Electric signal and molten pool image collaborative all-position welding penetration online detection method

Abstract

The invention discloses a full-position welding penetration online detection method based on cooperation of an electric signal and a molten pool image. A pipeline is subjected to full-position welding under cooperation of double robots; preliminary segmentation is conducted on the basis of the welding position, segmentation is conducted again at the angle of 5-10 degrees, a laser scanner is adopted to scan a pipeline groove to obtain the groove size of each segment, and self-adaptive setting of welding process parameters is achieved on the basis of a process knowledge base and a welding parameter adjusting model; fusing the molten pool image and the electric signal time domain characteristic value, and performing discrimination through a neural network prediction model; and obtaining weld pool size characteristics based on the weld pool image, calculating a reference weld pool oscillation frequency based on a reference weld pool oscillation frequency model, performing spectral analysis on an upper envelope signal of the voltage signal, calculating an actual weld pool oscillation frequency, and comparing the actual weld pool oscillation frequency with the reference weld pool oscillation frequency to judge whether the weld is penetrated or not. According to the invention, all-position welding forming defect monitoring of the robot is realized.

Description

Technical Field

[0001] The invention relates to the field of robot automatic welding, in particular to an all-position welding penetration online detection method in which an electric signal and a molten pool image are coordinated. Background Art

[0002] At present, the full-position welding of petrochemical process pipelines is mainly done manually or by automated special machines. However, this method requires experienced technical workers to adjust the welding process in real time according to the pipeline groove and position. Therefore, there are problems such as large workload, high labor cost, low welding efficiency and unstable welding quality. The advantages of welding robots, such as their ability to work for a long time, stable welding quality and high efficiency, have made it a major trend for welding robots to replace manual labor.

[0003] However, there are many challenges in the application of welding machine technology. Currently, pipeline welding is mainly carried out through manual teaching. However, this method has high requirements on the workpiece itself and processing, and it will not respond quickly to problems that arise during the welding process and make timely adjustments. Therefore, in the actual on-site construction of the full-position automatic welding of petrochemical process pipeline robots, due to the influence of external environment, assembly and other factors, it is easy to cause forming defects during the welding process. In order to solve the above problems, it is necessary to monitor and feedback the full-position automatic welding process of the petrochemical process pipeline robot in real time, and take corresponding control measures to ultimately improve welding efficiency and reduce the waste of time and materials.

[0004] Nowadays, people generally use molten pool vision and computers to carry out intelligent real-time monitoring of the welding process. Generally, a mapping model is established between welding defects and molten pool images directly through convolutional neural networks, so as to realize the monitoring of pipeline welding defects. However, this method is suitable for welding with unchanged welding position type and groove size, while the welding position type, groove morphology and molten pool force are constantly changing during the full-position welding of pipelines with low processing accuracy, resulting in different characteristics of the same welding defects in different positions. If the molten pool image is simply classified according to the welding defects and a pipeline full-position welding defect prediction model is established through a convolutional neural network, it is easy to lead to a low prediction accuracy of the prediction model, and thus it is impossible to effectively realize the monitoring of pipeline full-position welding defects. Summary of the invention

[0005] The object of the present invention is to provide an online detection method for all-position welding penetration by coordinating electrical signals with molten pool images, which is used for real-time monitoring of the automatic all-position welding process of a petrochemical pipeline robot.

[0006] The technical solution to achieve the purpose of the present invention is: a pipeline all-position welding defect monitoring method based on electrical signals and vision under the cooperation of dual robots, comprising the following steps:

[0007] Step 1: Before welding, input the pipe material, wall thickness and pipe diameter information, and start the welding robot equipped with a laser scanner to scan the pipe groove from bottom to top to obtain the original data of the groove profile at different positions.

[0008] Step 2: Perform preliminary segmentation of the pipeline groove based on the characteristics of overhead welding, flat welding and vertical welding, that is, 0-30° for flat welding, 30°-150° for vertical welding and 150°-180° for overhead welding. Then, perform 5-10° segmentation based on the preliminary segmentation, and match the corresponding welding process parameters from the process knowledge base according to the position and groove size of each segment.

[0009] Further, according to the position and groove size of each section, the corresponding welding process parameters are matched from the process knowledge base, including: the process knowledge base includes material, wall thickness, pipe diameter, position type, groove gap, groove misalignment, welding current, wire feeding speed, welding speed, swing length, swing amplitude, left dwell time and right dwell time; wherein the welding types include flat welding at 0° position, vertical welding at 90° position and overhead welding at 180°;

[0010] The process of obtaining welding process parameters for each section is as follows:

[0011] Firstly, the corresponding welding current, wire feeding speed, welding speed, swing length, swing amplitude, left dwell time and right dwell time are obtained according to the input material, wall thickness, pipe diameter, and welding position type, groove gap and groove misalignment obtained by segmentation and laser scanner;

[0012] Secondly, calculate the position of each groove. The specific formula is as follows:

[0013]

[0014] is the angle between the starting position of each segment and the vertical direction, is the angle between the end position of each segment and the vertical direction;

[0015] Finally, according to the position of each groove, some process parameters are adjusted;

[0016] When the section belongs to the flat welding position type, the welding current and wire feeding speed are adjusted. The specific formula is as follows:

[0017]

[0018] in, and To match the welding current and wire feeding speed in the process knowledge base, and is a constant;

[0019] When the section belongs to the vertical welding position type, the welding current and welding speed are adjusted. The specific formula is as follows:

[0020]

[0021]

[0022] in, To match the welding speed in the process knowledge base, , , and is a constant;

[0023] When the section belongs to the overhead welding position type, the wire feeding speed and welding speed are adjusted. The specific formula is as follows:

[0024]

[0025] in, and is a constant.

[0026] Step 3: Start two welding robots synchronously, and start the welding parameter acquisition box with an acquisition frequency of 150-200HZ and the industrial camera with an acquisition frequency of 30-50HZ to collect electrical signals and molten pool vision. At the same time, based on the symmetry of welding deformation and the mutual compensation characteristics of arc starting and arc extinction points, one welding robot welds clockwise from 180° on the left half to 3°-6° on the right half, and the other welding robot welds clockwise from 0° on the right half to 174°-177° on the right half.

[0027] Step 4: For the voltage signal collected in each section, a rectangular window with a length of 500-700 is used for windowing and framing, and the voltage signal in each frame is analyzed in the time domain. At the same time, the data of the time domain analysis, material, wall thickness, pipe diameter, groove position of the section, groove size, welding process parameters and the corresponding molten pool image in the rectangular window are input into the electrical signal-image collaborative perception burn-through defect prediction model; when the output result is burn-through, the welding is stopped; when the output result is unknown, step 5 is executed;

[0028] Furthermore, the electrical signal-image collaborative perception burn-through defect prediction model includes:

[0029] The material, wall thickness, pipe diameter, groove position of the section, groove size, and welding process parameters are normalized. The material uses 0.3, 0.6, and 0.9 to represent carbon steel, stainless steel, and nickel-based alloy. Other parameters are normalized using the following calculation formula:

[0030]

[0031] in, and They are the maximum and minimum values ​​of the same type of data respectively;

[0032] The time domain analysis of voltage signals includes mean, variance, peak-to-peak value and kurtosis. The calculation formula is as follows:

[0033]

[0034]

[0035]

[0036]

[0037] in, is the length of the rectangular window, is the mean, is the variance, is the peak-to-peak value and is a steep value.

[0038] The electrical signal-image collaborative perception burn-through defect prediction model adopts a two-branch prediction model, one of which uses a Resnet convolutional neural network with a structural backbone to extract molten pool image features, and the other branch uses an AlexNet convolutional neural network with a structural backbone to extract data features consisting of variance, peak-to-peak value and kurtosis, material, wall thickness, pipe diameter, groove position, groove size, and welding process parameters. Finally, feature fusion is performed through a fully connected layer.

[0039] Step 5: Perform frequency domain analysis on the collected voltage signal and obtain the envelope spectrum data of the voltage signal, that is, the actual molten pool oscillation frequency. At the same time, obtain the molten pool size based on the molten pool image and calculate the reference oscillation frequency of the molten pool in the melt-through state; when the actual molten pool oscillation frequency is greater than the reference oscillation frequency of the molten pool in the melt-through state, take measures to immediately stop welding; when the actual molten pool oscillation frequency is less than or equal to the reference oscillation frequency of the molten pool in the melt-through state, continue welding.

[0040] Furthermore, obtaining the envelope spectrum data of the voltage signal includes:

[0041] First, Fourier transform is used to convert the voltage signal in the time domain into the spectrum of the frequency domain signal;

[0042] Secondly, calculate the start and end frequencies of the intercepted frequency band according to the following formula:

[0043]

[0044] in, and They are the starting and ending frequencies of the intercepted frequency band with a pipe diameter of 150 mm, a wall thickness of 7 mm, a groove gap of 1.5 mm, and a groove misalignment of 1 mm; , , and They are groove gap, groove misalignment, wall thickness and pipe diameter; , , , , and is a constant.

[0045] Next, obtain the peak coordinates of the signal in the intercepted frequency band, sort the peak coordinates according to the size of the spectrum ordinate value, and save the two frequencies with the largest ordinate values. and ;

[0046] Then, and As the cut-off frequency of the band-pass filter, the collected voltage signal is filtered, and the upper envelope signal of the filtered voltage signal is obtained;

[0047] Finally, the upper envelope signal of the voltage signal is subjected to spectrum analysis to obtain the actual oscillation frequency of the molten pool.

[0048] Further, obtaining the molten pool size based on the molten pool image and calculating the reference oscillation frequency of the molten pool in the melt-through state include:

[0049] Firstly, the height of the molten pool and the length of the upper surface of the molten pool are obtained based on the molten pool image;

[0050] Secondly, based on the equivalent cylindrical film of molten metal in the melt-through state and the equal volume principle, the equivalent cylindrical diameter is obtained, and the following equation is obtained:

[0051]

[0052] in, is the height of the molten pool, is the length of the upper surface of the molten pool, is the equivalent melt pool radius.

[0053] At the same time, based on the fact that the molten pool is a convex truncated cone in the actual welding process, the above inequality is estimated to have a maximum value. Therefore, the above equation is simplified to:

[0054]

[0055] Thus, the equivalent molten pool diameter is obtained:

[0056]

[0057] Then, based on the oscillation model of the molten pool as a crescent-shaped molten pool, the wave number k is obtained, and the calculation formula is as follows:

[0058]

[0059] Finally, the oscillation frequency of the molten pool under the reference penetration state is calculated according to the following calculation formula:

[0060]

[0061] in, is the density of molten metal, and is a constant.

[0062] Compared with the prior art, the present invention has the following significant advantages: 1) The present invention selects welding process parameters based on the welding process adjustment model and the process knowledge base, so that the welding process parameters are more adapted to the groove size and position, which is beneficial to improving the welding quality; 2) The present invention performs different processing based on the different characteristics of burn-through defects and incomplete penetration defects in the molten pool visual and electrical signal performance, which is beneficial to eliminating interference factors and improving the accuracy of the prediction model; 3) The present invention considers the influence of the material and size of the pipeline itself, the welding position and the groove size on the welding defects, integrates these factors and fuses them with the visual and electrical signal characteristics, which is beneficial to enhancing the generalization ability and accuracy of the defect prediction model, thereby realizing automatic all-position welding defect monitoring of the pipeline robot. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a flow chart of the online detection method of all-position welding penetration by coordinating electrical signals with molten pool images of the present invention.

[0064] Figure 2 It is a schematic diagram of pipeline position division in the online detection method for all-position welding penetration by coordinating electrical signals with molten pool images of the present invention.

[0065] Figure 3 It is the electrical signal and molten pool image without defects at the 40°-50° position in the online detection method for all-position welding penetration by coordinating the electrical signal and the molten pool image of the present invention.

[0066] Figure 4 It is the electrical signal and molten pool image of the burn-through at the 40°-50° position in the online detection method for all-position welding penetration by coordinating the electrical signal and the molten pool image of the present invention.

[0067] Figure 5 It is the electrical signal and molten pool image without defects at the 50°-60° position in the online detection method for all-position welding penetration by coordinating the electrical signal and the molten pool image of the present invention.

[0068] Figure 6 It is the electrical signal and molten pool image of the burn-through at the 50°-60° position in the online detection method for all-position welding penetration by coordinating the electrical signal and the molten pool image of the present invention. DETAILED DESCRIPTION

[0069] The online detection method for all-position welding penetration by coordinating electrical signals with molten pool images of the present invention generally establishes a mapping model between welding formation defects and molten pool vision or electrical signals directly, thereby realizing the monitoring of pipeline welding formation defects. The actual quality of pipeline all-position welding formation is related to the groove size and welding position. Therefore, it is not possible to simply establish a mapping model between welding formation defects and molten pool images directly. Therefore, it is necessary to first establish a process knowledge base and a process parameter adjustment model according to the groove position and size to match and adjust the welding parameters; secondly, based on the different characteristics of different welding defects in different aspects of different information, different information source feature fusion is adopted, and different welding defect prediction models are established through convolutional neural networks, thereby realizing accurate judgment and monitoring of pipeline all-position welding formation defects.

[0070] By real-time monitoring of the automatic full-position welding process of petrochemical process pipelines, the problem of how welding robots can identify welding defects in real time is solved, thereby promoting the intelligent application of welding robot pipeline automatic welding. Therefore, the present invention proposes a mapping relationship between welding position, molten pool vision and electrical signals and welding defects, and establishes a pipeline robot automatic full-position welding defect prediction model, thereby realizing a robot full-position welding forming defect monitoring method.

[0071] The online detection method for all-position welding penetration by coordinating electrical signals and molten pool images of the present invention adopts dual robots to cooperate with all-position welding of pipelines based on the symmetry of welding deformation and the complementarity of arc starting and arc extinguishing points; for the setting of pipeline all-position welding process parameters, firstly, preliminary segmentation is performed based on the welding position and then segmented again at 5-10°, then a laser scanner is used to scan the pipeline groove to obtain the groove size of each section, and finally, the welding process parameters are adaptively set based on the process knowledge base and the welding parameter adjustment model; for the problem of pipeline burn-through defect identification, the molten pool image and the time domain eigenvalues ​​of the electrical signal are fused and distinguished through a neural network prediction model; for the problem of incomplete penetration identification, firstly, the molten pool size characteristics are obtained based on the molten pool image and the reference molten pool oscillation frequency is calculated based on the reference molten pool oscillation frequency model, then the upper envelope signal of the voltage signal is subjected to spectrum analysis and the actual molten pool oscillation frequency is calculated, and finally, whether the welding is penetrated is judged by comparing the actual and reference molten pool oscillation frequencies.

[0072] The present invention will be further described below in conjunction with the accompanying drawings.

[0073] The specific process is as follows Figure 1 shown.

[0074] Follow the steps below to implement it:

[0075] The present invention provides an online detection method for all-position welding penetration by coordinating electrical signals with molten pool images, comprising the following steps:

[0076] Step 1: Before welding, input the pipe material, wall thickness and pipe diameter information, and start the welding robot equipped with a laser scanner to scan the pipe groove from bottom to top to obtain the original data of the groove profile at different positions.

[0077] Step 2: If Figure 2 As shown in the figure, the pipeline groove is initially segmented based on the characteristics of overhead welding, flat welding and vertical welding, that is, 0-30° is flat welding, 30°-150° is vertical welding and 150°-180° is overhead welding. Then, 5-10° segmentation is performed on the basis of the initial segmentation. At the same time, the corresponding welding process parameters are matched from the process knowledge base shown in Table 1 according to the position and groove size of each segment.

[0078] Table 1 Process knowledge base in the online detection method of all-position welding penetration by coordinating electrical signals and molten pool images

[0079]

[0080] Furthermore, according to the position and groove size of each section, the corresponding welding process parameters are matched from the process knowledge base, including: the process knowledge base content is composed of material, wall thickness, pipe diameter, position type, groove gap, groove misalignment, welding current, wire feeding speed, welding speed, swing length, swing amplitude, left dwell time and right dwell time. The welding types include flat welding at 0° position, vertical welding at 90° position and overhead welding at 180°;

[0081] The process of obtaining welding process parameters for each section is as follows:

[0082] Firstly, the corresponding welding current, wire feeding speed, welding speed, swing length, swing amplitude, left dwell time and right dwell time are obtained according to the input material, wall thickness, pipe diameter, segmentation and position type, groove gap and groove misalignment obtained by laser scanner;

[0083] Secondly, calculate the position of each groove. The specific formula is as follows:

[0084]

[0085] is the angle between the starting position of each segment and the vertical direction, is the angle between the end position of each segment and the vertical direction;

[0086] Finally, according to the position of each groove, some process parameters are adjusted;

[0087] When the section belongs to the flat welding position type, the welding current and wire feeding speed are adjusted. The specific formula is as follows:

[0088]

[0089] in, and To match the welding current and wire feeding speed in the process knowledge base, and is a constant;

[0090] When the section belongs to the vertical welding position type, the welding current and welding speed are adjusted. The specific formula is as follows:

[0091]

[0092]

[0093] in, To match the welding speed in the process knowledge base, , , and is a constant;

[0094] When the section belongs to the overhead welding position type, the wire feeding speed and welding speed are adjusted. The specific formula is as follows:

[0095]

[0096] in, and is a constant.

[0097] Step 3: Start two welding robots synchronously, and start the welding parameter acquisition box with an acquisition frequency of 160Hz and the industrial camera with an acquisition frequency of 30Hz to collect electrical signals and molten pool vision. At the same time, based on the symmetry of welding deformation and the mutual compensation characteristics of arc starting and arc extinction points, one welding robot welds clockwise from 180° on the left half to 3° on the right half, and the other welding robot welds clockwise from 0° on the right half to 177° on the right half.

[0098] Step 4: Figure 3 As shown, the welding is flawless and the welding position is between 40° and 50°. The arc voltage fluctuates slightly and the arc voltage value is less than 12V. Figure 4As shown in the figure, there is a burn-through defect and the welding position is between 40° and 50°. The arc voltage fluctuates greatly and the maximum arc voltage value is higher than 30V. The molten pool image is incomplete. Therefore, for welding positions less than 50°, the burn-through defect can be analyzed by the time domain information of the arc voltage signal and the molten pool image. Figure 5 As shown, the welding is flawless and the welding position is between 50° and 60°. The arc voltage fluctuates slightly and the arc voltage value is between 25 and 26 V. Figure 6 As shown in the figure, there is a burn-through defect and the welding position is 50°-60°. The arc voltage shows a small fluctuation and the arc voltage value is between 25-26V, but the pool image is incomplete. The above shows that the burn-through defects at different positions show different characteristics in the electrical signal and the molten pool image. Therefore, it is necessary to combine the electrical signal, the molten pool image, the groove size and position and other information to predict the burn-through defect.

[0099] For the voltage signal collected in each section, a rectangular window with a length of 500-700 is used for windowing and framing. At the same time, the voltage signal in each frame is analyzed in the time domain. At the same time, the data of the time domain analysis, material, wall thickness, pipe diameter, groove position of the section, groove size, welding process parameters and the corresponding molten pool image in the rectangular window are input into the electrical signal-image collaborative perception burn-through defect prediction model. When the output result is burn-through, the welding is stopped.

[0100] Furthermore, the electrical signal-image collaborative perception burn-through defect prediction model includes:

[0101] The material, wall thickness, pipe diameter, groove position of the section, groove size, and welding process parameters are normalized. The material uses 0.3, 0.6, and 0.9 to represent carbon steel, stainless steel, and nickel-based alloy. Other parameters are normalized using the following calculation formula:

[0102]

[0103] in, and They are the maximum and minimum values ​​of the same type of data respectively.

[0104] The time domain analysis of voltage signals includes mean, variance, peak-to-peak value and kurtosis. The calculation formula is as follows:

[0105]

[0106]

[0107]

[0108]

[0109] in, is the length of the rectangular window, is the mean, is the variance, is the peak-to-peak value and is a steep value.

[0110] The electrical signal-image collaborative perception burn-through defect prediction model adopts a two-branch prediction model, one of which uses a Resnet convolutional neural network with a structural backbone to extract molten pool image features, and the other branch uses an AlexNet convolutional neural network with a structural backbone to extract data features consisting of variance, peak-to-peak value and kurtosis, material, wall thickness, pipe diameter, groove position, groove size, and welding process parameters. Finally, feature fusion is performed through a fully connected layer.

[0111] Step 5: When the output result of the electrical signal-image collaborative perception burn-through defect prediction model is unknown, the collected voltage signal is analyzed in the frequency domain and the envelope spectrum data of the voltage signal is obtained, that is, the actual molten pool oscillation frequency. At the same time, the molten pool size is obtained based on the molten pool image and the reference oscillation frequency of the molten pool in the melt-through state is calculated. When the actual molten pool oscillation frequency is greater than the reference oscillation frequency of the molten pool in the melt-through state, measures are taken to immediately stop welding. When the actual molten pool oscillation frequency is less than or equal to the reference oscillation frequency of the molten pool in the melt-through state, welding continues.

[0112] Furthermore, obtaining the envelope spectrum data of the voltage signal includes:

[0113] First, Fourier transform is used to convert the voltage signal in the time domain into the spectrum of the frequency domain signal;

[0114] Secondly, calculate the start and end frequencies of the intercepted frequency band according to the following formula:

[0115]

[0116] in, and They are the starting and ending frequencies of the intercepted frequency band with a pipe diameter of 150 mm, a wall thickness of 7 mm, a groove gap of 1.5 mm, and a groove misalignment of 1 mm; , , and They are groove gap, groove misalignment, wall thickness and pipe diameter; , , , , and is a constant.

[0117] Next, obtain the peak coordinates of the signal in the intercepted frequency band, sort the peak coordinates according to the size of the spectrum ordinate value, and save the two frequencies with the largest ordinate values. and ;

[0118] Then, and The voltage signal is filtered for the cutoff frequency of the bandpass filter, and the upper envelope signal of the filtered voltage signal is obtained.

[0119] Finally, the upper envelope signal of the voltage signal is subjected to spectrum analysis to obtain the actual oscillation frequency of the molten pool.

[0120] Further, obtaining the molten pool size based on the molten pool image and calculating the reference oscillation frequency of the molten pool in the melt-through state include:

[0121] Firstly, the height of the molten pool and the length of the upper surface of the molten pool are obtained based on the molten pool image;

[0122] Secondly, based on the equivalent cylindrical film of molten metal in the melt-through state and the equal volume principle, the equivalent cylindrical diameter is obtained, and the following equation is obtained:

[0123]

[0124] in, is the height of the molten pool, is the length of the upper surface of the molten pool, is the equivalent melt pool radius.

[0125] At the same time, based on the fact that the molten pool is a convex truncated cone in the actual welding process, the above inequality is estimated to have a maximum value. Therefore, the above equation is simplified to:

[0126]

[0127] Thus, the equivalent molten pool diameter is obtained:

[0128]

[0129] Then, based on the oscillation model of the molten pool as a crescent-shaped molten pool, the wave number k is obtained, and the calculation formula is as follows:

[0130]

[0131] Finally, the oscillation frequency of the molten pool under the reference penetration state is calculated according to the following calculation formula:

[0132]

[0133] in, is the density of molten metal, and is a constant.

Claims

1. An online detection method for all-position welding penetration by combining electrical signals with molten pool images, characterized in that: The following steps are involved: Step 1: Input the pipe material, wall thickness and pipe diameter information before welding, and start the welding robot equipped with a laser scanner to scan the pipe groove from bottom to top to obtain the original data of the groove profile at different positions; Step 2: Based on the characteristics of overhead welding, flat welding and vertical welding, the pipeline groove is initially segmented, that is, 0-30° is flat welding, 30°-150° is vertical welding, and 150°-180° is overhead welding. Then, based on the preliminary segmentation, 5-10° segmentation is performed, and the corresponding welding process parameters are matched from the process knowledge base according to the position and groove size of each segment; Step 3: Start two welding robots synchronously, and start the welding parameter acquisition box with an acquisition frequency of 150-200HZ and the industrial camera with an acquisition frequency of 30-50HZ to collect electrical signals and molten pool vision; one welding robot welds clockwise from 180° on the left half to 3°-6° on the right half, and the other welding robot welds clockwise from 0° on the right half to 174°-177° on the right half; Step 4: For the voltage signal collected in each section, a rectangular window with a length of 500-700 is used for windowing and framing, and the voltage signal in each frame is analyzed in the time domain. At the same time, the data of the time domain analysis, material, wall thickness, pipe diameter, groove position of the section, groove size, welding process parameters and the corresponding molten pool image in the rectangular window are input into the electrical signal-image collaborative perception burn-through defect prediction model; when the output result is burn-through, measures are taken to stop welding; when the output result is unknown, step 5 is executed; Step 5: Perform frequency domain analysis on the collected voltage signal and obtain the envelope spectrum data of the voltage signal, that is, the actual molten pool oscillation frequency. At the same time, obtain the molten pool size based on the molten pool image and calculate the reference oscillation frequency of the molten pool in the melt-through state; when the actual molten pool oscillation frequency is greater than the reference oscillation frequency of the molten pool in the melt-through state, take measures to immediately stop welding; when the actual molten pool oscillation frequency is less than or equal to the reference oscillation frequency of the molten pool in the melt-through state, continue welding.

2. The online detection method for all-position welding penetration by coordinating electrical signals with molten pool images according to claim 1 is characterized in that: Matching corresponding welding process parameters from the process knowledge base according to the position and groove size of each section described in step 2 includes: The process knowledge base includes material, wall thickness, pipe diameter, welding position type, groove gap, groove misalignment, welding current, wire feeding speed, welding speed, swing length, swing amplitude, left dwell time and right dwell time; the welding position types include 0° flat welding, 90° vertical welding and 180° overhead welding; The process of obtaining welding process parameters for each section is as follows: Firstly, the corresponding welding current, wire feeding speed, welding speed, swing length, swing amplitude, left dwell time and right dwell time are obtained according to the input material, wall thickness, pipe diameter, and welding position type, groove gap and groove misalignment obtained by segmentation and laser scanner; Secondly, calculate the position of each groove. The specific formula is as follows: ， is the angle between the starting position of each segment and the vertical direction, is the angle between the end position of each segment and the vertical direction; Finally, according to the position of each groove, some process parameters are adjusted: When the section belongs to the flat welding position type, the welding current and wire feeding speed are adjusted. The specific formula is as follows: ， in, and To match the welding current and wire feeding speed in the process knowledge base, and is a constant; When the section belongs to the vertical welding position type, the welding current and welding speed are adjusted. The specific formula is as follows: ， ， in, To match the welding speed in the process knowledge base, , , and is a constant; When the section belongs to the overhead welding position type, the wire feeding speed and welding speed are adjusted. The specific formula is as follows: ， in, and is a constant.

3. The online detection method for all-position welding penetration by coordinating electrical signals with molten pool images according to claim 1 is characterized in that: The electrical signal-image collaborative perception burn-through defect prediction model described in step 4 includes: The material, wall thickness, pipe diameter, groove position of the section, groove size, and welding process parameters are normalized. The material uses 0.3, 0.6, and 0.9 to represent carbon steel, stainless steel, and nickel-based alloy. Other parameters are normalized using the following calculation formula: ， in, and They are the maximum and minimum values ​​of the same type of data respectively; The time domain analysis of voltage signals includes mean, variance, peak-to-peak value and kurtosis. The calculation formula is as follows: ， ， ， ， in, is the length of the rectangular window, is the mean, is the variance, is the peak-to-peak value and is the steep value; The electrical signal-image collaborative perception burn-through defect prediction model adopts a two-branch prediction model, one of which uses a Resnet convolutional neural network with a structural backbone to extract molten pool image features, and the other branch uses an AlexNet convolutional neural network with a structural backbone to extract data features consisting of variance, peak-to-peak value and kurtosis, material, wall thickness, pipe diameter, groove position, groove size, and welding process parameters. Finally, feature fusion is performed through a fully connected layer.

4. The online detection method for all-position welding penetration by coordinating electrical signals with molten pool images according to claim 1 is characterized in that: The step 5 of obtaining the upper envelope spectrum data of the voltage signal includes: First, Fourier transform is used to convert the voltage signal in the time domain into the spectrum of the frequency domain signal; Secondly, calculate the start and end frequencies of the intercepted frequency band according to the following formula: ， in, and They are the starting and ending frequencies of the intercepted frequency band with a pipe diameter of 150 mm, a wall thickness of 7 mm, a groove gap of 1.5 mm, and a groove misalignment of 1 mm; , , and They are groove gap, groove misalignment, wall thickness and pipe diameter; , , , , and is a constant; Next, obtain the peak coordinates of the signal in the intercepted frequency band, sort the peak coordinates according to the size of the spectrum ordinate value, and save the two frequencies with the largest ordinate values. and ; Then, and As the cut-off frequency of the band-pass filter, the collected voltage signal is filtered, and the upper envelope signal of the filtered voltage signal is obtained; Finally, the upper envelope signal of the voltage signal is subjected to spectrum analysis to obtain the actual oscillation frequency of the molten pool.

5. The online detection method for all-position welding penetration by coordinating electrical signals with molten pool images according to claim 1 is characterized in that: The step 5 of obtaining the size of the molten pool based on the molten pool image and calculating the reference oscillation frequency of the molten pool under the melt penetration state includes: Firstly, the height of the molten pool and the length of the upper surface of the molten pool are obtained based on the molten pool image; Secondly, based on the equivalent cylindrical film of molten metal in the melt-through state and the equal volume principle, the equivalent cylindrical diameter is obtained, and the following equation is obtained: ， in, is the height of the molten pool, is the length of the upper surface of the molten pool, is the equivalent melt pool radius; At the same time, based on the fact that the molten pool is a convex truncated cone in the actual welding process, the above inequality is estimated to have a maximum value; therefore, the above equation is simplified to: ， Thus, the equivalent molten pool diameter is obtained: ， Then, based on the oscillation model of the molten pool as a crescent-shaped molten pool, the wave number is obtained , the calculation formula is as follows: ， Finally, the oscillation frequency of the molten pool under the reference penetration state is calculated according to the following calculation formula: ， in, is the density of molten metal, and is a constant.

Images