Three-dimensional mold surface defect real-time detection system based on deep learning

Abstract

The invention relates to the technical field of defect detection, and particularly discloses a three-dimensional mold surface defect real-time detection system based on deep learning, and the system comprises a data collection module which divides a mold region through a data recording unit, records point cloud and image data at each moment, and generates a historical data set; the preprocessing unit generates a geometric defect positioning curve by using the historical point cloud data, and generates an RGB defect positioning curve based on the image data; the data analysis module is used for determining a safety interval according to the geometric defect positioning curve, counting the defect frequency of each region and identifying a high-frequency defect region; positioning an RGB related area through the RGB defect positioning curve; the point cloud density correction module is used for determining an RGB sensitive area in combination with the RGB related area and setting an initial point cloud density; during detection, the RGB value of the RGB sensitive area is monitored in real time, and other areas are scanned at the initial density; after the safety interval is finished, the point cloud density of each area is dynamically adjusted according to the defect frequency, and the detection efficiency is optimized.

Description

Technical Field

[0001] The present invention relates to the field of defect detection technology, and in particular to a real-time detection system for three-dimensional mold surface defects based on deep learning. Background Art

[0002] In existing technology systems, surface inspection of 3D molds typically involves acquiring point cloud data from the mold using specialized measuring equipment or techniques. After acquiring the point cloud data, the 3D mold is modeled and the original model is compared with the real-time model. If deviations exceed a pre-defined tolerance, the mold is identified as having defects, such as surface wear, deformation, or scratches.

[0003] This modeling process involves more than simply piling up data; it requires the application of advanced mathematical models and algorithms to process and analyze massive amounts of point cloud data. Preprocessing operations such as filtering, noise reduction, and splicing remove noise and abnormal data. The discrete point cloud data is then fitted into a continuous surface model to accurately restore the true shape and surface features of the 3D mold.

[0004] However, in real-world production environments, 3D mold surface inspection often needs to be completed within a very short timeframe to identify problems and make adjustments promptly, ensuring production continuity and consistent product quality. However, high-precision 3D reconstruction algorithms are computationally complex, consuming significant computing resources and time. Even in environments equipped with high-performance computing equipment, these algorithms struggle to meet real-time requirements. Summary of the Invention

[0005] The purpose of the present invention is to provide a real-time detection system for three-dimensional mold surface defects based on deep learning to solve the following technical problems.

[0006] The purpose of the present invention can be achieved through the following technical solutions:

[0007] The deep learning-based real-time detection system for 3D mold surface defects includes:

[0008] Data collection module: a data recording unit and a pre-processing unit. The data recording unit divides the mold into several areas and is used to record the point cloud data and image data of the mold at each moment during its use cycle. The historical point cloud data and historical image data of the mold are obtained through the data recording unit.

[0009] The pre-processing unit generates a geometric defect location curve based on the historical point cloud data, and generates an RGB defect location curve based on the historical image data;

[0010] Data analysis module: determines the safe range based on the geometric defect location curve, and obtains the defect frequency and high-frequency defect area of ​​each area of ​​the mold; determines the RGB related area based on the RGB defect location curve;

[0011] Point cloud density correction module: obtains the RGB sensitive area based on the RGB related area and sets the initial density of the point cloud; when monitoring the surface defects of the mold, monitors the RGB value of the RGB sensitive area in real time, and monitors the point cloud data of the remaining areas in real time at the initial density within a safe range; and after the end of the safe range, adjusts the point cloud density of each area of ​​the mold according to the defect frequency of each area of ​​the mold.

[0012] As a further solution of the present invention: the usage cycle is the time period between the time when the mold is put into use and the time when it ends use, the time when the mold is put into use is the time when the mold is put into use, and the time when the mold ends use is the time when the mold ends use; and according to a preset time interval, a number of moments are equally selected within the usage cycle; the image data includes captured images of each surface of the mold.

[0013] As a further solution of the present invention: the process of obtaining the point cloud model and image data includes:

[0014] Establish a three-dimensional coordinate system, place the mold in the three-dimensional coordinate system, obtain the center point of each area of ​​the mold, and number each center point; and obtain the three-dimensional coordinates corresponding to each center point in the three-dimensional coordinate system to obtain point cloud data of the mold; the image data includes the RGB values ​​of each area of ​​the mold.

[0015] As a further solution of the present invention: the process of generating the geometric defect location curve includes:

[0016] When the surface of the mold is found to be free of defects, the three-dimensional coordinates of the center points of each area are obtained and recorded as standard coordinates; for the mold at any moment, the distance between the three-dimensional coordinates of each center point and the standard coordinates is obtained and recorded as the distance difference; a rectangular coordinate system is established with the number of the center point as the horizontal coordinate and the distance difference as the vertical coordinate, and the center points of each number and their corresponding distance differences are fitted in the rectangular coordinate system to obtain a geometric defect positioning curve.

[0017] As a further solution of the present invention: the process of determining the safety interval includes:

[0018] Obtaining the slope of each point on the geometric defect location curve at each moment during the use period, obtaining the accumulation of the slopes at each point, and obtaining a fluctuation value of the geometric defect location curve; recording the moment corresponding to the geometric defect location curve when the fluctuation value is not zero for the first time, and recording it as the defect moment;

[0019] The defect moments of all the abrasive tools during their service life are obtained to obtain a defect moment set, and the minimum value in the defect moment set is selected and recorded as the safety cutoff moment; the safety interval is obtained from the commissioning moment and the safety cutoff moment.

[0020] As a further solution of the present invention: the process of obtaining the defect frequency of each area of ​​the mold includes:

[0021] Obtain a geometric defect positioning curve of the mold at the end of use, which is recorded as the end-time curve. According to the historical point cloud data, obtain the end-time curves of all molds. For any numbered center point, obtain the number of end-time curves whose ordinate value of the center point is not 0 in all the end-time curves, and obtain the defect frequency f=n / N of the area corresponding to the center point, where n is the number of end-time curves whose ordinate value is not 0, and N is the total number of all end-time curves.

[0022] As a further solution of the present invention: the process of determining the RGB related area includes:

[0023] Obtain the moment corresponding to the geometric defect location curve whose ordinate value is not 0 for the first time in all geometric defect location curves of the high-frequency defect area in the usage cycle, and record it as the first defect moment; obtain the RGB defect location curve at the first defect moment, record it as the first curve, obtain the number of the center point of the high-frequency defect area, obtain the RGB value corresponding to the number on the first curve, and record it as the first RGB value; and obtain the RGB defect location curve at the moment before the first defect moment, record it as the second curve, obtain the RGB value corresponding to the number on the second curve, and record it as the second RGB value; if the first RGB value is not equal to the second RGB value, the high-frequency defect area is an RGB-related area.

[0024] As a further solution of the present invention: if there are several adjacent RGB-related areas, the areas occupied by these several RGB-related areas are recorded as an area set. If the area occupied by the area set exceeds a preset area threshold, the area set is recorded as an RGB-sensitive area.

[0025] Beneficial effects of the present invention:

[0026] This method dynamically adjusts the point cloud acquisition density of different areas based on defect frequencies and RGB-sensitive areas derived from historical data. High-density scanning is used in high-frequency defect areas or RGB-sensitive areas, while density is reduced in other areas to minimize computational effort and achieve a balance between real-time detection and resource consumption. During the initial mold use (within the safe range), only basic density monitoring is required, with targeted adjustments made later to avoid high-load computation throughout the entire process. High-precision scanning is performed only in critical areas (RGB-sensitive areas and high-frequency defect areas), reducing reliance on high-performance hardware. Furthermore, for areas where RGB values ​​change significantly when defects occur (such as scratches causing changes in reflection), this method directly uses image monitoring to avoid unnecessary high-density point cloud scanning. High-density point cloud detection is retained only in areas sensitive to geometric deformation (such as high-frequency defect areas), achieving precise distribution of computational load. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The present invention will be further described below with reference to the accompanying drawings.

[0028] Figure 1 It is a schematic diagram of the method of the real-time detection system of three-dimensional mold surface defects based on deep learning of the present invention. DETAILED DESCRIPTION

[0029] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0030] See also Figure 1 As shown, the present invention is a real-time detection system for three-dimensional mold surface defects based on deep learning, comprising:

[0031] Data collection module: a data recording unit and a pre-processing unit. The data recording unit divides the mold into several areas and is used to record the point cloud data and image data of the mold at each moment during its use cycle. The historical point cloud data and historical image data of the mold are obtained through the data recording unit.

[0032] The pre-processing unit generates a geometric defect location curve based on the historical point cloud data, and generates an RGB defect location curve based on the historical image data;

[0033] In a preferred embodiment of the present invention, the use cycle is a period of time between the time the mold is put into use and the time it ends use, wherein the time it is put into use is the time when the mold is put into use, and the time it ends use is the time when the mold ends use; and according to a preset time interval, a plurality of moments are equally selected within the use cycle to obtain; and the image data includes captured images of each surface of the mold;

[0034] In a preferred embodiment of the present invention, the process of obtaining the point cloud model and image data includes:

[0035] Establishing a three-dimensional coordinate system, placing the mold in the three-dimensional coordinate system, obtaining the center point of each area of ​​the mold, numbering each center point, and obtaining the three-dimensional coordinates corresponding to each center point in the three-dimensional coordinate system to obtain point cloud data of the mold; the image data includes RGB values ​​of each area of ​​the mold;

[0036] The mold surface is divided into several areas, and each area is assigned a center point number to establish a spatial mapping relationship in a three-dimensional coordinate system. The three-dimensional coordinates (x, y, z) of the center point of each area are obtained using a 3D scanner (such as lidar or structured light), and the geometric deformation is recorded to obtain point cloud data. The RGB image of the mold surface is collected using an industrial camera, and the color or texture changes are recorded to obtain image data.

[0037] As a preferred embodiment of the present invention, the process of generating the geometric defect location curve includes:

[0038] When the surface of the mold is found to be free of defects, the three-dimensional coordinates of the center points of each region are obtained and recorded as standard coordinates; for the mold at any moment, the distance between the three-dimensional coordinates of each center point and the standard coordinates is obtained and recorded as the distance difference; a rectangular coordinate system is established with the number of the center point as the abscissa and the distance difference as the ordinate, and the center points of each number and their corresponding distance differences are fitted in the rectangular coordinate system to obtain a geometric defect location curve;

[0039] Based on the standard coordinates when there is no defect, the deviation (distance difference) between the real-time point cloud coordinates and the standard value is calculated, and a curve is fitted to quantify the degree of deformation;

[0040] As a preferred embodiment of the present invention, the process of generating the RGB defect location curve includes:

[0041] According to the historical image data, the RGB value of each area of ​​the mold surface is obtained, recorded as the RGB value of the center point of the area, and the center point of each number and its corresponding RGB value are fitted to obtain an RGB defect location curve;

[0042] Extract the time-varying curve of the RGB values ​​of each region to capture color anomalies (such as color difference caused by oxidation and scratches);

[0043] Data analysis module: determines the safe range based on the geometric defect location curve, and obtains the defect frequency and high-frequency defect area of ​​each area of ​​the mold; determines the RGB related area based on the RGB defect location curve;

[0044] As a preferred embodiment of the present invention, the process of determining the safety interval includes:

[0045] Obtaining the slope of each point on the geometric defect location curve at each moment during the use period, obtaining the accumulation of the slopes at each point, and obtaining a fluctuation value of the geometric defect location curve; recording the moment corresponding to the geometric defect location curve when the fluctuation value is not zero for the first time, and recording it as the defect moment;

[0046] Obtaining defect moments of all the abrasive tools during their service life to obtain a defect moment set, and selecting the minimum value in the defect moment set as the safety cutoff moment; obtaining a safety interval based on the commissioning moment and the safety cutoff moment;

[0047] The cumulative fluctuation value of mold deformation is quantified by the slope change of each point on the geometric defect location curve. The moment when a non-zero fluctuation value first appears is marked as the defect starting time. The earliest defect time of all molds is statistically analyzed based on historical data to determine the safety interval.

[0048] In a preferred embodiment of the present invention, the process of obtaining the defect frequency of each area of ​​the mold includes:

[0049] Obtain a geometric defect location curve of the mold at the end of use, recorded as the end-time curve; obtain the end-time curves of all molds based on the historical point cloud data, and for any numbered center point, obtain the number of end-time curves whose ordinate value of the center point is not 0 among all the end-time curves, and obtain the defect frequency f=n / N of the area corresponding to the center point, where n is the number of end-time curves whose ordinate value is not 0, and N is the total number of all end-time curves;

[0050] Based on historical data, the number of defects in each area at the end of use is counted and the defect frequency is calculated. Areas where the frequency of high-frequency defects exceeds the threshold need to be monitored in a focused manner.

[0051] In a preferred embodiment of the present invention, the process of determining the high-frequency defect area includes:

[0052] Setting a frequency threshold, recording the area where the defect frequency exceeds the frequency threshold as a high-frequency defect area; otherwise, recording it as a low-frequency defect area;

[0053] In a preferred embodiment of the present invention, the process of determining the RGB related area includes:

[0054] Obtain the moment corresponding to the geometric defect location curve at which the ordinate value is not 0 for the first time among all geometric defect location curves of the high-frequency defect area during the usage period, and record it as the first defect moment; obtain the RGB defect location curve at the first defect moment, record it as the first curve, obtain the number of the center point of the high-frequency defect area, obtain the RGB value corresponding to the number on the first curve, and record it as the first RGB value; obtain the RGB defect location curve at the moment before the first defect moment, record it as the second curve, and obtain the RGB value corresponding to the number on the second curve, and record it as the second RGB value; if the first RGB value is not equal to the second RGB value, then the high-frequency defect area is an RGB-correlated area;

[0055] Compare the RGB values ​​of the high-frequency defect area at the first defect moment and the previous moment. If the RGB value changes suddenly, it is determined to be an RGB-related area.

[0056] Point cloud density correction module: derives an RGB sensitive area based on the RGB related area and sets the initial density of the point cloud; when monitoring surface defects of the mold, monitors the RGB values ​​of the RGB sensitive area in real time, and monitors the point cloud data of the remaining areas in real time at the initial density within a safe range; and after the safety range ends, adjusts the point cloud density of each area of ​​the mold based on the defect frequency of each area of ​​the mold;

[0057] It should be noted that image monitoring is used for RGB-sensitive areas, while point cloud monitoring is retained for geometrically sensitive areas. The hybrid detection architecture achieves data synchronization through the PCIe bus. The point cloud density is dynamically adjusted, with the initial density for the entire area within the safe zone and increased for high-frequency areas outside the safe zone. A layered architecture is also used, with the front end using FPGA for fast image RGB value comparison and the back end using GPU for accelerated point cloud feature extraction.

[0058] As a preferred embodiment of the present invention, if there are several adjacent RGB-related regions, the areas occupied by these several RGB-related regions are recorded as a region set. If the area occupied by the region set exceeds a preset area threshold, the region set is recorded as an RGB-sensitive region.

[0059] As a preferred embodiment of the present invention, the initial density of the point cloud is a preset number of point clouds per unit area of ​​the mold surface;

[0060] In a preferred embodiment of the present invention, the process of real-time monitoring of the RGB value of the RGB sensitive area includes:

[0061] If the RGB value of a region within the RGB sensitive region changes compared to the previous moment, a defect occurs on the surface of the mold in the region;

[0062] In a preferred embodiment of the present invention, the process of adjusting the point cloud density of each area of ​​the mold includes:

[0063] The initial density is recorded as Id, and the point cloud density of the area is obtained , where k is the preset correction coefficient and k>0;

[0064] The above is a detailed description of an embodiment of the present invention. However, the content described is only a preferred embodiment of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A real-time detection system for 3D mold surface defects based on deep learning, characterized by: include: Data collection module: a data recording unit and a pre-processing unit. The data recording unit divides the mold into several areas and is used to record the point cloud data and image data of the mold at each moment during its use cycle. The historical point cloud data and historical image data of the mold are obtained through the data recording unit. The pre-processing unit generates a geometric defect location curve based on the historical point cloud data, and generates an RGB defect location curve based on the historical image data; Data analysis module: determines the safe range based on the geometric defect location curve, and obtains the defect frequency and high-frequency defect area of ​​each area of ​​the mold; determines the RGB related area based on the RGB defect location curve; Point cloud density correction module: obtains the RGB sensitive area based on the RGB related area and sets the initial density of the point cloud; when monitoring the surface defects of the mold, monitors the RGB value of the RGB sensitive area in real time, and monitors the point cloud data of the remaining areas in real time at the initial density within a safe range; and after the end of the safe range, adjusts the point cloud density of each area of ​​the mold according to the defect frequency of each area of ​​the mold.

2. The deep learning-based real-time detection system for three-dimensional mold surface defects according to claim 1 is characterized in that: The usage cycle is the time period between the time the mold is put into use and the time it ends use, the time it is put into use is the time the mold is put into use, and the time it ends use is the time the mold ends use; and according to a preset time interval, several moments are equally selected within the usage cycle; the image data includes captured images of each surface of the mold.

3. The deep learning-based real-time detection system for three-dimensional mold surface defects according to claim 1 is characterized in that: The process of obtaining the point cloud model and image data includes: Establish a three-dimensional coordinate system, place the mold in the three-dimensional coordinate system, obtain the center point of each area of ​​the mold, and number each center point; and obtain the three-dimensional coordinates corresponding to each center point in the three-dimensional coordinate system to obtain point cloud data of the mold; the image data includes the RGB values ​​of each area of ​​the mold.

4. The deep learning-based real-time detection system for three-dimensional mold surface defects according to claim 1 is characterized in that: The generation process of the geometric defect location curve includes: When the surface of the mold is found to be free of defects, the three-dimensional coordinates of the center points of each area are obtained and recorded as standard coordinates; for the mold at any moment, the distance between the three-dimensional coordinates of each center point and the standard coordinates is obtained and recorded as the distance difference; a rectangular coordinate system is established with the number of the center point as the horizontal coordinate and the distance difference as the vertical coordinate, and the center points of each number and their corresponding distance differences are fitted in the rectangular coordinate system to obtain a geometric defect positioning curve.

5. The deep learning-based real-time detection system for three-dimensional mold surface defects according to claim 2 is characterized in that: The process of determining the safety interval includes: Obtaining the slope of each point on the geometric defect location curve at each moment during the use period, obtaining the accumulation of the slopes at each point, and obtaining a fluctuation value of the geometric defect location curve; recording the moment corresponding to the geometric defect location curve when the fluctuation value is not zero for the first time, and recording it as the defect moment; The defect moments of all the abrasive tools during their service life are obtained to obtain a defect moment set, and the minimum value in the defect moment set is selected and recorded as the safety cutoff moment; the safety interval is obtained from the commissioning moment and the safety cutoff moment.

6. The deep learning-based real-time detection system for three-dimensional mold surface defects according to claim 1 is characterized in that: The process of obtaining the defect frequency of each area of ​​the mold includes: Obtain a geometric defect positioning curve of the mold at the end of use, which is recorded as the end-time curve. According to the historical point cloud data, obtain the end-time curves of all molds. For any numbered center point, obtain the number of end-time curves whose ordinate value of the center point is not 0 in all the end-time curves, and obtain the defect frequency f=n / N of the area corresponding to the center point, where n is the number of end-time curves whose ordinate value is not 0, and N is the total number of all end-time curves.

7. The deep learning-based real-time detection system for three-dimensional mold surface defects according to claim 1 is characterized in that: The process of determining the RGB related area includes: Obtain the moment corresponding to the geometric defect location curve whose ordinate value is not 0 for the first time in all geometric defect location curves of the high-frequency defect area in the usage cycle, and record it as the first defect moment; obtain the RGB defect location curve at the first defect moment, record it as the first curve, obtain the number of the center point of the high-frequency defect area, obtain the RGB value corresponding to the number on the first curve, and record it as the first RGB value; and obtain the RGB defect location curve at the moment before the first defect moment, record it as the second curve, obtain the RGB value corresponding to the number on the second curve, and record it as the second RGB value; if the first RGB value is not equal to the second RGB value, the high-frequency defect area is an RGB-related area.

8. The deep learning-based real-time detection system for three-dimensional mold surface defects according to claim 1 is characterized in that: If there are several adjacent RGB related areas, the areas occupied by these several RGB related areas are recorded as an area set. If the area occupied by the area set exceeds a preset area threshold, the area set is recorded as an RGB sensitive area.

Images