A dynamic visual detection method for welding wire extension length

Abstract

The invention belongs to the field of welding technology, and discloses a wide dynamic visual detection method for the extension length of welding wire, which uses a wide dynamic range camera to collect welding images; the welding image is input into a WE-NET model composed of an SB module, a DB module, a polymerization layer and a WN module, and the welding image is divided into two parts after being input into the WE-NET model: one part enters the DB module, and the other part enters the SB module; then, the features obtained from the SB module and the DB module are input into the polymerization layer for fusion to determine the arc area, and then the WN module is input to calculate the positions of the welding wire end point and the nozzle edge point; the longitudinal coordinate of the welding wire end point is subtracted from the longitudinal coordinate of the nozzle edge point to obtain the extension length of the welding wire. The invention can accurately measure the extension length of the welding wire, which is helpful to maintain the stability of the welding process during welding and improve the welding quality.

Description

Technical Field

[0001] The invention belongs to the technical field of welding, and in particular relates to a wide dynamic visual detection method for the extension length of a welding wire. Background Art

[0002] Wide Dynamic Range (WDR) refers to the range of brightness signal values ​​that a camera can capture when there are both high-brightness and low-brightness areas in an image. Wide dynamic range technology enables the camera to capture details of both high-brightness and low-brightness areas in an image. In high-contrast scenes, ordinary cameras often cannot clearly capture areas with large brightness differences, resulting in loss of details. Wide dynamic range technology can capture these details and avoid overexposure or underexposure.

[0003] Cameras or vision systems with wide dynamic range capabilities can provide clearer images under extreme lighting conditions. By expanding the dynamic range of the camera, both very bright and very dark areas can be clearly captured and displayed in high-contrast scenes. In the field of welding, wide dynamic range cameras or vision systems help improve the visualization quality of the welding process, enhance the clarity of welding images and the ability to integrate information, thereby improving welding quality and efficiency.

[0004] The wire extension length, that is, the length of the wire extending from the end of the welding nozzle, directly affects the stability of the welding process and the welding quality. In actual welding operations, in order to ensure the stability of the welding process and the quality of the welding formation, the wire extension length is required to be kept within a certain range. If the extension length is too long or too short, it may lead to unstable welding, increase the risk of spatter and porosity, affect the formation of the weld and the penetration of the metal, and thus reduce the strength and corrosion resistance of the welded joint. Therefore, accurately controlling and detecting the extension length of the welding wire is of great significance to ensure the accuracy and consistency of the welding operation, improve production efficiency, and ensure the reliability of the final product.

[0005] Patent publication number CN118635625A discloses a method for measuring the dry extension of welding wire in real time by using a molten pool monitoring camera. The method uses machine vision to determine the position of the straight line at the bottom of the nozzle and the characteristic point at the lower end of the welding wire, and then calculates the dry extension of the welding wire. However, compared with the deep learning method, this method has poor anti-interference and generalization capabilities and low accuracy.

[0006] Patent publication number CN116511650A discloses a model-free dry extension length control method for gas shielded welding, which establishes a regression equation through welding current, and compares the current parameters at different positions with the current collected during the automatic welding process as the reference value, so as to achieve the purpose of controlling the dry extension length. However, the decision of this method depends on indirect current parameters, rather than actual visual data, and it is difficult to accurately measure the extension length of the welding wire. Summary of the invention

[0007] In order to ensure the stability of the welding process and the quality of the welding formation, the present invention provides a wide dynamic visual detection method for the extension length of the welding wire.

[0008] The invention provides a wide dynamic visual detection method for welding wire extension length, which uses a wide dynamic range camera to collect welding images; the welding images are input into a WE-NET model composed of an SB module, a DB module, an aggregation layer and a WN module, wherein the DB module is composed of three parts connected in sequence, each of which includes a convolution layer, a batch normalization layer and an activation function layer; the SB module is composed of a downsampling module, a depthwise separable convolution module and an activation function in sequence, and the depthwise separable convolution module includes a convolution layer and a depthwise separable convolution module; the welding image is divided into two parts after being input into the WE-NET model, one part of which enters the DB module, and the other part enters the SB module; then, the features obtained from the SB module and the DB module are input into the aggregation layer for fusion to determine the arc area, and then input into the WN module to calculate the positions of the welding wire end point and the nozzle edge point;

[0009] The WN module processes the acquired arc area; first, the centroid of the arc area is calculated, and then, starting from the centroid, rows containing zero elements are searched row by row along the negative direction of the Y axis. When row R containing zero elements is found, the midpoint of the row is the welding wire end point; starting from the welding wire end point, the grayscale value change is analyzed along the negative direction of the Y axis, and the roof edge model feature is searched according to the grayscale value change. The midpoint of the roof edge model feature is the nozzle edge point; the ordinate of the welding wire end point is subtracted from the ordinate of the nozzle edge point to obtain the welding wire extension length.

[0010] Further preferably, the welding image is preprocessed before being input into the WE-NET model, and the preprocessing includes one or both of denoising and data enhancement.

[0011] Further preferably, the formula for calculating the centroid of the arc area is as follows:

[0012] ;

[0013] ;

[0014] in, and are the horizontal and vertical coordinates of the center of mass, is the pixel coordinate, It is a pixel classification function used to segment pixels.

[0015] Further preferably, the coordinates of the welding wire end point are:

[0016] ;

[0017] ;

[0018] in, is the horizontal coordinate of the welding wire end point and is the ordinate of the wire end point, The vertical coordinate of the row.

[0019] Further preferably, the nozzle edge point coordinates are:

[0020] ;

[0021] ;

[0022] in, and are the boundaries of the roof edge model features, is a binary function, is the abscissa of the nozzle edge point, is the ordinate of the nozzle edge point, h is the pixel position iteration variable when traversing the roof edge model features, is the starting gray value.

[0023] The method of the present invention accurately divides the arc area through the WE-NET model, determines the end point of the welding wire and the edge point of the welding gun nozzle, and then calculates the extension length of the welding wire. The method has the following advantages: by accurately calculating the extension length of the welding wire and then adjusting the CTWD, the consistency of the arc length and the heat input can be ensured when welding the workpiece surface with height changes, thereby maintaining the stability of the welding process and improving the welding quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a flow chart of the present invention.

[0025] Figure 2 This is a structural diagram of the WE-NET model of the present invention.

[0026] Figure 3 This is a structural diagram of the DB module of the present invention.

[0027] Figure 4 This is a structural diagram of the SB module of the present invention.

[0028] Figure 5 Schematic diagram of welding wire extension length. Specific implementation methods

[0029] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is described clearly and completely below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0030] A method for dynamic visual detection of the welding wire extension length in this embodiment is shown in the flowchart as follows: Figure 1 As shown in the figure. First, the welding image during the welding process is collected by a wide dynamic range camera. Then, the collected welding image is preprocessed; the welding image is input into the WE-NET model composed of the SB module, DB module, polymerization layer and WN module to determine the precise position of the welding wire end point and the nozzle edge point. Finally, according to the precise position of the determined welding wire end point and the nozzle edge point, the straight-line distance from the welding wire end point to the nozzle edge point, i.e., the welding wire extension length, is calculated.

[0031] The following is a detailed description of each part:

[0032] 1. Collect welding images during the welding process through a wide dynamic range camera, and pre-process the collected welding images: A large number of experiments have shown that in GMAW, there is a certain relationship between the illumination of the arc and the welding current; the greater the welding current, the stronger the arc and the greater the illumination. Therefore, it is difficult for ordinary industrial cameras to capture clear images in a high-contrast welding environment, and they cannot provide information such as the arc and molten pool status. Therefore, it is necessary to use a wide dynamic range camera for welding image acquisition. Since the image contains interference such as arc light and sputtering, the welding image is first denoised by filtering to gradually reduce interference factors. Then, the welding image is data enhanced to enhance the generalization ability.

[0033] 2. If Figure 2 As shown, the welding image is input into the WE-NET model composed of the SB module, DB module, polymerization layer and WN module to determine the precise positions of the wire end point and the nozzle edge point: after being input into the WE-NET model, the welding image is divided into two parts, one part enters the DB module and the other part enters the SB module.

[0034] like Figure 3 As shown in the figure, the DB module consists of three parts connected in sequence, each of which includes a convolution layer, a batch normalization layer, and an activation function layer. In each part, the input features first undergo a convolution operation:

[0035] ;

[0036] in, is the output of the convolutional layer, is the input feature, is a 3×3 convolution kernel, is the offset, is the pixel horizontal coordinate, is the pixel ordinate, Represents a convolution operation.

[0037] Next, the output of the convolutional layer is batch normalized to improve standardization, improve the convergence speed and stability of the model. The formula is as follows:

[0038] ;

[0039] in, is the normalized output, is the scaling factor, is the data mean, is the data variance, is a constant.

[0040] Finally, the ReLU activation function is used to introduce nonlinear features to enhance the model's ability to learn complex features. The stride of the convolutional layer in the first part is set to 2, and the stride of the rest is 1, so that the DB module can better capture the low-level details in the image.

[0041] like Figure 4 As shown in the figure, the SB module is composed of a downsampling module, a depthwise separable convolution module, and an activation function. The depthwise separable convolution module includes a convolution layer and a depthwise separable convolution module. The input features are first downsampled:

[0042] ;

[0043] in, is the input feature, represents the convolution kernel, is the convolution kernel sliding interval, and are the row index and column index of the convolution kernel at the current position respectively. After the downsampling operation, the convolution layer is used to obtain the compressed feature map, and then a depth-separable convolution is performed to obtain the restored feature map. The formula is:

[0044] ;

[0045] in, is the restored feature map output by the depthwise separable convolution, and is the stride, is the convolution kernel of the depthwise separable convolution. Finally, after being fused with the forward features, the features are output through the ReLU activation function.

[0046] Among them, the DB module can be used to extract low-level detail information in welding images and generate high-resolution feature representations due to its wide channel characteristics; the SB module can obtain high-level contextual information due to its narrow channel characteristics, and has a faster operation speed due to its small channel capacity and fast sampling speed.

[0047] Since the features obtained by the DB module and the SB module are isolated from each other, the aggregation layer is required to complement the complementary detail information and semantic information. Semantic information provides attention to different weights for detail information; detail information makes up for the loss of its own detail information for semantic information. The two types of information are exchanged frequently through the aggregation layer to obtain the precise arc area.

[0048] After obtaining the precise arc area, the WN module processes the obtained precise arc area. First, the centroid of the arc area is calculated using the following formula:

[0049] ;

[0050] ;

[0051] in, and are the horizontal and vertical coordinates of the center of mass, is the pixel coordinate, It is a pixel classification function used to segment pixels.

[0052] Next, starting from the centroid, search for rows containing zero elements along the negative direction of the Y axis. When a row R containing zero elements is found, the midpoint of the row is calculated to be the welding wire end point to be found. The coordinates of the welding wire end point are:

[0053] ;

[0054] ;

[0055] in, is the horizontal coordinate of the welding wire end point and is the ordinate of the wire end point, The vertical coordinate of the row.

[0056] Since the nozzle surface has obvious metallic luster in the acquired welding image, under arc lighting conditions, the arc reflected from the edge of the nozzle will form a curve with a higher grayscale value, and its grayscale value shows a trend of increasing first, then decreasing, and finally tending to the original grayscale value. Based on this characteristic, starting from the end point of the welding wire, the grayscale value change is first analyzed along the negative direction of the Y axis:

[0057] ;

[0058] in, Indicates Row and The grayscale difference of the row, For the The gray value of the row, For the Next, search for the roof edge model feature based on the change in gray value. After finding the roof edge model feature, calculate the midpoint coordinates of the roof edge model feature, which is the nozzle edge point coordinates:

[0059] ;

[0060] ;

[0061] in, and are the boundaries of the roof edge model features, is a binary function, is the abscissa of the nozzle edge point, is the ordinate of the nozzle edge point, h is the pixel position iteration variable when traversing the roof edge model features, is the starting gray value.

[0062] 3. By Figure 5 It can be seen that after obtaining the coordinates of the welding wire end point 1 and the nozzle edge point 2, the vertical coordinate of the welding wire end point The vertical coordinate of the nozzle edge point Subtract to get the wire extension length :

[0063] ;

[0064] in, is the ordinate of the wire end point, is the ordinate of the nozzle edge point.

[0065] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0066] 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 fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for dynamic visual detection of welding wire extension length, characterized in that: A wide dynamic range camera is used to collect welding images; the welding images are input into a WE-NET model composed of an SB module, a DB module, an aggregation layer and a WN module, wherein the DB module is composed of three parts connected in sequence, each of which includes a convolution layer, a batch normalization layer and an activation function layer; the SB module is composed of a downsampling module, a depthwise separable convolution module and an activation function in sequence, and the depthwise separable convolution module includes a convolution layer and a depthwise separable convolution module; after the welding image is input into the WE-NET model, it is divided into two parts: one part enters the DB module and the other part enters the SB module; then, the features obtained from the SB module and the DB module are input into the aggregation layer for fusion to determine the arc area, and then input into the WN module to calculate the positions of the wire end point and the nozzle edge point; The WN module processes the obtained arc area; first, the centroid of the arc area is calculated, and then, starting from the centroid, rows containing zero elements are searched row by row along the negative direction of the Y axis. When row R containing zero elements is found, the midpoint of the row is the welding wire end point; starting from the welding wire end point, the grayscale value change is analyzed along the negative direction of the Y axis, and the roof edge model feature is searched according to the grayscale value change. The midpoint of the roof edge model feature is the nozzle edge point; the ordinate of the welding wire end point is subtracted from the ordinate of the nozzle edge point to obtain the welding wire extension length.

2. A method for dynamic visual detection of welding wire extension length according to claim 1, characterized in that: Before the welding image is input into the WE-NET model, the welding image is preprocessed, and the preprocessing includes one or both of denoising and data enhancement.

3. A method for dynamic visual detection of welding wire extension length according to claim 1, characterized in that: The formula for calculating the centroid of the arc area is as follows: ； ； in, and are the horizontal and vertical coordinates of the center of mass, is the pixel coordinate, It is a pixel classification function used to segment pixels.

4. A method for dynamic visual detection of welding wire extension length according to claim 3, characterized in that: The coordinates of the wire end point are: ； ； in, is the horizontal coordinate of the welding wire end point and is the ordinate of the wire end point, The vertical coordinate of the row.

5. A method for dynamic visual detection of welding wire extension length according to claim 4, characterized in that: The coordinates of the nozzle edge point are: ； ； in, and are the boundaries of the roof edge model features, is a binary function, is the abscissa of the nozzle edge point, is the ordinate of the nozzle edge point, h is the pixel position iteration variable when traversing the roof edge model features, is the starting gray value.

6. A method for dynamic visual detection of welding wire extension length according to claim 1, characterized in that ,The convolutional layer processing process of the DB module is as follows: ; in, is the output of the convolutional layer, is the input feature, is a 3×3 convolution kernel, is the offset, is the pixel horizontal coordinate, is the pixel ordinate, Represents a convolution operation.

7. A method for dynamic visual detection of welding wire extension length according to claim 6, characterized in that ,In the SB module, the input features are first subjected to downsampling operation: ； in, is the input feature, represents the convolution kernel, is the convolution kernel sliding interval, and are the row index and column index of the convolution kernel at the current position respectively; after the downsampling operation, the convolution layer is used to obtain the compressed feature map, and then a depth-separable convolution is performed to obtain the restored feature map, the formula of which is: ； in, is the restored feature map output by the depthwise separable convolution, and is the stride, It is the convolution kernel of the depthwise separable convolution; finally, after being fused with the forward features, the features are output through the ReLU activation function.

Images