A welding apparatus and a welding method

Abstract

This invention discloses a welding device and method, belonging to the field of nuclear power technology. The device's welding head is equipped with a light source and a camera capable of emitting a linear light band. The light source illuminates the weld before welding, and the plane of the emitted linear light is parallel to the axis of the heat exchange tube. At least two cameras are used; at least one camera captures an image of the weld before welding to predict the welding trajectory; during welding, the two cameras capture short-wavelength and long-wavelength weld images respectively to correct the welding trajectory. The welding device and method disclosed in this invention offer strong automatic welding stability, can automatically adapt to different defects at the welding position, and achieve high welding precision.

Description

Technical Field

[0001] This invention relates to a welding equipment and welding method, belonging to the field of nuclear power technology. Background Technology

[0002] Steam generators have a large number of heat exchange tubes, with some nuclear power projects having more than 10,000 tubes. The tube diameter is relatively small, and the welding process between the tubes and the tube sheet holes mostly relies on manual welding. However, manual welding is inefficient and the welding yield fluctuates greatly.

[0003] Due to the small pipe diameter and high welding precision requirements, most existing automatic welding equipment cannot directly reach the welding position. Although some welding equipment can reach the welding position, the welding yield is low, and phenomena such as incomplete penetration are prone to occur, resulting in a high rework rate after welding.

[0004] In addition, although existing automatic welding methods have automatic calibration functions, most of these functions are based on PID calibration of the welding torch position, speed and angle. Their main function is to correct mechanical control deviations, but they cannot calibrate actual welding deviations, resulting in welding defects still occurring from time to time.

[0005] Therefore, it is necessary to conduct in-depth research on existing steam generator tube sheet and heat exchanger tube welding devices and methods to solve the above problems. Summary of the Invention

[0006] To overcome the aforementioned problems, in-depth research was conducted, and a welding device was designed, including a welding power source, a system controller, a wire feeding mechanism, and a movable welding head. The welding head has a tungsten electrode holder with a tungsten electrode mounted on it. The welding head is also equipped with a light source capable of emitting a linear light band and a camera.

[0007] The light source is used to irradiate the weld before welding, and the plane of the linear light emitted is parallel to the axis of the heat exchange tube.

[0008] The camera has at least two cameras. Before welding, at least one camera captures an image of the weld seam to predict the welding trajectory. During welding, the two cameras capture short-wavelength channel weld seam images and long-wavelength channel weld seam images respectively to correct the welding trajectory.

[0009] In a preferred embodiment, a temperature measuring module is also provided on the welding head for monitoring the temperature of the weld pool.

[0010] In a preferred embodiment, a short-wavelength bandpass filter and a long-wavelength bandpass filter are respectively installed in front of the two camera lenses. The camera with the long-wavelength bandpass filter is used to capture images of the 800-900nm long-wavelength light channel, and the camera with the short-wavelength bandpass filter is used to capture images of the 400-500nm short-wavelength light channel.

[0011] In a preferred embodiment, the welding head is further provided with a dichroic mirror to separate long-wavelength and short-wavelength light for separate imaging.

[0012] The present invention also discloses a welding method using the above-mentioned welding equipment, comprising the following steps:

[0013] S1. Before welding, the weld is scanned and photographed by a light source to obtain the scanned image. A trajectory prediction network is set up to obtain the first welding trajectory based on the scanned image.

[0014] S2. Welding is performed based on the first welding trajectory. During the welding process, welding images and the temperature of the weld pool are continuously acquired.

[0015] S3. Set up a welding feedback network to correct the first welding trajectory in real time based on the welding image and the temperature of the weld pool, so as to achieve stable welding.

[0016] In a preferred embodiment, in S1, the trajectory prediction network includes an image branch subnetwork, a parameter branch subnetwork, a feature fusion subnetwork, and a process decision subnetwork, wherein,

[0017] The input to the image branch subnetwork is a sequence of image frames captured by the camera, and its output is the depth, width, and heat map of the weld edge.

[0018] The parameter branch sub-network maps the relevant parameters of the input heat exchanger tube to be welded to the feature space;

[0019] The feature fusion subnetwork is used to concatenate the outputs of the image branch subnetwork and the parameter branch subnetwork to generate a comprehensive feature vector.

[0020] The process decision sub-network generates the first welding trajectory based on the comprehensive feature vector.

[0021] In a preferred embodiment, in S2, differential imaging processing is performed on the welding images captured by the two cameras to obtain the final welding image, in order to reduce arc interference, including the following steps:

[0022] S21. Establish a shortwave channel noise field model, expressed as follows:

[0023]

[0024] in, This represents the estimated noise field. An image representing a short-wavelength light channel. Indicates the Gaussian filter kernel;

[0025] S22. Establish dual-band noise correlation;

[0026]

[0027] in, This represents the predicted noise field in the long-wavelength ray channel image. This represents the radiative transfer coefficient. This represents the nonlinear correction factor;

[0028] S23. The processed image is obtained through dynamic difference, and is represented as follows:

[0029]

[0030] in, This represents the processed image. An image representing a long-wavelength light channel. These are dynamic weighting coefficients.

[0031] In a preferred embodiment, in step S3, the welding feedback network input layer includes an image input layer and a temperature input layer.

[0032] The input to the image input layer is the welding image at the current moment and welding images from a previous period of time;

[0033] The input to the temperature input layer is the current molten pool temperature and the temperature sequence over a previous period.

[0034] In a preferred embodiment, the welding feedback network has a feature extraction layer, which includes an image stream extraction layer and a temperature stream extraction layer. The image stream extraction layer is connected to the image input layer to extract the contour spatial features of the weld.

[0035] The temperature flow extraction layer is connected to the temperature input layer, encoding the temperature change trend and heat accumulation state.

[0036] In a preferred embodiment, the welding feedback network has a feature extraction layer and a multi-task prediction head.

[0037] The fusion layer connects the image stream extraction layer and the temperature stream extraction layer, and fuses their outputs to obtain a fused feature vector.

[0038] The multi-task prediction head decouples the fused features and outputs the welding head position offset, welding speed, and welding current.

[0039] The beneficial effects of this invention include:

[0040] (1) It can achieve automatic welding with high welding precision and extremely high welding qualification rate;

[0041] (2) Automatic welding has strong stability and can automatically adapt to different defects in the welding position, including weld roundness deviation, poor weld uniformity, etc., and has high welding standards. Attached Figure Description

[0042] Figure 1 A schematic diagram of a portion of the structure of a welding device according to a preferred embodiment of the present invention is shown;

[0043] Figure 2 A schematic diagram of a portion of the structure of a welding device according to a preferred embodiment of the present invention is shown;

[0044] Figure 3 A schematic diagram of the twisted structure of the linear light strip after the light source illuminates the weld seam according to a preferred embodiment of the present invention is shown.

[0045] Figure 4 This illustrates the relationship between the linear light band of the welding equipment and the weld seam according to a preferred embodiment of the present invention.

[0046] Figure 5 A schematic diagram of a welding method according to a preferred embodiment of the present invention is shown.

[0047] Explanation of icon numbers

[0048] 1-Light Source

[0049] 11-Linear light;

[0050] 2-Camera;

[0051] 3-Dichroic mirror

[0052] 4-Tungsten electrode;

[0053] 5-Weld. Detailed Implementation

[0054] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Through these descriptions, the features and advantages of the present invention will become clearer and more apparent.

[0055] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments. Although various aspects of embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless specifically indicated otherwise.

[0056] The welding position of the steam generator tube sheet and heat exchange tubes is annular.

[0057] According to the present invention, a welding device is similar to a conventional welding machine in that it has a welding power source, a system controller, a wire feeding mechanism, and a movable welding head.

[0058] The welding power source is used to provide the electrical energy required for welding.

[0059] The system controller is used to control the movement of various parts of the welding machine;

[0060] The wire feeding mechanism is used to uniformly deliver the welding wire to the welding area;

[0061] The welding head has a tungsten electrode clip, and a tungsten electrode 4 is disposed on the tungsten electrode clip.

[0062] Unlike traditional welding equipment, such as Figure 1 , Figure 2 As shown, a welding head is provided, on which a light source 1 capable of emitting a linear light band and a camera 2 are mounted. The light source 1 is used to illuminate the weld before welding, and the plane of the emitted linear light 11 is parallel to the axis of the heat exchange tube, such that the linear light is perpendicular to the projection of the weld 5 on the plane of the linear light. Figure 4 As shown,

[0063] The camera has at least two cameras. Before welding, at least one camera captures an image of the weld seam, and during welding, the two cameras capture short-wave channel weld seam images and long-wave channel weld seam images, respectively.

[0064] According to the present invention, before welding, the weld position and weld depth are determined by the bending of the light band based on an image of a linear light band illuminating the heat exchange tube.

[0065] While directly photographing the weld seam with a camera can identify the weld seam, its accuracy is low, and the information obtained, such as the depth and width of the weld seam, will have a large deviation, resulting in poor automatic welding effect.

[0066] In this invention, a linear light strip is used to illuminate the weld 5. When the linear light strip illuminates the weld, it will distort. This distortion not only characterizes the planar position of the weld but also reflects its depth. Figure 3 As shown, the linear light band is more sensitive to irregularities at the weld edge, thus more accurately reflecting the morphology of the position to be welded, providing a basis for fine welding.

[0067] According to a preferred embodiment of the present invention, the emitted light rays in the plane containing the weld are not collinear with any radius in the circular cross-section of the heat exchanger tube, thereby ensuring that the light band exhibits a curvature change at the weld position. According to the present invention, when the linear light emitted by the light source is collinear with the radius in the circular cross-section of the heat exchanger tube, the curvature change of the light band at the weld position is small, and the accuracy of obtaining weld depth information decreases significantly.

[0068] In a preferred embodiment, the light source is a hybrid light source, that is, it includes long-spectrum light and short-spectrum light. During welding, a short-wave bandpass filter and a long-wave bandpass filter are respectively installed in front of the two camera lenses to capture the long-wave and short-wave light separately, thereby using differential imaging to suppress welding arc interference.

[0069] Unlike traditional automatic calibration in welding, this invention employs computer vision for automatic welding calibration. However, abnormally bright welding arcs can occur during the welding process, making it difficult to accurately determine the actual welding state at the welding location in the captured images. In this invention, two cameras are used to capture images in long-wavelength and short-wavelength channels respectively. Differential imaging is used to suppress welding arc interference, thereby allowing the welding state to be presented more clearly, which is then used to correct the welding process.

[0070] According to the present invention, the long-wavelength light is light with a wavelength of 800-900nm; the short-wavelength light source is light with a wavelength of 400-500nm.

[0071] In this invention, a camera equipped with a long-wavelength bandpass filter is called a long-wavelength camera, and a camera equipped with a short-wavelength bandpass filter is called a short-wavelength camera. That is, the long-wavelength camera is used to capture images of long-wavelength light channels of 800-900nm, and the short-wavelength camera is used to capture images of short-wavelength light channels of 400-500nm.

[0072] Preferably, the shortwave camera captures images with a main wavelength of 450nm, and the longwave camera captures images with a main wavelength of 850nm.

[0073] Analysis revealed that the long-wavelength (850nm) channel mainly contains arc noise and weld information, while the short-wavelength (450nm) channel mainly contains arc noise. Differential imaging of the two can purify the weld signal and suppress arc interference signals.

[0074] In a preferred embodiment, both the light source and the camera are located on the plane of the weld.

[0075] In a preferred embodiment, both the light source and the camera are fixedly mounted on the welding head, so that the positions of the light source and the camera, as well as the two cameras, are relatively fixed, ensuring image consistency when taking pictures of different positions of the weld before welding and during welding.

[0076] According to the present invention, the welding head can rotate about the heat exchange tube axis as the center line, thereby photographing and welding the annular weld.

[0077] In this invention, the specific relative positions between the light source and the camera are not limited, and those skilled in the art can freely set them according to actual needs.

[0078] In this invention, the specific structure of the welding head is not limited. Those skilled in the art can freely set it according to actual needs, as long as it can fix the light source, camera and tungsten electrode clamp, and rotate with the heat exchange tube axis as the center line.

[0079] According to the present invention, a trajectory prediction network and a welding feedback network are provided in the system controller.

[0080] The trajectory prediction network predicts the welding trajectory based on the weld image before welding.

[0081] The welding feedback network corrects the welding trajectory based on the weld seam image during welding.

[0082] According to the present invention, before welding, an image of the bending of the light strip at the weld is captured by any camera, and scanned images are obtained by sequentially capturing images of all welds. The trajectory prediction network obtains the welding trajectory based on the scanned images.

[0083] During welding, two cameras are used to capture welding images in the long-wavelength and short-wavelength light channels, respectively. The welding feedback network monitors and corrects the welding process based on the weld images.

[0084] In a preferred embodiment, a temperature measuring module is also provided on the welding head for monitoring the temperature of the weld pool.

[0085] According to the present invention, the tungsten electrode clamp is a six-degree-of-freedom moving platform, which allows the tungsten electrode to move freely inside the heat exchange tube to achieve welding at different angles and positions.

[0086] Furthermore, before welding, the light source and camera are working, and the tungsten electrode holder moves the tungsten electrode outside the light path of the light source to avoid the tungsten electrode blocking the light; during welding, the tungsten electrode holder moves the tungsten electrode to the weld position for welding.

[0087] In a preferred embodiment, the welding head is further provided with a dichroic mirror 3, which separates long-wavelength and short-wavelength light for separate imaging.

[0088] The dichroic mirror is a commonly used device in laser optical paths. Its characteristic is that it transmits light of a certain wavelength almost completely, while reflecting light of other wavelengths almost completely. Therefore, placing it in the optical path can achieve the separation of long-wavelength and short-wavelength light.

[0089] Furthermore, one camera is positioned in the projection direction of the dichroic mirror, and the other camera is positioned in the reflection direction of the dichroic mirror, so that the two cameras can simultaneously capture long-wavelength and short-wavelength channel images of the same area, which facilitates subsequent image processing.

[0090] In a preferred embodiment, the camera is provided with a cooling water channel to cool the camera and prevent the large amount of heat generated during the welding process from damaging the camera.

[0091] Obviously, similar to traditional welding heads, the welding head is also equipped with necessary units such as protective gas nozzles. In this invention, the specific structure of these units is not limited, and those skilled in the art can freely set them according to actual needs.

[0092] This invention discloses a welding method, such as... Figure 5 As shown, it includes the following steps:

[0093] S1. Before welding, the weld is scanned and photographed by a light source to obtain the scanned image. A trajectory prediction network is set up to obtain the first welding trajectory based on the scanned image.

[0094] S2. Welding is performed based on the first welding trajectory. During the welding process, welding images and the temperature of the weld pool are continuously acquired.

[0095] S3. Set up a welding feedback network to correct the first welding trajectory in real time based on the welding image and the temperature of the weld pool, so as to achieve stable welding.

[0096] According to the present invention, the welding trajectory includes a welding head trajectory curve and welding current and welding speed at different positions on the curve.

[0097] Furthermore, in S1, during the scanning process, any camera can be used to capture the image. Preferably, a camera equipped with a long-wavelength bandpass filter is used.

[0098] In S1, the trajectory prediction network includes an image branch subnetwork, a parameter branch subnetwork, a feature fusion subnetwork, and a process decision subnetwork, wherein,

[0099] The input to the image branch subnetwork is an image frame captured by the camera, and its output is the depth and width of the weld. Preferably, it also includes a heat map of the weld edge.

[0100] Preferably, the image frame is processed into grayscale before input.

[0101] In a preferred embodiment, the image branch subnetwork includes a feature extraction unit and a geometric feature decoder arranged sequentially.

[0102] The feature extraction unit is used to extract features from the image frame to obtain a feature map, and the geometric feature decoder is used to map the extracted features into weld depth and width information.

[0103] In a preferred embodiment, the feature extraction unit includes a multi-layer convolutional structure and a spatial attention mechanism. The multi-layer convolutional structure can extract image features from image frames, and the spatial attention mechanism can focus on the extracted image features to obtain a weighted feature map.

[0104] According to the present invention, by setting a spatial attention mechanism in the feature extraction unit, the grating distortion features of the weld area are focused, background noise is suppressed, thereby adaptively enhancing the response weight of the weld edge area and improving feature quality.

[0105] The multi-layer convolutional structure and spatial attention mechanism are commonly used structures in neural networks. In this invention, their specific parameters are not limited, and those skilled in the art can set them freely based on experience.

[0106] Preferably, the geometric feature decoder includes a depth prediction subunit and a width prediction subunit.

[0107] The depth prediction subunit predicts information related to weld depth based on extracted feature maps. Preferably, it has a series of dilated convolutional layers, a global max pooling layer, and a regression layer connected in sequence.

[0108] Among them, multiple convolutional layers can expand the receptive field to capture large-scale light band distortion, global max pooling can transform distortion features into feature vectors, and regression layers transform feature vectors into depth scalars.

[0109] The width prediction subunit predicts information related to the weld width based on the extracted feature map. Preferably, it has a series of dilated convolutional layers, a global average pooling layer, and a regression layer connected in sequence.

[0110] Among them, multi-layer dilated convolutional layers can expand the receptive field, global average pooling layers can comprehensively evaluate the overall distortion region to obtain feature vectors, and regression layers convert feature vectors into width scalars.

[0111] Preferably, the geometric feature decoder further includes an edge position prediction subunit, which predicts a probability heatmap of the weld edge based on the extracted feature map.

[0112] Preferably, the edge location prediction subunit has a transposed convolutional layer, a coordinate enhancement layer, a cascaded convolutional layer, and a heatmap generation layer.

[0113] The transposed convolutional layer recovers spatial details through learnable upsampling.

[0114] The coordinate enhancement layer is used to generate two matrices with the same size as the feature map output by the transposed convolutional layer: a horizontal coordinate matrix and a vertical coordinate matrix. The values ​​in each column of the horizontal coordinate matrix change linearly from -1 to 1, and the values ​​in each row of the vertical coordinate matrix change linearly from -1 to 1. The horizontal coordinate matrix and the vertical coordinate matrix are concatenated to the feature map output by the transposed convolutional layer and then output.

[0115] The use of coordinate enhancement layers can solve the translation invariance defect of CNN network structure, enabling the network to perceive absolute position and improve edge localization accuracy.

[0116] Cascaded convolutional layers are used to fuse the channel features output by the coordinate enhancement layer, thereby reducing the number of channels and parameters.

[0117] The heatmap generation layer performs 1×1 convolution compression and bilinear upsampling on the output of the cascaded convolutional layer to generate a single-channel heatmap. Each pixel value in the heatmap represents the probability that the pixel is an edge (the pixel value is 0 or 1, 0 indicates non-edge, and 1 indicates edge), thereby achieving sub-pixel weld edge localization.

[0118] The parameter branch sub-network maps the relevant parameters of the input heat exchanger tube to be welded to the feature space, and adopts a parameter embedding layer structure.

[0119] Preferably, the relevant parameters of the heat exchange tube to be welded include the inner diameter of the tube and the material thickness.

[0120] The feature fusion subnetwork is used to concatenate the outputs of the image branch subnetwork and the parameter branch subnetwork to generate a comprehensive feature vector.

[0121] The process decision sub-network generates the first welding trajectory based on the comprehensive feature vector.

[0122] Preferably, the process decision sub-network includes a two-stage LSTM network and an output layer.

[0123] The first-level LSTM network takes the comprehensive feature vector as input and outputs a short-term spatiotemporal sequence to capture the continuity features of the weld.

[0124] The second-stage LSTM network takes short-term spatiotemporal sequences as input and outputs long-term dependency sequences.

[0125] The output layer includes a shared feature processing unit and multiple parallel task output branch units.

[0126] The shared feature processing unit is a fully connected layer that takes long-term dependency sequences as input and outputs enhanced feature vectors.

[0127] The multiple parallel task output branch units include a trajectory offset output branch unit, a welding speed branch unit, and a welding current branch unit.

[0128] Preferably, the trajectory offset output branch unit, welding speed branch unit, and welding current branch unit all adopt a fully connected layer and a ReLU activation function, taking the enhanced feature vector as input, and output the welding trajectory curve coordinates, the corresponding welding speed, and the welding current, respectively.

[0129] In this invention, a two-level LSTM network is used to handle the continuous changes in the circumferential weld, making the predicted trajectory smooth and continuous, and avoiding welding abnormalities caused by abrupt changes.

[0130] According to the present invention, the trajectory offset output branch unit outputs the welding head movement trajectory, the welding speed branch unit outputs the corresponding welding speed, and the welding current branch unit outputs the corresponding welding current.

[0131] Although the trajectory prediction network can output a relatively accurate welding trajectory, in the actual welding process, due to material reasons, welding rod reasons, welding machine vibration and other reasons, there may still be minor defects after welding according to the first welding trajectory. Since there are a large number of heat exchange tubes, some projects have more than 10,000, these minor defects will cause some heat exchange tubes to need to be reworked. How to further reduce welding abnormalities and improve the welding qualification rate is the difficulty of this invention.

[0132] According to the present invention, by real-time monitoring of the welding process and correction of the welding process based on the monitoring results, the welding qualification rate is improved.

[0133] Preferably, the monitoring is based on welding images and the temperature of the weld pool.

[0134] In this invention, the welding trajectory is corrected in real time during the welding process, using the real-time temperature of the weld pool as a reference, thereby further improving the welding effect.

[0135] In S2, the temperature of the weld pool is obtained by detecting it using an infrared temperature measurement module.

[0136] According to the present invention, in S2, it is no longer necessary to use a linear light strip to illuminate the weld; it is only necessary for the welding position to have a natural light environment.

[0137] The inventors discovered that images directly captured by the camera contain a large number of electric arcs, which severely affect the monitoring of the welding process and render the welding images unusable for correcting the welding process.

[0138] In this invention, a long-wave camera and a short-wave camera are used to take pictures of the welding area. The welding images taken by the two cameras are processed by differential imaging to obtain the final welding image, so as to reduce arc interference and enable the subsequent feedback network to obtain a more accurate welding image, thereby correcting the welding process.

[0139] In S2, the differential imaging processing includes the following steps:

[0140] S21. Establish a shortwave channel noise field model, expressed as follows:

[0141]

[0142] in, This represents the estimated noise field. An image representing a short-wavelength light channel. This represents the Gaussian filter kernel. It is the standard deviation parameter of the Gaussian function;

[0143] According to the present invention, the noise field characterizes the principal component of the electric arc interference.

[0144] S22. Establish dual-band noise correlation;

[0145]

[0146] in, This represents the predicted noise field in the long-wavelength ray channel image. This represents the radiative transfer coefficient. This represents the nonlinear correction factor. 、 The specific values ​​can be freely set by those skilled in the art according to actual needs, and are not limited in this invention.

[0147] S23. The processed image is obtained through dynamic difference, and is represented as follows:

[0148]

[0149] in, This represents the processed image. An image representing a long-wavelength light channel. These are dynamic weighting coefficients.

[0150] Preferably, dynamic weighting coefficient α Set to:

[0151]

[0152] in, This represents the maximum grayscale value of the image.

[0153] In this invention, dynamic weighting coefficients It achieves adaptive noise reduction intensity, with higher weighting for stronger electric arcs.

[0154] According to the present invention, since the shortwave channel is more sensitive to arc noise, the noise field of the longwave image can be predicted based on the shortwave image, which can more accurately identify the arc in the longwave image, thereby removing the arc noise in the longwave image and obtaining a welding image with less arc interference.

[0155] In S3, the welding feedback network includes an input layer, a feature extraction layer, a fusion layer, a multi-task prediction head, and an output layer.

[0156] The input layer includes an image input layer and a temperature input layer.

[0157] The input to the image input layer is the welding image at the current moment and the welding images from a previous period, preferably the current frame image and the previous two frames. The welding image is obtained in S2.

[0158] The input to the temperature input layer is the current molten pool temperature and the temperature sequence of a previous period, preferably the current molten pool temperature and the sampled temperature sequence within the previous second.

[0159] The feature extraction layer includes an image stream extraction layer and a temperature stream extraction layer. The image stream extraction layer is connected to the image input layer and extracts the spatial features of the weld contour.

[0160] Preferably, the image stream extraction layer uses a lightweight CNN network, such as MobileNetV3, to reduce computational load and improve real-time performance.

[0161] The temperature flow extraction layer is connected to the temperature input layer, encoding the temperature change trend and heat accumulation state.

[0162] Preferably, the temperature flow extraction layer employs a temporal convolutional network.

[0163] The fusion layer connects the image stream extraction layer and the temperature stream extraction layer, and fuses their outputs to obtain a fused feature vector.

[0164] Preferably, the fusion layer includes channel attention, which dynamically weights the importance of image channels based on temperature features to obtain the fused features, expressed as follows:

[0165]

[0166] in, The output of the image stream extraction layer, Indicates global average pooling. This represents the image stream after pooling. This indicates the output of the temperature flow extraction layer. express and Vector concatenation Indicates the weights of the fully connected layer. Indicates the bias term. It is the ReLU activation function. It is the Sigmoid activation function. This is the weight matrix. This indicates multiplication by channel. This indicates the features after fusion.

[0167] In this invention, a fusion layer with channel attention is used to dynamically fuse temperature features and image features, making the subsequent correction trajectory highly correlated with the temperature of the weld pool, thereby making the corrected welding effect more stable.

[0168] The multi-task prediction head decouples the fused features and outputs the welding head position offset, welding speed, and welding current.

[0169] Preferably, the multi-task prediction head includes a position correction head, a velocity correction head, and a current correction head, all of which adopt a fully connected network structure.

[0170] Independent position correction head, speed correction head, and current correction head can prevent trajectory correction from interfering with speed decision-making and reduce the risk of oscillation.

[0171] In a preferred embodiment, the multi-task prediction head also has a temperature prediction head for predicting the temperature of the weld pool at the next moment, and the temperature prediction head adopts a fully connected network structure.

[0172] Furthermore, the temperature prediction head is enabled during training and disabled after formal deployment. By using temperature prediction as a self-supervised signal, the network is constrained to understand the thermodynamic process, forcing it to learn physical laws (such as heat conduction hysteresis) rather than simply fitting data. This enables early triggering of current adjustments and avoids parameter abrupt overshoot, thereby improving welding stability.

[0173] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," "outer," "front," and "rear," etc., indicate the orientation or positional relationship based on the orientation or positional relationship in the working state of this invention, and are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention. Furthermore, the terms "first," "second," "third," and "fourth" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0174] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0175] The present invention has been described above with reference to preferred embodiments; however, these embodiments are merely exemplary and illustrative. Various substitutions and modifications can be made to the present invention based on these embodiments, all of which fall within the scope of protection of the present invention.

Claims

1. A welding apparatus, comprising a welding power source, a system controller, a wire feeding mechanism, and a movable welding head, wherein the welding head has a tungsten electrode holder, and a tungsten electrode is disposed on the tungsten electrode holder, characterized in that, The welding head is equipped with a light source and a camera capable of emitting linear light bands. The light source is used to irradiate the weld before welding, and the plane of the linear light emitted is parallel to the axis of the heat exchange tube. The camera has at least two cameras. Before welding, at least one camera captures a weld image to predict the welding trajectory. During welding, the two cameras capture short-wave channel weld images and long-wave channel weld images respectively to correct the welding trajectory. A short-wavelength bandpass filter and a long-wavelength bandpass filter are installed in front of the two camera lenses, respectively. The camera with the long-wavelength bandpass filter is used to capture images of the 800-900nm long-wavelength light channel, and the camera with the short-wavelength bandpass filter is used to capture images of the 400-500nm short-wavelength light channel.

2. The welding equipment according to claim 1, characterized in that, A temperature measuring module is also provided on the welding head to monitor the temperature of the weld pool.

3. The welding equipment according to claim 2, characterized in that, The welding head is also equipped with a dichroic mirror, which separates long-wave and short-wave light for separate imaging.

4. A welding method, characterized in that, The welding is performed using the welding equipment described in any one of claims 1-3, and includes the following steps: S1. Before welding, the weld is scanned and photographed by a light source to obtain the scanned image. A trajectory prediction network is set up to obtain the first welding trajectory based on the scanned image. S2. Welding is performed based on the first welding trajectory. During the welding process, welding images and the temperature of the weld pool are continuously acquired. S3. Set up a welding feedback network to correct the first welding trajectory in real time based on the welding image and the temperature of the weld pool, so as to achieve stable welding.

5. The welding method according to claim 4, characterized in that, In S1, the trajectory prediction network includes an image branch subnetwork, a parameter branch subnetwork, a feature fusion subnetwork, and a process decision subnetwork, wherein, The input to the image branch subnetwork is a sequence of image frames captured by the camera, and its output is the depth, width, and heat map of the weld edge. The parameter branch sub-network maps the relevant parameters of the input heat exchanger tube to be welded to the feature space; The feature fusion subnetwork is used to concatenate the outputs of the image branch subnetwork and the parameter branch subnetwork to generate a comprehensive feature vector. The process decision sub-network generates the first welding trajectory based on the comprehensive feature vector.

6. The welding method according to claim 4, characterized in that, In S2, differential imaging processing is performed on the welding images captured by the two cameras to obtain the final welding image, in order to reduce arc interference. This includes the following steps: S21. Establish a shortwave channel noise field model, expressed as follows: ， in, This represents the estimated noise field. An image representing a short-wavelength light channel. Indicates the Gaussian filter kernel; S22. Establish dual-band noise correlation; ， in, This represents the predicted noise field in the long-wavelength ray channel image. This represents the radiative transfer coefficient. This represents the nonlinear correction factor; S23. The processed image is obtained through dynamic difference, and is represented as follows: ， in, This represents the processed image. An image representing a long-wavelength light channel. These are dynamic weighting coefficients.

7. The welding method according to claim 4, characterized in that, In S3, the welding feedback network input layer includes an image input layer and a temperature input layer. The input to the image input layer is the welding image at the current moment and welding images from a previous period of time; The input to the temperature input layer is the current molten pool temperature and the temperature sequence over a previous period.

8. The welding method according to claim 7, characterized in that, The welding feedback network has a feature extraction layer, which includes an image stream extraction layer and a temperature stream extraction layer. The image stream extraction layer is connected to the image input layer to extract the contour spatial features of the weld. The temperature flow extraction layer is connected to the temperature input layer, encoding the temperature change trend and heat accumulation state.

9. The welding method according to claim 8, characterized in that, The welding feedback network has a feature extraction layer, a fusion layer, and a multi-task prediction head. The fusion layer connects the image stream extraction layer and the temperature stream extraction layer, and fuses their outputs to obtain a fused feature vector. The multi-task prediction head decouples the fused features and outputs the welding head position offset, welding speed, and welding current.

Images