A weld surface quality detection method and a computer readable storage medium

Abstract

The application discloses a kind of weld surface quality detection method and computer readable storage medium, belong to the field of welding quality nondestructive testing, comprising: (S1) the profile height data of each weld section collected using line structured light camera along the length direction of weld is preprocessed;Preprocessing includes removing 0, outlier and Gaussian noise;(S2) the profile height data after preprocessing is input to the depth confidence network pre-trained, to judge whether each weld section exists defect;(S3) the weld section that exists defect and continuous distribution is identified as the distribution range of same defect, and the profile height data after preprocessing in each defect distribution range is spliced as corresponding defect three-dimensional point cloud data;(S4) each defect three-dimensional point cloud data is input to the weld defect detection network trained, and the type of each defect of weld surface is obtained.The application optimizes the welding quality detection process, improves the efficiency and accuracy of weld surface quality detection.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of non-destructive testing of welding quality, and more particularly, relates to a welding seam surface quality detection method and a computer readable storage medium. BACKGROUND

[0002] Welding technology is widely used in the fields of automobile manufacturing and rail train manufacturing, and due to the interference in the welding process and the instability of the welding environment, accidental defects such as undercut, porosity, incomplete fusion, inclusions, porosity, cracks, etc. may occur during the welding process. The welding seams with welding defects will cause the quality and performance of the welded workpiece to be reduced and cannot meet the application requirements. Due to the safety of the service object of the automobile and the rail train and the particularity of the application scenario, the quality of the welding seam needs to be ensured. Under normal circumstances, the vehicle production workshop will detect the quality of the welded seam after the workpiece is welded, and comprehensively detect the forming size of the welding seam and the defects existing in the welding seam. If the welding seam quality is unqualified, it will be repaired or supplemented, etc. to improve the service life, yield and good rate of the workpiece. However, the welding seam quality detection process of most factories is completed by manual operation, which requires a quality inspector who has received professional training to detect the welding seam by the naked eye and professional detection equipment, which not only consumes time and effort, but also has strong subjectivity in the detection results.

[0003] To solve the problems of low efficiency and subjective detection results in manual quality inspection, researchers have proposed solutions. In Chinese invention patent CN115229374A, a welding seam quality detection method for automobile body based on deep learning is disclosed. The method divides the collected three-dimensional point cloud of the automobile body into welding seams, and then inputs the segmented welding seam point cloud into a trained deep learning model to detect the welding seam quality. In Chinese invention patent specification CN115290656A, a welding seam surface defect detection and classification method based on line structured light is disclosed. The method filters and fits the original contour curve of the welding seam collected by the line structured light camera to judge whether there is a defect at the current position of the welding seam and classify and calculate the defect size. In Chinese invention patent specification CN114740006A, a welding piece surface defect online detection method based on 3D vision sensing is disclosed. The method obtains the coordinate data of the welding piece surface and then divides it into welding seams and components. Subsequently, the defects of the welding piece surface are identified according to the cross-sectional morphology of the welding seam.

[0004] The above post-weld weld surface defect detection scheme and technology obtains the picture or 3D coordinate information of the workpiece weld through a line structured light camera, a plane structured light camera or a CCD camera, detects the weld surface quality through a non-contact method, improves production efficiency, reduces production cost, and provides an effective solution for intelligent production of a factory. However, the existing technology performs simple processing and fitting according to the obtained weld picture or weld contour data, or directly pours the collected data into a neural network, so that the accuracy and precision of the detection result cannot be guaranteed. SUMMARY

[0005] In view of the defects of the prior art and the improvement demand, the present application provides a weld surface quality detection method and a computer readable storage medium, which aims to optimize the weld quality detection process by combining laser vision and deep learning, thereby improving the detection efficiency and effectively improving the accuracy of the weld surface quality detection result.

[0006] To achieve the above-mentioned purpose, according to one aspect of the present application, a weld surface quality detection method is provided, comprising the following steps:

[0007] (S1) Preprocessing the contour height data at each weld cross section collected by a line structured light camera along the length direction of the weld; the preprocessing includes removing 0 values, outliers and Gaussian noise;

[0008] (S2) Inputting the preprocessed contour height data into a pre-trained deep belief network to determine whether there is a defect at each weld cross section;

[0009] (S3) Identifying the weld cross sections with defects and continuous distribution as the distribution range of the same defect, and splicing the preprocessed contour height data in each defect distribution range into corresponding defect 3D point cloud data;

[0010] (S4) Inputting each defect 3D point cloud data into a trained weld defect detection network to obtain the type of each defect on the weld surface;

[0011] The deep belief network is used for a deep learning network, which is used to determine whether the weld cross section corresponding to the contour height data has a defect; the weld defect detection network is a deep learning network, which is used to predict the type of the defect according to the defect 3D point cloud data.

[0012] Further, in step (S1), the preprocessing further includes data alignment.

[0013] The data alignment includes calculating the motion vector of the contour height data of the current weld cross section relative to the contour height data of the previous weld cross section V Mand offset the profile height data of the current weld cross section V M align the cross section height data of the current weld with the height data of the last weld cross section.

[0014] Further, in step (S1), the preprocessing further includes: weld area cutting;

[0015] The weld area cutting includes: taking the midpoint of the left and right endpoints of the weld cross section as the midpoint of the weld area, cutting a region with a width of w in the corresponding profile height data as the weld area at the current weld cross section.

[0016] wherein, w = εw 0, w 0 represents the standard weld fusion width, ε and is a coefficient greater than 1.

[0017] Further, in step (S3), before splicing the profile height data in each defect distribution range after preprocessing into corresponding defect three-dimensional point cloud data, it further includes: removing defects with a length direction size less than a minimum detection standard.

[0018] Further, in step (S3), the weld cross sections with defects and continuous distribution are identified as the distribution range of the same defect, including:

[0019] According to L i = nδ calculate the defect length L i ;

[0020] Determine the center position of the defect along the weld length direction distribution P i ;

[0021] Determine as the distribution range of the defect along the weld length direction;

[0022] wherein, n represents the number of weld cross sections with defects and continuous distribution, δ represents the distance between the two adjacent weld cross sections; Δ L represents a preset length.

[0023] Further, the type of defect predicted by the weld defect detection network according to the defect three-dimensional point cloud data includes: surface porosity, surface depression, incomplete fusion, undercut, crack, weld bead and false defect; the false defect is used to represent that there is actually no defect in the current input defect three-dimensional point cloud data.

[0024] Further, the weld defect detection network is a PointNet++ model.

[0025] Further, the deep belief network comprises an input layer, a plurality of restricted Boltzmann machine layers and a back propagation neural network layer connected in sequence.

[0026] Further, after step (S1), the method further comprises: calculating the contour feature size of the weld according to the profile height data of the weld section after the preprocessing;

[0027] And, after step (S3), the method further comprises: calculating the three-dimensional scale information of each defect according to the three-dimensional point cloud data of the defect;

[0028] And, after step (S4), the method further comprises: judging whether the current weld quality is qualified or needs to be repaired in combination with the contour feature size of the weld, the three-dimensional scale information of each defect and the type of each defect.

[0029] According to another aspect of the present application, a computer readable storage medium is provided, comprising a stored computer program; the computer program is executed by a processor to control the device where the computer readable storage medium is located to execute the weld surface quality detection method provided by the present application.

[0030] Overall, the above technical solutions conceived by the present application can achieve the following beneficial effects:

[0031] (1) The weld surface quality detection method provided by the present application first uses the deep belief network to judge the weld section with defects from the profile height data collected by the line structured light camera, and further accurately extracts the three-dimensional point cloud of each defect from the weld profile data based on the defect judgment result, and then uses the weld defect detection network to detect the specific defect type according to the three-dimensional point cloud of each defect. Since the three-dimensional point cloud data input into the weld defect detection network removes the data less associated with the defect on the basis of containing complete defect information, the weld defect detection network can accurately obtain the specific features of the defect, effectively improve the detection accuracy of the weld defect type, and at the same time, reduce the amount of data processed by the network, effectively improve the detection efficiency. Overall, the present application optimizes the weld quality detection process by combining laser vision and deep learning, which effectively improves the detection efficiency and the accuracy of the weld surface quality detection result.

[0032] (2) In the process of collecting weld contour height data, the line structure light camera is often mounted on the welding robot. Since the data acquisition frequency of the line structure light camera is high, the end of the welding robot will shake during the movement, causing the collected data to fluctuate regularly in the height and width directions. In the preferred embodiment of the present invention, before using the depth belief network contour height data to determine the weld section with defects, 0 values, outliers and Gaussian noise will be removed in the preprocessing process. Data alignment will also be used to ensure that the contour height data of adjacent weld sections are aligned, which effectively solves the problem of regular fluctuation in the height and width directions of the collected data caused by the shaking of the welding robot end, ensuring the accurate extraction of the three-dimensional point cloud data of defects and further improving the accuracy of defect type detection.

[0033] (3) Most of the weld profile height data actually collected is from the base material area. Directly extracting the three-dimensional point cloud of the defect will result in too much information unrelated to the defect type being input into the weld defect detection network. In the preferred embodiment of the present invention, the preprocessing operation of the weld profile height data also includes weld area truncation. Specifically, only the weld area and a small amount of base material area data are truncated for subsequent detection analysis. On the basis of ensuring the integrity of the defect data, the characteristics of the weld area are more significant, and the accuracy of the three-dimensional point cloud of the defect is further improved.

[0034] (4) In a preferred embodiment of the present invention, after the defect is identified, defects with small dimensions in the length direction (less than the minimum inspection standard) are removed. These defects are often negligible in the weld quality inspection process. Removing them can make the subsequent inspection more accurate and efficient.

[0035] (5) In the preferred embodiment of the present invention, when determining the distribution length of the defect, a certain distance is extended to both ends along the weld length direction based on the cross-sectional distribution range, thereby ensuring that the complete three-dimensional point cloud of the defect is extracted, and further ensuring the accuracy of subsequent detection.

[0036] (6) In a preferred embodiment of the present invention, the defect type predicted by the weld defect detection network includes false defects, which are used to indicate that there are actually no defects in the currently input three-dimensional point cloud data of defects. This can prevent the weld area without defects from being judged as a defect by the depth belief network, thereby affecting the final accuracy of weld surface quality detection and improving the accuracy and robustness of weld surface quality detection.

[0037] (7) When performing defect detection, the present invention also calculates the external feature size of the weld and the three-dimensional dimensional information of the defect, and combines the defect type detection results to make a comprehensive judgment on the weld quality. It can comprehensively consider the specific requirements of different defects in actual applications and more accurately determine whether the weld quality is qualified. Attached Figure Description

[0038] Figure 1 A flow chart of the weld surface quality detection method provided by the embodiment of the present application is shown in the figure.

[0039] Figure 2 The profile height data collected by the line structured light camera at a certain weld section in the embodiment of the present application is shown in the figure.

[0040] Figure 3 The profile height data after preprocessing is shown in the figure. Figure 2

[0041] Figure 4 A schematic diagram of the deep belief network provided by the embodiment of the present application is shown in the figure.

[0042] Figure 5 The profile model of the weld shape feature size provided by the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0044] In the present application, the terms "first", "second", etc. (if any) in the present application and the drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence.

[0045] In order to solve the technical problem that the existing weld surface quality detection method is simple processing and fitting on the weld picture or profile data, or simply pouring the collected data into a neural network, resulting in low accuracy of the weld surface quality detection result, the present application provides a weld surface quality detection method and a computer readable storage medium, the overall idea of which is to optimize the weld quality detection process by combining laser vision and deep learning, first use a deep learning model to detect defects completely to accurately extract three-dimensional point cloud data of defects in the weld, and then use another deep learning model to accurately detect the type of defects according to the three-dimensional point cloud data of the defects, thereby improving the detection efficiency and effectively improving the accuracy of the weld surface quality detection result.

[0046] The following is an embodiment.

[0047] Embodiment 1

[0048] A weld surface quality detection method, as shown in the figure, includes the following steps: Figure 1 ​​

[0049] (S1) The contour height data of each weld section collected by the line structured light camera along the weld length direction are preprocessed;

[0050] (S2) Input the preprocessed contour height data into the pre-trained depth confidence network to determine whether there are defects at each weld section;

[0051] (S3) Identify the weld sections with defects that are continuously distributed as the distribution range of the same defect, and stitch together the pre-processed contour height data within each defect distribution range to form the corresponding three-dimensional point cloud data of the defect.

[0052] (S4) Input the three-dimensional point cloud data of each defect into the trained weld defect detection network to obtain the type of each defect on the weld surface;

[0053] Among them, the deep belief network is a deep learning network used to determine whether there are defects in the weld section corresponding to the contour height data; the weld defect detection network is a deep learning network used to predict the type of defect based on the 3D point cloud data of the defect.

[0054] In this embodiment, a line structured light camera for acquiring weld cross-sectional profile height data is integrated and fixed to the end of the welding robot, and installed within the optimal installation height range of the welding robot. This ensures that the data acquired by the line structured light camera during weld surface profile height acquisition is not incomplete due to the shaking of the welding robot. The line structured light camera can communicate with the welding robot in real time. The communication information includes camera connection signals, camera disconnection signals, camera start data acquisition signals, and stop data acquisition signals. All of these signals are sent from the welding robot's operating end to the detection system for automatic control.

[0055] Based on the above measurement system, the specific process for acquiring weld cross-sectional profile height data is as follows: the welding robot drives the line structured light camera to move along the weld length direction of the welded workpiece, and during the movement, the line structured light acquires the profile height data of the weld surface in real time. Figure 2 The image shows a schematic diagram of the profile height data at a weld section; the distance between two adjacent weld sections is also shown. δ linear speed of welding robots v m and the scanning frequency of the line structured light camera f Related, specifically:

[0056]

[0057] It is easy to understand that, in the actual measurement process, the linear running speed of the welding robot... v mAnd the scanning frequency of the line structured light camera f Should be controlled within a suitable range to avoid too large spacing between adjacent weld cross sections, which results in failure to collect complete profile data, and too small spacing between adjacent weld cross sections, which affects the efficiency of data collection; preferably, in the embodiment, v m And the rate f The setting is such that the spacing between adjacent weld cross sections is between 0.02mm and 0.04mm.

[0058] Due to environmental factors, line structured light camera and problems of the welding robot itself, there are abnormal values and noises such as 0 value and outliers in the collected original profile height data, as shown in Figure 2 Therefore, in the embodiment, the preprocessing includes:

[0059] (1) removing 0 value, which aims to remove a few points with height value of 0 in the data caused by problems of the line structured light camera itself;

[0060] (2) removing outliers, which aims to remove outliers outside the weld cross section profile line caused by environmental factors or interference such as spatter;

[0061] (3) removing Gaussian noise, which aims to remove Gaussian noise in the weld surface profile height data to make the data smoother.

[0062] Optionally, in the embodiment, the removal of 0 value and outliers is realized by a bilateral filter, and after removing 0 value and outliers, filling will be performed to ensure the integrity of the profile data, specifically, the filling value at the removed 0 value point and outlier point Is calculated by the following formula:

[0063]

[0064] Wherein, is the height value of the first non-abnormal point in the positive direction of the current filtered point, is the height value of the first non-abnormal point in the negative direction of the current filtered point, n And m All vary according to actual conditions, k is the serial number of the weld cross section profile height data collected by the line structured light camera, j represents the k th data in the j th profile height data, j The value of Q , QThe value of the parameter is related to the hardware parameters and parameter settings of the line structured light camera, and is generally not less than 1400; in the process of calculating the filling value, if the first point in the current weld cross section profile height data is a 0 value point or an outlier point, the height calculation formula of the first point is n . n The height calculation formula of the first point is ; if the current weld cross section profile height data from the m point is followed by 0 value points or outlier points, the height calculation formula of each point after the point is .

[0065] Optionally, in the embodiment, the Gaussian noise is removed by a low-pass Gaussian filter, and the specific parameters of the filter should be determined according to the actual filtering effect, so that the profile curve after filtering is as smooth as possible without affecting the overall trend of the profile.

[0066] In the process of collecting the weld profile height data, due to the high data collection frequency of the line structured light camera, the end of the welding robot will vibrate during the movement, resulting in regular fluctuations in the collected data in the height and width directions; in order to solve this problem, as an optional embodiment, the pre-processing in step (S1) of the embodiment further includes data alignment.

[0067] The data alignment includes calculating the motion vector V M of the profile height data of the current weld cross section relative to the profile height data of the last weld cross section, and offsetting the profile height data of the current weld cross section by V M , so that the height data of the current weld cross section is aligned with the height data of the last weld cross section.

[0068] Optionally, in the embodiment, the motion vector V M of the profile height data of the current weld cross section relative to the profile height data of the last weld cross section is calculated by a hierarchical search algorithm, which can be specifically represented as , , respectively, are the distances by which the profile height data of the current weld cross section is offset relative to the profile height data of the last weld cross section. The data alignment method of the embodiment is a data alignment method based on a motion estimation algorithm. The premise of realizing data alignment by using the motion estimation algorithm is that the difference between the adjacent two profile data under the high scanning frequency of the line structured light camera is not large, the target function for calculating the motion vector is the root mean square error RMSE of the profile height data, and the motion vector at the time of the minimum root mean square error RMSE is the optimal result.

[0069] Most of the actually collected weld cross-section profile height data is in the base material region, and directly extracting the three-dimensional point cloud of defects will cause too much information irrelevant to the defect type in the weld defect detection network. To solve this problem, in step (S1) of the embodiment, the preprocessing further includes: weld region intercepting;

[0070] The weld region intercepting includes: taking the midpoint of the left and right endpoints of the weld cross-section as the midpoint of the weld region, intercepting a region with a width of w in the corresponding profile height data as the weld region at the current weld cross-section;

[0071] wherein, w = εw 0, w 0 represents the standard weld width, ε is a coefficient greater than 1; in the embodiment, the width of the intercepted region is 1.5 times the standard weld width, i.e. 10mm~25mm; through this operation, only the weld region and a small amount of base material region data are intercepted for subsequent detection and analysis, which further improves the accuracy of the defect three-dimensional point cloud on the basis of ensuring the integrity of the defect data.

[0072] Figure 2 After the weld cross-section profile height data shown in Figure 3 undergoes the above preprocessing, the obtained profile height data is as shown in Figure 3 It can be clearly seen that the preprocessed weld cross-section profile height data has significantly reduced interference and noise compared to the original weld cross-section profile height data, and the characteristics of the weld region are more prominent, which can greatly reduce the learning difficulty of the defect detection network and the defect type detection network, and improve the detection ability and robustness of the detection method.

[0073] Optionally, in step (S3) of the embodiment, the selected deep belief network is as shown in Figure 4 , which includes an input layer, multiple restricted Boltzmann machine layers (RBM) and a back propagation neural network layer (BPNN), and the output vector is the result of binary classification, [0, 1] indicating the presence of defects and [1, 0] indicating the absence of defects; wherein the number of neuron nodes of the input layer is the total number of data points τ intercepted, which is related to the width w of the intercepted weld region, if the X-direction resolution of the selected line structured light camera is λ , then the total number of data points is τ : ; the number of restricted Boltzmann machine layers is preferably 3, i.e. RBM1, RBM2 and RBM3 in Figure 4 , and the result of the deep belief network is preferably [ , ~ , , ,10,2];

[0074] It is easy to understand that the deep belief network can be used for detection only after being trained by the real weld contour data set. Figure 4 The structure shown is only one preferred embodiment of the present application and should not be understood as the only limitation of the present application. The specific number of the restricted Boltzmann machine layer in the model can be adjusted according to actual needs. In some other embodiments of the present application, other binary classification networks trained can also be used as deep belief networks for detecting defects.

[0075] After scanning the complete weld, according to the principle that the weld defects are continuous in time and space, the weld sections with defects and continuous distribution belong to the same defect, so all defects in the entire weld can be detected, and the corresponding defect length is calculated as: L i = nδ ; n represents the number of weld sections with defects and continuous distribution, δ represents the distance between the adjacent two weld sections; the center of the defect along the length direction is the center position of the defect along the length direction of the weld P i In practical applications, defects with a length smaller than the minimum detection standard can be ignored, and after being removed, the subsequent detection can be more accurate and efficient. Therefore, in step (S3) of the embodiment, before splicing the pre-processed contour height data in the defect distribution range into corresponding defect three-dimensional point cloud data, it further includes: removing defects with a length smaller than the minimum detection standard.

[0076] In order to ensure that the complete defect three-dimensional point cloud is extracted, in the embodiment, after determining the size of the defect along the length direction of the weld, the extracted point cloud will extend a certain distance to both ends, and correspondingly, the extracted point cloud is data within ;

[0077] Optionally, in the embodiment, Δ L is specifically 10 mm.

[0078] Since the point cloud data of the defect is accurately extracted in the embodiment, and the factors that will affect the accuracy of the defect type detection are eliminated to the greatest extent, therefore, in the embodiment, the detection accuracy of the defect type is greatly improved.

[0079] Optionally, in the embodiment, the weld defect detection network is specifically a PointNet++ model; similarly, the PointNet++ needs to be trained by a real weld contour data set before being used for defect classification, and the data set is composed of defect three-dimensional point cloud and corresponding defect type.

[0080] It should be noted that in some other embodiments of the present application, other point cloud classification models can also be used after training as a defect type detection network.

[0081] In the embodiment, the defect types to be detected include six typical weld defects, i.e., surface porosity, surface depression, incomplete fusion, undercut, crack, and weld bead, and also include false defects. The false defects are used to indicate that there is actually no defect in the input defect three-dimensional point cloud data. The false defects are set to prevent the deep belief network from determining the weld area without defects as defects and inputting the three-dimensional point cloud of the weld area into the PointNet++ model for further judgment of the weld defects by the PointNet++ model. The addition of the false defects can greatly improve the accuracy of detecting the weld surface defects and the robustness of the method provided by the weld surface quality detection method. It should be noted that in actual application, in addition to the false defects, the remaining defect types can also be adjusted according to specific detection requirements.

[0082] Weld surface quality detection often needs to comprehensively detect defects and the forming size of the weld, and therefore, the embodiment further includes calculating the contour feature size of the weld after the contour height data of the weld section after preprocessing in step (S1).

[0083] Figure 5 The standard weld feature size contour model is shown, based on which the contour feature size of the weld calculated by the embodiment includes weld width , weld excess height , and misalignment amount ; generally, the highest point of the weld , the left end point of the weld area , and the right end point of the weld area are selected as the weld feature points, and the coordinates in the weld surface contour are respectively represented as ( , ), ( , ), and ( , ). The specific calculation method of the contour feature size of the weld is as follows:

[0084]

[0085]

[0086]

[0087] And, the embodiment further comprises: calculating three-dimensional scale information of each defect according to the three-dimensional point cloud data of the defect, including defect length, defect width, defect height / depth, defect volume, defect maximum cross-sectional area, etc., after step (S3).

[0088] And, the embodiment further comprises: judging whether the current weld quality is qualified or needs to be repaired in combination with the contour characteristic size of the weld, the three-dimensional scale information of each defect, and the type of each defect, after step (S4).

[0089] The specific detection index depends on the current welding object and the production detection index, and the determination result can also be determined by comprehensively considering multiple indexes. For example, in the scene of arc welding of an aluminum alloy roof of a high-speed rail train, in addition to paying attention to defects such as pores, collapse, and welding bumps existing on the weld surface during the detection process, the weld reinforcement is also required to be not too high, and the part of the weld exceeding the reinforcement threshold range needs to be polished and repaired. In the scene of brazing of a vehicle body, more attention is paid to the morphology and surface defects of the weld surface, and the splashing defect is not sensitive.

[0090] Based on the above detection, the embodiment can more accurately determine whether the weld quality is qualified by comprehensively considering the specific requirements of actual applications for different defects.

[0091] The weld surface quality detection method provided by the embodiment can accurately complete surface quality detection for welds under any metal base material, any butt joint form of weld, and any welding method, one of the most common scenes being an aluminum alloy butt joint arc welding weld.

[0092] In general, the embodiment realizes the detection of the workpiece weld surface quality by integrating the line structured light camera at the end of the welding robot to collect the weld surface profile height information in real time, thereby providing an automatic, efficient and low-cost detection means for the post-weld quality detection process, which can avoid the problems of inaccurate detection results, strong subjectivity of detection results and low detection efficiency caused by manual quality inspection, and has the advantages of high detection efficiency, good robustness and adaptability to different types of weld detection requirements. The embodiment optimizes the weld surface quality detection process. Specifically, the weld surface profile height data collected by the line structured light camera is input into the depth confidence network after preprocessing, and whether the current weld cross section profile has defects is judged separately. Then, the detection result is corrected according to the time and spatial continuity of the defects in the data. Finally, the defect area is extracted and input into the PointNet++ model for defect classification. The accuracy of the defect detection result obtained through this process is significantly improved, and the defect feature information obtained is more rich, which is beneficial to use as data basis for subsequent welding process optimization.

[0093] Embodiment 2

[0094] A computer readable storage medium comprising a stored computer program; the computer program is executed by a processor to control the device where the computer readable storage medium is located to execute the weld surface quality detection method provided in Embodiment 1.

[0095] Those skilled in the art will readily understand that the above description is only the preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for inspecting the surface quality of welds, characterized in that, Includes the following steps: (S1) The contour height data of each weld section collected by the line structured light camera along the weld length direction are preprocessed; the preprocessing includes removing 0 values, outliers and Gaussian noise; (S2) Input the preprocessed contour height data into the pre-trained depth confidence network to determine whether there are defects at each weld section; (S3) Identify the weld sections with defects that are continuously distributed as the distribution range of the same defect, and stitch together the pre-processed contour height data within each defect distribution range to form the corresponding three-dimensional point cloud data of the defect. (S4) Input the three-dimensional point cloud data of each defect into the trained weld defect detection network to obtain the type of each defect on the weld surface; The deep belief network is a deep learning network used to determine whether there are defects in the weld section corresponding to the contour height data; the weld defect detection network is a deep learning network used to predict the type of defect based on the three-dimensional point cloud data of the defect. In step (S1), the preprocessing further includes: data alignment; The data alignment includes: calculating the motion vector of the current weld section profile height data relative to the previous weld section profile height data. V M And offset the current weld section profile height data. V M This aligns the current weld section height data with the previous weld section height data. In step (S1), the preprocessing further includes: cutting off the weld area; The weld area segmentation includes: taking the midpoint of the left and right endpoints of the weld section as the midpoint of the weld area, and segmenting the corresponding contour height data with a width of... w The area is designated as the weld area at the current weld section. in, w = εw 0, w 0 indicates the standard weld width. ε A coefficient greater than 1.

2. The weld surface quality inspection method as described in claim 1, characterized in that, In step (S3), before stitching together the preprocessed contour height data within each defect distribution range into the corresponding three-dimensional point cloud data of the defect, the method further includes: removing defects whose length direction dimension is smaller than the minimum detection standard.

3. The weld surface quality inspection method as described in claim 2, characterized in that, In step (S3), identifying weld sections with defects that are continuously distributed as the distribution range of the same defect includes: according to L i = nδ Calculate the length of the defect L i ; The center position of the defect distribution along the weld length direction is determined by the center of the weld section containing the continuous defect. P i ; Will The distribution range of defects along the length of the weld was determined. in, n This indicates the number of weld sections that are defective and continuously distributed. δ Δ represents the distance between two adjacent weld sections. L This indicates the preset length.

4. The weld surface quality inspection method as described in claim 1, characterized in that, The weld defect detection network predicts the types of defects based on the three-dimensional point cloud data, including: surface porosity, surface depression, lack of fusion, undercut, cracks, weld beads, and false defects; the false defects are used to indicate that there are actually no defects in the currently input three-dimensional point cloud data.

5. The weld surface quality inspection method as described in claim 1, characterized in that, The weld defect detection network is a PointNet++ model.

6. The weld surface quality inspection method as described in claim 1, characterized in that, The deep belief network comprises an input layer, multiple restricted Boltzmann machine layers, and a backpropagation neural network layer connected in sequence.

7. The weld surface quality inspection method as described in claim 1, characterized in that, After step (S1), the method further includes: calculating the external feature dimensions of the weld based on the contour height data of the pre-processed weld section; Furthermore, after step (S3), the method further includes: calculating the three-dimensional scale information of each defect based on the three-dimensional point cloud data of the defects; Furthermore, after step (S4), the method further includes: combining the external shape and size of the weld, the three-dimensional dimensional information of each defect, and the type of each defect to determine whether the current weld quality is qualified or needs to be repaired.

8. A computer-readable storage medium, characterized in that, The device includes a stored computer program; when executed by a processor, the computer program controls the device containing the computer-readable storage medium to perform the weld surface quality inspection method according to claims 1 to 7.

Images