voxel transparency value, which can be obtained from the attribute information; ∈ is a constant to prevent division by zero.

[0053] The initial standard deviation of the Gaussian function can be the initial setting value of the Gaussian function, determining its diffusion range. It can be determined experimentally, typically as a fixed value or an initial estimate based on voxel characteristics. For example, the standard deviation can be set to values ​​such as 5 or 10, depending on application requirements and voxel properties.

[0054] The preset standard deviation update formula can be:

[0055] Where σ is the updated standard deviation of the Gaussian function; σ0 is the initial standard deviation of the Gaussian function; W D This represents the voxel density weighting value. Higher density results in a smaller standard deviation and a narrower function.

[0056] Based on preset weighting rules and the attribute information of each voxel, the system calculates the specific voxel density weight, voxel color weight, and voxel transparency weight values ​​for each voxel within a preset Gaussian function. Then, it retrieves the initial Gaussian function standard deviation from the database, combines it with the voxel density weight values, and applies a preset standard deviation update formula to calculate the updated Gaussian function standard deviation. Next, it obtains the center point position of each voxel and determines the distance x between the voxel center point and the camera based on the camera pose. Finally, it performs Gaussian sputtering based on the updated standard deviation, voxel attribute weights, distance variable values, and the preset Gaussian function, calculating the voxel's contribution to the light beam and applying this contribution value to the light propagation path. This Gaussian distribution overlays the voxel's influence onto the final rendered 3D voxel model.

[0057] In this scheme, the Gaussian sputtering process is ensured to accurately reflect the property information of voxels, and the scattering and absorption of light are simulated in the rendering to generate a realistic 3D voxel model after 3D rendering.

[0058] S104, determine the mask of each workpiece image in the workpiece image set according to the preset mask extraction model, and divide the workpiece image into workpiece structure image and background image according to the mask.

[0059] The preset mask extraction model can be a pre-trained deep learning model used to automatically extract the mask of the region of interest from the image. Specifically, it can be trained using a convolutional neural network or other image segmentation algorithms to identify the workpiece part in the image and distinguish it from the background.

[0060] A mask can be a binary image where regions with a pixel value of 1 represent the workpiece portion of interest, while regions with a pixel value of 0 represent the background. Mask images are used to separate the workpiece from the image, allowing subsequent analysis and processing to focus solely on the workpiece portion.

[0061] A workpiece structure image can be an image containing only a portion of the workpiece, with the background removed. By applying a mask, background pixels in the original image are masked or removed, leaving only the structural information of the workpiece, such as its shape, color, and texture.

[0062] A background image can be an image containing only the background portion, obtained by removing the workpiece portion using a mask. Background images are typically used to analyze background lighting conditions, noise levels, etc.

[0063] A pre-defined mask extraction model can be used to process each image in the workpiece image set. The model automatically identifies the workpiece region and generates a corresponding mask image. Each image generates a corresponding mask that separates the workpiece from the background. The generated mask is then multiplied pixel-by-pixel with the original image to obtain the workpiece structure image and the background image. The workpiece region in the mask is preserved, while the background region is masked (pixel values ​​are set to 0 or transparent). The resulting workpiece structure image and background image are used for subsequent workpiece analysis and background noise processing, respectively.

[0064] S105. Determine the workpiece point cloud of the current workpiece based on the three-dimensional voxel model after Gaussian sputtering and the segmented workpiece image. Calculate the similarity between the workpiece point cloud and the preset workpiece standard point cloud. If the similarity exceeds the preset similarity threshold, determine that the current workpiece is a qualified workpiece.

[0065] A workpiece point cloud is a three-dimensional data structure composed of a large set of points, each with specific coordinates (x, y, z) in three-dimensional space, which can be used to represent the surface shape and structure of an object. A workpiece point cloud can represent the set of all points on the workpiece surface.

[0066] The preset workpiece standard point cloud can be a pre-generated point cloud model of a standard workpiece, representing the ideal shape and structure of the workpiece. It is used to compare with the detected workpiece point cloud to evaluate the quality of the current workpiece.

[0067] Similarity can be a numerical measure of the degree of similarity between two point clouds. Commonly used methods include those based on statistical discrete metrics, such as structural similarity scoring, mean absolute deviation (μADA), and median absolute deviation (mADA). These methods can be used to assess the similarity between two point clouds in terms of geometric shape and surface features.

[0068] The preset similarity threshold can be a pre-defined value used to determine whether the similarity between the workpiece point cloud and the standard point cloud is high enough. If the similarity exceeds this threshold, the current workpiece is considered to meet the quality requirements and is thus a qualified workpiece; otherwise, it is considered an unqualified workpiece.

[0069] For each voxel, its coordinates in 3D space can be determined. After Gaussian sputtering, the voxel's properties are weighted and diffused, but its position coordinates remain unchanged. Then, the center point coordinates of effective voxels (i.e., those with non-zero properties) are extracted from the 3D voxel model. These coordinates form the workpiece's point cloud. Specifically, all voxels in the 3D voxel model can be traversed, and for each effective voxel (with non-zero density or other properties), its center point's 3D coordinates are recorded. These coordinate points are then collected to form the workpiece's point cloud data. Information from the workpiece structure image (such as contours and surface features) is then used to further correct and refine the point cloud data. Specifically, the point cloud in the voxel model can be compared with the edges and features in the workpiece structure image to ensure that the point cloud data accurately reflects the actual shape of the workpiece. Image segmentation and edge detection are performed on the image to extract important feature points, and these feature points are mapped into the point cloud to increase the point cloud's detail. Finally, a standard point cloud model corresponding to the current workpiece type is retrieved from the database. This model, generated during the production process design phase, represents the ideal 3D shape of the workpiece. Preprocessing is performed on both point clouds to ensure the accuracy of the comparison. Duplicate points can be eliminated through point fusion, and surface properties such as normals and curvature can be estimated. Local regions are formed around each point in the point clouds using k-nearest neighbor or range search, and statistical features are calculated. Various statistical discrete measures, such as median (mA), variance (σ^2A), and mean absolute deviation (μADA), are used to compare the local region features between the workpiece point cloud and the standard point cloud. A structural similarity score between the point clouds is calculated based on the relative differences of these feature values. The calculated similarity score is compared with a preset similarity threshold. If the similarity exceeds the threshold, the current workpiece is determined to be a qualified workpiece; otherwise, it is a non-qualified workpiece.

[0070] Compared to existing technologies, this embodiment acquires workpiece images from at least two angles, generates a workpiece image set based on these images, and determines the camera pose and image depth information based on the workpiece image set. A three-dimensional voxel model of the current workpiece is then established based on the workpiece image set, camera pose, and image depth information. The attribute information of the current workpiece is acquired and allocated to each voxel in the three-dimensional voxel model. Gaussian sputtering is performed on each voxel based on its attribute information, camera pose, and a preset Gaussian function. The attribute information includes density, color, and transparency. A mask is determined for each workpiece image in the workpiece image set based on a preset mask extraction model, and the workpiece images are divided into workpiece structure images and background images based on the masks. The workpiece point cloud is determined based on the Gaussian sputtered three-dimensional voxel model and the divided workpiece images. The similarity between the workpiece point cloud and a preset standard workpiece point cloud is calculated. If the similarity exceeds a preset similarity threshold, the current workpiece is determined to be a qualified workpiece. Through this AI-based workpiece processing quality inspection method, the automated process can significantly reduce manual intervention and improve inspection speed and consistency. By comparing the similarity of point clouds, defects in workpieces can be effectively identified, and errors in detection results can be reduced. Combining image depth information, camera pose, and voxel models allows the system to comprehensively consider information from different data sources, thereby providing more accurate detection results.

[0071] The second embodiment of this application relates to a workpiece machining quality inspection method based on artificial intelligence. The second embodiment is an improvement upon the first embodiment, and is largely the same as the first embodiment, with the main difference being that it includes the following steps:

[0072] S201, acquire workpiece images from at least two angles of the current workpiece, generate a workpiece image set based on the workpiece images, and determine the camera pose and image depth information based on the workpiece image set.

[0073] S202, establish a three-dimensional voxel model of the current workpiece based on the workpiece image set, camera pose, and image depth information.

[0074] S203: Obtain the attribute information of the current workpiece, assign the attribute information to each voxel of the 3D voxel model, and perform Gaussian sputtering on each voxel according to the attribute information of each voxel, the camera pose, and the preset Gaussian function; wherein, the attribute information includes density, color, and transparency.

[0075] S204, determine the mask of each workpiece image in the workpiece image set according to the preset mask extraction model, and divide the workpiece image into workpiece structure image and background image according to the mask.

[0076] S205. Based on the three-dimensional voxel model after Gaussian sputtering and the segmented workpiece image, determine the workpiece point cloud of the current workpiece, calculate the similarity between the workpiece point cloud and the preset workpiece standard point cloud, and if the similarity does not exceed the preset similarity threshold, determine that the current workpiece is a defective workpiece.

[0077] S206, input the workpiece point cloud and the preset workpiece standard point cloud into the preset defect detection model, determine the defect type and repair method of the current workpiece, and send the defect type and repair method to the worker's handheld terminal.

[0078] In this solution, the pre-set defect detection model can be a trained machine learning model used to identify defects in the workpiece and suggest corresponding repair methods. This model is trained based on point cloud data of historical non-conforming workpieces and standard point cloud data.

[0079] Defect types can be categorized based on potential problems that may occur during the production, processing, or use of a workpiece. Specifically, they can include: Surface defects: Scratches: Obvious scratches or dents on the surface. Cracks: Tiny cracks on or inside the material. Bubbles: Bubbles present inside or on the surface of the material. Deformation: The workpiece shape deviates from design specifications, such as bending or twisting.

[0080] Dimensional deviation: Too large or too small: The dimensions of the workpiece exceed or fail to meet the design specifications. Uneven: The dimensions of different parts are inconsistent.

[0081] Material defects:

[0082] Impurities: The presence of unwanted impurities or foreign objects in the material. Uneven hardness: Uneven hardness distribution in the workpiece, resulting in some parts being too hard or too soft.

[0083] Assembly defects: Misalignment: Components are not properly aligned or assembled. Loosening: Fasteners are not securely fastened, causing parts to become loose.

[0084] Coating issues: Peeling: The coating peels or flakes off. Uneven color: The color distribution of the coating is uneven.

[0085] Repair methods can be corrective or remedial measures provided for detected defects to bring the workpiece back to design or usage standards. Specifically, these can include: grinding and polishing: grinding surface scratches, cracks, etc., and then polishing to restore smoothness and finish.

[0086] Welding repair: Welding is used to repair cracks or fractures, followed by grinding and heat treatment.

[0087] Size adjustment: For size deviations, the dimensions are corrected through machining (such as turning and grinding).

[0088] Heat treatment: adjusting the hardness of a material or removing internal stress by heating or cooling.

[0089] Coating recoating: removing the problematic coating and recoating it to meet the standard requirements.

[0090] Replacement of parts: If the defect cannot be repaired, consider replacing the faulty component.

[0091] Reassembly: For assembly defects, realignment and fixation are performed to ensure correct assembly.

[0092] Handheld terminals can be devices carried by workers (such as smartphones or dedicated industrial-grade handheld devices) to receive information on the type of defects in workpieces and repair methods. This device can connect to the company's internal network or cloud system to receive and process feedback from the defect detection model.

[0093] If the similarity between the current workpiece's point cloud and the standard point cloud does not reach a preset threshold, the workpiece is considered unqualified. The point cloud data of the unqualified workpiece and the standard point cloud data are input into a preset defect detection model. This model analyzes the point cloud data and identifies the defect type of the workpiece. Based on the identified defect type, the model generates a corresponding repair method. This information (defect type and repair method) is then transmitted to the worker's handheld terminal via wireless communication technology to help them make quick decisions. For example, if the similarity does not meet the requirements, the system inputs the current workpiece's point cloud data and the standard point cloud data into the preset defect detection model. The model analyzes the local differences in these point cloud data, combines them with pre-labeled defect types and repair method tags, and draws the following conclusions:

[0094] Defect type: Surface scratches are present on the upper surface of the workpiece near the right edge.

[0095] Based on the output of the defect detection model, it is recommended to use grinding and polishing methods to repair this surface scratch. The model can also recommend specific grinding tools and process parameters, such as abrasive grit size and polishing time, based on the depth and length of the scratch.

[0096] It is easy to see that in the embodiments of this application, the model automatically identifies surface scratches and defects of the workpiece by comparing point cloud data with standard point clouds, and provides specific repair methods, thereby improving the efficiency and accuracy of quality inspection and reducing errors in manual judgment.

[0097] Based on the above technical solution, in an exemplary embodiment, after sending the defect type and repair method to the worker's handheld terminal, the method further includes:

[0098] The number of defective workpieces is obtained in real time. If the number of workpieces exceeds a preset threshold for defective workpieces, the workpiece image, structural parameters, attribute information and current production parameters of each defective workpiece are obtained.

[0099] Input the workpiece image, structural parameters, attribute information, and current production parameters of each defective workpiece into the preset production parameter optimization model to determine the production parameter adjustment information;

[0100] Update the current production parameters based on the production parameter adjustment information.

[0101] In this solution, the number of workpieces can be the number of workpieces that are detected and determined to be defective on the production line. The system counts the number of these defective workpieces in real time and compares them with a preset threshold.

[0102] Structural parameters can be the geometric and physical structural information of a workpiece, such as dimensions, shape, surface roughness, hole size, and wall thickness. These parameters directly affect the manufacturing accuracy and final quality of the workpiece.

[0103] Current production parameters can be various control parameters in the current production process, such as temperature, pressure, feed rate, tool speed, and material ratio. The settings of these parameters will affect the final quality and performance of the workpiece.

[0104] The pre-defined production parameter optimization model can be a machine learning-trained model used to predict and optimize production parameters based on images, structural parameters, attribute information, and actual parameters during the production process of defective workpieces. By learning the characteristics of defective workpieces and their corresponding production parameters from historical data, the model can automatically generate new production parameter adjustment suggestions to reduce or avoid the production of defective workpieces.

[0105] Production parameter adjustment information can be suggestions output by the model, including recommendations for adjusting current production parameters. For example, the model might suggest lowering the temperature, adjusting the feed rate, or changing the proportions of certain materials to improve product quality.

[0106] The system can record the number of defective workpieces detected on the production line in real time using sensors and detection devices. Each time a workpiece is marked as defective, the system updates the current number of defective workpieces. If the number of defective workpieces exceeds a preset threshold, the system collects detailed data for each defective workpiece, including workpiece images, structural parameters, attribute information, and the current production parameters of the production line. This data collection is achieved through automated detection and sensor networks, ensuring real-time data accuracy. The system inputs the collected data on defective workpieces into a preset production parameter optimization model. This model, trained based on historical data, can identify factors that may lead to defective products under current production conditions. By analyzing the input workpiece images, structural parameters, attribute information, and production parameters, the model generates production parameter adjustment information. The model then outputs a set of new production parameter suggestions aimed at optimizing the production process and reducing the generation of defective workpieces. Production parameter adjustment information may include adjusting temperature, speed, pressure, and material ratios to ensure that the produced workpieces meet quality standards. Finally, based on the production parameter adjustment information provided by the model, the system automatically updates the current production parameters of the production line. This can be done automatically by the control system or manually confirmed and implemented by operators. For example, a workpiece on a production line develops surface defects due to excessive temperature. After detecting a certain number of defective workpieces, the system collects information such as images, dimensional data, and the current production temperature of the relevant workpieces. After analysis by a preset production parameter optimization model, it recommends reducing the furnace temperature by 5 degrees Celsius and slowing down the feed rate. The system automatically adjusts the production parameters based on these recommendations and then continues to monitor workpiece quality to ensure the problem is resolved.

[0107] In this solution, the real-time detection and response to the occurrence of defective workpieces, and the timely adjustment of production parameters, can avoid the large-scale generation of defective workpieces, thereby reducing the scrap rate and improving production efficiency.

[0108] Based on the above technical solution, in an exemplary embodiment, the training process of the preset defect detection model includes:

[0109] Obtain defect point clouds and standard point clouds of historical non-conforming workpieces, label the defect point clouds with defect type labels and repair method labels, and create a second dataset based on the defect point clouds, standard point clouds, defect type labels and repair method labels;

[0110] Construct a defect detection model and train the defect detection model based on the second dataset until the defect detection model reaches the preset defect detection model training standard.

[0111] In this solution, the defect point cloud of historically defective workpieces can be point cloud data of historically defective workpieces obtained through 3D scanning or other measurement techniques, representing the three-dimensional structure of the workpiece surface or interior, including defect areas. Each point cloud point typically contains location information (x, y, z) and other attributes (such as color, density, etc.).

[0112] The standard point cloud of a historically non-conforming workpiece can be point cloud data of similar workpieces under ideal conditions, typically generated from a design model or a qualified workpiece that has passed quality inspection. The standard point cloud serves as a benchmark for comparison and is used to measure the degree of defect in non-conforming workpieces.

[0113] Defect type labels can be used to classify various defect regions in a defect point cloud. Each label represents a specific defect type, such as cracks, porosity, or deformation. These labels need to be annotated by professionals or automated labeling tools based on the characteristics of the defects.

[0114] Repair method labels can be repair strategy labels corresponding to defect type labels. Each defect type has a corresponding repair method, such as welding repair, surface grinding, and material replenishment. Repair method labels are used to guide the repair operations for non-conforming workpieces during production or maintenance.

[0115] The second dataset consists of four parts: defect point clouds, standard point clouds, defect type labels, and repair method labels. It serves as the foundational data for training the defect detection model, containing defect information about the workpiece and its corresponding repair strategies. Each sample in the dataset corresponds to one or more defects and their corresponding repair methods.

[0116] Preset training standards for defect detection models can be target standards set during the training process, and may include evaluation metrics such as accuracy, recall, precision, and F1-score. These standards are used to measure the model's performance and ensure that the model achieves the expected results in defect detection and repair suggestions.

[0117] Point cloud data of historical defective workpieces can be collected to ensure the accuracy and completeness of the point cloud data. Standard point cloud data can be obtained from design models or qualified workpieces. Then, defect type labeling is performed on the point clouds of historical defective workpieces, marking each defective region and attaching a corresponding defect type label. Based on the defect type, a corresponding repair method label is assigned to each defective point cloud region. The defective point cloud, standard point cloud, defect type labels, and repair method labels are integrated to create a second dataset. The data is cleaned and normalized to ensure the consistency and accuracy of the training data. Then, a defect detection model is built using deep learning or machine learning methods. The model input is the defective point cloud and the standard point cloud, and the output is the defect type and the corresponding repair method. The second dataset is input into the model for iterative training, and model parameters are adjusted to improve performance. The model is evaluated using preset defect detection model training standards. Based on the model's performance on the validation set, it is decided whether to continue training or adjust the model structure; training stops if the standard is met.

[0118] In this solution, the pre-set defect detection model can accurately identify the types of defects in the workpiece and provide repair suggestions, thereby improving production quality and efficiency.

[0119] Based on the above technical solution, in an exemplary embodiment, the training process of the preset production parameter optimization model includes:

[0120] Obtain historical defective workpiece images, label the structural parameter labels, attribute information labels, and production parameter labels of the historical defective workpiece images, and create a third dataset based on the historical defective workpiece images, structural parameter labels, attribute information labels, and production parameter labels;

[0121] Construct a production parameter optimization model, and train the production parameter optimization model based on the third dataset until the production parameter optimization model reaches the preset production parameter optimization model training standard.

[0122] In this approach, historical images of defective workpieces refer to images of defective workpieces captured during the production process using a vision inspection system, industrial camera, or other imaging equipment. These images record the workpiece's appearance, surface defects, shape anomalies, and other details, which are used for subsequent analysis and model training.

[0123] Structural parameter labels describe structural information such as the workpiece's geometry, dimensions, and morphological characteristics. Examples include the workpiece's length, width, height, angles, and curvature. Structural parameter labels help in understanding shape changes during manufacturing and their correlation with non-conformities.

[0124] Attribute information labels can include physical or chemical properties such as material properties, color, surface smoothness, and hardness. For example, a workpiece may have a specific material composition, color uniformity, or a specific surface roughness grade. These attribute information labels help identify the physical characteristics of the workpiece and the relationship between these characteristics and nonconformities.

[0125] Production parameter labels record the process parameters of a workpiece during production. These include, for example, temperature, pressure, processing speed, and tool wear. These parameters are directly related to the production equipment and processes, and can help analyze which production parameters might lead to workpiece defects.

[0126] The third dataset can be a comprehensive dataset formed by combining historical images of defective workpieces with corresponding structural parameter labels, attribute information labels, and production parameter labels. This dataset is used to train the production parameter optimization model, where each data point (sample) contains an image and its associated label information. The third dataset reflects the performance of different workpieces under different production conditions and serves as the foundation for training the production parameter optimization model.

[0127] The preset production parameter optimization model training standards include the performance indicators that the model needs to achieve during the training and validation phases. Specifically, these may include: Accuracy: the proportion of production parameter adjustment suggestions that the model correctly predicts.

[0128] Loss value: The loss value of the model during training should be lower than a certain set threshold.

[0129] Model convergence: The model converges to a stable performance within a number of training cycles, and no longer exhibits significant performance fluctuations.

[0130] Validation set performance: The model's performance on the validation set should meet specific requirements, such as achieving the expected accuracy and recall.

[0131] Image data of historically defective workpieces can be extracted from production lines or inspection systems. Based on images or CAD models, experts or automated tools annotate the workpieces' geometric parameters, such as dimensions and shape. Physical attribute information is also annotated based on the workpiece's material and surface features. Relevant process parameters of the workpieces during production are collected and matched with corresponding images for annotation. Image data is integrated with various label data to form a third dataset. Missing and outlier values ​​in the dataset are handled, data formats are standardized, and data normalization is performed. Then, algorithms such as neural networks, decision trees, and random forests are used to build a production parameter optimization model. The model input is image and label data, and the output is optimized production parameter suggestions. The model is trained using the third dataset, and model parameters are adjusted to improve model performance. A validation set is used to evaluate model performance and check if it meets the preset training standards. If the standards are met, the trained and compliant model is deployed to the production system to analyze new data in real time and provide production parameter adjustment suggestions.

[0132] In this solution, the production parameter optimization model can effectively identify and analyze the production parameters that cause workpiece defects, and provide optimization suggestions, thereby improving the overall efficiency of the production line and product quality.

[0133] The third embodiment of this application relates to a workpiece machining quality inspection method based on artificial intelligence. The third embodiment is an improvement upon the first embodiment, specifically including the following steps:

[0134] The training process of the pre-defined mask extraction model includes:

[0135] S301, acquire historical workpiece images, label the historical workpiece images with mask tags, and create a first dataset based on the historical workpiece images and the mask tags.

[0136] Historical workpiece images can refer to digital images of workpieces preserved during past production or inspection processes. These images are typically captured using high-resolution cameras or industrial video systems, covering different angles and appearance features of the workpiece.

[0137] Mask labels can be annotation data corresponding to historical workpiece images, defining which pixels in the image belong to the workpiece itself (foreground) and which pixels belong to the background. The mask can be represented as a binary image, with the foreground (workpiece portion) represented by one color (e.g., white) and the background represented by another color (e.g., black).

[0138] The first dataset can be a training dataset consisting of historical workpiece images and corresponding mask labels. This dataset can be used to train a mask extraction model, enabling it to automatically identify and segment the foreground and background in workpiece images.

[0139] A large number of workpiece images can be collected from the factory's production or inspection history, ensuring that the images cover different types, shapes, colors, and angles of the workpieces. The collected images are preprocessed to remove noise and adjust brightness and contrast, ensuring image quality suitable for subsequent annotation and model training. Then, professional image annotation tools, such as LabelMe, Labelbox, or CVAT, are used to manually or semi-automatically annotate the workpiece regions in the images. Based on the workpiece's contour and shape, corresponding mask labels are created on each image. Annotations should be as accurate as possible to ensure the model learns effective information. Each historical workpiece image is paired with its corresponding mask label, ensuring a one-to-one correspondence between image and mask. The paired dataset is then divided into training, validation, and test sets for model training, validation, and testing. Typically, the training set comprises 70-80% of the data, while the validation and test sets each comprise 10-15%.

[0140] S302, Construct a mask extraction model and train the mask extraction model based on the first dataset until the mask extraction model reaches the preset mask extraction model training standard.

[0141] Preset training criteria define the metrics and conditions required for a model to reach a satisfactory state during training. Specifically, these may include: Accuracy: the degree of matching between the mask extraction model's mask prediction results on the test set and the actual labels. A high accuracy (e.g., above 90%) is typically required.

[0142] IoU: Cross-Union Ratio between the predicted mask and the actual mask label, used to measure the segmentation performance of the model. Preset criteria can require IoU to reach a certain threshold (e.g., 0.85 or higher).

[0143] Loss function: During model training, a specific loss function (such as cross-entropy loss or Dice coefficient loss) is used to measure the degree of error in the model. Training criteria may require the loss value to decrease to a preset range (e.g., less than 0.1).

[0144] Recall and precision: These measures the model’s ability to identify the positive class (foreground) and the correct recognition rate. Typically, these metrics are required to reach a certain standard (e.g., above 90%).

[0145] Training epochs: Based on the training results, training can be stopped when the model performance stabilizes within a certain number of epochs and no longer improves significantly. The preset standard is usually set to stop training when the model's performance on the validation set no longer improves within a certain number of consecutive epochs.

[0146] Depending on the application scenario and dataset characteristics, a suitable model architecture can be selected, specifically including U-Net, Mask R-CNN, SegNet, etc. The model is built using a deep learning framework (such as TensorFlow or PyTorch). Define the input layer, encoder (extracting features), decoder (restoring spatial information), and output layer (generating the mask). Then, initialize the parameters of each layer, using random initialization or a pre-trained model for transfer learning. Load the first dataset into the model and divide it into training and validation sets. Ensure the diversity and balance of the dataset to avoid overfitting. To improve the model's generalization ability, data augmentation (such as rotation, scaling, flipping, and lighting adjustment) can be applied to the training data. Choose an appropriate loss function, such as Cross-Entropy Loss or Dice Loss, to measure the difference between the model's prediction and the actual mask. Select an optimizer, such as Adam or SGD, to update the model parameters and minimize the loss function. Then begin training the model, continuously adjusting the model parameters during training to gradually reduce the loss function. After each training epoch, the model performance is evaluated using a validation set. At the end of each epoch, the model's precision, IoU, recall, and other metrics are evaluated using the validation set and compared to preset standards. Based on the evaluation results, the model's hyperparameters (such as learning rate and batch size) are adjusted, or regularization techniques (such as Dropout and weight decay) are added to improve model performance. During training, the loss and metric performance on the validation set are monitored in real time. Training is stopped when the validation loss stabilizes and the validation metrics reach or exceed preset standards.

[0147] It is not difficult to see that, in the embodiments of this application, by setting strict training standards, it can be ensured that the model can accurately extract the mask in practical applications, thereby improving the reliability of workpiece recognition and segmentation.

[0148] Based on the above technical solution, in an exemplary embodiment, the mask of each workpiece image in the workpiece image set is determined according to a preset mask extraction model, and the workpiece image is divided into a workpiece structure image and a background image according to the mask, including:

[0149] The mask of each workpiece image in the workpiece image set is determined according to the preset mask extraction model. If the masks of each workpiece image are the same, the workpiece image is divided into a workpiece structure image and a background image according to the mask.

[0150] Accordingly, after determining the mask for each workpiece image in the workpiece image set based on a preset mask extraction model, the method further includes:

[0151] If the masks of the workpiece images are inconsistent, mask labels are added to each workpiece image in the workpiece image set, and each workpiece image and its corresponding mask label are added to the first dataset.

[0152] The mask extraction model is retrained based on the first dataset until it reaches the preset mask extraction model training standard.

[0153] The mask for each workpiece image in the workpiece image set is re-determined based on the retrained mask extraction model. If the masks for each workpiece image are consistent, the workpiece image is divided into a workpiece structure image and a background image based on the mask.

[0154] In this scheme, each workpiece image in the workpiece image set can be input into a preset mask extraction model. The model generates a mask based on image features, representing the foreground and background portions of the workpiece in the image. Then, the masks of each workpiece image are compared to see if they are consistent in shape, position, and structure. If the masks match at these key points, they are considered consistent; otherwise, they are considered inconsistent. If the masks generated by the model are consistent, these masks are used to divide the workpiece image into a workpiece structure image and a background image. The workpiece structure image represents the foreground portion of the workpiece, and the background image represents the background portion outside the workpiece. Specifically, the image can be segmented using masks to obtain two image regions. Typically, the mask is a binary image, with the foreground (workpiece) portion being 1 and the background portion being 0. Based on the mask, the portion of the original image corresponding to a mask value of 1 is extracted as the workpiece structure image, and the portion corresponding to a mask value of 0 is extracted as the background image. For workpiece images with inconsistent masks, mask labels can be manually or automatically annotated. These annotated workpiece images and their corresponding mask labels are added to the first dataset to enrich the model's training data. Specifically, if the output of the mask extraction model does not match the expected result, the mask needs to be manually corrected or other algorithms can be used to generate more accurate mask labels. The annotation process can be performed using image annotation tools to ensure that the mask labels accurately reflect the boundary between the workpiece and the background. Then, the mask extraction model is retrained using the updated first dataset (containing the newly annotated mask labels), and the model is continuously trained until its performance reaches the preset training standard. Finally, the retrained mask extraction model is used to extract masks again from each image in the workpiece image set, and the consistency of the newly generated masks is checked. If they are consistent, the workpiece image is re-divided into workpiece structure images and background images based on the masks. If the masks are still inconsistent after one retraining, the annotation, training, and segmentation process needs to be repeated until the model can consistently generate correct masks and complete the segmentation of the workpiece image.

[0155] In this solution, through continuous annotation and model optimization, the segmentation effect will gradually improve, eventually achieving a stable and high-precision mask extraction effect to meet the needs of subsequent workpiece analysis and processing.

[0156] The steps of the various methods described above are only for clarity. In practice, they can be combined into one step or some steps can be split into multiple steps. As long as they include the same logical relationship, they are all within the scope of protection of this patent. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, but without changing the core design of the algorithm and process, are also within the scope of protection of this patent.

[0157] The fourth embodiment of this application relates to an artificial intelligence-based workpiece machining quality inspection system, which specifically includes the following:

[0158] The image acquisition module 401 is used to acquire workpiece images from at least two angles of the current workpiece, generate a workpiece image set based on the workpiece images, and determine the camera pose and image depth information based on the workpiece image set.

[0159] The model building module 402 is used to build a three-dimensional voxel model of the current workpiece based on the workpiece image set, camera pose and image depth information.

[0160] The Gaussian sputtering module 403 is used to acquire the attribute information of the current workpiece, distribute the attribute information to each voxel of the three-dimensional voxel model, and perform Gaussian sputtering on each voxel according to the attribute information of each voxel, the camera pose, and the preset Gaussian function; wherein, the attribute information includes density, color, and transparency.

[0161] The mask determination module 404 is used to determine the mask of each workpiece image in the workpiece image set according to the preset mask extraction model, and divide the workpiece image into a workpiece structure image and a background image according to the mask.

[0162] The workpiece detection module 405 is used to determine the workpiece point cloud of the current workpiece based on the three-dimensional voxel model after Gaussian sputtering and the segmented workpiece image, calculate the similarity between the workpiece point cloud and the preset workpiece standard point cloud, and determine the current workpiece as a qualified workpiece if the similarity exceeds the preset similarity threshold.

[0163] It is worth noting that all modules involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. Furthermore, to highlight the innovative aspects of this application, this embodiment does not introduce units that are not closely related to solving the technical problem proposed in this application; however, this does not mean that other units are absent from this embodiment.

[0164] The flowcharts or block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of devices, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-specific system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0165] The scope of this application is defined by the appended claims rather than the foregoing description, and is therefore intended to encompass all variations falling within the meaning and scope of equivalents of the claims. No reference numerals in the claims should be construed as limiting the scope of the claims. Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices recited in a device claim may also be implemented by a single unit or device in software or hardware. Terms such as "first," "second," etc., are used only for distinguishing descriptions and do not indicate any particular order, nor should they be construed as indicating or implying relative importance.

[0166] Those skilled in the art will understand that the above embodiments are specific embodiments for implementing this application, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of this application.

Claims

1. A workpiece machining quality inspection method based on artificial intelligence, wherein, The method includes: The process involves acquiring workpiece images from at least two angles, generating a workpiece image set based on these images, and determining the camera pose and image depth information based on the workpiece image set. A three-dimensional voxel model of the current workpiece is then established based on the workpiece image set, camera pose, and image depth information. The attribute information of the current workpiece is acquired and distributed to each voxel in the three-dimensional voxel model. Gaussian sputtering is then performed on each voxel based on its attribute information, camera pose, and a preset Gaussian function. The attribute information includes density, color, and transparency. A mask is determined for each workpiece image in the workpiece image set based on a preset mask extraction model, and the workpiece images are divided into workpiece structure images and background images based on the masks. The workpiece point cloud of the current workpiece is determined based on the Gaussian sputtered three-dimensional voxel model and the divided workpiece images. The similarity between the workpiece point cloud and a preset standard workpiece point cloud is calculated. If the similarity exceeds a preset similarity threshold, the current workpiece is determined to be a qualified workpiece.

2. The method as described in claim 1, wherein, The preset Gaussian function is: Where G(x, D, C, T) is the voxel contribution value; x is the distance variable, representing the distance between the voxel center point and the camera; D is the voxel density, i.e., the material concentration of the voxel in three-dimensional space; C is the voxel color, which can represent color intensity or color information; T is the voxel transparency, i.e., the transparency of the voxel; σ is the standard deviation of the Gaussian function, which determines the degree of sputtering diffusion; W D Voxel density weights reflect the influence of voxel density on the Gaussian function; W C Voxel color weights reflect the influence of voxel colors on the Gaussian function; W T The voxel transparency weights reflect the influence of voxel transparency on the Gaussian function; The effect of voxels on light was calculated; is the normalization factor of the Gaussian function, ensuring that the integral of the entire function is 1.

3. The method as described in claim 1, wherein, Before performing Gaussian sputtering on each voxel based on its attribute information, camera pose, and a preset Gaussian function, the method further includes: The system acquires preset weighting rules, initial Gaussian function standard deviation, and preset standard deviation update formula. Correspondingly, it performs Gaussian sputtering on each voxel based on its attribute information, camera pose, and preset Gaussian function. This includes: determining the voxel density weight, voxel color weight, and voxel transparency weight values ​​for each voxel within the preset Gaussian function based on the preset weighting rules and voxel attribute information; updating the Gaussian function standard deviation based on the voxel density weight values, initial Gaussian function standard deviation, and preset standard deviation update formula; acquiring the center point position of each voxel; determining the distance variable value of each voxel based on its center point position and camera pose; and performing Gaussian sputtering on each voxel based on the voxel density weight values, voxel color weight values, voxel transparency weight values, updated Gaussian function standard deviation, distance variable value, and preset Gaussian function.

4. The method of claim 1, wherein, After calculating the similarity between the workpiece point cloud and a preset standard workpiece point cloud, the method further includes: If the similarity does not exceed the preset similarity threshold, the current workpiece is determined to be a defective workpiece; the workpiece point cloud and the preset workpiece standard point cloud are input into the preset defect detection model to determine the defect type and repair method of the current workpiece, and the defect type and repair method are sent to the worker's handheld terminal.

5. The method of claim 4, wherein, After sending the defect type and repair method to the worker's handheld terminal, the method further includes: The system acquires the number of defective workpieces in real time. If the number of workpieces exceeds a preset threshold for defective workpieces, it acquires the workpiece image, structural parameters, attribute information, and current production parameters of each defective workpiece. The system inputs the workpiece image, structural parameters, attribute information, and current production parameters of each defective workpiece into a preset production parameter optimization model to determine production parameter adjustment information. The system updates the current production parameters based on the production parameter adjustment information.

6. The method of claim 1, wherein, The training process of the pre-defined mask extraction model includes: Acquire historical workpiece images, label the historical workpiece images with mask tags, and create a first dataset based on the historical workpiece images and the mask tags; construct a mask extraction model, and train the mask extraction model based on the first dataset until the mask extraction model reaches a preset mask extraction model training standard.

7. The method of claim 6, wherein, The mask for each workpiece image in the workpiece image set is determined according to a preset mask extraction model, and the workpiece images are divided into workpiece structure images and background images according to the masks, including: The method further includes the following steps: First, if the masks of each workpiece image in the workpiece image set are consistent, the method determines the masks of each workpiece image according to a preset mask extraction model. If the masks of each workpiece image are consistent, the workpiece image is divided into a workpiece structure image and a background image based on the masks. Then, after determining the masks of each workpiece image in the workpiece image set according to the preset mask extraction model, the method further includes: if the masks of each workpiece image are inconsistent, labeling each workpiece image in the workpiece image set with mask labels and adding each workpiece image and its corresponding mask label to a first dataset; retraining the mask extraction model based on the first dataset until the mask extraction model reaches a preset mask extraction model training standard; and re-determining the masks of each workpiece image in the workpiece image set based on the retrained mask extraction model. If the masks of each workpiece image are consistent, the workpiece image is divided into a workpiece structure image and a background image based on the masks.

8. The method of claim 4, wherein, The training process of the pre-defined defect detection model includes: Obtain defect point clouds and standard point clouds of historical non-conforming workpieces, label the defect point clouds with defect type labels and repair method labels, and create a second dataset based on the defect point clouds, standard point clouds, defect type labels, and repair method labels; construct a defect detection model, and train the defect detection model based on the second dataset until the defect detection model reaches the preset defect detection model training standard.

9. The method of claim 5, wherein, The training process of the pre-defined production parameter optimization model includes: Acquire historical defective workpiece images, and label the structural parameter labels, attribute information labels, and production parameter labels of the historical defective workpiece images. Create a third dataset based on the historical defective workpiece images, structural parameter labels, attribute information labels, and production parameter labels. Construct a production parameter optimization model, and train the production parameter optimization model based on the third dataset until the production parameter optimization model reaches the preset production parameter optimization model training standard.

10. A workpiece machining quality inspection system based on artificial intelligence, characterized in that, The system includes: The system includes the following modules: an image acquisition module for acquiring workpiece images from at least two angles, generating a workpiece image set based on these images, and determining camera pose and image depth information based on the workpiece image set; a model building module for building a 3D voxel model of the current workpiece based on the workpiece image set, camera pose, and image depth information; a Gaussian sputtering module for acquiring the attribute information of the current workpiece, distributing the attribute information to each voxel in the 3D voxel model, and performing Gaussian sputtering on each voxel based on the attribute information of each voxel, camera pose, and a preset Gaussian function; wherein the attribute information includes density, color, and transparency; a mask determination module for determining the mask for each workpiece image in the workpiece image set based on a preset mask extraction model, and dividing the workpiece image into a workpiece structure image and a background image based on the mask; and a workpiece detection module for determining the workpiece point cloud of the current workpiece based on the Gaussian sputtered 3D voxel model and the divided workpiece images, calculating the similarity between the workpiece point cloud and a preset workpiece standard point cloud, and determining the current workpiece as a qualified workpiece if the similarity exceeds a preset similarity threshold.

Images