Defect visual classification detection method and system oriented to numerical control system

Abstract

The invention discloses a numerical control system-oriented defect visual classification detection method and system, and the method comprises the steps: obtaining an original defect sample covering a numerical control system part, carrying out the multi-modal image collection, carrying out the weighted fusion of visible light, near-infrared light and ultraviolet light images in an original sample library through a multispectral fusion algorithm based on the original defect sample, and carrying out the detection of the defect visual classification. The method comprises the following steps: generating an original defect sample, generating an enhanced sample, acquiring three-dimensional geometric information of the surface of a part by utilizing a 3D laser scanning device, transmitting and receiving ultrasonic waves through an ultrasonic detection module, detecting defects in the part, generating a virtual sample based on the original defect sample, the enhanced sample and a 3D model sample, and combining the original defect sample, the enhanced sample, the 3D model sample and the virtual sample to obtain a three-dimensional image. And training a defect classification model based on the mixed sample library. Defect evolution under different working conditions can be simulated, virtual defect samples conforming to real physical laws are generated, the defect state coverage range is wide, and defect data which are difficult to obtain by real samples are supplemented.

Description

Technical Field

[0001] This invention relates to the field of visual defect detection in CNC systems, specifically to a visual defect classification and detection method and system for CNC systems. Background Technology

[0002] As the core equipment of modern manufacturing, the operational stability and machining accuracy of CNC systems directly determine product quality and production efficiency. Traditional manual visual inspection and offline testing methods suffer from low efficiency, high missed detection rate, and difficulty in quantification. Automated testing technologies based on a single modality (such as visible light imaging only) are limited by detection depth and feature extraction capabilities, and cannot detect hidden defects such as internal cracks and stress concentrations. With the development of industry and intelligent manufacturing, the industry has put forward higher requirements for defect detection of CNC systems, requiring the realization of full life cycle monitoring of multiple types of defects (surface / internal / functional), real-time online detection on the production line, and defect adaptability detection under new materials and processes. Summary of the Invention

[0003] The purpose of this invention is to provide a visual classification and detection method and system for defects in CNC systems, so as to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method and system for visual classification and detection of defects in CNC systems, comprising:

[0005] S1: Obtain original defect samples covering CNC system components;

[0006] S2: Multimodal image acquisition, which uses a multispectral camera system to acquire multidimensional images of CNC system components, including visible light, near-infrared light and ultraviolet light images. The acquired multimodal images are then standardized to establish an original sample library.

[0007] S3: Based on the original defect samples, a multispectral fusion algorithm is used to perform weighted fusion of visible light, near-infrared light and ultraviolet light images in the original sample library to generate enhanced samples containing multiple features;

[0008] S4: Use 3D laser scanning equipment to obtain three-dimensional geometric information of the component surface, and at the same time, use ultrasonic testing module to emit and receive ultrasonic waves to detect defects inside the component. Register and fuse the 3D laser scanning data and ultrasonic testing data to construct a 3D model sample containing surface and internal defect information.

[0009] S5: Based on original defect samples, enhanced samples, and 3D model samples, a physical model of the component is constructed using finite element analysis to simulate the evolution of defects under different working conditions and generate virtual samples. An intelligent agent is introduced into the physical model environment to simulate dynamic defects by adjusting equipment parameters and generate defect evolution sequence samples. Using the original defect samples as training data, a generative adversarial network is built. The generator learns the distribution characteristics of real defects, and the discriminator optimizes the generator by comparing it with real samples, and finally generates virtual samples.

[0010] S6: Merge the original defect samples, enhanced samples, 3D model samples and virtual samples to form a hybrid sample library;

[0011] S7: Based on a hybrid sample library, a defect classification model is trained, and the trained model is deployed to an industrial inspection system to perform real-time image acquisition and analysis of CNC system components on the production line, outputting defect type and location, and periodically collecting misjudged samples that occur during the inspection process to supplement the original defect samples.

[0012] Furthermore, the specific processing method of the multispectral fusion algorithm is as follows:

[0013] Pixel-level alignment of visible light, near-infrared light, and ultraviolet light images is performed based on spatiotemporal synchronization parameters.

[0014] Surface defect contour features were extracted from visible light images, internal crack features were extracted from near-infrared light images, and stress concentration fluorescence information was extracted from ultraviolet light images.

[0015] The fusion weights are dynamically set according to the defect type. Through matrix weighting operations, the feature matrices of the three types of images are superimposed according to the weights to generate an enhanced sample containing surface, internal and stress features.

[0016] Histogram equalization and contrast stretching techniques are used to optimize the visual effect of the fused image.

[0017] Furthermore, the process of constructing the 3D model sample is as follows:

[0018] The transformation relationship between the coordinate system of the 3D laser scanning equipment and the coordinate system of the ultrasonic testing module is established through a calibration procedure to ensure that the two modal data are in a unified spatial reference.

[0019] By using a feature point matching algorithm, the same geometric feature points are identified in the surface point cloud data acquired by the 3D laser scanning device and the tomographic image data of the ultrasonic detection module. Data registration is then performed based on these feature points to eliminate deviations caused by differences in acquisition position and angle.

[0020] By employing voxel fusion technology, the scanning data from the registered 3D laser scanning device is fused with the data acquired by the ultrasonic testing module. This process fills in and integrates surface and internal defect information in three-dimensional space, generating a 3D model sample that includes the surface morphology, dimensional information, and internal defect distribution of the component.

[0021] Furthermore, the hybrid sample library is generated in the following way:

[0022] Based on the defect feature consistency standard, the virtual sample is matched with the original defect sample. If the match is consistent, the virtual sample is retained.

[0023] Based on the standard of scene diversity, virtual samples with different lighting conditions, different wear levels, and different component states are obtained;

[0024] The selection criteria are adjusted based on the set proportions of the original defect samples, enhanced samples, 3D model samples, and virtual samples, and the total number of samples is not less than the set value. If the proportion requirement is not met or the total number is insufficient, the previous steps are returned to adjust the screening criteria or supplement the samples until the conditions are met, and finally a mixed sample library is formed.

[0025] Furthermore, the standardization process includes:

[0026] Geometric correction: Eliminates lens distortion through calibration procedures and performs perspective transformation on the image;

[0027] Noise suppression: Nonlocal mean filtering is used to remove Gaussian noise and salt-and-pepper noise while preserving defect edge details;

[0028] Data archiving: Store data by component type and establish an original sample library.

[0029] Furthermore, the process of training the defect classification model is as follows:

[0030] A deep learning backbone network is used to extract features from mixed samples, capturing the shape, texture, and multimodal fusion features of defects;

[0031] The model is trained using the original sample library, and the defect categories and confidence levels are output. The model is then tested using a set mixed sample to ensure that the model's accuracy in identifying defects reaches the preset threshold standard.

[0032] This invention also provides a defect visual classification and detection system for CNC systems, used in the application of the aforementioned defect visual classification and detection method. The system includes a database, a multispectral camera system, a scanning and detection unit, a 3D model building unit, a virtual sample generation unit, a sample integration unit, and a model optimization unit. The database is used for acquiring and storing original defect samples. The multispectral camera system is used for acquiring and fusing multimodal images to form enhanced samples. The scanning and detection unit is used to acquire three-dimensional geometric information using a scanning device. The 3D model building unit is used to construct 3D model samples based on the three-dimensional geometric information. The virtual sample generation unit is used to generate virtual samples. The sample integration unit is used to merge the original defect samples, enhanced samples, 3D model samples, and virtual samples to form a hybrid sample library. The model optimization unit is used to train a defect classification model based on the hybrid sample library.

[0033] Furthermore, the multispectral camera system includes an imaging unit, a control module, and a data processing module. The imaging unit is used to detect and image target objects from different spectral dimensions to acquire visual information. The imaging unit includes a visible light imaging module, a near-infrared light imaging module, and an ultraviolet light imaging module. The control module is used to coordinate the synchronous triggering of the three imaging modules, adjust exposure parameters, and control the image acquisition timing. The data processing module has a built-in multispectral image fusion algorithm for weighted fusion of visible light, near-infrared light, and ultraviolet light images.

[0034] Furthermore, the data processing module is also used for the standardization processing of multimodal images.

[0035] Furthermore, the scanning and detection unit includes a 3D laser scanning device and an ultrasonic detection module. The 3D laser scanning device is used to acquire three-dimensional geometric information of the component surface, and the ultrasonic detection module is used to emit and receive ultrasonic waves to detect defects inside the component.

[0036] Compared with the prior art, the beneficial effects of the present invention are:

[0037] The imaging unit uses three independent imaging modules—visible light, near-infrared light, and ultraviolet light—to capture surface defects, internal cracks, and stress concentration areas, respectively, enabling multi-dimensional data acquisition of CNC system components. It can comprehensively acquire information on different types of surface, internal, and stress defects, avoiding the limitations of single-mode detection and providing a comprehensive data foundation for subsequent analysis.

[0038] Based on a multispectral fusion algorithm, the visible light, near-infrared light, and ultraviolet light images are first aligned at the pixel level, and surface defect contours, internal cracks, and stress concentration fluorescence information are extracted respectively. The fusion weights are dynamically set according to the defect type, and the feature matrix is ​​fused through matrix weighted operation. Histogram equalization and contrast stretching techniques are used to optimize the visual effect, give full play to the advantages of different modal detection, fuse multi-source information, and generate enhanced samples containing surface, internal, and stress features, enriching the feature dimensions of the samples. Dynamic weight allocation makes the fused samples more prominent in key defect features, improves sample quality, optimizes visual effect, makes defect details more obvious, and facilitates model learning.

[0039] For complex components such as curved guide rails, the surface three-dimensional geometric information obtained by 3D laser scanning and the internal defect data of ultrasonic testing are integrated into a complete 3D model sample through voxel fusion technology. This enriches the spatial dimension features of the defects, provides more comprehensive information for defect detection of complex components, makes up for the inadequacy of two-dimensional images in expressing complex structures, and improves the accuracy and reliability of detection.

[0040] Based on original defect samples, enhanced samples, and 3D model samples, a physical model of the component is constructed using finite element analysis to simulate the evolution of defects under different working conditions. This generates virtual defect samples that conform to real physical laws, covering the state of defect evolution, supplementing defect data under extreme working conditions that are difficult to obtain from real samples, increasing the number and diversity of samples, and providing richer data for model training.

[0041] By merging real and virtual samples and filtering them according to the criteria of defect feature consistency and scenario diversity, a hybrid sample library is formed. The samples cover multiple scenarios, and the reasonable sample ratio allocation fully leverages the advantages of real and virtual samples, providing sufficient and high-quality data for model training. A deep learning backbone network is used to extract defect shape, texture, and multimodal fusion features to train an end-to-end classification model. The model is tested using 10% of the hybrid samples. The powerful feature extraction capability of deep learning can effectively capture complex defect features. The end-to-end training method simplifies the model building process and improves training efficiency. Attached Figure Description

[0042] Figure 1 This is a flowchart of a defect visual classification and detection method for CNC systems according to an embodiment of the present invention.

[0043] Figure 2 This is a diagram illustrating the composition of a defect visual classification and detection system for CNC systems according to an embodiment of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Please see Figure 1 As shown, this embodiment of the invention provides a defect visual classification and detection method for CNC systems, including:

[0046] S1: Obtain original defect samples covering CNC system components;

[0047] S2: Multimodal image acquisition, which uses a multispectral camera system to acquire multidimensional images of CNC system components, including visible light, near-infrared light and ultraviolet light images. The acquired multimodal images are then standardized to establish an original sample library.

[0048] S3: Based on the original defect samples, a multispectral fusion algorithm is used to perform weighted fusion of visible light, near-infrared light and ultraviolet light images in the original sample library to generate enhanced samples containing multiple features;

[0049] S4: Use 3D laser scanning equipment to obtain three-dimensional geometric information of the component surface, and at the same time, use ultrasonic testing module to emit and receive ultrasonic waves to detect defects inside the component. Register and fuse the 3D laser scanning data and ultrasonic testing data to construct a 3D model sample containing surface and internal defect information.

[0050] S5: Based on original defect samples, enhanced samples, and 3D model samples, a physical model of the component is constructed using finite element analysis to simulate the evolution of defects under different working conditions, generating virtual samples from wear to fracture. An intelligent agent is introduced into the physical model environment to simulate dynamic defects of the motor rotor caused by changes in working conditions, such as imbalance and winding short circuit, by adjusting parameters such as current and speed. Defect evolution sequence samples are generated. Using the original defect samples as training data, a generative adversarial network is built. The generator learns the distribution characteristics of real defects, such as the edge burrs of scratches and the random distribution of bubbles in solder joints. The discriminator optimizes the generator by comparing it with real samples, and finally generates virtual samples.

[0051] S6: Merge the original defect samples, enhanced samples, 3D model samples and virtual samples to form a hybrid sample library;

[0052] S7: Based on a hybrid sample library, a defect classification model is trained, and the trained model is deployed to an industrial inspection system to perform real-time image acquisition and analysis of CNC system components on the production line, outputting defect type and location, and periodically collecting misjudged samples that occur during the inspection process to supplement the original defect samples.

[0053] The standardization process mentioned above includes:

[0054] Geometric correction: Eliminate lens distortion through calibration procedures and perform perspective transformation on images of complex components such as curved guide rails to ensure that defect shapes are not distorted;

[0055] Noise suppression: Nonlocal mean filtering is used to remove Gaussian noise and salt-and-pepper noise while preserving defect edge details;

[0056] Data archiving: Data is categorized and stored according to component type (spindle / guide rail / motor) and preliminary defect type (surface / internal / stress) to establish an original sample library.

[0057] The specific processing method of the multispectral fusion algorithm is as follows: Based on the spatiotemporal synchronization parameters, the visible light, near-infrared light, and ultraviolet light images are aligned at the pixel level to ensure accurate matching of defect locations in different modal images; surface defect contour features are extracted from the visible light image, internal crack features are extracted from the near-infrared light image, and stress concentration fluorescence information is extracted from the ultraviolet light image; fusion weights are dynamically set according to the defect type. For example, for samples dominated by surface defects, the weight of the visible light image is increased; for internal defects, the weight of the near-infrared light image is enhanced; when stress concentration exists, the weight of the ultraviolet light image is highlighted. Through matrix weighting operations, the feature matrices of the three types of images are superimposed according to their weights to generate an enhanced sample containing surface, internal, and stress features; histogram equalization and contrast stretching techniques are used to optimize the visual effect of the fused image and highlight defect details.

[0058] It can fully leverage the advantages of different modal detection methods, fuse multi-source information, and generate enhanced samples containing surface, internal, and stress features. This enriches the feature dimensions of the samples, and the dynamic weight allocation makes the fused samples more prominent in terms of key defect features, improving sample quality, optimizing visual effects, making defect details more obvious, and facilitating model learning.

[0059] The process of constructing the 3D model sample is as follows: A calibration procedure is used to establish the transformation relationship between the coordinate system of the 3D laser scanning device and the coordinate system of the ultrasonic testing module, ensuring that the two modal data are on a unified spatial reference. Using a feature point matching algorithm, identical geometric feature points are identified in the surface point cloud data acquired by the 3D laser scanning device and the tomographic image data from the ultrasonic testing module. Data registration is performed based on these feature points to eliminate deviations caused by differences in acquisition position and angle. Voxel fusion technology is used to fuse the registered scanning data from the 3D laser scanning device and the data from the ultrasonic testing module, filling and integrating surface and internal defect information in three-dimensional space to generate a 3D model sample containing the surface morphology, dimensional information, and internal defect distribution of the component.

[0060] The specific fusion process of voxel fusion technology is as follows:

[0061] Voxelization: The 3D laser scanning data and ultrasonic detection data are voxelized separately, that is, the three-dimensional space is divided into regular cubic voxels, each voxel corresponds to a position in the space and is assigned a corresponding attribute value;

[0062] Weighting: Based on the characteristics of the data source, weights are assigned to data of different modalities. In the expression of surface features, the weight of 3D laser scanning data is set to 0.7, and the weight of ultrasonic testing data is 0.3. In the expression of internal defect information, the weight of ultrasonic testing data is increased to 0.8, and the weight of 3D laser scanning data is adjusted to 0.2. Through this differentiated weighting, it is ensured that the fused data can truly reflect the characteristics of the component.

[0063] Fusion calculation: Traverse all voxels. For each voxel in the overlapping region, according to the pre-set weights, use a weighted average algorithm to fuse the attribute values ​​corresponding to the 3D laser scanning data and the ultrasonic detection data to generate new attribute values.

[0064] Hole filling and smoothing: During the fusion process, check for missing data or void areas, fill the voids using interpolation algorithms or based on information from adjacent voxels, and smooth the fused voxel data to eliminate discontinuities or abrupt changes caused by data fusion, making the generated 3D model smoother and more natural.

[0065] For complex components such as curved guide rails, the surface three-dimensional geometric information obtained by 3D laser scanning and the internal defect data of ultrasonic testing are integrated into a complete 3D model sample through voxel fusion technology. This enriches the spatial dimensional features of defects, provides more comprehensive information for defect detection of complex components, makes up for the inadequacy of two-dimensional images in expressing complex structures, and improves the accuracy and reliability of detection.

[0066] The hybrid sample library is generated as follows: Based on the defect feature consistency standard, that is, matching and judging according to defect type and defect size, virtual samples are matched with original defect samples. If they match, the virtual sample is retained. For example, for scratch defects, if the scratch length of the original defect sample is in the range of 1-10mm, it is necessary to ensure that the scratch length of the virtual sample also covers this range, and samples that do not meet the requirements are removed. According to the scene diversity standard, virtual samples with different lighting conditions (such as top light, side light), different wear levels (such as light threshold, heavy threshold), and different component states (new equipment, old equipment) are obtained. The samples are classified and statistically analyzed to ensure that the original defect samples account for 20%, the enhanced samples and 3D model samples each account for 15%, and the virtual samples account for 50%, and the total number of samples is not less than 100,000. If the proportion requirement is not met or the total number is insufficient, the previous steps are returned to adjust the screening criteria or supplement the samples until the conditions are met, and finally the hybrid sample library is formed.

[0067] By merging real and virtual samples and filtering them according to the criteria of consistency of defect features and diversity of scenarios, a hybrid sample library is formed. The samples cover a variety of scenarios, and the reasonable distribution of sample ratios fully leverages the advantages of real and virtual samples, providing sufficient and high-quality data for model training.

[0068] The process of training the defect classification model is as follows:

[0069] A deep learning backbone network is used to extract features from mixed samples, capturing the shape (area, perimeter), texture (gray-level co-occurrence matrix), and multimodal fusion features (surface + internal defect correlation) of defects.

[0070] The model is trained using the original sample database, and the defect categories and confidence scores are output. A 10% mixed sample is used for testing to ensure that the model's accuracy in identifying common defects is ≥90%.

[0071] A deep learning backbone network is used to extract defect shape, texture, and multimodal fusion features. An end-to-end classification model is trained and tested with 10% mixed samples. The powerful feature extraction capability of deep learning can effectively capture complex defect features. The end-to-end training method simplifies the model building process and improves training efficiency.

[0072] Please see Figure 2As shown in the figure, this embodiment of the invention provides a defect visual classification and detection system for CNC systems, comprising a database, a multispectral camera system, a scanning detection unit, a 3D model building unit, a virtual sample generation unit, a sample integration unit, and a model optimization unit. The database is used for acquiring and storing original defect samples. The multispectral camera system is used for acquiring and fusing multimodal images to form enhanced samples. The scanning detection unit is used to acquire three-dimensional geometric information using a scanning device. The 3D model building unit is used to construct 3D model samples based on the three-dimensional geometric information. The virtual sample generation unit is used to generate virtual samples. The sample integration unit is used to merge the original defect samples, enhanced samples, 3D model samples, and virtual samples to form a hybrid sample library. The model optimization unit is used to train a defect classification model based on the hybrid sample library.

[0073] The multispectral camera system includes an imaging unit, a control module, and a data processing module. The imaging unit is used to detect and image target objects from different spectral dimensions to acquire visual information. The imaging unit includes a visible light imaging module, a near-infrared light imaging module, and an ultraviolet light imaging module. The control module is used to coordinate the synchronous triggering of the three imaging modules, adjust exposure parameters, and control the image acquisition timing. The data processing module has a built-in multispectral image fusion algorithm for weighted fusion of visible light, near-infrared light, and ultraviolet light images.

[0074] The data processing module is also used for the standardization processing of multimodal images;

[0075] The scanning and detection unit includes a 3D laser scanning device and an ultrasonic detection module. The 3D laser scanning device is used to acquire three-dimensional geometric information of the component surface, and the ultrasonic detection module is used to emit and receive ultrasonic waves to detect defects inside the component.

[0076] By using a multispectral camera system to acquire and fuse multimodal images, combined with 3D laser scanning and ultrasonic testing, sample information can be obtained from multiple dimensions, including surface geometry, internal defects, and multispectral vision. Virtual sample generation and sample integration further expand the sample library, providing rich and diverse data for defect classification model training. This helps improve the accuracy, comprehensiveness, and robustness of visual defect classification and detection in CNC systems, overcomes the limitations of single detection methods, and achieves more efficient and accurate defect detection and classification.

[0077] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A defect visual classification and detection method for CNC systems, characterized in that, include: S1: Obtain original defect samples covering CNC system components; S2: Multimodal image acquisition, which uses a multispectral camera system to acquire multidimensional images of CNC system components, including visible light, near-infrared light and ultraviolet light images. The acquired multimodal images are then standardized to establish an original sample library. S3: Based on the original defect samples, a multispectral fusion algorithm is used to perform weighted fusion of visible light, near-infrared light and ultraviolet light images in the original sample library to generate enhanced samples containing multiple features; S4: Use 3D laser scanning equipment to obtain three-dimensional geometric information of the component surface, and at the same time, use ultrasonic testing module to emit and receive ultrasonic waves to detect defects inside the component. Register and fuse the 3D laser scanning data and ultrasonic testing data to construct a 3D model sample containing surface and internal defect information. S5: Based on original defect samples, enhanced samples, and 3D model samples, a physical model of the component is constructed using finite element analysis to simulate the evolution of defects under different working conditions and generate virtual samples. An intelligent agent is introduced into the physical model environment to simulate dynamic defects by adjusting equipment parameters and generate defect evolution sequence samples. Using the original defect samples as training data, a generative adversarial network is built. The generator learns the distribution characteristics of real defects, and the discriminator optimizes the generator by comparing it with real samples, and finally generates virtual samples. S6: Merge the original defect samples, enhanced samples, 3D model samples and virtual samples to form a hybrid sample library; S7: Based on a hybrid sample library, a defect classification model is trained, and the trained model is deployed to an industrial inspection system to perform real-time image acquisition and analysis of CNC system components on the production line, outputting defect type and location, and periodically collecting misjudged samples that occur during the inspection process to supplement the original defect samples.

2. The defect visual classification and detection method for CNC systems according to claim 1, characterized in that: The specific processing method of the multispectral fusion algorithm is as follows: Pixel-level alignment of visible light, near-infrared light, and ultraviolet light images is performed based on spatiotemporal synchronization parameters. Surface defect contour features were extracted from visible light images, internal crack features were extracted from near-infrared light images, and stress concentration fluorescence information was extracted from ultraviolet light images. The fusion weights are dynamically set according to the defect type. Through matrix weighting operations, the feature matrices of the three types of images are superimposed according to the weights to generate an enhanced sample containing surface, internal and stress features. Histogram equalization and contrast stretching techniques are used to optimize the visual effect of the fused image.

3. The defect visual classification and detection method for CNC systems according to claim 1, characterized in that: The process of constructing the 3D model sample is as follows: The transformation relationship between the coordinate system of the 3D laser scanning equipment and the coordinate system of the ultrasonic testing module is established through a calibration procedure to ensure that the two modal data are in a unified spatial reference. By using a feature point matching algorithm, the same geometric feature points are identified in the surface point cloud data acquired by the 3D laser scanning device and the tomographic image data of the ultrasonic detection module. Data registration is then performed based on these feature points to eliminate deviations caused by differences in acquisition position and angle. By employing voxel fusion technology, the scanning data from the registered 3D laser scanning device is fused with the data acquired by the ultrasonic testing module. This process fills in and integrates surface and internal defect information in three-dimensional space, generating a 3D model sample that includes the surface morphology, dimensional information, and internal defect distribution of the component.

4. The defect visual classification and detection method for CNC systems according to claim 1, characterized in that: The hybrid sample library is generated in the following way: Based on the defect feature consistency standard, the virtual sample is matched with the original defect sample. If the match is consistent, the virtual sample is retained. Based on the standard of scene diversity, virtual samples with different lighting conditions, different wear levels, and different component states are obtained; The selection criteria are adjusted based on the set proportions of the original defect samples, enhanced samples, 3D model samples, and virtual samples, and the total number of samples is not less than the set value. If the proportion requirement is not met or the total number is insufficient, the previous steps are returned to adjust the screening criteria or supplement the samples until the conditions are met, and finally a mixed sample library is formed.

5. The defect visual classification and detection method for CNC systems according to claim 1, characterized in that: The standardization process includes: Geometric correction: Eliminates lens distortion through calibration procedures and performs perspective transformation on the image; Noise suppression: Nonlocal mean filtering is used to remove Gaussian noise and salt-and-pepper noise while preserving defect edge details; Data archiving: Store data by component type and establish an original sample library.

6. The defect visual classification and detection method for CNC systems according to claim 1, characterized in that: The process of training the defect classification model is as follows: A deep learning backbone network is used to extract features from mixed samples, capturing the shape, texture, and multimodal fusion features of defects; The model is trained using the original sample library, and the defect categories and confidence levels are output. The model is then tested using a set mixed sample to ensure that the model's accuracy in identifying defects reaches the preset threshold standard.

7. A defect visual classification and detection system for CNC systems, used in the application of the defect visual classification and detection method according to any one of claims 1-7, characterized in that: The system comprises a database, a multispectral camera system, a scanning and detection unit, a 3D model building unit, a virtual sample generation unit, a sample integration unit, and a model optimization unit. The database is used for acquiring and storing original defect samples. The multispectral camera system is used for acquiring and fusing multimodal images to form enhanced samples. The scanning and detection unit is used to acquire three-dimensional geometric information using scanning equipment. The 3D model building unit is used to construct 3D model samples based on the three-dimensional geometric information. The virtual sample generation unit is used to generate virtual samples. The sample integration unit is used to merge original defect samples, enhanced samples, 3D model samples, and virtual samples to form a hybrid sample library. The model optimization unit is used to train a defect classification model based on the hybrid sample library.

8. A defect visual classification and detection system for CNC systems according to claim 7, characterized in that: The multispectral camera system includes an imaging unit, a control module, and a data processing module. The imaging unit is used to detect and image target objects from different spectral dimensions to acquire visual information. The imaging unit includes a visible light imaging module, a near-infrared light imaging module, and an ultraviolet light imaging module. The control module is used to coordinate the synchronous triggering of the three imaging modules, adjust exposure parameters, and control the image acquisition timing. The data processing module has a built-in multispectral image fusion algorithm for weighted fusion of visible light, near-infrared light, and ultraviolet light images.

9. A defect visual classification and detection system for CNC systems according to claim 8, characterized in that: The data processing module is also used for the standardization processing of multimodal images.

10. A defect visual classification and detection system for CNC systems according to claim 7, characterized in that: The scanning and detection unit includes a 3D laser scanning device and an ultrasonic detection module. The 3D laser scanning device is used to acquire three-dimensional geometric information of the component surface, and the ultrasonic detection module is used to emit and receive ultrasonic waves to detect defects inside the component.

Images