3D non-destructive detection method for precise metalwork and based on embedded ai platform

Abstract

A 3D non-destructive detection method for precise metalwork. The detection method comprises: fixing a laser line scanning sensor on a support; placing a conveyor belt below the laser line scanning sensor, and placing precise metalwork on the conveyor belt; providing an embedded AI platform to receive image data of the precise metalwork that is transmitted by the laser line scanning sensor; and the laser line scanning sensor collecting data, i.e. laser line scanning data, and the embedded AI platform performing depth map generation on data transmitted by the laser line scanning sensor, and performing defect detection on the basis of the depth map, wherein the precise metalwork should be placed in the middle region of a laser line; defects comprise warping and cracks; the embedded AI platform is deployed at a PC end; and the embedded AI platform is loaded with a 3D non-destructive detection model for precise metalwork.

Description

A precision metalworking 3D nondestructive testing method based on an embedded AI platform Technical Field

[0001] The present invention belongs to the field of industrial detection technology, and in particular relates to a precision metalworking 3D non-destructive detection method based on an embedded AI platform. Background Art

[0002] Precision metalworking, such as metal domes, is widely used in the electronics industry. Circuit boards are a crucial component in electronic product manufacturing, and metal domes are used to implement various circuit board functions, such as jumper functionality, connectors, and conductivity. These require extremely high precision, with the overall dome height tolerance within 0.1mm.

[0003] Summary of the Invention

[0004] One of the embodiments of the present disclosure is a precision metalworking 3D non-destructive testing method, which includes:

[0005] Fix the laser line scan sensor on the bracket;

[0006] A conveyor belt is arranged below the laser line scanning sensor, and the precision metalworking tool is placed on the conveyor belt;

[0007] An embedded AI platform is configured to receive the image data of precision metalworking transmitted by the laser line scan sensor;

[0008] The laser line scan sensor collects laser line scan data, and the embedded AI platform generates a depth map based on the data transmitted by the laser line scan sensor and performs defect detection based on the depth map.

[0009] The embedded AI platform is loaded with a precision metalworking 3D non-destructive testing model.

[0010] The method for constructing the detection model includes: collecting height data of precision metalworking to generate a depth map; performing threshold segmentation on the depth map; performing bicubic interpolation super-resolution processing, and using an AI model for training. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily apparent by reading the following detailed description with reference to the accompanying drawings, in which several embodiments of the present invention are shown by way of example and not limitation, in which:

[0012] FIG1 is a schematic diagram of a method for constructing a precision metalworking 3D non-destructive testing model according to one embodiment of the present invention.

[0013] FIG2 is a schematic diagram of a method for constructing a precision metalworking 3D non-destructive testing model according to one embodiment of the present invention. DETAILED DESCRIPTION

[0014] Inspection solutions for precision metalworking include 2D camera solutions, 3D line scan solutions, and PC-based 3D visual inspection with AI. The 2D camera solution primarily uses an LED light source to illuminate the surface to be inspected, then uses a 2D camera at a fixed position to take photos and upload the images to a computer, providing the results of defect detection items. The 3D line scan solution uses a laser line scan sensor to obtain height data, generally generating a point cloud image, and using point cloud tools to configure the algorithm to obtain the inspection results. The PC-based 3D visual inspection method with AI combines a laser line scan sensor and AI. This method is generally deployed on the PC side, where AI inspection is performed.

[0015] Through testing of various solutions, it can be found that the 2D camera solution has the following disadvantages:

[0016] 1. 2D camera inspection requires a strict light source to provide uniform lighting levels, which places high demands on the technicians' lighting design. This is especially true for shrapnel of different shapes, where the need to adjust the light source angle greatly increases the complexity.

[0017] 2. There is reflection on the surface of precision metalworking, which will result in poor imaging effect in the camera.

[0018] 3. The camera image is a 2D image that does not contain height information and cannot detect height defects.

[0019] The shortcomings of the 3D point cloud algorithm solution include:

[0020] 1. The point cloud tool configuration algorithm is used, which has high user learning costs.

[0021] 2. The development of point cloud tools is difficult.

[0022] The disadvantages of the PC-based 3D visual inspection method with AI are that the PC requires a high-performance computer to support AI's large-model recognition, which is costly and takes up a large installation space.

[0023] To address the above issues, the present disclosure proposes a non-destructive detection method for precision metalworking defects, which utilizes a high-precision laser line scanning sensor to generate a depth map from 3D scanning data for AI analysis on an embedded platform to solve the above technical problems.

[0024] According to one or more embodiments, a method for 3D nondestructive testing of precision metalworking involves a laser line scan sensor and an embedded AI platform. The precision metalworking is placed on a conveyor belt and scanned in conjunction with the control of the conveyor belt. The sensor is used to collect laser line scan data, and the embedded AI platform generates a depth map of the data transmitted by the sensor. Based on the depth map, the high-computing power embedded AI platform performs model analysis to detect defects. The following is a further detailed description of the embodiments of the present disclosure.

[0025] The laser line scan sensor is mounted on a fixed bracket, with the sensor head positioned above the conveyor belt. The scanning line is always aligned with the precision metalwork, with the laser line perpendicular to the conveyor belt. The precision metalwork should be positioned in the center of the laser line. The movement of the precision metalwork completes the 3D scanning of the laser line scan sensor, processes the height data, and transmits it to the embedded platform to generate a depth map for AI model training and analysis, as shown in Figures 1 and 2.

[0026] The disclosed embodiments can also be used to detect irregularly shaped shrapnel. By continuously optimizing the model in later production, continuous, accurate, and efficient intelligent detection can be achieved.

[0027] According to one or more embodiments, a method for constructing a 3D non-destructive testing model for precision metalworking is shown in FIG1 . Among them, threshold segmentation is used for segmentation of the depth map. For example, the qualified height range of the sample is 1 to 2 mm, and the tilting height exceeds 0.1 mm and is considered unqualified. The scanned height data is converted into a grayscale value corresponding to the depth map of 0 to 255, and the conversion is performed at a ratio of 1:100. The grayscale value of the qualified range is 0 to 200, and the grayscale value exceeding 210 is considered unqualified. The threshold range is set to 205 to 255, and graphic segmentation is performed to achieve defect extraction. This method greatly improves the speed of defect location.

[0028] In the embodiment of the present disclosure, the grayscale value range of the segmented depth map is 50, and the remapped grayscale value is 0-255. This method will cause the depth map to be distorted and the resolution will be reduced. Therefore, bicubic interpolation in super-resolution is used for processing. The specific implementation process is as follows:

[0029] a. Convert the low-resolution image into floating-point representation and determine the interpolation point location.

[0030] b. For each interpolation point, select 16 sampling points (4x4 grid) around it to calculate the difference.

[0031] c. Use the bicubic interpolation function to weight the sampling points of the selected area and calculate the grayscale value of the interpolation point.

[0032] This method can greatly improve the resolution and clarity of the image. Through the above image processing algorithms, defects can be located faster, contrast can be increased, accuracy can be improved, and AI analysis can be made more efficient.

[0033] Figure 2 shows the AI ​​model training process. When initializing YOLO-v5, the learning rate hyperparameter is set. The loss function is defined as a weighted sum of localization loss, classification loss, and confidence loss. The localization loss uses the squared error loss function; the classification loss uses the binary cross entropy loss function (BCE); and the confidence loss uses the sigmoid function. Each loss has a corresponding weight, which is adjusted based on experimental results.

[0034] When initializing the Adam optimizer, the learning rate, beta1, and beta2 are set.

[0035] When training and evaluating the model, the yolo model is adjusted according to the experimental results.

[0036] The learning rate or the adjustment of Adam hyperparameters.

[0037] The disclosed embodiment imports the trained model into the embedded device, collects and processes the depth map, and performs AI defect analysis to obtain the results.

[0038] In summary, the beneficial effects brought about by the technical solution of the present disclosure include:

[0039] 1. The present invention uses a laser line scan sensor to collect height information. Compared with a 2D camera, it eliminates the need for lighting and is not easily affected by the environment.

[0040] 2. Line scan laser has higher detection accuracy than 2D camera. It can generate point cloud map and 3D depth map by using height data.

[0041] 3. Use AI to analyze depth maps, and only need model training to automatically detect and improve efficiency.

[0042] 4. The embedded platform is cost-effective and flexible for deployment in confined spaces. It combines with laser line scan sensors for 3D scanning, is not restricted by ambient light, and uses height data to detect defects.

[0043] 5. Segment the depth map by setting a threshold based on the depth value, narrowing the defect range and locating the defect more accurately. Expand the defect mapping range and increase the contrast.

[0044] 6. Perform bicubic interpolation super-resolution processing to improve resolution, facilitate AI analysis, and improve accuracy.

[0045] 7. Use Adam to optimize the YOLO-v5 model, set the learning rate and beta1 and beta2 hyperparameters to improve detection efficiency.

[0046] The embodiments of the present disclosure are not limited to non-destructive detection of defects in precision metalworking including metal shrapnel, but can also be used in application scenarios such as missing parts, solder joints and damage on circuit boards.

[0047] It is worth noting that although the foregoing content has described the spirit and principles of the present invention with reference to several specific embodiments, it should be understood that the present invention is not limited to the disclosed specific embodiments, and the division into various aspects does not mean that the features of these aspects cannot be combined. Such division is merely for the convenience of expression. The present invention is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A precision metalworking 3D nondestructive testing method, It is characterized in that The detection method includes, Fix the laser line scan sensor on the bracket; A conveyor belt is arranged below the laser line scanning sensor, and the precision metalwork is placed on the conveyor belt; An embedded AI platform is configured to receive the image data of precision metalworking transmitted by the laser line scanning sensor; Laser line scan sensor collects data Laser line scan data; The embedded AI platform generates a depth map from the data transmitted by the laser line scan sensor and performs defect detection based on the depth map.

2. The detection method according to claim 1, It is characterized in that The position of precision metalworking should be placed in the middle area of ​​the laser line.

3. The detection method according to claim 1, It is characterized in that The defects include warping and cracks.

4. The detection method according to claim 1, It is characterized in that The embedded AI platform is deployed on the PC side.

5. The detection method according to claim 1, It is characterized in that The embedded AI platform is loaded with a precision metalworking 3D non-destructive testing model.

6. The detection method according to claim 5, It is characterized in that The method for constructing the detection model includes: collecting height data of precision metalworking to generate a depth map; performing threshold segmentation on the depth map; performing bicubic interpolation super-resolution processing, and training using an AI model.

7. The detection method according to claim 6, It is characterized in that The method for constructing the detection model also includes: Initialize YOLO-v5 and set the learning rate hyperparameters; Load the pre-trained YOLO model and pre-processed depth map dataset; Define the loss function according to task requirements; When initializing the Adam optimizer, set the learning rate hyperparameter; When training and evaluating the model, the hyperparameters are adjusted based on the training performance.

8. The detection method according to claim 7, It is characterized in that The loss function is defined as a weighted sum of positioning loss, classification loss, and confidence loss.

9. The detection method according to claim 1, It is characterized in that The precision metalworking includes metal shrapnel.

10. The detection method according to claim 1, It is characterized in that This detection method is used to detect missing parts, solder leaks and damage on circuit boards.

Images