Aluminum alloy variable polarity TIG deep penetration welding penetration depth prediction method and system
By combining classification and semantic segmentation convolutional neural networks with an ensemble learning model, the problem of real-time monitoring and accurate prediction of weld formation in variable polarity TIG deep penetration welding was solved, achieving efficient and low-cost penetration depth prediction for aluminum alloy welding.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIHANG UNIV
- Filing Date
- 2024-05-31
- Publication Date
- 2026-06-26
AI Technical Summary
Existing technologies struggle to monitor and accurately predict weld formation in real-time for variable polarity TIG deep penetration welding, especially in aluminum alloy welding. Cathode spots blur the molten pool image, and uncertainties in the welding environment lead to inaccurate data. Traditional models require a large amount of data and cannot handle uncertainties, resulting in high welding costs.
We employ a classification-based convolutional neural network and a semantic segmentation-based convolutional neural network to identify welding states, and combine them with an ensemble learning model to predict weld depth. By processing images frame by frame, we extract features of the molten pool and keyhole, and then use the ensemble learning model to predict weld depth.
Real-time prediction of weld penetration depth in blind-hole deep-penetration welding of aluminum alloys was achieved, reducing prediction errors, improving the model's generalization ability, and lowering welding costs.
Smart Images
Figure CN118587489B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of welding technology, and more specifically to a method and system for predicting the penetration depth of aluminum alloy variable polarity TIG deep penetration welding. Background Technology
[0002] Aluminum alloys are widely used in aerospace, transportation and construction due to their lightweight, high strength and good corrosion resistance. Achieving deep penetration welding of medium and thick aluminum alloy plates can help improve the production efficiency of aluminum alloy products and enhance corporate benefits.
[0003] Currently, during welding, using a heat source with high energy density to stimulate deep penetration pinholes helps to increase the weld penetration depth. Deep penetration pinholes include through-hole type and blind hole type, both of which have certain applications.
[0004] However, the blind hole deep penetration welding process of aluminum alloys is complex. Due to the high heat sensitivity of aluminum alloys, the welding quality is easily affected by energy input. Excessive energy input may cause the molten pool to collapse, while insufficient energy input may cause incomplete penetration, both of which will seriously affect the mechanical properties and structural integrity of the weld. For through hole deep penetration welding (such as plasma welding), the forming prediction technology has been well developed and applied. However, for variable polarity TIG deep penetration welding, there is no similar technology yet.
[0005] Predicting the penetration depth of variable polarity TIG deep penetration welding presents the following challenges:
[0006] For the negative polarity stage current mode of variable polarity TIG deep penetration welding, cathode spots are formed on the surface of the aluminum alloy base material as the cathode to remove the oxide film on the surface of the aluminum alloy. However, the appearance of cathode spots will cause the molten pool image to be blurred and the molten pool features cannot be effectively identified.
[0007] During the negative polarity period, the polarity change frequency is usually around 100Hz, the molten pool morphology is constantly fluctuating violently, and the camera light source (arc light) is constantly fluctuating. The frame rate of ordinary welding cameras is generally between 30 and 60 frames, so the frequency of molten pool fluctuation and arc light flicker is higher than the sampling frequency, resulting in different light intensity and arc state in each image captured by the camera. It will be difficult to extract the molten pool width and keyhole features using traditional fixed parameter image processing techniques.
[0008] Uncertainties such as welding environment, initial conditions, arc dwell time before movement, and arc ignition position have a certain impact on welding results, leading to inaccurate experimental data. Furthermore, due to the inconsistency of camera equipment parameters and installation positions, the results of molten pool feature recognition may be biased. In addition, welding experiments are costly and cannot use a large amount of data to build a melt depth prediction model. Traditional deterministic prediction models, such as support vector machines and backpropagation neural networks, not only require a large amount of modeling data as support, but also cannot take into account various uncertainties.
[0009] Therefore, for variable polarity TIG deep penetration welding, how to develop a technology that can monitor and accurately predict weld formation in real time in order to improve the quality and efficiency of aluminum alloy deep penetration welding is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0010] In view of this, the present invention provides a method and system for predicting the penetration depth of aluminum alloy variable polarity TIG deep penetration welding to solve some of the technical problems mentioned in the background art.
[0011] To achieve the above objectives, the present invention adopts the following technical solution:
[0012] A method for predicting the penetration depth of TIG deep penetration welding of aluminum alloys with variable polarity includes the following steps:
[0013] S1. Acquire images of the weld pool and input the images into a pre-trained classification convolutional neural network to identify the state of the weld pool, which includes positive polarity stage and negative polarity stage.
[0014] S2. If it is a negative polarity stage, the processing of the current frame image ends and the acquisition of the next frame of weld pool image continues. If it is a positive polarity stage, proceed to step S3.
[0015] S3. Input the image into a pre-trained semantic segmentation convolutional neural network to label the molten pool region and the shadow region caused by the small hole, and obtain the segmentation result image of the molten pool feature region and the small hole feature region;
[0016] S4. Calculate the feature dimensions of the molten pool and the orifice based on the segmentation result image. The feature dimensions include the width of the molten pool and the approximate depth of the orifice.
[0017] S5. Input the molten pool width and the approximate depth of the pinhole into the pre-trained ensemble learning prediction model to predict the current molten depth.
[0018] Preferably, the semantic segmentation convolutional neural network in step S3 includes, but is not limited to, U-Net, SegNet, or DeepLab networks.
[0019] Preferably, in step S4, the method for calculating the molten pool width is as follows:
[0020] (1) Extract the boundary of the melt pool segmentation result and obtain the coordinates of the boundary in the image;
[0021] (2) Calculate the Euclidean distance between points on the boundary;
[0022] (3) Compare the size of each distance, and the largest distance is the width of the molten pool.
[0023] Preferably, in step S4, the method for calculating the approximate depth of the hole is as follows:
[0024] (1) Search for each column of pixels in the segmentation result image of the small hole feature region from top to bottom. When the first marked point is found, stop searching the current column and record the coordinates of the point to obtain the upper boundary of the small hole shadow region.
[0025] (2) Divide the extracted boundary into left and right contours, find the highest and lowest points on the left and right contours respectively, calculate the Euclidean distance between the highest points of the left and right contours, and obtain the aperture of the small hole.
[0026] (3) Calculate the slope of the straight line passing through the left and right contours respectively, and then obtain the cone angle of the small hole by the two slopes;
[0027] (4) Approximate the shape of the small hole as an inverted cone, and the height of the cone is approximately the depth of the small hole. Then, calculate the approximate depth of the small hole by using the hole diameter and the cone angle.
[0028] The preferred method for calculating the approximate depth of the pinhole is as follows:
[0029] The slope k of the straight line passing through the left profile l for:
[0030]
[0031] The slope k of the straight line passing through the right profile r for:
[0032]
[0033] Among them, (x tl ,y tl (x) is the highest point of the left contour. bl ,y bl (x) is the lowest point of the left contour. tr ,y tr (x) is the highest point of the right contour. br ,y br () represents the lowest point of the right contour;
[0034] Small hole cone angle ak for:
[0035] a k =arctan(k l )-arctan(k r )
[0036] The approximate depth h of the hole is:
[0037]
[0038] Among them, L k This refers to the aperture of the small hole.
[0039] Preferably, the ensemble learning prediction model includes categorical prediction models and regression prediction models;
[0040] The penetration state of the weld is predicted using a classification prediction model. The input of the prediction model is the weld pool width and the keyhole depth, and the output is the current penetration state of the weld, including: incomplete penetration, penetration, and over-penetration.
[0041] A regression prediction model is used to estimate the specific value of the weld penetration depth. The input of the prediction model is the weld pool width and the keyhole depth, and the output is the weld penetration depth.
[0042] A system for predicting the weld depth of aluminum alloy variable polarity TIG deep penetration welding, based on the aforementioned method for predicting the weld depth of aluminum alloy variable polarity TIG deep penetration welding, includes: an image acquisition device, a weld pool state identification module, a feature region segmentation module, a feature size calculation module, and a weld depth prediction module.
[0043] Image acquisition device, used to acquire images of the weld pool;
[0044] The melt pool state identification module is used to input the image into a pre-trained classification convolutional neural network to identify the melt pool state, which includes a positive polarity phase and a negative polarity phase.
[0045] The feature region segmentation module is used to input the image into a pre-trained semantic segmentation convolutional neural network when the molten pool is in the positive polarity stage, and to mark the molten pool region and the shadow region caused by the small hole, so as to obtain the segmentation result image of the molten pool feature region and the small hole feature region.
[0046] The feature size calculation module is used to calculate the feature size of the molten pool and the orifice based on the segmentation result image. The feature size includes the width of the molten pool and the approximate depth of the orifice.
[0047] The melt depth prediction module is used to input the melt pool width and the approximate depth of the pinhole into a pre-trained ensemble learning prediction model to predict the current melt depth.
[0048] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for predicting the penetration depth of aluminum alloy variable polarity TIG deep penetration welding.
[0049] A processing terminal includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the computer program, it implements the aforementioned method for predicting the penetration depth of aluminum alloy variable polarity TIG deep penetration welding.
[0050] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method and system for predicting the penetration depth of aluminum alloy variable polarity TIG deep penetration welding, which has the following advantages:
[0051] First, the images acquired during the welding process are processed frame by frame. The molten pool images during the clear periods of deep-melted small holes that occur periodically are selected using a classification convolutional neural network. This solves the problem that the appearance of cathode spots during the negative polarity stage can cause the molten pool image to be blurred and the molten pool features to be unable to be effectively identified.
[0052] In addition, the semantic segmentation convolutional neural network extracts the shadow areas caused by the molten pool and the deep molten hole, and then calculates the width of the molten pool and the depth of the hole. Compared with traditional image processing techniques, the semantic segmentation convolutional neural network is more robust. It is reasonable and necessary for this invention to use the semantic segmentation convolutional neural network to mark the shadow areas caused by the molten pool and the hole.
[0053] Finally, based on the weld pool width and pinhole depth, an ensemble learning model is used to predict the current weld penetration depth. Ensemble learning improves the model's performance by building and combining multiple learners, thus mitigating the impact of uncertainties, reducing prediction errors, and enhancing the model's generalization ability. Furthermore, using ensemble learning to build a prediction model does not require a large amount of data.
[0054] This invention enables real-time prediction of the weld penetration depth of blind-hole deep-penetration welding of aluminum alloys, providing support for energy control and process optimization in the welding process. Attached Figure Description
[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0056] Figure 1 A flowchart of a method for predicting the penetration depth of aluminum alloy variable polarity TIG deep penetration welding provided by the present invention;
[0057] Figure 2 Schematic diagrams of molten pool images under different current modes and states provided by this invention.
[0058] Figure 3 A schematic diagram of the segmentation results of the feature region provided by the present invention;
[0059] Figure 4 This is a schematic diagram illustrating the relationship between the morphology of the molten pool and the depth of the melt, provided by the present invention.
[0060] Figure 5 A schematic diagram illustrating the method for calculating the approximate depth of the molten pool width and the keyhole provided by this invention;
[0061] Figure 6 This is a schematic diagram showing the equipment installation location and connection of the prediction system provided by the present invention. Detailed Implementation
[0062] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0063] This invention discloses a method for predicting the penetration depth of variable polarity TIG deep penetration welding of aluminum alloys, such as... Figure 1 This includes the following steps:
[0064] S1. Acquire images of the weld pool and input the images into a pre-trained classification convolutional neural network to identify the state of the weld pool, which includes positive polarity stage and negative polarity stage.
[0065] S2. If it is a negative polarity stage, the processing of the current frame image ends and the acquisition of the next frame of weld pool image continues. If it is a positive polarity stage, proceed to step S3.
[0066] S3. Input the image into a pre-trained semantic segmentation convolutional neural network to label the molten pool region and the shadow region caused by the small hole, and obtain the segmentation result image of the molten pool feature region and the small hole feature region;
[0067] S4. Calculate the feature dimensions of the molten pool and the orifice based on the segmentation result image. The feature dimensions include the width of the molten pool and the approximate depth of the orifice.
[0068] S5. Input the molten pool width and the approximate depth of the pinhole into the pre-trained ensemble learning prediction model to predict the current molten depth.
[0069] In practical applications, such as Figure 2Aluminum alloy welding generally uses alternating current and is divided into two stages: the positive polarity stage is when the tungsten electrode is connected to the negative and the workpiece to the positive polarity stage; the negative polarity stage is when the tungsten electrode is connected to the positive polarity and the workpiece to the negative polarity stage. In the positive polarity stage, the molten pool image is clear, and the morphological features of the molten pool can be effectively identified. In the negative polarity stage, the oxide film on the surface of the workpiece can be removed. During the removal process, many bright spots will appear on the surface of the workpiece. At the same time, the arc light intensity is different between the positive and negative polarities. The change in light intensity and the appearance of spots work together to make the molten pool image in the negative polarity stage unclear and difficult to extract effective features. Therefore, if it is the negative polarity stage, the processing of the current frame image is ended and the acquisition of the next frame of the welding molten pool image continues. If it is the positive polarity stage, step S3 is performed.
[0070] To further implement the above technical solution, the semantic segmentation convolutional neural network in step S3 includes, but is not limited to, U-Net, SegNet, or DeepLab networks, such as... Figure 3 This is the segmentation effect of the melt pool and pinhole feature regions based on the U-Net network.
[0071] In this embodiment, welding tests verified a strong correlation between weld penetration depth and keyhole depth and weld pool width, such as... Figure 4 As shown, there is an approximately linear relationship between the orifice depth and the melt depth. While there is no obvious quantifiable relationship between the melt pool width and the melt depth, there is a certain correlation. For example, when the melt width is small (around 13 mm), the melt depth is small; when the melt width is large (around 15 to 16 mm), the melt depth increases; and when the melt width is 14 to 15 mm, the melt depth is the largest. Therefore, it is reasonable and necessary to use the melt pool width and orifice depth as inputs to the prediction model.
[0072] To further implement the above technical solution, in step S4, as follows: Figure 5 The specific method for calculating the width of the molten pool is as follows:
[0073] (1) Extract the boundary of the feature region segmentation result of the molten pool and obtain the coordinates of the boundary in the image;
[0074] (2) Calculate the Euclidean distance between points on the boundary;
[0075] (3) Compare the magnitudes of each distance; the largest distance is the width W of the molten pool. p .
[0076] To further implement the above technical solution, in step S4, as follows: Figure 5 The specific method for calculating the approximate depth of the pinhole is as follows:
[0077] (1) Search for each column of pixels in the segmentation result image of the small hole feature region from top to bottom. When the first marked point is found, stop searching the current column and record the coordinates of the point to obtain the upper boundary of the small hole shadow region.
[0078] (2) Divide the extracted boundary into left and right contours, find the highest and lowest points on the left and right contours respectively, calculate the Euclidean distance between the highest points of the left and right contours, and obtain the aperture L of the small hole. k ;
[0079] (3) Calculate the slopes of the straight lines that pass through the left and right contours respectively, and then obtain the cone angle α of the small hole using the two slopes. k ;
[0080] (4) Approximate the shape of the small hole as an inverted cone, and the height of the cone is approximately the depth of the small hole. Then, calculate the approximate depth of the small hole by using the hole diameter and the cone angle.
[0081] To further implement the above technical solution, the specific method for calculating the approximate depth of the pinhole is as follows:
[0082] The slope k of the straight line passing through the left profile l for:
[0083]
[0084] The slope k of the straight line passing through the right profile r for:
[0085]
[0086] Among them, (x tl ,y tl (x) is the highest point of the left contour. bl ,y bl (x) is the lowest point of the left contour. tr ,y tr (x) is the highest point of the right contour. br ,y br () represents the lowest point of the right contour;
[0087] Small hole cone angle a k for:
[0088] a k =arctan(k l )-arctan(k r )
[0089] The approximate depth h of the hole is:
[0090]
[0091] Among them, L k This refers to the aperture of the small hole.
[0092] To further implement the above technical solutions, ensemble learning prediction models include categorical prediction models and regression prediction models, and ensemble learning prediction models include random forest, AdaBoost, etc.
[0093] The penetration state of the weld is predicted using a classification prediction model. The input of the prediction model is the weld pool width and the keyhole depth, and the output is the current penetration state of the weld, including: incomplete penetration, penetration, and over-penetration.
[0094] A regression prediction model is used to estimate the specific value of the weld penetration depth. The input of the prediction model is the weld pool width and the keyhole depth, and the output is the weld penetration depth.
[0095] In practical applications, the complexity of the welding process, the differences in welding environment, materials and equipment, and the initial conditions of welding are difficult to quantify, which can lead to inaccurate modeling data. Using conventional machine learning methods can result in overfitting. At the same time, the cost of weld penetration data is high, and the available modeling data is often limited. Ensemble learning can provide a more accurate estimate of the prediction target with less training data.
[0096] If the welding process is more concerned with whether the weld is in a good penetration state, such as butt welding of flat plates without a bottom support or lock bottom structure, the prediction model is trained into a classification model: the model takes the weld pool width and keyhole depth as input and the current penetration state as output for training and prediction. The penetration state includes three states: incomplete penetration, penetration, and over-penetration. It can also be further subdivided, such as into far incomplete penetration, incomplete penetration, near penetration, penetration, over-penetration, collapse, or blow-through, etc. The state classification is set according to the control accuracy requirements of the welding task for the penetration state.
[0097] When welding processes are more concerned with the exact weld penetration, such as lock-bottom structure welding (where the penetration needs to be controlled within the lock-bottom thickness) or when complete penetration is not required, the prediction model is trained as a regression model. The model takes the weld pool width and keyhole depth as input and the exact weld penetration as output for training and prediction. To ensure the accuracy of the prediction, the regression prediction model requires more training data and higher data quality. When creating the training dataset, welding current or other current parameters related to weld penetration can be used as design variables. Welding experiments can be conducted using orthogonal experimental design or Latin hypercube sampling. During the experiment, a CCD camera is used to record images of the weld pool. The keyhole depth and weld pool width recorded in the images are calculated as input data for the prediction model. Then, the weld is cut open, the cut surface is polished and etched, and the weld penetration is observed as output data, forming a dataset to train the model.
[0098] In this embodiment, the ensemble learning prediction model is constructed using the random forest method, specifically:
[0099] Input the original dataset D = {(x1, y1), (x2, y2), ..., (x...} m y m )}, where x i It is a d-dimensional feature vector, y i Here, is the corresponding label, m is the number of samples, and d is the number of features; the number of base learners is T, and the size of the feature subset is k.
[0100] For t = 1, 2, ..., T, repeat the following steps.
[0101] Use the bootstrap sampling method to generate a sample set D of size m from the original dataset D. t ;
[0102] The random subspace method is used to randomly select k features from d features as a feature subset F. t ;
[0103] Based on sample set D t and feature subset F t Use the CART algorithm to build a decision tree h t During the splitting process at each node, only the feature subset F is considered. t The characteristics of the [branch] are such that no pruning is performed;
[0104] The T decision trees are combined to form a random forest model H = {h1, h2, ..., h} T};
[0105] For a new sample x, it is input into each decision tree, resulting in T predictions h1(x), h2(x), ..., h T (x);
[0106] For classification problems, a voting method is used to determine the final category, where the prediction result of each tree is counted as a vote, and the category with the most votes is the final output. For regression problems, an averaging method is used to determine the final output, where the arithmetic mean of the prediction results of each tree is the final output.
[0107] A system for predicting the penetration depth of variable polarity TIG deep penetration welds in aluminum alloys, based on a method for predicting the penetration depth of variable polarity TIG deep penetration welds in aluminum alloys, such as... Figure 6 It includes: an image acquisition device, a molten pool state identification module, a feature region segmentation module, a feature size calculation module, and a molten depth prediction module;
[0108] Image acquisition device, used to acquire images of the weld pool;
[0109] The melt pool state identification module is used to input the image into a pre-trained classification convolutional neural network to identify the melt pool state, which includes a positive polarity phase and a negative polarity phase.
[0110] The feature region segmentation module is used to input the image into a pre-trained semantic segmentation convolutional neural network when the molten pool is in the positive polarity stage, and to mark the molten pool region and the shadow region caused by the small hole, so as to obtain the segmentation result image of the molten pool feature region and the small hole feature region.
[0111] The feature size calculation module is used to calculate the feature size of the molten pool and the orifice based on the segmentation result image. The feature size includes the width of the molten pool and the approximate depth of the orifice.
[0112] The melt depth prediction module is used to input the melt pool width and the approximate depth of the pinhole into a pre-trained ensemble learning prediction model to predict the current melt depth.
[0113] In this embodiment, the image acquisition device is a CCD camera with a light-reducing lens. The camera faces the tungsten electrode tip and is at a certain angle to the horizontal plane. It acquires images of the weld pool from directly behind the welding torch (the direction of the welding torch's movement is directly in front). The camera communicates with the computer and can transmit the acquired images to the computer.
[0114] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for predicting the penetration depth of aluminum alloy variable polarity TIG deep penetration welding.
[0115] A processing terminal includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the computer program, it implements a method for predicting the penetration depth of aluminum alloy variable polarity TIG deep penetration welding.
[0116] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.
[0117] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for predicting the penetration depth of TIG deep penetration welding of aluminum alloys with variable polarity, characterized in that, Includes the following steps: S1. Acquire images of the weld pool and input the images into a pre-trained classification convolutional neural network to identify the state of the weld pool, which includes positive polarity stage and negative polarity stage. S2. If it is a negative polarity stage, the processing of the current frame image ends and the acquisition of the next frame of weld pool image continues. If it is a positive polarity stage, proceed to step S3. S3. Input the image into a pre-trained semantic segmentation convolutional neural network to label the molten pool region and the shadow region caused by the small hole, and obtain the segmentation result image of the molten pool feature region and the small hole feature region; S4. Calculate the feature dimensions of the molten pool and the orifice based on the segmentation result image. The feature dimensions include the width of the molten pool and the approximate depth of the orifice. S5. Input the molten pool width and the approximate depth of the pinhole into the pre-trained ensemble learning prediction model to predict the current molten depth; In step S4, the method for calculating the approximate depth of the hole is as follows: (1) Search for each column of pixels in the segmentation result image of the small hole feature region from top to bottom. When the first marked point is found, stop searching the current column and record the coordinates of the point to obtain the upper boundary of the small hole shadow region. (2) Divide the extracted boundary into left and right contours, find the highest and lowest points on the left and right contours respectively, calculate the Euclidean distance between the highest points of the left and right contours, and obtain the aperture of the small hole. (3) Calculate the slope of the straight line passing through the left and right contours respectively, and then obtain the cone angle of the small hole through the two slopes; (4) Approximate the shape of the small hole as an inverted cone, and the height of the cone is approximately the depth of the small hole. Then, calculate the approximate depth of the small hole by using the hole diameter and the cone angle.
2. The method for predicting the penetration depth of aluminum alloy variable polarity TIG deep penetration welding according to claim 1, characterized in that, The semantic segmentation convolutional neural network in step S3 includes, but is not limited to, U-Net, SegNet, or DeepLab networks.
3. The method for predicting the penetration depth of aluminum alloy variable polarity TIG deep penetration welding according to claim 1, characterized in that, In step S4, the specific method for calculating the molten pool width is as follows: (1) Extract the boundary of the melt pool segmentation result and obtain the coordinates of the boundary in the image; (2) Calculate the Euclidean distance between points on the boundary; (3) Compare the size of each distance, and the largest distance is the width of the molten pool.
4. The method for predicting the penetration depth of aluminum alloy variable polarity TIG deep penetration welding according to claim 1, characterized in that, The specific method for calculating the approximate depth of the pinhole is as follows: The slope of the straight line passing through the left profile k l for: ； The slope of the straight line passing through the right profile k r for: ； in,( x tl , y tl ) is the highest point of the left contour, ( x bl , y bl ) is the lowest point of the left contour, ( x tr , y tr ) is the highest point of the right contour, ( x br , y br () represents the lowest point of the right contour; Small hole cone angle a k for: a k =arctan ( k l ) -arctan ( k r ) Approximate depth of the pinhole h for: ； in, L k This refers to the aperture of the small hole.
5. The method for predicting the penetration depth of aluminum alloy variable polarity TIG deep penetration welding according to claim 1, characterized in that, Ensemble learning prediction models include categorical prediction models and regression prediction models; The penetration state of the weld is predicted using a classification prediction model. The input of the prediction model is the width of the weld pool and the approximate depth of the keyhole. The output is the current penetration state of the weld, including: incomplete penetration, penetration, and over-penetration. A regression prediction model is used to estimate the specific value of the weld penetration depth. The input of the prediction model is the weld pool width and the approximate depth of the keyhole, and the output is the weld penetration depth.
6. A system for predicting the penetration depth of aluminum alloy variable polarity TIG deep penetration welding, characterized in that, A method for predicting the penetration depth of aluminum alloy variable polarity TIG deep penetration welding according to any one of claims 1-5, comprising: an image acquisition device and a computing device; The computing device includes a molten pool state identification module, a feature region segmentation module, a feature size calculation module, and a molten depth prediction module; Image acquisition device, used to acquire images of the weld pool; The molten pool state identification module is used to input the acquired image into a pre-trained classification convolutional neural network to identify the molten pool state, which includes a positive polarity phase and a negative polarity phase. The feature region segmentation module is used to input the image into a pre-trained semantic segmentation convolutional neural network when the molten pool is in the positive polarity stage, and to mark the molten pool region and the shadow region caused by the small hole, so as to obtain the segmentation result image of the molten pool feature region and the small hole feature region. The feature size calculation module is used to calculate the feature size of the molten pool and the orifice based on the segmentation result image. The feature size includes the width of the molten pool and the approximate depth of the orifice. The melt depth prediction module is used to input the melt pool width and the approximate depth of the pinhole into a pre-trained ensemble learning prediction model to predict the current melt depth.
7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the method for predicting the penetration depth of aluminum alloy variable polarity TIG deep penetration welding as described in any one of claims 1-5.
8. A processing terminal, comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, characterized in that, When the processor executes the computer program, it implements the method for predicting the penetration depth of aluminum alloy variable polarity TIG deep penetration welding as described in any one of claims 1-5.