A metal junction box stamping quality detection method and device

Abstract

The present application provides a metal junction box stamping quality inspection method and device, which relates to the field of stamping quality inspection technology. The method includes: if the size, shape and position deviation sets all meet the predetermined deviation threshold, the service life is predicted based on the actual scene data set, the surface defect set and the internal defect set, and the operation failure is predicted based on the actual scene data set, the size, shape and position deviation set, and the predicted service life and predicted failure frequency are generated; if the predicted service life and the predicted failure frequency meet the expected indicators, the stamping quality inspection is qualified, and a stamping quality inspection report is generated. This application can solve the technical problems that the existing metal junction box stamping quality inspection methods often tend to static inspection, and the detection means have a low degree of intelligence, resulting in poor accuracy, comprehensiveness and practicality of the detection results. It can reduce the probability of false detection and missed detection, improve the accuracy, comprehensiveness and practicality of detection, and improve detection efficiency.

Description

Technical Field

[0001] The present application relates to the technical field of stamping quality detection, and in particular to a stamping quality detection method and device for a metal junction box. Background Art

[0002] Metal junction boxes are devices used for electrical wiring and routing, usually made of metal materials such as stainless steel, aluminum or copper. They are widely used in construction, industry and residential areas to connect and protect wires and cables, ensuring the safety and reliability of electrical systems. In the stamping manufacturing process of metal junction boxes, quality inspection is a crucial link, which directly affects the performance and safety of the product.

[0003] At present, the existing metal junction box stamping quality inspection methods tend to tend to static inspections such as appearance defects and dimensions. At the same time, the detection means have a low degree of intelligence, resulting in a high probability of false detection and missed detection, resulting in poor accuracy, comprehensiveness and practicality of the detection results, and low technical problems such as low detection efficiency. Summary of the invention

[0004] The purpose of this application is to provide a metal junction box stamping quality inspection method and device to solve the technical problems that the existing metal junction box stamping quality inspection methods often tend to static inspections such as appearance defects and dimensions, and the detection means have a low degree of intelligence, resulting in a high probability of false detection and missed detection, resulting in poor accuracy, comprehensiveness and practicality of the detection results, and low detection efficiency.

[0005] In view of the above problems, the present application provides a method and device for detecting the stamping quality of a metal junction box.

[0006] In a first aspect, the present application provides a method for detecting the stamping quality of a metal junction box, which is implemented by a metal junction box stamping quality detection device, wherein the method includes: using a dual-source sensing device group to perform multi-angle acquisition of a target junction box to obtain a surface image set, an internal image set and a point cloud data set; performing image restoration on the surface image set and the internal image set based on an image preprocessing strategy to obtain a standard surface image set and a standard internal image, and inputting the standard surface image set and the standard internal image into a defect recognition channel, and outputting a surface defect set and an internal defect set, wherein the standard surface image set includes junction box images of multiple different planes; performing contour extraction on the standard surface image set and the standard internal image according to a predetermined contour extraction strategy, and generating a target three-dimensional contour based on the fusion of contour extraction results, wherein the predetermined contour extraction strategy includes a convolutional network and a watershed. Algorithm; taking the target three-dimensional contour as a constraint, performing fitting analysis on the point cloud data set, and performing simulation modeling based on the point cloud fitting result to obtain a target three-dimensional model; in a three-dimensional comparison space, performing deviation calculation on the target three-dimensional model based on a standard three-dimensional model to obtain a deviation data set, wherein the deviation data set includes a size deviation set, a shape deviation set and a position deviation set; if the size deviation set, the shape deviation set and the position deviation set all meet a predetermined deviation threshold, performing a service life prediction based on an actual scene data set, the surface defect set and the internal defect set, performing an operation fault prediction based on the actual scene data set, the size deviation set, the shape deviation set and the position deviation set, and generating a predicted service life and a predicted fault frequency; if the predicted service life and the predicted fault frequency all meet the expected indicators, the stamping quality inspection of the target junction box is qualified, and a stamping quality inspection report is generated.

[0007] In the second aspect, the present application also provides a metal junction box stamping quality inspection device for executing a metal junction box stamping quality inspection method as described in the first aspect, including: a multi-angle acquisition module, used to use a dual-source sensing device group to perform multi-angle acquisition on the target junction box to obtain a surface image set, an internal image set and a point cloud data set; a defect recognition module, used to restore the surface image set and the internal image set based on an image preprocessing strategy to obtain a standard surface image set and a standard internal image, and input the standard surface image set and the standard internal image into a defect recognition channel, and output a surface defect set and an internal defect set, wherein the standard surface image set includes junction box images of multiple different planes; a contour extraction module, used to perform contour extraction on the standard surface image set and the standard internal image according to a predetermined contour extraction strategy, and generate a target three-dimensional contour based on the fusion of contour extraction results, wherein the predetermined contour extraction strategy includes a convolutional network and a watershed algorithm; a point cloud simulation module; A fitting module is used to perform fitting analysis on the point cloud data set with the target three-dimensional contour as a constraint, and perform simulation modeling based on the point cloud fitting result to obtain a target three-dimensional model; a deviation comparison module is used to perform deviation calculation on the target three-dimensional model based on a standard three-dimensional model in a three-dimensional comparison space to obtain a deviation data set, wherein the deviation data set includes a size deviation set, a shape deviation set and a position deviation set; a performance prediction module is used to predict the service life based on the actual scene data set, the surface defect set and the internal defect set if the size deviation set, the shape deviation set and the position deviation set all meet the predetermined deviation threshold, and to predict the operation fault based on the actual scene data set, the size deviation set, the shape deviation set and the position deviation set, and generate a predicted service life and a predicted fault frequency; a test report generation module is used to determine that the stamping quality test of the target junction box is qualified if the predicted service life and the predicted fault frequency both meet the expected indicators, and generate a stamping quality test report.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0009] The target junction box is collected at multiple angles by using a dual-source sensing device group to obtain a surface image set, an internal image set and a point cloud data set; the surface image set and the internal image set are restored based on an image preprocessing strategy to obtain a standard surface image set and a standard internal image, and the standard surface image set and the standard internal image are input into a defect recognition channel to output a surface defect set and an internal defect set, wherein the standard surface image set includes a plurality of junction box images on different planes; contour extraction is performed on the standard surface image set and the standard internal image according to a predetermined contour extraction strategy, and a target three-dimensional contour is generated based on the fusion of contour extraction results, wherein the predetermined contour extraction strategy includes a convolutional network and a watershed algorithm; the point cloud data set is fitted and classified with the target three-dimensional contour as a constraint. Analysis is performed, and simulation modeling is performed based on the point cloud fitting result to obtain a target three-dimensional model; in the three-dimensional comparison space, deviation calculation is performed on the target three-dimensional model based on the standard three-dimensional model to obtain a deviation data set, wherein the deviation data set includes a size deviation set, a shape deviation set and a position deviation set; if the size deviation set, the shape deviation set and the position deviation set all meet the predetermined deviation threshold, a service life prediction is performed based on the actual scene data set, the surface defect set and the internal defect set, and an operation fault prediction is performed based on the actual scene data set, the size deviation set, the shape deviation set and the position deviation set, and a predicted service life and a predicted fault frequency are generated; if the predicted service life and the predicted fault frequency both meet the expected indicators, the stamping quality inspection of the target junction box is qualified, and a stamping quality inspection report is generated. That is to say, by collecting images and point clouds of the target junction box, an image data set and a point cloud data set are obtained; then a defect recognition channel is constructed to perform intelligent defect recognition on the image data set to obtain a surface defect set and an internal defect set; on the other hand, contour extraction is performed on the image data set according to a predetermined contour extraction strategy, and a target three-dimensional contour is generated based on the fusion of the contour extraction results; then the point cloud data set is fitted with the target three-dimensional contour as a constraint to construct a target three-dimensional model; further, in the three-dimensional comparison space, the deviation of the target three-dimensional model is calculated based on the standard three-dimensional model to obtain a size deviation set, a shape deviation set and a position deviation set; if the size deviation set, the shape deviation set and the position deviation set all meet the predetermined deviation threshold, it indicates that the target junction box meets the static quality inspection requirements, and then the service life is predicted based on the actual scene data set, the surface defect set and the internal defect set to generate a predicted service life; based on the actual scene data set, the size deviation set, the shape deviation set and the position deviation set, the operation fault prediction is performed to generate a predicted fault frequency; finally, if the predicted service life and the predicted fault frequency both meet the expected indicators, it indicates that the stamping quality inspection of the target junction box is qualified, and a stamping quality inspection report is generated at the same time.By combining machine vision technology and artificial intelligence algorithms, the technical goals of automated and intelligent detection of metal junction box stamping quality can be achieved, thereby reducing the probability of false detection and missed detection, improving the accuracy, comprehensiveness and practicality of stamping quality detection, and at the same time improving the technical effect of detection efficiency.

[0010] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented according to the contents of the specification, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are specifically cited below. It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present application, nor is it intended to limit the scope of the present application. Other features of the present application will become easy to understand through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solutions in the present application or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0012] Figure 1 A schematic diagram of a process for detecting the stamping quality of a metal junction box according to the present application;

[0013] Figure 2 A schematic diagram of a process for constructing a defect identification channel in a metal junction box stamping quality inspection method of the present application;

[0014] Figure 3 This is a schematic structural diagram of a metal junction box stamping quality inspection device according to the present application.

[0015] Description of reference numerals:

[0016] Multi-angle acquisition module 11, defect recognition module 12, contour extraction module 13, point cloud fitting module 14, deviation comparison module 15, performance prediction module 16, and detection report generation module 17. DETAILED DESCRIPTION

[0017] This application provides a metal junction box stamping quality detection method and device to solve the existing metal junction box stamping quality detection methods, which often tend to be static detection of appearance defects, dimensions, etc., and the detection means have a low degree of intelligence, resulting in a high probability of false detection and missed detection, resulting in poor accuracy, comprehensiveness and practicality of the detection results, and low detection efficiency. The technical goal of automated and intelligent detection of metal junction box stamping quality can be achieved by combining machine vision technology and artificial intelligence algorithms, thereby reducing the probability of false detection and missed detection, improving the accuracy, comprehensiveness and practicality of stamping quality detection, and improving the technical effect of detection efficiency.

[0018] Below, the technical solutions in the present application will be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments of the present application. It should be understood that the present application is not limited to the example embodiments described herein. Based on the embodiments of the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present application. It should also be noted that, for the convenience of description, only the parts related to the present application are shown in the accompanying drawings, rather than all of them.

[0019] Embodiment 1

[0020] Please see attached Figure 1 The present application provides a metal junction box stamping quality detection method, wherein the method is applied to a metal junction box stamping quality detection device, and the method specifically comprises the following steps:

[0021] Step 1: Use the dual-source sensing equipment group to collect multi-angle images of the target junction box to obtain a surface image set, an internal image set, and a point cloud data set.

[0022] Specifically, a dual-source sensing device group is constructed, wherein the dual-source sensing device group includes multiple image sensors (such as industrial cameras) and multiple laser point cloud scanning devices (such as laser radar or structured light scanners), each of which has a different monitoring angle, so as to capture the surface and internal features of the junction box from multiple perspectives. Then, the dual-source sensing device group is used to collect multi-angle data of the target junction box to obtain a surface image set, an internal image set and a point cloud data set, wherein the surface image refers to the outer shell surface image of the target junction box, and the internal image refers to the surface image of the internal structure of the target junction box. Then, the surface image set, the internal image set and the point cloud data set are transmitted to the intelligent detection platform, wherein the dual-source sensing device group is connected to the intelligent detection platform in communication.

[0023] Step 2: Perform image restoration on the surface image set and the internal image set based on the image preprocessing strategy to obtain a standard surface image set and a standard internal image, and input the standard surface image set and the standard internal image into a defect recognition channel to output a surface defect set and an internal defect set, wherein the standard surface image set includes multiple junction box images of different planes.

[0024] Specifically, in the intelligent detection platform, the surface image set and the internal image set are restored based on the image preprocessing strategy to obtain a standard surface image set and a standard internal image; the image preprocessing strategy includes image cleaning, image enhancement and image stitching, wherein image cleaning is used to remove noise interference and redundant information in the image, such as removing spots, stripes and other noise in the image through a filter, and removing background noise through background subtraction technology; image enhancement includes improving the contrast of the image to make the edges and details of the image clearer; adjusting the brightness and saturation of the image to improve the visibility of the image, etc.; image stitching is used to restore the image with occlusion based on the existing image to form a continuous View to overcome incomplete image information caused by occlusion, viewing angle limitation or device movement. For example, when some parts of an image are blocked, the blocked area can be located and filled by comparing and matching images at different angles to restore the complete image. This includes image alignment, image fusion and other technologies, such as using feature matching, optical flow or other alignment algorithms to align images from different viewing angles to the same coordinate system, while ensuring that the key feature points in the image maintain consistent positions in the stitched image; selecting appropriate fusion algorithms, such as weighted averaging, pixel-level fusion, etc., to merge the aligned images to ensure that the stitched image has a smooth transition in visual effect without obvious seams.

[0025] The standard surface image set includes multiple junction box images of different planes, such as: the front image, side image or bottom image of the junction box; then a defect recognition channel is constructed based on a convolutional neural network, wherein the convolutional neural network is a deep learning model with powerful feature learning capabilities. It can automatically extract useful features from the original image and use them for further image processing, including convolutional layers, pooling layers and fully connected layers; by constructing a defect recognition channel based on a convolutional neural network, automatic recognition of defects in metal junction box images can be achieved, and the accuracy and efficiency of defect recognition can be improved.

[0026] Then, the standard surface image set and the standard internal image are input into the defect recognition channel for defect recognition to obtain a surface defect set and an internal defect set, wherein the surface defect set includes surface defect features and position coordinates of the surface defect features in the target junction box, and the surface defect features include features such as flatness, smoothness, scratches, dents and cracks; the internal defect set includes internal defect features and position coordinates of the internal defect features in the target junction box, and the internal defect features include features such as dents, cracks, thread damage, edge material accumulation and burrs.

[0027] Step three: perform contour extraction on the standard surface image set and the standard internal image according to a predetermined contour extraction strategy, and generate a target three-dimensional contour based on the fusion of contour extraction results, wherein the predetermined contour extraction strategy includes a convolutional network and a watershed algorithm.

[0028] Specifically, a predetermined contour extraction strategy is configured, and the predetermined contour extraction strategy includes a contour extraction strategy based on a convolutional network, a contour extraction strategy based on a watershed algorithm, and a contour fitting strategy; the standard surface image set and the standard internal image are subjected to contour extraction by the contour extraction strategy based on the convolutional network to obtain a first surface contour set and a first internal contour; the standard surface image set and the standard internal image are subjected to contour extraction by the contour extraction strategy based on the watershed algorithm to obtain a second surface contour set and a second internal contour; finally, the first surface contour set and the second surface contour set, the first internal contour and the second internal contour are subjected to weighted fitting calculation according to the contour fitting strategy to obtain a surface contour set and an internal contour. Then, a fusion connection is performed based on the surface contour set and the internal contour, that is, a feature matching algorithm (such as SIFT, SURF, etc.) is used to perform feature matching on the surface contour set and the internal contour, and the matched features are spatially transformed so that they overlap in three-dimensional space; then, the registered surface contour and internal contour are fused and connected to generate a target three-dimensional contour.

[0029] Step 4: Taking the target three-dimensional contour as a constraint, performing fitting analysis on the point cloud data set, and performing simulation modeling based on the point cloud fitting result to obtain a target three-dimensional model.

[0030] Specifically, the point cloud data set is then fitted and analyzed based on the point cloud fitting algorithm with the target three-dimensional contour as a constraint to obtain a point cloud fitting result that satisfies the preset fitting constraint. Then, simulation modeling is performed based on the point cloud fitting result in the three-dimensional simulation platform to generate a target three-dimensional model. By performing point cloud fitting with the target three-dimensional contour as a constraint, the accuracy and reliability of the point cloud fitting result can be improved, thereby improving the accuracy of the target three-dimensional model construction.

[0031] Step 5: In the three-dimensional comparison space, the deviation of the target three-dimensional model is calculated based on the standard three-dimensional model to obtain a deviation data set, wherein the deviation data set includes a size deviation set, a shape deviation set and a position deviation set.

[0032] Specifically, a three-dimensional comparison space is constructed, where the three-dimensional comparison space is a coordinate space with a grid size, where the grid size can be set according to the actual situation, for example, the grid size is set to 0.01 mm to improve the precision of the comparison. Then, in the three-dimensional comparison space, the deviation of the target three-dimensional model is calculated based on the standard three-dimensional model, where the standard three-dimensional model is constructed based on product production indicators, and the deviation comparison dimensions include size dimension, shape dimension and position dimension, where the size dimension includes length, width, height, aperture, etc.; the shape dimension includes angle, curvature, flatness, etc. Based on the deviation comparison results, a size deviation set, a shape deviation set and a position deviation set are obtained, where the deviation characteristics include the deviation type and the specific deviation value, for example, the length deviation is 0.05 mm.

[0033] Step six: If the size deviation set, shape deviation set and position deviation set all meet the predetermined deviation threshold, a service life prediction is performed based on the actual scene data set, the surface defect set and the internal defect set, and an operation failure prediction is performed based on the actual scene data set, the size deviation set, shape deviation set and position deviation set to generate a predicted service life and a predicted failure frequency.

[0034] Specifically, a predetermined deviation threshold is obtained, and the predetermined deviation threshold is a deviation allowable range, which can be set according to actual conditions, including a size deviation threshold, a shape deviation threshold, and a position deviation threshold. If the size deviation set is less than the size deviation threshold, the shape deviation set is less than the shape deviation threshold, and the position deviation set is less than the position deviation threshold, it indicates that the target junction box meets the static quality inspection requirements.

[0035] Then, the actual scene data set of the target junction box is obtained. The actual scene data set is set based on the actual application scenario of the target junction box, including information such as temperature, humidity, and corrosion. Further, with the actual scene data set as a constraint, the sample training data for service life prediction is retrieved to supervise the training of the BP neural network and build a service life prediction model; and the service life of the surface defect set and the internal defect set is predicted through the service life prediction model to generate a predicted service life; then, with the actual scene data set as a constraint, the sample data set for operation fault prediction is retrieved to supervise the training of the BP neural network, and a fault prediction model is built to predict the operation fault of the size deviation set, shape deviation set, and position deviation set, and generate a predicted fault frequency. The predicted service life is proportional to the stamping quality. The higher the predicted service life, the better the stamping quality; the predicted fault frequency is inversely proportional to the stamping quality. The lower the predicted fault frequency, the better the stamping quality.

[0036] Step 7: If the predicted service life and the predicted failure frequency both meet the expected indicators, the stamping quality test of the target junction box is qualified, and a stamping quality test report is generated.

[0037] Specifically, the predicted service life and the predicted failure frequency are judged based on expected indicators, wherein the expected indicators include the expected service life and the expected failure frequency. If the predicted service life is greater than the expected service life and the predicted failure frequency is less than the expected failure frequency, then it indicates that the target junction box meets the dynamic quality inspection requirements, and the stamping quality inspection of the target junction box is qualified, and a stamping quality inspection report is generated, wherein the stamping quality inspection report includes a surface defect set, an internal defect set, a size deviation set, a shape deviation set, a position deviation set, a predicted service life and a predicted failure frequency, thereby improving the accuracy and comprehensiveness of the stamping quality inspection report, and at the same time providing data support for the subsequent optimization of the stamping process.

[0038] The metal junction box stamping quality detection method is applied to a metal junction box stamping quality detection device, which can solve the existing metal junction box stamping quality detection method, which often tends to static detection of appearance defects, dimensions, etc., and the detection means have a low degree of intelligence, resulting in a high probability of false detection and missed detection, resulting in poor accuracy, comprehensiveness and practicality of the detection results, and low detection efficiency. Technical problems. First, a dual-source sensor device group is used to collect multi-angle data of the target junction box to obtain a surface image set, an internal image set and a point cloud data set. Then, based on the image preprocessing strategy, the surface image set and the internal image set are image restored to obtain a standard surface image set and a standard internal image, and the standard surface image set and the standard internal image are input into the defect recognition channel, and the surface defect set and the internal defect set are output, wherein the standard surface image set includes multiple junction box images of different planes. Then, the standard surface image set and the standard internal image are contour extracted according to a predetermined contour extraction strategy, and the target three-dimensional contour is generated based on the fusion of the contour extraction results, wherein the predetermined contour extraction strategy includes a convolutional network and a watershed algorithm. Next, with the target three-dimensional contour as a constraint, the point cloud data set is fitted and analyzed, and simulation modeling is performed based on the point cloud fitting result to obtain the target three-dimensional model. Then, in the three-dimensional comparison space, the target three-dimensional model is subjected to deviation calculation based on the standard three-dimensional model to obtain a deviation data set, wherein the deviation data set includes a size deviation set, a shape deviation set and a position deviation set. In addition, if the size deviation set, the shape deviation set and the position deviation set all meet the predetermined deviation threshold, the service life is predicted based on the actual scene data set, the surface defect set and the internal defect set, and the operation fault prediction is performed based on the actual scene data set, the size deviation set, the shape deviation set and the position deviation set, and the predicted service life and predicted fault frequency are generated. Finally, if the predicted service life and the predicted fault frequency both meet the expected indicators, the stamping quality inspection of the target junction box is qualified, and a stamping quality inspection report is generated. By combining machine vision technology and artificial intelligence algorithms, the technical goals of automated and intelligent stamping quality inspection of metal junction boxes can be achieved, thereby reducing the probability of false detection and missed detection, improving the accuracy, comprehensiveness and practicality of stamping quality inspection, and improving the technical effect of inspection efficiency.

[0039] Further, construct a defect identification channel, as shown in the attached Figure 2 As shown, step 2 of this application includes:

[0040] Taking the target junction box and preset surface defect features as constraints, sample surface image sets and sample surface defect feature sets are retrieved online as sample surface training sets, wherein the preset surface defect features include at least flatness, smoothness, scratches, dents and cracks; the sample surface training set is divided into a surface training set and a surface verification set according to a preset ratio, and the convolutional neural network is supervised trained and feedback verified through the surface training set and the surface verification set to obtain a surface defect recognition unit that meets the expected convergence constraints; taking the target junction box and preset internal defect features as constraints, sample internal image sets and sample internal defect feature sets are retrieved online, and the convolutional neural network is supervised trained and feedback verified to obtain an internal defect recognition unit, wherein the preset internal defect features include at least dents, cracks, thread damage, edge material accumulation and burrs; a defect recognition channel is constructed based on the surface defect recognition unit and the internal defect recognition unit.

[0041] Specifically, first, with the attribute information and preset surface defect features of the target junction box as constraints, an online search is performed based on big data technology, where the preset surface defect features refer to the high-frequency surface defect features of the target junction box, including at least flatness, smoothness, scratches, dents and cracks, etc., and technical personnel in this field can also add according to actual conditions; the attribute information of the target junction box includes information such as the type, size and material of the junction box; big data is a series of technologies and methods that enable the value contained in massive data to be mined and displayed, including data planning, data collection, analysis and mining and other technical means; a sample surface image set and a sample surface defect feature set are obtained, and the sample surface image set and the sample surface defect feature set are used as a sample surface training set.

[0042] Then, the sample surface training set is divided into a surface training set and a surface verification set according to a preset ratio. For example, the training set usually accounts for 85% and the verification set accounts for 15%. Custom divisions can also be made according to actual conditions. Then, an initial surface defect recognition unit is constructed based on a convolutional neural network, where the input data of the initial surface defect recognition unit is a surface image and the output data is a surface defect feature. The initial surface defect recognition unit is supervised and trained through the surface training set. When the output result of the initial surface defect recognition unit tends to be stable, the initial surface defect recognition unit is then feedback verified through the surface verification set until the output result meets the expected convergence constraint, and the trained surface defect recognition unit is output. The expected convergence constraint is the expected output accuracy, which can be set according to actual needs.

[0043] On the other hand, with the attribute information of the target junction box and the preset internal defect features as constraints, the sample internal image set and the sample internal defect feature set are retrieved online, wherein the preset internal defect features include at least dents, cracks, thread damage, edge material accumulation and burrs; then, an initial internal defect recognition unit is constructed based on a convolutional neural network, wherein the input of the initial internal defect recognition unit is the internal image, and the output is the internal defect feature. The sample internal image set and the sample internal defect feature set are used as training data, and the initial internal defect recognition unit is supervised trained and feedback verified to obtain an internal defect recognition unit. Finally, a defect recognition channel is constructed based on the surface defect recognition unit and the internal defect recognition unit.

[0044] Further, contour extraction is performed on the standard surface image set and the standard internal image according to a predetermined contour extraction strategy. Step three of the present application includes:

[0045] The predetermined contour extraction strategy includes a convolution extraction strategy, a watershed extraction strategy and a contour fitting strategy; a contour extraction model is constructed based on the convolution extraction strategy, and contour extraction is performed on the standard surface image set and the standard internal image through the contour extraction model to obtain a first surface contour set and a first internal contour; contour extraction is performed on the standard surface image set and the standard internal image based on the watershed extraction strategy to obtain a second surface contour set and a second internal contour; fitting analysis is performed on the first surface contour set and the second surface contour set, the first internal contour and the second internal contour according to the contour fitting strategy to obtain a surface contour set and an internal contour, wherein the contour fitting strategy includes a first trusted weight and a second trusted weight, and the trusted weight is negatively correlated with the contour extraction error rate.

[0046] Specifically, the predetermined contour extraction strategy includes a convolution extraction strategy, a watershed extraction strategy and a contour fitting strategy. First, a contour extraction model is constructed based on a convolutional neural network, wherein the input data of the contour extraction model is the image data of the junction box, and the output data is the contour data. The sample data set can be retrieved by taking the attribute information of the target junction box as a constraint to perform supervised training on the contour extraction model; then, the standard surface image set and the standard internal image are input into the contour extraction model for contour extraction, and a first surface contour set and a first internal contour are output. On the other hand, the standard surface image set and the standard internal image are subjected to contour extraction based on the watershed algorithm to obtain a second surface contour set and a second internal contour.

[0047] Based on the retrieved sample data, the convolution extraction strategy and the watershed extraction strategy are analyzed for the contour extraction error rate, where the error rate refers to the ratio of the number of contour extraction errors to the total number of contour extractions. The higher the contour extraction error rate, the worse the accuracy of the contour extraction. Then, based on the coefficient of variation method, the first credible weight and the second credible weight are set according to the contour extraction error rate, where the credible weight is inversely proportional to the contour extraction error rate, that is, the higher the contour extraction error rate, the lower the credible weight. Finally, based on the first credible weight and the second credible weight, the first surface contour set and the second surface contour set, the first internal contour and the second internal contour are weighted fitting calculated, and the surface contour set and the internal contour are obtained based on the weighted fitting results.

[0048] Further, based on the watershed extraction strategy, contour extraction is performed on the standard surface image set and the standard internal image, and the present application further includes the following steps:

[0049] Grayscale processing is performed on the standard surface image set and the standard internal image to obtain a surface grayscale image set and an internal grayscale image; a first surface grayscale image is selected, and water is filled with water using the minimum grayscale value of the first surface grayscale image as a starting point, and the first grayscale value that the water surface is submerged after water filling is set as the first-stage grayscale value; it is determined whether the deviation value between the first-stage grayscale value and the minimum grayscale value meets a preset grayscale difference, and if so, the simulated water filling is continued; if not, a first watershed is set at the first-stage grayscale value, and the simulated water filling is continued until the water surface submerges the maximum grayscale value of the first surface grayscale image, and then the water filling is stopped to obtain multiple watersheds; the first surface grayscale image is divided using multiple watersheds, and contour extraction is performed based on the division result to obtain a first surface grayscale contour, which is added to the second surface contour set.

[0050] Specifically, firstly, grayscale processing is performed on the standard surface image set and the standard internal image, wherein grayscale processing refers to an image processing technique for converting a color image into a grayscale image, which converts the color information of each pixel in the image into a grayscale value, usually represented by an integer between 0 (black) and 255 (white), to obtain a surface grayscale image set and an internal grayscale image. Then, a first surface grayscale image is randomly selected from the surface grayscale image set, and the first surface grayscale image is any one of the surface grayscale image set.

[0051] Then, water filling is simulated with the minimum grayscale value of the first surface grayscale image as the starting point, and the first grayscale value that the water surface covers after water filling is set as the first-stage grayscale value; then the grayscale deviation value between the first-stage grayscale value and the minimum grayscale value is calculated, and it is determined whether the grayscale deviation value is less than a preset grayscale difference value, and the preset grayscale difference value can be set according to actual conditions; if the grayscale deviation value is less than the preset grayscale difference value, the simulated water filling is continued; if the grayscale deviation value is greater than or equal to the preset grayscale difference value, the first watershed is set at the first-stage grayscale value, and the simulated water filling is continued; the simulated water filling is continuously iterated until the water surface covers the maximum grayscale value of the first surface grayscale image, then the water filling is stopped, and multiple watersheds are obtained.

[0052] Finally, multiple watersheds are used to divide the first surface grayscale image, and contour extraction is performed based on the division results. For example, in each connected area, feature points such as edges and corners are extracted, and these feature points constitute the contour of each area; then the contours of all connected areas are merged to form a complete contour, and it is ensured that the merged contour contains all contour information of the first surface grayscale image to obtain the first surface grayscale contour, and the first surface grayscale contour is added to the second surface contour set, where the extraction method of the second internal contour is the same as the extraction method of the second surface contour set, which will not be explained in detail here.

[0053] Further, taking the target three-dimensional contour as a constraint, the point cloud data set is subjected to fitting analysis. Step 4 of the present application includes:

[0054] Taking the target three-dimensional contour as a constraint, randomly fitting the point cloud dataset is performed to obtain a first point cloud fitting result, and a first fitting accuracy is calculated, wherein the first fitting accuracy is the ratio of the number of point clouds falling within the target three-dimensional contour to the total number of point clouds in the point cloud dataset; again taking the target three-dimensional contour as a constraint, randomly fitting the point cloud dataset is performed to obtain a second point cloud fitting result and a second fitting accuracy; iterative fitting is performed until a preset fitting number constraint is met, and a point cloud fitting result with the highest fitting accuracy is output.

[0055] Specifically, with the target three-dimensional contour as a constraint, the point cloud data set is randomly fitted, that is, the point cloud data in the point cloud data set is randomly placed in the target three-dimensional contour to obtain a first point cloud fitting result, and the first fitting accuracy of the first point cloud fitting result is calculated, wherein the first fitting accuracy is the ratio of the number of point clouds falling into the target three-dimensional contour to the total number of point clouds in the point cloud data set, wherein the greater the first fitting accuracy, the better the point cloud fitting quality. Then, with the target three-dimensional contour as a constraint, the point cloud data set is randomly fitted to obtain a second point cloud fitting result and a second fitting accuracy; the same method is used to iterate the fitting until the preset fitting times constraint is met, wherein the preset fitting times constraint can be set based on the point cloud fitting accuracy requirement, wherein the higher the point cloud fitting accuracy requirement, the greater the preset fitting times, and the point cloud fitting result with the highest fitting accuracy is output at this time.

[0056] Furthermore, step six of this application includes:

[0057] The predetermined deviation thresholds include a size deviation threshold, a shape deviation threshold and a position deviation threshold; if the size deviation set does not satisfy the size deviation threshold and / or the shape deviation set does not satisfy the shape deviation threshold and / or the position deviation set does not satisfy the position deviation threshold, the stamping quality inspection of the target junction box fails.

[0058] Specifically, the predetermined deviation threshold includes a size deviation threshold, a shape deviation threshold and a position deviation threshold, which can be set according to the stamping quality inspection index; if the size deviation set is greater than or equal to the size deviation threshold and / or the shape deviation set is greater than or equal to the shape deviation threshold and / or the position deviation set is greater than or equal to the position deviation threshold, it indicates that the target junction box does not meet the static quality inspection requirements, and the stamping quality inspection result of the target junction box is unqualified.

[0059] Further, based on the actual scene data set, the surface defect set and the internal defect set, the service life prediction is performed, and step six of the present application includes:

[0060] Acquire an actual scenario data set of a target junction box, wherein the actual scenario data set includes at least ambient temperature, ambient humidity, dust density and air corrosion; retrieve a sample data set through an Internet search with the target junction box and the actual scenario data set as constraints, wherein the sample data set includes a sample surface defect set, a sample internal defect set and a sample service life set; divide the sample data set into Q equal parts, randomly select Q times without replacement to construct a first data set, iterate and select Q times to obtain Q data sets; use the Q data sets to perform supervised training on the BP neural network respectively to obtain Q convergence life prediction units, and generate a life prediction model based on the Q convergence life prediction units, wherein the output of the life prediction model is the mode of the output results of the Q convergence life prediction units; input the surface defect set and the internal defect set into the life prediction model, and output the predicted service life.

[0061] Specifically, first, the actual scene data set of the target junction box is collected and obtained, and the actual scene data set includes parameters such as ambient temperature, ambient humidity, dust density and air corrosion. Then, the attribute information of the target junction box and the actual scene data set are used as constraints, and the sample data set is retrieved online to obtain the sample data set, wherein the sample data set includes a sample surface defect set, a sample internal defect set and a sample service life set, and the sample surface defects, sample internal defects and sample service life correspond to each other.

[0062] Then, the sample data set is divided into Q parts, where Q is a positive integer greater than 3 and can be set according to the actual data volume. In the Q data sets, Q parts are randomly selected without replacement to construct the first data set; then, the same method is used to iteratively select Q times to obtain Q data sets. Then, a life prediction unit is constructed based on the BP neural network. The life prediction unit is a neural network model that can be iteratively optimized in the BP neural network. The input data of the life prediction unit are surface defects and internal defects, and the output data is the service life; then, the life prediction unit is supervised and cross-validated using the Q data sets to obtain Q convergent life prediction units that meet the expected accuracy index; further, a life prediction model is constructed based on the Q convergent life prediction units, wherein the output of the life prediction model is the mode of the output results of the Q convergent life prediction units; finally, the surface defect set and the internal defect set are input into the life prediction model for prediction, and the predicted service life is output. By constructing a life prediction model in parallel based on Q convergent life prediction units, the prediction error of the service life can be reduced, and the accuracy and reliability of the service life prediction can be improved.

[0063] The method for predicting operational faults based on the actual scene data set, the size deviation set, the shape deviation set and the position deviation set is the same as the service life prediction method, that is, taking the attribute information of the target junction box and the actual scene data set as constraints, retrieve and obtain a fault prediction sample data set, which includes a sample size deviation set, a sample shape deviation set, a sample position deviation set and a sample fault frequency, and then build a fault prediction model based on the fault prediction sample data set to perform fault prediction and obtain the predicted fault frequency.

[0064] Further, step seven of this application includes:

[0065] The expected indicators include an expected service life and an expected failure frequency; if the predicted service life does not meet the expected service life and / or the predicted failure frequency does not meet the expected failure frequency, the stamping quality test of the target junction box fails.

[0066] Specifically, the expected indicators include expected service life and expected failure frequency, which can be set according to the actual use requirements of the target junction box. If the predicted service life is less than or equal to the expected service life and / or the predicted failure frequency is greater than or equal to the expected failure frequency, it indicates that the target junction box does not meet the dynamic quality inspection requirements, and the stamping quality inspection result of the target junction box is unqualified.

[0067] In summary, the metal junction box stamping quality detection method provided by the present application has the following technical effects:

[0068] By collecting images and point clouds of the target junction box, an image data set and a point cloud data set are obtained; then a defect recognition channel is constructed to perform intelligent defect recognition on the image data set to obtain a surface defect set and an internal defect set; on the other hand, contour extraction is performed on the image data set according to a predetermined contour extraction strategy, and a target three-dimensional contour is generated based on the fusion of the contour extraction results; then the point cloud data set is fitted with the target three-dimensional contour as a constraint to construct a target three-dimensional model; further, in the three-dimensional comparison space, the deviation of the target three-dimensional model is calculated based on the standard three-dimensional model to obtain a size deviation set, a shape deviation set and a position deviation set; if the size deviation set, the shape deviation set and the position deviation set all meet the predetermined deviation threshold, it indicates that the target junction box meets the static quality inspection requirements, and then the service life is predicted based on the actual scene data set, the surface defect set and the internal defect set to generate a predicted service life; based on the actual scene data set, the size deviation set, the shape deviation set and the position deviation set, the operation fault prediction is performed to generate a predicted fault frequency; finally, if the predicted service life and the predicted fault frequency both meet the expected indicators, it indicates that the stamping quality inspection of the target junction box is qualified, and a stamping quality inspection report is generated at the same time. By combining machine vision technology and artificial intelligence algorithms, the technical goals of automated and intelligent detection of metal junction box stamping quality can be achieved, thereby reducing the probability of false detection and missed detection, improving the accuracy, comprehensiveness and practicality of stamping quality detection, and at the same time improving the technical effect of detection efficiency.

[0069] Embodiment 2

[0070] Based on the same inventive concept as the metal junction box stamping quality detection method in the above embodiment, the present application also provides a metal junction box stamping quality detection device, see the attached Figure 3 ,include:

[0071] The multi-angle acquisition module 11 is used to perform multi-angle acquisition on the target junction box using a dual-source sensing device group to obtain a surface image set, an internal image set and a point cloud data set.

[0072] The defect recognition module 12 is used to perform image restoration on the surface image set and the internal image set based on the image preprocessing strategy to obtain a standard surface image set and a standard internal image, and input the standard surface image set and the standard internal image into the defect recognition channel, and output a surface defect set and an internal defect set, wherein the standard surface image set includes multiple junction box images of different planes.

[0073] The contour extraction module 13 is used to perform contour extraction on the standard surface image set and the standard internal image according to a predetermined contour extraction strategy, and generate a target three-dimensional contour based on the fusion of contour extraction results, wherein the predetermined contour extraction strategy includes a convolutional network and a watershed algorithm.

[0074] The point cloud fitting module 14 is used to perform fitting analysis on the point cloud data set with the target three-dimensional contour as a constraint, and perform simulation modeling based on the point cloud fitting result to obtain a target three-dimensional model.

[0075] The deviation comparison module 15 is used to calculate the deviation of the target three-dimensional model based on the standard three-dimensional model in the three-dimensional comparison space to obtain a deviation data set, wherein the deviation data set includes a size deviation set, a shape deviation set and a position deviation set.

[0076] The performance prediction module 16 is used to predict the service life based on the actual scene data set, the surface defect set and the internal defect set if the size deviation set, the shape deviation set and the position deviation set all meet the predetermined deviation threshold, and to predict the operation failure based on the actual scene data set, the size deviation set, the shape deviation set and the position deviation set, and generate the predicted service life and predicted failure frequency.

[0077] The test report generating module 17 is used to determine if the predicted service life and the predicted failure frequency meet the expected indicators, that the stamping quality test of the target junction box is qualified, and to generate a stamping quality test report.

[0078] Furthermore, the defect identification module 12 in the device is also used for:

[0079] The sample surface training set is divided into a surface training set and a surface verification set according to a preset ratio, and the convolutional neural network is supervised trained and feedback verified through the surface training set and the surface verification set to obtain a surface defect recognition unit that meets the expected convergence constraints; with the target junction box and the preset internal defect features as constraints, the sample internal image set and the sample internal defect feature set are retrieved online to obtain the convolutional neural network through supervised training and feedback verification to obtain the internal defect recognition unit, wherein the preset internal defect features include at least dents, cracks, thread damage, edge material accumulation and burrs; a defect recognition channel is constructed based on the surface defect recognition unit and the internal defect recognition unit.

[0080] Furthermore, the contour extraction module 13 in the device is also used for:

[0081] The predetermined contour extraction strategy includes a convolution extraction strategy, a watershed extraction strategy and a contour fitting strategy; a contour extraction model is constructed based on the convolution extraction strategy, and contour extraction is performed on the standard surface image set and the standard internal image through the contour extraction model to obtain a first surface contour set and a first internal contour; contour extraction is performed on the standard surface image set and the standard internal image based on the watershed extraction strategy to obtain a second surface contour set and a second internal contour; fitting analysis is performed on the first surface contour set and the second surface contour set, the first internal contour and the second internal contour according to the contour fitting strategy to obtain a surface contour set and an internal contour, wherein the contour fitting strategy includes a first trusted weight and a second trusted weight, and the trusted weight is negatively correlated with the contour extraction error rate.

[0082] Furthermore, the contour extraction module 13 in the device is also used for:

[0083] Grayscale processing is performed on the standard surface image set and the standard internal image to obtain a surface grayscale image set and an internal grayscale image; a first surface grayscale image is selected, and water is filled with water using the minimum grayscale value of the first surface grayscale image as a starting point, and the first grayscale value that the water surface is submerged after water filling is set as the first-stage grayscale value; it is determined whether the deviation value between the first-stage grayscale value and the minimum grayscale value meets a preset grayscale difference, and if so, the simulated water filling is continued; if not, a first watershed is set at the first-stage grayscale value, and the simulated water filling is continued until the water surface submerges the maximum grayscale value of the first surface grayscale image, and then the water filling is stopped to obtain multiple watersheds; the first surface grayscale image is divided using multiple watersheds, and contour extraction is performed based on the division result to obtain a first surface grayscale contour, which is added to the second surface contour set.

[0084] Furthermore, the point cloud fitting module 14 in the device is also used for:

[0085] Taking the target three-dimensional contour as a constraint, randomly fitting the point cloud dataset is performed to obtain a first point cloud fitting result, and a first fitting accuracy is calculated, wherein the first fitting accuracy is the ratio of the number of point clouds falling within the target three-dimensional contour to the total number of point clouds in the point cloud dataset; again taking the target three-dimensional contour as a constraint, randomly fitting the point cloud dataset is performed to obtain a second point cloud fitting result and a second fitting accuracy; iterative fitting is performed until a preset fitting number constraint is met, and a point cloud fitting result with the highest fitting accuracy is output.

[0086] Furthermore, the performance prediction module 16 in the device is also used for:

[0087] The predetermined deviation thresholds include a size deviation threshold, a shape deviation threshold and a position deviation threshold; if the size deviation set does not satisfy the size deviation threshold and / or the shape deviation set does not satisfy the shape deviation threshold and / or the position deviation set does not satisfy the position deviation threshold, the stamping quality inspection of the target junction box fails.

[0088] Furthermore, the performance prediction module 16 in the device is also used for:

[0089] Acquire an actual scenario data set of a target junction box, wherein the actual scenario data set includes at least ambient temperature, ambient humidity, dust density and air corrosion; retrieve a sample data set through an Internet search with the target junction box and the actual scenario data set as constraints, wherein the sample data set includes a sample surface defect set, a sample internal defect set and a sample service life set; divide the sample data set into Q equal parts, randomly select Q times without replacement to construct a first data set, iterate and select Q times to obtain Q data sets; use the Q data sets to perform supervised training on the BP neural network respectively to obtain Q convergence life prediction units, and generate a life prediction model based on the Q convergence life prediction units, wherein the output of the life prediction model is the mode of the output results of the Q convergence life prediction units; input the surface defect set and the internal defect set into the life prediction model, and output the predicted service life.

[0090] Furthermore, the detection report generating module 17 in the device is also used for:

[0091] The expected indicators include an expected service life and an expected failure frequency; if the predicted service life does not meet the expected service life and / or the predicted failure frequency does not meet the expected failure frequency, the stamping quality test of the target junction box fails.

[0092] Each embodiment in this specification is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The metal junction box stamping quality detection method and specific examples in the aforementioned embodiment 1 are also applicable to a metal junction box stamping quality detection device in this embodiment. Through the aforementioned detailed description of a metal junction box stamping quality detection method, those skilled in the art can clearly understand a metal junction box stamping quality detection device in this embodiment, so for the sake of brevity of the specification, it will not be described in detail here. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0093] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

[0094] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application belong to the scope of the present application and its equivalent technology, the present application is also intended to include these modifications and variations.

Claims

1. A method for detecting the stamping quality of a metal junction box, characterized in that: include: The dual-source sensing device group is used to collect multi-angle images of the target junction box to obtain surface image sets, internal image sets and point cloud data sets; Performing image restoration on the surface image set and the internal image set based on an image preprocessing strategy to obtain a standard surface image set and a standard internal image, and inputting the standard surface image set and the standard internal image into a defect recognition channel, and outputting a surface defect set and an internal defect set, wherein the standard surface image set includes multiple junction box images of different planes; Performing contour extraction on the standard surface image set and the standard internal image according to a predetermined contour extraction strategy, and generating a target three-dimensional contour based on the fusion of contour extraction results, wherein the predetermined contour extraction strategy includes a convolutional network and a watershed algorithm; Taking the target three-dimensional contour as a constraint, performing fitting analysis on the point cloud data set, and performing simulation modeling based on the point cloud fitting result to obtain a target three-dimensional model; In the three-dimensional comparison space, the target three-dimensional model is subjected to deviation calculation based on the standard three-dimensional model to obtain a deviation data set, wherein the deviation data set includes a size deviation set, a shape deviation set and a position deviation set; If the size deviation set, shape deviation set and position deviation set all meet the predetermined deviation threshold, a service life prediction is performed based on the actual scene data set, the surface defect set and the internal defect set, and an operation failure prediction is performed based on the actual scene data set, the size deviation set, shape deviation set and the position deviation set to generate a predicted service life and a predicted failure frequency; If the predicted service life and the predicted failure frequency both meet the expected indicators, the stamping quality test of the target junction box is qualified, and a stamping quality test report is generated; Among them, building a defect identification channel includes: Taking the target junction box and the preset surface defect features as constraints, a sample surface image set and a sample surface defect feature set are retrieved online as a sample surface training set, wherein the preset surface defect features include at least flatness, smoothness, scratches, dents and cracks; The sample surface training set is divided into a surface training set and a surface verification set according to a preset ratio, and the convolutional neural network is supervised trained and feedback verified through the surface training set and the surface verification set to obtain a surface defect recognition unit that meets the expected convergence constraints; Taking the target junction box and the preset internal defect features as constraints, the sample internal image set and the sample internal defect feature set are retrieved online, and the convolutional neural network is supervised trained and feedback verified to obtain the internal defect recognition unit, wherein the preset internal defect features include at least dents, cracks, thread damage, edge material accumulation and burrs; Building a defect recognition channel based on the surface defect recognition unit and the internal defect recognition unit; The service life prediction based on the actual scene data set, the surface defect set and the internal defect set includes: Acquire an actual scene data set of a target junction box, wherein the actual scene data set at least includes ambient temperature, ambient humidity, dust density, and air corrosion; With the target junction box and the actual scene dataset as constraints, a sample dataset is retrieved online to obtain the sample dataset, wherein the sample dataset includes a sample surface defect set, a sample internal defect set, and a sample service life set; The sample data set is equally divided into Q parts, and randomly selected without replacement Q times to construct a first data set, and iteratively selected Q times to obtain Q data sets; Using Q data sets to supervise the BP neural network training, Q convergence prediction units are obtained, and a prediction model is generated based on the Q convergence prediction units, wherein the output of the prediction model is the mode of the output results of the Q convergence prediction units; The surface defect set and the internal defect set are input into a service life prediction model, and a predicted service life is output.

2. A metal junction box stamping quality inspection method according to claim 1, characterized in that: Performing contour extraction on the standard surface image set and the standard internal image according to a predetermined contour extraction strategy includes: The predetermined contour extraction strategy includes a convolution extraction strategy, a watershed extraction strategy and a contour fitting strategy; A contour extraction model is constructed based on the convolution extraction strategy, and contour extraction is performed on the standard surface image set and the standard internal image through the contour extraction model to obtain a first surface contour set and a first internal contour: Performing contour extraction on the standard surface image set and the standard internal image based on the watershed extraction strategy to obtain a second surface contour set and a second internal contour; According to the contour fitting strategy, a first surface contour set and a second surface contour set, a first internal contour and a second internal contour are fitted and analyzed to obtain a surface contour set and an internal contour, wherein the contour fitting strategy includes a first credible weight and a second credible weight, and the credible weight is negatively correlated with the contour extraction error rate.

3. A metal junction box stamping quality inspection method according to claim 2, characterized in that: Performing contour extraction on the standard surface image set and the standard internal image based on the watershed extraction strategy includes: Performing grayscale processing on the standard surface image set and the standard internal image to obtain a surface grayscale image set and an internal grayscale image; Select the first surface grayscale image, take the minimum grayscale value of the first surface grayscale image as the starting point for water filling, and set the first grayscale value covered by the water surface after water filling as the grayscale value of the first stage; Determine whether the deviation between the grayscale value in the first stage and the minimum grayscale value meets the preset grayscale difference. If so, continue to simulate watering; If not, the first watershed is set at the grayscale value of the first stage, and the simulated watering is continued until the water surface exceeds the maximum grayscale value of the grayscale image of the first surface, then the watering is stopped, and multiple watersheds are obtained; The first surface grayscale image is divided by using multiple watersheds, and contours are extracted based on the division results to obtain the first surface grayscale contour, which is added to the second surface contour set.

4. A metal junction box stamping quality inspection method according to claim 1, characterized in that: Taking the target three-dimensional contour as a constraint, performing fitting analysis on the point cloud data set, including: Taking the target three-dimensional contour as a constraint, randomly fitting the point cloud data set to obtain a first point cloud fitting result, and calculating a first fitting accuracy, wherein the first fitting accuracy is a ratio of the number of point clouds falling within the target three-dimensional contour to the total number of point clouds in the point cloud data set; Again, taking the target three-dimensional contour as a constraint, randomly fitting the point cloud data set to obtain a second point cloud fitting result and a second fitting accuracy; Perform iterative fitting until the preset fitting number constraint is met, and output the point cloud fitting result with the highest fitting accuracy.

5. A metal junction box stamping quality inspection method according to claim 1, characterized in that: The method further comprises: The predetermined deviation thresholds include a size deviation threshold, a shape deviation threshold and a position deviation threshold; If the size deviation set does not satisfy the size deviation threshold and / or the shape deviation set does not satisfy the shape deviation threshold and / or the position deviation set does not satisfy the position deviation threshold, the stamping quality inspection of the target junction box fails.

6. A metal junction box stamping quality inspection method according to claim 1, characterized in that: The method further comprises: The expected indicators include expected service life and expected failure frequency; If the predicted service life does not meet the expected service life and / or the predicted failure frequency does not meet the expected failure frequency, the stamping quality test of the target junction box fails.

7. A metal junction box stamping quality detection device, characterized in that: The steps for implementing the method according to any one of claims 1 to 6 include: A multi-angle acquisition module is used to perform multi-angle acquisition of the target junction box using a dual-source sensor device group to obtain a surface image set, an internal image set and a point cloud data set; a defect recognition module, configured to perform image restoration on the surface image set and the internal image set based on an image preprocessing strategy to obtain a standard surface image set and a standard internal image, input the standard surface image set and the standard internal image into a defect recognition channel, and output a surface defect set and an internal defect set, wherein the standard surface image set includes a plurality of junction box images of different planes; A contour extraction module, used for performing contour extraction on the standard surface image set and the standard internal image according to a predetermined contour extraction strategy, and generating a target three-dimensional contour based on the fusion of contour extraction results, wherein the predetermined contour extraction strategy includes a convolutional network and a watershed algorithm; A point cloud fitting module is used to perform fitting analysis on the point cloud data set with the target three-dimensional contour as a constraint, and to perform simulation modeling based on the point cloud fitting result to obtain a target three-dimensional model; A deviation comparison module, used for performing deviation calculation on the target three-dimensional model based on the standard three-dimensional model in the three-dimensional comparison space to obtain a deviation data set, wherein the deviation data set includes a size deviation set, a shape deviation set and a position deviation set; A performance prediction module, for predicting the service life based on the actual scene data set, the surface defect set and the internal defect set, and for predicting operational failures based on the actual scene data set, the size deviation set, the shape deviation set and the position deviation set, if the size deviation set, the shape deviation set and the position deviation set all meet the predetermined deviation threshold, and for generating a predicted service life and a predicted failure frequency; The test report generation module is used to determine if the predicted service life and the predicted failure frequency meet the expected indicators, then the stamping quality test of the target junction box is qualified, and generate a stamping quality test report.

Images