An online detection method and detection device for laser deep penetration welding defects

Abstract

The present invention relates to an online detection method for laser deep fusion welding defects. During the welding process, the image of the shape contour of the small holes and through holes in the molten pool formed during welding is extracted and stored in real time by an imaging system, and based on the pre-trained relationship model between the image contour features of the small holes and through holes and the fusion welding defects, the online detection of the fiber laser welding quality is realized. The establishment of the relationship model between the image contour features of the small holes and through holes and the welding quality is based on the detection of a large amount of sample data, and the characteristic data of the small holes and through holes are trained using a hidden Markov process. The present invention also relates to a detection device applied to the above-mentioned online detection method. The present invention realizes online defect detection and quality analysis in the process of fiber laser welding, which is convenient for discovering various defects that may affect the welding quality during the welding process, and reduces the time and cost of later manual detection or X-ray detection.

Description

Technical Field

[0001] The present invention relates to the technical field of laser deep penetration welding defect detection, and in particular to an online detection method and a detection device for laser deep penetration welding defects. Background Art

[0002] As one of the most ideal heat sources for high-speed deep-melting welding, laser has developed rapidly in industrial processing and modern intelligent manufacturing. Laser welding is one of the important aspects of the application of laser material processing technology. Compared with arc welding, the advantage of laser welding is that the laser beam can be focused by the lens to obtain a small point, so as to accurately locate the welding position. As an efficient and precise welding method, it has been widely used in the welding process of precision workpieces. In addition, it also has the advantages of large depth and width, high energy density, welding of heterogeneous materials and easy control. However, due to the cross-action of various process parameters, defects are inevitable in the process of laser welding, especially incomplete penetration and penetration. It is generally believed that keyhole behavior and instability are caused by the abnormal dynamic flow of metal around the keyhole, which is one of the reasons for welding quality defects. Therefore, monitoring the welding process through machine vision and image processing technology and extracting important information for diagnosing welding quality during the welding process are also research hotspots in the field of automatic welding control. The common method adopted by the prior art is to calculate the total area or approximate diameter of the keyholes by extracting the keyholes as an evaluation criterion to distinguish welding defects. This method does not take into account the complexity of the small hole contour structure, and its accuracy and robustness are difficult to meet the requirements of actual detection. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide an online detection method for laser deep penetration welding defects, so as to realize the requirement of discovering various welding defects during the welding process.

[0004] In order to solve the above problems, the technical solution adopted by the present invention is as follows:

[0005] An online detection method for laser deep penetration welding defects comprises the following steps:

[0006] S1. During the welding process, the image of the molten pool holes and penetration holes formed during welding is extracted and stored in real time through the imaging system;

[0007] S2. By collecting welding images when different types of welding defects occur, extracting their contour features to form different data sets, calculating the curvature descriptor of the boundary contour as the feature of welding quality, and after collecting enough welding defect data, obtaining the required process parameters through data set training, and establishing a relationship model between contour features and welding quality defects;

[0008] S3. Based on the relationship model established in step S2, the quality of welding is judged, potential defects are discovered, and online detection of fiber laser welding quality is achieved.

[0009] Furthermore, in step S1 of the above scheme, the extraction of the shape contour image of the molten pool small hole and the through hole includes the following steps:

[0010] S11, start and initialize the device;

[0011] S12, collecting welding images through an imaging system;

[0012] S13, preprocessing the collected images;

[0013] S14, segmenting the welding pool image into a background area, a keyhole area and an area to be processed by grayscale projection;

[0014] S15, using Poisson extinction method to separate the contour boundaries of small holes and penetration holes in the area to be processed;

[0015] S16, binarizing the target image;

[0016] S17, extracting the outlines of small holes and through holes.

[0017] Further, the process of collecting welding images by the imaging system in step S12 includes: first, taking the leftmost end point of the penetration hole as the starting point, extracting points in a clockwise direction, and then extracting the molten pool holes in the same way to form a point list;

[0018] The process of preprocessing the collected image in step S13 includes filtering the coordinates using a Gaussian filter to reduce abnormal burrs.

[0019] Furthermore, in step S2, the characteristic data of the small holes and through holes in the molten pool are trained using a hidden Markov process to establish a relationship model, and the proportion of small holes and through holes in distinguishing welding defects is included in the training process.

[0020] The present invention also provides a detection device for an online detection method of laser deep penetration welding defects, comprising a mechanical arm, a laser welding head, a focusing lens or a beam expander, a first laser, a second laser, an image acquisition system, a computer, a transmission device, a welding base material, a gas nozzle, a gas cylinder and a controller;

[0021] The laser welding head is installed on the mechanical arm, the focusing lens or beam expander is installed on the laser welding head to control the size of the laser spot, and the first laser is connected to the laser welding head through an optical fiber;

[0022] The image acquisition system includes a high-speed camera, a light attenuation plate, a filter and a protective lens installed on the lens of the high-speed camera, and the welding base material is clamped and controlled by a transmission device;

[0023] The second laser, high-speed camera and gas nozzle are all located near the welding base material;

[0024] The mechanical arm and the laser welding head are connected to a controller, and the first laser and the high-speed camera and the controller are connected to a computer.

[0025] Furthermore, a camera bracket mounting seat is fixed on the laser welding head, one end of the camera bracket mounting seat is fixedly connected with an inverted L-shaped arm, the horizontal part of the L-shaped arm is connected to the camera bracket mounting seat, and the high-speed camera is installed at the lower end of the vertical part of the L-shaped arm.

[0026] Furthermore, in order to obtain clear images of the molten pool holes and penetration holes and reduce the interference of welding photo-induced plasma / metal vapor above the holes, and because the photosensitive chip of the shooting camera is more sensitive in this band, the second laser is a semiconductor laser source with a central wavelength of 808nm.

[0027] Furthermore, the first laser is a 1064nm continuous wave fiber laser.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] 1. The welding images collected by the present invention can quickly and accurately separate the shapes of small holes and through holes by using grayscale projection distribution and Poisson extinction method; the small hole contour features extracted by the curvature descriptor can simultaneously characterize the geometric features and geometric lengths of the small holes and through holes in the molten pool, thereby improving the accuracy of welding defect prediction; the proposed hidden Markov model based on contour curvature can simultaneously predict the welding quality and welding defect type according to the extracted small hole and through hole features, thereby improving the accuracy of welding defect prediction.

[0030] 2. The detection method and device of the present invention realize online defect detection and quality analysis in the process of fiber laser welding, and can be applied to the online detection of non-contact machine vision welding quality in the case of deep penetration welding, which is convenient for finding various defects that may affect the welding quality during the welding process, including pores, melt-through, incomplete welding, etc., reducing the time and cost of later manual inspection or X-ray inspection. At the same time, the test method can understand the matching relationship between the molten pool holes and penetration holes and welding defects, which helps to explain the abnormal metal flow in and out of the holes when defects occur, the behavior of the holes, and the instability of the holes.

[0031] 3. In the detection device of the present invention, the laser welding head and the high-speed camera are both installed on the camera adjustment bracket, so that the image acquisition and welding can be performed synchronously.

[0032] In summary, compared with the existing technical solutions, the present invention can improve the accuracy and robustness of defect detection, which is of great significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a schematic diagram of the positions of the molten pool A, the molten pool hole B and the penetration hole C in the collected image involved in the present invention.

[0034] Figure 2 It is a schematic diagram of the detection device involved in the present invention.

[0035] Figure 3 The present invention is a flowchart for extracting the contour images of the molten pool holes and penetration holes during the welding process.

[0036] Figure 4a is an example picked from the dataset used to train the model.

[0037] Figure 4b is an example selected from the test dataset.

[0038] Figure 4c Schematic diagram of using the model to diagnose welding defects.

[0039] Figure 5 Schematic diagram of the structure of the camera adjustment bracket.

[0040] Figure 6 It is a schematic diagram of the decomposition of the image acquisition system. DETAILED DESCRIPTION

[0041] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. According to the following description, the purpose, technical solutions and advantages of the present invention will become clearer. It should be noted that the described embodiments are preferred embodiments of the present invention, rather than all embodiments.

[0042] An online detection method for laser deep penetration welding defects comprises the following steps:

[0043] S1. During the welding process, the imaging system is used to monitor the Figure 1 The shape contour images of the molten pool small hole B and the penetration hole C shown are extracted and stored in real time;

[0044] S2. By collecting welding images when different types of welding defects occur, extracting their contour features to form different data sets, calculating the curvature descriptor of the boundary contour as the feature of welding quality, and after collecting enough welding defect data, obtaining the required process parameters through data set training, and establishing a relationship model between contour features and welding quality defects;

[0045] S3. Based on the relationship model established in step S2, the quality of welding is judged, potential defects are discovered, and online detection of fiber laser welding quality is achieved.

[0046] In step S1 of the above technical solution, the extraction of the shape contour image of the molten pool small hole and the penetration hole includes the following steps:

[0047] S11, start and initialize the device;

[0048] S12, collecting welding images through an imaging system;

[0049] S13, preprocessing the collected images;

[0050] S14, segmenting the welding pool image into a background area, a keyhole area and an area to be processed by grayscale projection;

[0051] S15, using Poisson extinction method to separate the contour boundaries of small holes and penetration holes in the area to be processed;

[0052] S16, binarizing the target image;

[0053] S17, extracting the outlines of small holes and through holes.

[0054] After obtaining the contour coordinates of the small holes and penetration holes, training and diagnosis are performed. By collecting welding images when different types of welding defects occur, their contour features are extracted to form different data sets, and the curvature descriptor of the boundary contour is calculated as the feature of welding quality. After collecting enough welding defect data, the required process parameters are obtained through data set training to establish a model. In the subsequent welding process, this model can be used to judge the quality of welding and discover potential defects.

[0055] First, starting from the leftmost end of the penetration hole, points are extracted in a clockwise direction, and then the small holes are extracted in the same way to form a point list. Then the coordinates are filtered using a Gaussian filter to reduce abnormal burrs. The curvature descriptor can be calculated based on the relationship between adjacent points in the table. By taking three adjacent coordinate points X on the point list i-1 , X i and X i+1 , defined in X i The curvature descriptor at is:

[0056]

[0057] After calculating the curvature sequence, the dense sequence can be sampled to reduce the complexity of the model. Sampling can be done by using spatial equidistant sampling or local extreme density method. Through experiments, these sampling methods retain the characteristic information of the high curvature parts of small holes and through holes, and have little effect on the classification results as long as the appropriate parameters are selected.

[0058] The model of the relationship between the contour features and the welding quality defects is established by training the feature data of the keyhole and its penetration hole using the hidden Markov process, which includes the following definitions:

[0059] Hidden state set S = {S1, S2, ..., S N}, N represents the number of states.

[0060] 1) State transition probability matrix A = {a ij}, indicating the state S i Transition to state S j probability.

[0061] 2) Observation sequence set O = O1O2…O T , T represents the observation sequence number. The observation sequence is extracted from the rate sequence by sampling or directly generated from the curvature sequence.

[0062] 3) Observation symbol matrix B = {b j (k)}, b j (k) is in the system state S j The probability of the sequence k observed under this condition t. j (k) = P[O k at t|q t =S j ], 1≤j≤N, where b j (k)≥0 and

[0063] 4) Initial state probability distribution π = {π i}, π i is the model λ in the system state S i The probability of i =P[q1=S i ],1≤i≤N

[0064] For simplicity of expression and uniqueness of the model, the triple λ=(A, B, π) is used to represent the hidden Markov model.

[0065] In order to apply the model to actual monitoring, some standard algorithms are used directly in the specific operation process. With the help of forward variables, the observation sequence O = O1O2…O is calculated. TThe probability of occurrence of . The theory of dynamic programming algorithm is used to deal with the state sequence Q=q1q2…q when P(O,Q|λ) is the maximum value. T . However, the shape characteristics of keyholes and through holes formed during the welding process have different weights for different defects. Moreover, some parts of the contour features in keyholes and through holes are more important than other parts in defect classification and quality evaluation. Generally speaking, in a stable welding process, the shape of the through hole will not change significantly, while the keyhole may experience normal fluctuations. Therefore, the tolerance of the two to shape changes is also different. In order to solve the problem in this model, the solution adopted by the present invention is to embed the required global parameters into the hidden Markov model. The embedding is completed by adopting a weighted likelihood algorithm with minimum error classification, which achieves the embedding of the required variables. The weight parameters can be iterated using the generalized probability descent (GPD) theory. For the observation sequence O=O1O2…O T And the state sequence Q = q1q2…q T , the probability of O can be modeled using the triple λ = (A, B, π) of the Hidden Markov Model:

[0066]

[0067] Each observation is weighted separately, the maximum likelihood criterion is expressed as a forward variable, and then the maximum likelihood inference is obtained by performing logarithmic operations on both sides of the equation:

[0068]

[0069] The next step is to t The weight ratio w j Introduced into the formula. j Represents C j The weight function of the class provides additional recognition of the classification. The new discriminant function is defined as:

[0070]

[0071] From the above function, it can be deduced that the weight of each state in the defect feature classification needs to be calculated. Considering the weight estimation of each state sequence, the amount of calculation is too large for any theory using weight estimation, and it requires unrealistic iterations and time to complete. Here, the present invention introduces a window function to segment the contours of the full penetration hole and the keyhole. Assuming that t=1 to t1 in the vector sequence belongs to the penetration hole, and t1 to T belongs to the small hole, the new discriminant function can be expressed as:

[0072]

[0073] The coefficient h i,j Indicates height, μ i,jAdjustment position, s i,j Indicates j th Class Window th The generalized probability descent method (GPD) is used to iterate the three parameters to minimize the classification error.

[0074] Figure 4a is an example selected from the dataset used to train the model. The hollow dots are the number of states. Figure 4b is an example selected from the test data set, and the hollow circle is the result obtained by matching the training data set; Figure 4a and Figure 4b As shown, from 4a and Figure 4b The starting point starts at the left end of the full penetration hole and grows in a clockwise direction, and the small holes are followed in the same way. Figure 4c is the matching result with the test sample. It can be seen from the figure that when the similarity is low, local shape changes will occur between the two.

[0075] Combination Figure 2 As shown, a detection device for laser deep penetration welding defects is applied to the above-mentioned online detection method, including a robot arm 1, a laser welding head 2, a focusing lens or a beam expander 3, an optical fiber 4, a first laser 5, a second laser 6, an image acquisition system, a computer 10, a transmission device 11, a welding base material 12, a gas nozzle 13, a gas cylinder 14 and a controller 15.

[0076] The laser welding head 2 is mounted on the robot arm 1, the focusing lens or beam expander 3 is mounted on the laser welding head 2 to control the size of the laser spot, and the first laser 5 is connected to the laser welding head 2 via an optical fiber 4. Figure 6 As shown, the image acquisition system includes a high-speed camera 7, a light attenuation sheet 7a installed on the lens of the high-speed camera, a filter 8 and a protective lens 9. The welding base material 12 is clamped and controlled by a transmission device 11; the second laser 6 provides lighting for image acquisition, and the gas nozzle 13 and the gas cylinder 14 are responsible for providing protective gas for the welding process. The second laser 6, the high-speed camera 7 and the gas nozzle 13 are all located near the welding base material 12; the mechanical arm 1, the first laser 5 and the controller 15 are connected, and the high-speed camera 7 and the controller 15 are connected to the computer 10.

[0077] In order to obtain clear images of the molten pool holes and penetration holes and reduce the interference of welding photo-induced plasma / metal vapor above the holes, and because the photosensitive chip of the shooting camera is more sensitive in this band, the second laser 6 is a semiconductor laser source with a central wavelength of 808nm.

[0078] Combination Figure 5As shown, in order to realize the synchronization of image acquisition and welding, the laser welding head 2 is installed on the robot arm 1, and a camera bracket mounting seat 16a is fixed on the laser welding head 2. An inverted L-shaped arm 16b is fixed to one end of the camera bracket mounting seat 16a, and the horizontal part of the L-shaped arm 16a is connected to the camera bracket mounting seat 16a. The high-speed camera 7 is installed at the lower end of the vertical part of the L-shaped arm 16b.

[0079] refer to Figure 6 As shown, the image acquisition system includes a high-speed camera 7, a light attenuation plate 7a installed on the lens of the high-speed camera, a filter 8 and a protective mirror 9. Among them, the light attenuation plate is a full-band attenuation, which is used to reduce the overall light intensity of the plasma entering the camera's photosensitive components and protect the photosensitive components from being damaged by excessive light intensity. The filter is selected based on the auxiliary light source; this device uses an 808nm semiconductor laser source to provide illumination. In order to filter out the interference of plasma / metal vapor strong light radiation, the half-wave width should be small, the peak transmittance should be high, the cut-off range should be wide, and the cut-off depth should be large. The filter uses a narrow-band filter with a central wavelength of 808nm. The protective mirror 9 is used to protect the inner optical lens to prevent it from being damaged by splashing.

[0080] The present invention can adopt the following equipment in the implementation process:

[0081] (1) One desktop host: Dell / Dell D24M, Intel platform, Windows 10, large chassis, chipset: Z370, independent graphics card, integrated sound card, 1000Mbps Ethernet card. CPU speed: 4.7GHz, 9th generation Intel Core i7 processor, eight cores, L2 cache: 12MB, graphics chip: GTX1070, video memory capacity: independent 8GB. Memory: capacity 16GB, speed: DDR4; number of slots 4, maximum supported capacity 32GB. Hard disk: capacity 1TB, type SATA serial, speed 7200 rpm. Image acquisition card: CoaXPress AS-FBD-4XCXP6-2PE8; bandwidth: 25Gbps.

[0082] (2) 1064nm continuous wave fiber laser: IPG YLS-6000K 1064nm.

[0083] (3) Laser welding head: 1 IPG P30-010595 FLW D50.

[0084] (4) Protective gas: Argon

[0085] (5) One CoaXPress CP70-12-M-167 black and white CCD high-speed camera, resolution: 1280x960, shutter speed: 50 μs, frame rate: 4000fps, interface: FD / FL. 808nm filter 1 piece, lens protection sheet 1 piece.

[0086] (6) One 808nm auxiliary lighting source, power: 10W.

[0087] (7) One set of large robotic arm and control platform.

[0088] (8) One set of welding base material clamping and transmission device.

[0089] (9) A number of focusing lenses, a number of beam expanders, a gas nozzle and a number of fixing devices.

[0090] When the testing device of the present invention is applied to the online detection of laser deep penetration welding defects, the robot arm 1 is adjusted to locate the welding position. First, a sufficient number of welding images, including images when defects are formed, are collected by a high-speed camera 7 and a computer 10 with a high-speed acquisition card. Then, the image is effectively segmented by the grayscale projection feature of the welding image, and the contour boundaries of the small holes and the penetration holes are separated by the Poisson extinction method. By collecting welding images when different types of welding defects are generated, their contour features are extracted to form different data sets, and the curvature descriptor of the boundary contour is calculated as a feature of the welding quality. After collecting enough welding defect data, the required process parameters are obtained through data set training to establish a model. In the subsequent welding process, this model can be used to judge the quality of the welding and find potential defects.

[0091] The above is only a description of the preferred embodiment of the present invention, and is not intended to limit the scope of protection of the present invention. Obviously, any technician familiar with the field can easily think of replacements or changes to obtain other embodiments based on the above embodiments, which should all be included in the scope of protection of the present invention.

Claims

1. An online detection method for laser deep penetration welding defects, characterized in that: The following steps are involved: S1. During the welding process, the image of the molten pool holes and penetration holes formed during welding is extracted and stored in real time through the imaging system; S2. By collecting welding images when different types of welding defects occur, extracting their contour features to form different data sets, calculating the curvature descriptor of the boundary contour as the feature of the welding quality, and after collecting enough welding defect data, obtaining the required process parameters through data set training, and establishing a relationship model between contour features and welding quality defects. Specifically, the characteristic data of small holes and through holes in the molten pool are trained using a hidden Markov process to establish a relationship model, and the training process includes the proportion of small holes and through holes in the identification of welding defects; S3. Based on the relationship model established in step S2, the quality of welding is judged, potential defects are discovered, and online detection of fiber laser welding quality is achieved.

2. The online detection method for laser deep penetration welding defects according to claim 1 is characterized in that: In step S1, the extraction of the shape contour image of the molten pool small hole and the penetration hole includes the following steps: S11, start and initialize the device; S12, collecting welding images through an imaging system; S13, preprocessing the collected images; S14, segmenting the welding pool image into a background area, a keyhole area and an area to be processed by grayscale projection; S15, using Poisson extinction method to separate the contour boundaries of small holes and penetration holes in the area to be processed; S16, binarizing the target image; S17, extracting the outlines of small holes and through holes.

3. The online detection method for laser deep penetration welding defects according to claim 2 is characterized in that: The process of collecting welding images by the imaging system in step S12 includes: First, take the leftmost end of the penetration hole as the starting point and extract points in a clockwise direction. Then, the small holes in the molten pool are extracted in the same way to form a point list. The process of preprocessing the collected image in step S13 includes filtering the coordinates using a Gaussian filter to reduce abnormal burrs.

Images