Vacuum laser welding full gesture adaptive regulation method and system

Abstract

This invention discloses a method and system for adaptive control of the entire attitude in vacuum laser welding. The method includes: acquiring basic data of the workpiece to be welded and the robot; extracting three-dimensional trajectory data of the weld and discretizing it to generate multiple weld state nodes; collecting real-time welding data and mapping it to the current weld state node and its neighboring weld state nodes to form a node feature matrix and feature confidence; physically coupling the weld state nodes to generate an adjacency matrix and edge features; calculating attention weights and gating propagation coefficients in a physically informed gated graph attention network and updating the nodes to obtain the predicted welding state value for the next moment; and generating control adjustment quantities by combining preset control parameters to achieve adaptive control of the entire attitude in vacuum laser welding. This invention integrates multi-dimensional node features and utilizes a gated graph attention mechanism that introduces physical balance residual penalties for node updates and state prediction, achieving real-time and precise closed-loop control of welding quality under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of advanced welding manufacturing technology, and more specifically to a method and system for adaptive control of all postures in vacuum laser welding. Background Technology

[0002] Vacuum laser welding, as a high-energy beam precision joining technology, has been widely used in aerospace, nuclear industry, and high-end equipment manufacturing due to its significant advantages such as deep penetration, low heat input, and high weld purity. Especially in the welding of complex components in all postures, the welding quality is easily affected by changes in the direction of gravity vector and the unique hydrodynamic behavior of the molten pool in a vacuum environment, resulting in defects such as porosity, collapse, or keyhole instability. Therefore, real-time monitoring and adaptive control of the welding process are crucial to ensuring the reliability of high-end equipment.

[0003] Existing welding quality control methods mainly rely on offline process parameter optimization or simple feedback control based on traditional sensors. However, these methods have significant limitations when facing complex welding scenarios with varying attitudes and working conditions. On the one hand, while traditional numerical simulation methods can reveal physical mechanisms, they are computationally expensive and cannot meet the millisecond-level real-time control requirements. On the other hand, while purely data-driven deep learning models (such as CNNs and RNNs) have fast inference speeds, they are often considered "black boxes," lacking explicit constraints on physical laws such as energy conservation and momentum balance. This results in poor generalization ability outside the training data distribution, and the prediction results may violate physical common sense.

[0004] Furthermore, the weld pool is a typical multiphysics coupled dynamic system, with highly nonlinear and spatiotemporally correlated thermal-fluid-force interactions within it. While existing graph neural network methods can handle non-Euclidean data, they often neglect the physical equilibrium residuals between nodes during feature aggregation, making it difficult to accurately capture the transient changes in weld pool flow under vacuum conditions.

[0005] Therefore, there is an urgent need for an adaptive control method that can both guarantee real-time performance and deeply integrate physical mechanisms to solve the technical problems of inaccurate prediction of welding status and lag in control response under complex working conditions. Summary of the Invention

[0006] In view of the above problems, this invention is proposed to provide a vacuum laser welding full-attitude adaptive control method and system that overcomes or at least partially solves the above problems. The method clarifies the three-dimensional trajectory and spatial nodes of the weld seam in the pre-welding stage, collects and maps real-time data on the molten pool, keyhole, temperature, attitude, energy field and vacuum environment during the welding process, and performs state prediction and control quantity generation through the information propagation mechanism of physical equilibrium residual constraints, so as to solve the welding stability control problem under the coupled influence of attitude change, energy field shift, keyhole fluctuation and vacuum environment.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] In a first aspect, embodiments of the present invention provide a method for adaptive control of all postures in vacuum laser welding, comprising the following steps: S1. Obtain basic data of the workpiece to be welded and the welding robot, and extract the three-dimensional trajectory data of the weld; the three-dimensional trajectory data of the weld includes the spatial position data of the weld and the local geometric state data. S2. Discretize the three-dimensional trajectory data of the weld according to a preset spatial step size to generate multiple weld state nodes; wherein, each weld state node is bound to the corresponding weld spatial position data and local geometric state data; S3. During vacuum laser welding, the current weld state node is determined based on the current position of the welding head; real-time welding data is collected and mapped to the current weld state node and its adjacent weld state nodes to form a node feature matrix and feature confidence. S4. Physically couple the weld state node, attitude node, laser energy field node and material environment node, and generate an adjacency matrix and edge features based on the weld spatial adjacency relationship, heat conduction coupling relationship, attitude gravity projection relationship, laser energy input relationship and vacuum environment constraint relationship. S5. Calculate the physical balance residual and the physical balance residual between nodes based on the node feature matrix, adjacency matrix, edge features and multiple physical constraints. S6. In the physically informed gating graph attention network, the attention weights and gating propagation coefficients are calculated based on the node feature matrix, edge features and physical balance residuals between nodes, and the nodes are updated based on the attention weights and gating propagation coefficients to obtain the predicted welding state value at the next moment. S7. Based on the predicted welding state value at the next moment and the preset control parameters, generate a control adjustment amount, and adjust the laser power, scanning speed, focal position, spot shape, attitude compensation amount and local vacuum degree according to the control adjustment amount to realize full attitude adaptive control of vacuum laser welding.

[0009] Secondly, embodiments of the present invention provide a vacuum laser welding all-attitude adaptive control system, employing the method described in any one of the first aspects, comprising the following modules: Pre-welding data acquisition module: used to acquire basic data of the workpiece to be welded and the welding robot, and extract three-dimensional trajectory data of the weld; the three-dimensional trajectory data of the weld includes spatial position data of the weld and local geometric state data; Weld joint node generation module: used to discretize the three-dimensional trajectory data of the weld joint according to a preset spatial step size to generate multiple weld joint state nodes; wherein, each weld joint state node is bound to the corresponding weld joint spatial position data and local geometric state data; Feature extraction and confidence calibration module: used to determine the current weld state node based on the current position of the welding head during vacuum laser welding; collect real-time welding data and map the real-time welding data to the current weld state node and its neighboring weld state nodes to form a node feature matrix and feature confidence; Physical coupling construction module: used to physically couple the weld state node, attitude node, laser energy field node and material environment node, and generate adjacency matrix and edge features based on weld spatial adjacency relationship, heat conduction coupling relationship, attitude gravity projection relationship, laser energy input relationship and vacuum environment constraint relationship; Physical balance residual calculation module: used to calculate physical balance residual and inter-node physical balance residual based on the node feature matrix, adjacency matrix, edge features and multiple physical constraints; Physically informed gating graph attention network prediction module: In the physically informed gating graph attention network, it calculates attention weights and gating propagation coefficients based on the node feature matrix, edge features and physical balance residuals between nodes, and updates nodes based on the attention weights and gating propagation coefficients to obtain the predicted welding state value at the next moment. Scene-specific control output module: It is used to generate control adjustment amount based on the predicted value of welding state at the next moment and preset control parameters, and adjust the laser power, scanning speed, focal position, spot shape, attitude compensation amount and local vacuum degree according to the control adjustment amount, so as to realize the full attitude adaptive control of vacuum laser welding.

[0010] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a vacuum laser welding full-attitude adaptive control method and system, which has the following beneficial effects: First, the present invention discretizes the three-dimensional trajectory of the weld into weld state nodes and maps real-time molten pool, keyhole, temperature, attitude, energy field and vacuum environment data to the corresponding nodes, which solves the problem that welding process data and weld spatial position are not clearly bound in the existing methods.

[0011] Second, this invention physically couples the weld state node, attitude node, laser energy field node, and material environment node, enabling the unified expression of weld spatial continuity, attitude gravity projection, laser energy input, material thermal properties, and vacuum environment constraints.

[0012] Third, this invention introduces energy balance residual, heat conduction residual, molten pool / keyhole stability residual, attitude gravity projection residual, and vacuum environment residual into attention weights and gating propagation coefficients, so that network prediction not only depends on data correlation, but is also constrained by the physical laws of vacuum laser welding.

[0013] Fourth, this invention reduces the information propagation intensity of abnormal nodes corresponding to splashing, reflection, keyhole collapse, sensor jitter, or attitude change by using a residual suppression gating mechanism, thereby improving the stability of cross-attitude prediction.

[0014] Fifth, the present invention outputs attitude compensation control quantity or energy field / vacuum degree compensation control quantity according to whether the workpiece is easy to adjust its attitude, which can adapt to different engineering scenarios such as small adjustable attitude workpieces, large difficult-to-adjust attitude workpieces, curved surface welds and circumferential welds.

[0015] Sixth, this invention uses the oscillation parameter as an optional laser energy field adjustment quantity, rather than a single core control method. Therefore, it can avoid the control strategy being limited to oscillation beam adjustment and is more suitable for all-attitude vacuum laser welding processes where attitude change, vacuum environment and molten pool / keyhole stability are the main constraints. Attached Figure Description

[0016] 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.

[0017] Figure 1 This is a flowchart of the vacuum laser welding full-attitude adaptive control method provided in the embodiments of the present invention; Figure 2 This is the physical awareness gating graph attention network propagation logic diagram provided in the embodiments of the present invention; Figure 3 This is a diagram showing the correspondence between scenario-based control outputs and actuators provided in this embodiment of the invention. Figure 4 This is a block diagram of the vacuum laser welding all-attitude adaptive control system architecture provided in this embodiment of the invention; Figure 5 This is a schematic diagram of the architecture of the vacuum laser welding all-attitude adaptive control system provided in this embodiment of the invention. Detailed Implementation

[0018] 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.

[0019] Example 1 This invention discloses a method for adaptive control of all postures in vacuum laser welding, referring to... Figure 1 As shown, it includes the following steps: S1. Obtain the basic data of the workpiece to be welded and the welding robot, and extract the three-dimensional trajectory data of the weld; the three-dimensional trajectory data of the weld includes the spatial position data of the weld and the local geometric state data. S2. Discretize the three-dimensional trajectory data of the weld according to the preset spatial step size to generate multiple weld status nodes; wherein, each weld status node is bound to the corresponding weld spatial position data and local geometric state data; S3. During vacuum laser welding, determine the current weld state node based on the current position of the welding head; collect real-time welding data and map the real-time welding data to the current weld state node and its adjacent weld state nodes to form a node feature matrix and feature confidence. S4. Physically couple the weld state node, attitude node, laser energy field node, and material environment node, and generate the adjacency matrix and edge features based on the weld spatial adjacency relationship, heat conduction coupling relationship, attitude gravity projection relationship, laser energy input relationship, and vacuum environment constraint relationship. S5. Calculate the physical balance residual and the physical balance residual between nodes based on the node feature matrix, adjacency matrix, edge features and multiple physical constraints. S6. In the physically informed gating graph attention network, the attention weights and gating propagation coefficients are calculated based on the node feature matrix, edge features and physical balance residuals between nodes. The nodes are then updated based on the attention weights and gating propagation coefficients to obtain the predicted welding state value at the next moment. S7. Based on the predicted welding state value at the next moment and the preset control parameters, generate control adjustment quantities, and adjust the laser power, scanning speed, focal position, spot shape, attitude compensation amount and local vacuum degree according to the control adjustment quantities to achieve full attitude adaptive control of vacuum laser welding.

[0020] This embodiment applies to vacuum laser welding of circumferential welds in large fuel tanks in the aerospace manufacturing field. The fuel tanks are composed of multiple cylindrical sections, characterized by large diameters, long welds, and extremely high sealing requirements. To ensure the weld's depth-to-width ratio and purity, deep-penetration welding using a high-energy laser in a vacuum environment is necessary. Due to the large size of the workpiece, a method of "workpiece rotation and welding head fixation" or "gantry-driven welding head movement" is typically employed. During this process, the welding head undergoes various attitude changes (such as gradually transitioning from a flat welding position to a vertical or overhead welding position, or moving along complex curved surfaces).

[0021] In this embodiment, the 3D trajectory of the weld is first extracted from the 3D model of the tank to be welded, welding drawings, robot teaching trajectory, offline programming trajectory, or 3D scan point cloud, and the weld centerline is discretized into weld state nodes. Subsequently, during the vacuum laser welding process, the molten pool image, keyhole image, temperature field, attitude data, laser energy field parameters, and local vacuum degree are mapped to the corresponding nodes, and combined with material thermophysical parameters to form a node feature matrix, forming a physically coupled heterogeneous graph composed of weld state nodes, attitude nodes, laser energy field nodes, and material environment nodes. Then, the energy balance residual, heat conduction residual, molten pool / keyhole stability residual, attitude gravity projection residual, and vacuum environment residual are further calculated, and these residuals are introduced into the graph attention weight and gating propagation coefficient to suppress the information propagation of nodes corresponding to abnormal molten pools, keyhole fluctuations, sensor interference, or attitude abrupt changes. Based on the next-moment welding state prediction value output by the physically informed gating graph attention network, laser power, scanning speed, focal position, spot shape, attitude compensation amount, and local vacuum degree adjustment amount are generated and sent to the welding actuator through closed-loop control. This enables the improvement of welding process stability, energy field adaptability, and weld formation consistency in all posture vacuum laser welding scenarios, including flat welding, vertical welding, overhead welding, circumferential welding, and curved surface welding.

[0022] The implementation steps of this embodiment include: S1: Pre-welding basic data acquisition and weld 3D trajectory extraction. Acquire at least one of the following data: 3D model of the workpiece to be welded, welding drawings, robot teaching trajectory, offline programming trajectory or 3D scanning point cloud. Extract the weld centerline coordinate sequence, weld tangent vector, weld normal vector, local curvature, plate thickness, groove parameters and material parameters to form weld 3D trajectory data. S2: Weld discretization and weld state node initialization: Discretize the three-dimensional trajectory data of the weld according to the preset spatial step size to generate multiple weld state nodes, and bind each weld state node with the corresponding spatial coordinates, weld tangent vector, weld normal vector, local lateral quantity, local curvature, plate thickness, bevel parameters and material thermophysical parameters. S3: Real-time welding data acquisition and node mapping. During vacuum laser welding, the current weld state node is determined based on the current position of the welding head. Real-time welding data is acquired and mapped to the current weld state node and its neighboring weld state nodes to form a node feature matrix and feature confidence. The real-time welding data includes molten pool image, keyhole image, temperature field, attitude data, laser energy field parameters, and local vacuum degree. The material thermophysical parameters obtained in step S1 are used to jointly form the node feature matrix. S4: Physical coupling heterogeneous graph construction, constructing a physical coupling heterogeneous graph including the weld state node, attitude node, laser energy field node and material environment node, and generating an adjacency matrix and edge features based on the weld spatial adjacency relationship, heat conduction coupling relationship, attitude gravity projection relationship, laser energy input relationship and vacuum environment constraint relationship; S5: Physical balance residual calculation: Based on the node feature matrix, the adjacency matrix, the edge features, energy balance constraints, heat conduction constraints, molten pool / keyhole stability constraints, attitude gravity projection constraints, and vacuum environment constraints, calculate the physical balance residual and the physical balance residual between nodes. S6: Physically informed gating graph attention prediction. In the physically informed gating graph attention network, attention weights and gating propagation coefficients are calculated based on node hidden features, node feature correlation scores, edge features, and physical balance residuals between nodes. Nodes are updated based on the attention weights and gating propagation coefficients to obtain the predicted welding state value at the next moment. S7: Scene-specific control output and closed-loop execution. Based on the predicted welding state value at the next moment, the preset welding quality target, the current control quantity, the quality deviation weight matrix, the control smoothing weight matrix, the predicted physical balance residual, the physical balance residual penalty weight, the lower limit of the control quantity, and the upper limit of the control quantity, a control adjustment quantity is generated. Based on the control adjustment quantity, the laser power, scanning speed, focal position, spot shape, attitude compensation amount, and / or local vacuum degree are adjusted to achieve full attitude adaptive control of vacuum laser welding.

[0023] The specific implementation steps are described in detail below.

[0024] S1: First, obtain the pre-welding basic data of the workpiece to be welded. This pre-welding basic data is not from a single source, but rather comprises data from geometric sources, process sources, motion sources, measured sources, and material sources. The sources and functions of the different data items are as follows.

[0025] For workpieces with complete 3D models, the weld centerline can be extracted from the weld boundary line, joint assembly line, or welding process layer in the 3D model; the workpiece surface normal, local curvature, and partial plate thickness information can be obtained from the surface geometry calculation of the 3D model. For workpieces with 2D welding drawings, plate thickness, bevel angle, bevel type, weld gap, weld type, and material grade can be read from the welding drawings or process cards. For workpieces whose paths are obtained through robot teaching, the welding head teaching trajectory is converted into an approximate spatial trajectory of the weld centerline after calibration in the robot base coordinate system, tool coordinate system, and workpiece coordinate system. For offline programming trajectories, the welding head path, welding speed, and posture sequence are read from the CAM or robot offline programming file, and the weld trajectory is obtained through the offset relationship between the tool center point and the weld centerline. For workpieces with large assembly deviations or complex surfaces, the joint boundary points are extracted through 3D scanning point cloud, and the actual weld centerline is obtained through curve fitting. At the same time, the measured normal and local gap can be obtained by locally fitting a plane or surface from the point cloud. Material parameters such as density, thermal conductivity, specific heat capacity, melting point, surface tension, coefficient of thermal expansion, and laser absorptivity are retrieved from a material database or process database.

[0026] Let the weld centerline be the arc length parameter curve. ;in, The arc length parameter along the weld centerline has a range of values. , This represents the total length of the weld centerline. The weld centerline is sampled according to a certain spatial step size to obtain the number of weld sampling points:

[0027] in, Indicates the number of weld condition nodes; Indicates the total length of the weld centerline; This indicates the distance between the weld seam and the walkway; This indicates rounding up to the nearest integer. The arc length position of the original sampling point of each weld is The corresponding spatial coordinates are:

[0028] in, Indicates the first The three-dimensional coordinates of the original sampling points of the weld in the workpiece coordinate system; , and These represent the three coordinate components of the sampling point in the workpiece coordinate system; This indicates transpose.

[0029] No. ' Unit tangent vector at each original sampling point of the weld It can be obtained by the difference between adjacent sampling points:

[0030] in, Indicates the first ' The unit tangent vector at each original sampling point of the weld. For the endpoints and The unit tangent vector can be obtained by using forward difference and backward difference respectively.

[0031] weld normal vector The source is determined based on the data conditions: when a three-dimensional model exists, Obtained from the unit normal of the workpiece surface where the original sampling point of the weld is located; when a three-dimensional scan point cloud exists, The normal vector of the plane obtained from the least-squares fitting of the point cloud of the original sampling points is used; when only welding drawings and teaching trajectories are available, It is obtained from the joint section normal or welding tool posture transformation in the welding procedure document. The local transverse quantity is defined as:

[0032] in, Indicates the first Local transverse measurement at the original sampling point of each weld; This represents the cross product of vectors. (From...) , and It can form the first The local coordinate system at the original sampling point of each weld is used for subsequent gravity projection, laser incident angle and attitude compensation calculations.

[0033] Local curvature It can be estimated from the changes in adjacent tangent vectors:

[0034] in, Indicates the first Scalar curvature at the original sampling point of each weld. Plate thickness. , bevel angle weld gap The material grade is obtained as follows: if the corresponding parameters are contained in the 3D model or drawing, they are read directly; if the 3D scanned point cloud can reflect the assembly boundary, the plate thickness and gap are calculated from the distance between the two boundary surfaces; if they cannot be obtained directly from the point cloud, they are given by the welding process document or material process card. Therefore, the output of S1 is a set of 3D weld trajectory data:

[0035] in, This represents the set of 3D trajectory data for the weld. This set is entered into S2 to initialize the weld state node.

[0036] S2: Weld discretization and weld state node initialization.

[0037] This embodiment is based on the weld seam three-dimensional trajectory data set output by S1. Generate weld state nodes. Sample the weld centerline according to the preset spatial step size. Each weld condition node is denoted as superscript This indicates the weld condition node type. The initial pre-weld characteristics of this node are represented as follows:

[0038] in, Indicates the first Initial pre-welding characteristics of each weld seam state node; This represents a vector of material thermophysical parameters obtained from a table based on the material grade, including at least density, thermal conductivity, specific heat capacity, melting point, and surface tension. The output of S2 is a set of weld state nodes.

[0039] in, This represents the set of weld state nodes. This set serves as both the spatial index for the S3 real-time data mapping and the source of weld state nodes in the S4 physical coupling heterogeneous diagram.

[0040] S3: Real-time welding data acquisition and node mapping.

[0041] During the welding process, this embodiment collects real-time welding data in each control cycle. The real-time welding data for each control cycle is recorded as follows: It includes images of the molten pool. Keyhole Images Infrared temperature field Current position of the welding head Workpiece posture Welding head posture Laser power Scanning speed Focus position morphology of light spots Local vacuum degree And sensor timestamps. In this embodiment... Indicates the sampling time or the control cycle number corresponding to the sampling time.

[0042] This embodiment first determines the weld state node corresponding to the current welding joint. Let the current position of the welding joint be... The current node index is defined as:

[0043] in, Indicates the first The current weld status node number corresponding to the current position of the welding head in each control cycle; Indicates the weld joint at the 1st Spatial location of each control cycle; Indicates the first Spatial coordinates of each weld condition node.

[0044] To reflect the continuous propagation of heat effects and molten pool condition along the weld path, real-time welding data is not only assigned to the current weld condition node. It is also assigned to its neighboring weld state nodes. The set of neighboring nodes is defined as follows:

[0045] in, Indicates the first The set of nodes affected by real-time welding data in a control cycle; Indicates the index of adjacent weld status nodes; Indicates the first The arc length position of each weld condition node; This indicates the radius of influence of the real-time welding data along the weld direction.

[0046] for The node's real-time data mapping weight is defined as follows:

[0047] in, Indicates the first In the first control cycle, real-time welding data is allocated to the first... Mapping weights for each weld state node; Represents the set of neighboring nodes The node indices used for summation and normalization. Mapping weights. This is used to distribute sensor data from the current control cycle to the current node and neighboring nodes, thereby forming node characteristics.

[0048] For the molten pool image, first, based on the tangent vector of the current node... and weld normal vector After identifying the region of interest, grayscale conversion, filtering, thresholding, and candidate contour extraction are performed. The grayscale conversion formula is:

[0049] in, Represents the pixel coordinates of the image; Represents pixel coordinates The grayscale value at that location; , and These represent the red, green, and blue channel values ​​of the original image at that pixel, respectively. This indicates a green channel.

[0050] After extracting the effective molten pool contour, this embodiment obtains the molten pool width. Length of molten pool molten pool area Aspect ratio of the molten pool and the offset of the center of the molten pool .

[0051] For the keyhole image, this embodiment obtains the keyhole area. Equivalent diameter of keyhole and keyhole eccentricity ,in:

[0052] in, Indicates the first The weld condition node at the [number]th weld [location] Equivalent diameter of keyhole per control cycle; Indicates the area of ​​the keyhole region; It represents pi (π).

[0053] For the infrared temperature field, extract the temperature peak. Average temperature Temperature gradient and temperature field asymmetry coefficient .

[0054] This embodiment will use the first Initial pre-welding characteristics of each weld condition node Combined with the mapped real-time features, the node at the [missing information] level is obtained. Original node characteristics for each control cycle:

[0055] in, Indicates the first The weld condition node at the [number]th weld [location] Original node characteristics of each control cycle; This represents the real-time molten pool feature vector obtained from the molten pool image; This represents the real-time keyhole feature vector obtained from the keyhole image; This represents the real-time temperature feature vector obtained from the temperature field; Represents the real-time attitude feature vector; Represents the eigenvectors of the real-time laser energy field; This represents the feature vector of the real-time vacuum environment.

[0056] To reduce the impact of splashing, reflections, image blurring, and instantaneous jumps between adjacent frames on node input, feature confidence is calculated:

[0057] in, Indicates the first The weld condition node at the [number]th weld [location] Characteristic confidence level of each control cycle; Indicates the degree of splash interference; Indicates the degree of reflectivity and interference; Indicates image blur; Indicates the instantaneous jump degree of features between adjacent frames; , , and These represent the confidence attenuation coefficients for the corresponding interference terms.

[0058] The normalized node features are:

[0059] in, Indicates the first The node at the th Node features of the input graph model for each control cycle; This represents the normalization function. Arrange the node features of all nodes row-wise to form a node feature matrix:

[0060] in, Indicates the first The node feature matrix of PI-GGAT is input for each control cycle; Indicates the first The total number of graph nodes in each control cycle. Therefore, the output of S3 is the node feature matrix. and feature confidence set Both are then used in S4 to construct a physically coupled heterogeneous graph.

[0061] S4: Construction of physically coupled heterogeneous graphs.

[0062] This embodiment is based on the weld state node set obtained in S2 and the node feature matrix obtained in S3. and feature confidence set Construct a physically coupled heterogeneous graph. The specific output of this step includes a set of nodes. Edge set Adjacency matrix Edge feature set .

[0063] Node set There are four types of nodes. The first type is weld condition nodes. The first category consists of discrete weld nodes in S2 and the molten pool, keyhole, temperature, and confidence features mapped to those nodes in S3. The second category consists of attitude nodes. The values ​​originate from the robot's end-effector pose, workpiece attitude sensor, welding head attitude feedback, and local coordinate system transformation results of the weld. Their characteristics include attitude angle, axis angle, laser incident angle, attitude change rate, and gravity projection components. The third category consists of laser energy field nodes. The feedback originates from the laser controller, motion controller, focusing mechanism, and spot shaping mechanism, and its characteristics include laser power, scanning speed, focal position, spot diameter, spot aspect ratio, and energy bias. The fourth category is material environment nodes. The data is derived from the S1 material database and a real-time vacuum sensor, and its characteristics include material thermal conductivity, density, specific heat capacity, melting point, surface tension, and local vacuum level. and vacuum degree change rate .

[0064] Adjacency Matrix It is a matrix used to represent the connection strength between nodes in a graph, not a single edge weight. Its... Line number Column elements Indicates the first Nodes in each control cycle To the node Edge weights during information propagation. When a node... With nodes When there is no physical coupling, When the two are spatially adjacent, thermally coupled, have attitude interactions, are constrained by laser energy input, or are in a vacuum environment, .

[0065] For nodes that have connections and nodes The edge feature vector is defined as:

[0066] in, Indicates the first Nodes in each control cycle To the node Edge features during information propagation; Represents a node With nodes The equivalent distance of heat conduction between them; Indicates the influence factor of attitude gravity projection; Indicates the laser energy coupling factor; Indicates the factors affecting the vacuum environment; and Representing nodes respectively and nodes The feature confidence level.

[0067] The elements of the adjacency matrix can be determined by the following formula:

[0068] in, Indicates the distance attenuation coefficient; Indicates the attitude influence coefficient; Indicates the energy coupling coefficient; This represents the influence coefficient of the vacuum environment. This formula illustrates that S4 does not yield a single computational quantity, but rather a complete graph structure, namely the set of nodes, the set of edges, the adjacency matrix, and the set of edge features.

[0069] The attitude gravity projection influence factor is determined by the gravity direction and the local coordinate system of the weld. Let the unit gravity direction vector be... , No. The unit normal vector of each weld state node is Then it is acceptable:

[0070] in, Represents a node To the node The influence factor of the gravity projection of the edge's attitude during information propagation; This represents the vector dot product. This factor reflects the intensity of the effect of gravity on molten pool collapse, flow deviation, and spreading under different welding postures. (S4 output) , and Enter S5 to calculate the physical equilibrium residuals.

[0071] S5: Calculation of physical equilibrium residuals.

[0072] Reference Figure 2 As shown, the physical equilibrium residual calculation step occurs after the construction of the physically coupled heterogeneous graph and before the propagation prediction of the Physically Informed Gated Graph Attention Network (PI-GGAT). Step S4 provides the node feature matrix. Adjacency matrix Sum of edge features Step S5 uses these inputs to calculate the node physical equilibrium residuals. Physical balance residuals between nodes .in, Describe whether the state of a single node satisfies the physical constraints for stable vacuum laser welding; Description node To the node Whether the information propagation relationship is reliable. The output of S5 is not directly used as a control variable, but is fed into S6 to modulate the attention weights and gating propagation coefficients.

[0073] No. The node at the th The physical balance residual for each control cycle is defined as:

[0074] in, Indicates the first The node at the th Physical balance residuals for each control cycle; Represents the energy balance residual; Indicates the residual thermal conductivity; Indicates the molten pool / keyhole stability residual; This represents the attitude gravity projection residual; Indicates the residual in a vacuum environment; , , , and These represent the weighting coefficients of the corresponding residuals.

[0075] The energy balance residual can be expressed as:

[0076] in, Indicates laser absorption efficiency; Indicates the first Laser power per control cycle; Indicates the duration of the control cycle; Indicates the first Each node corresponds to a region with heat conduction output energy; Indicates radiative heat dissipation energy; This indicates the energy required for the material to melt; This indicates a positive number that prevents the denominator from being zero.

[0077] The thermal conductivity residual can be expressed as:

[0078] in, This represents the temperature gradient measured by the sensor or estimated by the model. This represents the reference temperature gradient estimated from the material's thermal conductivity, thermal diffusivity, and welding heat input.

[0079] The molten pool / keyhole stability residual can be expressed as:

[0080] in, This indicates the number of molten pool / keyhole features involved in the stability evaluation; Indicates the feature number; Indicates the first The node at the th The first control cycle Individual molten pool / keyhole features; This represents the target value of the feature; This indicates the allowable fluctuation range of the feature; This represents the weighting coefficient of the feature.

[0081] The attitude gravity projection residual and the vacuum environment residual are expressed as follows:

[0082]

[0083] in, Indicates the first The node at the th The gravity projection effect value for each control cycle; Indicates the first Each node corresponds to a gravity projection reference value under stable welding conditions; Indicates the first Local vacuum level for each control cycle; Indicates the first Each node corresponds to the target vacuum level under the operating conditions; Indicates the rate of change of local vacuum level; This represents the weight of the rate of change of vacuum degree.

[0084] The physical equilibrium residual between nodes is defined as:

[0085] in, Represents a node To the node In the The physical balance residuals between nodes corresponding to the information propagation relationship of each control cycle; and Representing nodes respectively With nodes The physical equilibrium residual; Let S represent the adjacency matrix elements obtained in S4. This formula indicates that if there is strong physical coupling between two nodes and their states deviate from the stability condition, then the propagation relationship needs to be more strongly suppressed in S6.

[0086] S6: Attention Prediction for Physically Informed Gated Graphs Reference Figure 2 As shown, the input for prediction by the Physically Informed Gated Graph Attention Network (PI-GGAT) comes from the preceding step: the node feature matrix. From S3; adjacency matrix Sum of edge features From S4; Physical balance residuals between nodes From S5. The above inputs respectively enter Figure 2 The system consists of a physical informed input layer, a residual modulation attention computation unit, and a residual suppression gating unit.

[0087] First, in this embodiment, node features are mapped to hidden features at layer 0:

[0088] in, Indicates the first The node at the th The 0th hidden feature before entering PI-GGAT in each control cycle; Indicates the first The input features of each node; Represents a node The node type; Indicates node type The corresponding initial mapping matrix; This represents the corresponding bias vector; This represents a non-linear activation function.

[0089] For nodes connected by edges and nodes Calculate node feature-related scores:

[0090] in, Represents a node To the node In the In the first control cycle Layer node feature-related scoring; Represents the attention parameter vector; Represents the hidden feature transformation matrix of a node; Represents the edge feature transformation matrix; and They represent the first Layer nodes and nodes Hidden features; This represents the edge features obtained from S4.

[0091] Attention weights are determined jointly by node feature correlation scores and physical balance residuals between nodes:

[0092] in, Indicates the first Layer nodes To the node Attention weights during information dissemination; This represents the residual penalty coefficient; Represents a node The set of adjacent nodes; Represents a node The summation index in the set of adjacent nodes. This formula shows that the larger the physical balance residual between nodes, the more likely the node... To the node The lower the attention weight.

[0093] The gating propagation coefficient is calculated as follows:

[0094] in, Indicates the first Layer nodes To the node Gating propagation coefficient during information transmission; Represents the gate weight matrix; Represents the gate bias vector; This represents the residual suppression coefficient. This formula is related to the one above. , The meanings are consistent; both indicate the first. The nodes in the layers have hidden features, the only difference being that one is used for attention weight calculation and the other is used for gating propagation coefficient calculation.

[0095] The node's hidden features are updated as follows:

[0096] in, Indicates the first The node at the th In the first control cycle Hidden features of layers; Indicates the first The feature transformation matrix of the layer. This formula clearly shows that the propagation of node information in S6 is simultaneously affected by the attention weights. and gate propagation coefficient Both are jointly controlled by the S5 output and are subject to the inter-node physical balance residual. constraint.

[0097] The node hidden features updated after the last propagation layer are input into a fully connected regression layer, and the predicted welding state value for the next time step is obtained through linear transformation and mapping with a nonlinear activation function. This embodiment has been implemented... After layer propagation, output the predicted welding state value for the next moment:

[0098] in, Indicates the first Predicted welding status for each control cycle; Indicates the current weld status node On the last floor Hidden features; This represents the output layer weight matrix; This represents the output layer bias vector. The predicted weld condition values ​​include at least the predicted weld pool width, keyhole stability, temperature gradient, weld formation deviation, and predicted physical equilibrium residuals.

[0099] S7: Scene-specific control output and closed-loop execution This embodiment is based on the predicted welding state value at the next moment. and target welding state Generate control adjustment quantities. The control variable vector is defined as follows:

[0100] in, Indicates the first The current control quantity for each control cycle; Indicates laser power; Indicates the scanning speed; Indicates the focal position; Indicates the morphological parameters of the light spot; Indicates the attitude compensation amount; Indicates the degree of local vacuum.

[0101] The control adjustment amount can be obtained through constraint optimization:

[0102] And satisfy:

[0103] in, Indicates the first The control adjustment amount generated in each control cycle; This represents the candidate control adjustment amount during the constraint optimization process; Indicates the target welding state; This represents the predicted welding status at the next moment, as output by PI-GGAT. This represents the quality deviation weight matrix; This represents the control smoothing weight matrix; This represents the penalty weight for the predicted physical equilibrium residual; This represents the predicted physical equilibrium residual; and These represent the lower limit and upper limit of the control quantity, respectively.

[0104] Reference Figure 3 As shown, when the workpiece to be welded is an easily adjustable workpiece, according to Figure 3 The left-side process outputs adjustments for workpiece tilt angle, workpiece flip angle, laser incident angle, focus offset, laser power, and scanning speed, which are then distributed to the workpiece attitude adjustment mechanism, welding head attitude adjustment mechanism, focusing mechanism, laser, and motion platform, respectively. When the workpiece to be welded is a large workpiece with difficult attitude adjustment, according to... Figure 3 The right-side process outputs adjustments for local vacuum level, laser power, scanning speed, focal position, beam aspect ratio, and energy bias, which are then sent to the vacuum pump / vacuum valve, laser, motion platform, focusing mechanism, and beam shaping mechanism.

[0105] To reduce actuator hysteresis and prediction errors, this embodiment can also use incremental PID for engineering correction based on the constraint optimization results:

[0106] in, Indicates the first The final control command is issued to the actuator in each control cycle; This indicates the execution control command of the previous control cycle; , and These represent the proportional, integral, and differential gain matrices, respectively. Indicates the first Feedback error per control cycle; This indicates the current welding status as observed by the sensor.

[0107] This embodiment relates to intelligent control of vacuum laser welding, welding physics-aware deep learning, and all-attitude welding process optimization technology. It clarifies the three-dimensional trajectory and spatial nodes of the weld seam before welding, and collects and maps real-time data on the molten pool, keyhole, temperature, attitude, energy field, and vacuum environment during the welding process. It then uses a physical equilibrium residual constraint information propagation mechanism for state prediction and control quantity generation to address the welding stability control problem under the coupled influence of attitude changes, energy field shifts, keyhole fluctuations, and the vacuum environment. This solves the problems in existing vacuum laser welding control methods, such as the lack of node-level binding between the weld seam trajectory and real-time state, difficulty in quantifying the impact of attitude changes, easy propagation of abnormal information in the molten pool / keyhole, difficulty in incorporating vacuum environment constraints into the model inference process, and difficulty in compensating for large workpieces through overall attitude adjustment.

[0108] Example 2 This invention discloses a vacuum laser welding all-attitude adaptive control system, referring to... Figure 4 As shown, it includes the following modules: Pre-welding data acquisition module: used to acquire basic data of the workpiece to be welded and the welding robot, and extract the three-dimensional trajectory data of the weld; the three-dimensional trajectory data of the weld includes the spatial position data of the weld and the local geometric state data; Weld Node Generation Module: This module is used to discretize the three-dimensional trajectory data of the weld according to a preset spatial step size, and generate multiple weld status nodes. Each weld status node is bound to the corresponding weld spatial location data and local geometric state data. Feature extraction and confidence calibration module: used to determine the current weld state node based on the current position of the welding head during vacuum laser welding; to collect real-time welding data and map the real-time welding data to the current weld state node and its neighboring weld state nodes to form a node feature matrix and feature confidence; Physical Coupling Module: Used to physically couple weld state nodes, attitude nodes, laser energy field nodes, and material environment nodes, and generate adjacency matrices and edge features based on weld spatial adjacency relationships, thermal conduction coupling relationships, attitude gravity projection relationships, laser energy input relationships, and vacuum environment constraint relationships; Physical balance residual calculation module: used to calculate physical balance residuals and inter-node physical balance residuals based on node feature matrix, adjacency matrix, edge features and multiple physical constraints; The Physically Informed Gated Graph Attention Network Prediction Module is used to calculate attention weights and gating propagation coefficients in the Physically Informed Gated Graph Attention Network based on the node feature matrix, edge features, and physical balance residuals between nodes, and to update nodes based on the attention weights and gating propagation coefficients to obtain the predicted welding state value at the next moment. Scene-specific control output module: It is used to generate control adjustment quantities based on the predicted welding status value at the next moment and the preset control parameters, and adjust the laser power, scanning speed, focal position, spot shape, attitude compensation amount and local vacuum degree according to the control adjustment quantities, so as to realize the full attitude adaptive control of vacuum laser welding.

[0109] This embodiment is applied to a vacuum laser welding system for the circumferential weld seam of a large fuel tank in the aerospace manufacturing field; see reference. Figure 5 As shown, the system includes a pre-weld data acquisition module, a weld node generation module, a feature extraction and confidence calibration module, a physical coupling construction module, a physical equilibrium residual calculation module, a physical informed gating graph attention network prediction module, a scenario-specific control output module, and a real-time feedback and model update module; the data relationships between these modules are illustrated. Pre-weld basic data includes geometric, process, and assembly data required for weld 3D trajectory extraction; the material database outputs material thermophysical parameters; real-time welding data includes data on the molten pool, keyholes, temperature, attitude, energy field, and vacuum environment output during the welding process. The pre-weld data acquisition module obtains the weld centerline and its geometric properties based on the pre-weld basic data. The weld node generation module generates weld state nodes based on the weld centerline. The feature extraction and confidence calibration module transforms real-time welding data into calculable molten pool, keyhole, temperature, attitude, energy field, and environmental features. The physical coupling construction module organizes the above nodes and features into node feature matrices, adjacency matrices, and edge features. The physical equilibrium residual calculation unit uses the node feature matrices, adjacency matrices, and edge features to calculate the node physical equilibrium residuals and the physical equilibrium residuals between nodes. The Physically Informed Gated Graph Attention Network (PI-GGAT) state prediction module utilizes graph structure, node features, and physical residuals for node propagation and updates, outputting the predicted welding state for the next moment. The scenario-specific control output module generates control adjustment values ​​based on the predicted values ​​and sends them to the corresponding welding actuators.

[0110] This embodiment also includes a real-time feedback and model update module, which is responsible for collecting the actual result data after welding is performed, comparing it with the predicted value of the PI-GGAT model, calculating the deviation, and using these deviation data to fine-tune the model parameters online or update them offline, thereby eliminating the error between the model prediction and the actual physical process and improving the accuracy of subsequent predictions.

[0111] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0112] 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 adaptive control of all postures in vacuum laser welding, characterized in that, Includes the following steps: S1. Obtain basic data of the workpiece to be welded and the welding robot, and extract the three-dimensional trajectory data of the weld; the three-dimensional trajectory data of the weld includes the spatial position data of the weld and the local geometric state data. S2. Discretize the three-dimensional trajectory data of the weld according to a preset spatial step size to generate multiple weld state nodes; wherein, each weld state node is bound to the corresponding weld spatial position data and local geometric state data; S3. During vacuum laser welding, the current weld state node is determined based on the current position of the welding head; real-time welding data is collected and mapped to the current weld state node and its adjacent weld state nodes to form a node feature matrix and feature confidence. S4. Physically couple the weld state node, attitude node, laser energy field node and material environment node, and generate an adjacency matrix and edge features based on the weld spatial adjacency relationship, heat conduction coupling relationship, attitude gravity projection relationship, laser energy input relationship and vacuum environment constraint relationship. S5. Calculate the physical balance residual and the physical balance residual between nodes based on the node feature matrix, adjacency matrix, edge features and multiple physical constraints. S6. In the physically informed gating graph attention network, the attention weights and gating propagation coefficients are calculated based on the node feature matrix, edge features and physical balance residuals between nodes, and the nodes are updated based on the attention weights and gating propagation coefficients to obtain the predicted welding state value at the next moment. S7. Based on the predicted welding state value at the next moment and the preset control parameters, generate a control adjustment amount, and adjust the laser power, scanning speed, focal position, spot shape, attitude compensation amount and local vacuum degree according to the control adjustment amount to realize full attitude adaptive control of vacuum laser welding.

2. The method as described in claim 1, characterized in that, In step S1, the basic data includes the basic data of the workpiece to be welded, which comes from the three-dimensional model of the workpiece to be welded, the three-dimensional scan point cloud, the welding drawings, the welding process card and the material database, and is used to determine the weld centerline, plate thickness, groove parameters, weld gap and material thermophysical parameters. The basic data also includes the basic data of the welding robot, which comes from the robot's teaching trajectory and / or offline programming trajectory, and is used to determine the motion path of the welding head and the coordinate transformation relationship. The three-dimensional trajectory data of the weld includes: the coordinate sequence of the weld centerline, the weld tangent vector, the weld normal vector, the local transverse quantity, the local curvature, the plate thickness, the bevel parameters, the weld gap, and the material thermophysical parameters.

3. The method as described in claim 2, characterized in that, In step S3, determining the current weld state node based on the current position of the weld joint specifically includes: During the vacuum laser welding process, the first Within each control cycle, the nearest weld condition node is determined based on the current position of the weld joint; expressed by the formula: in, Indicates the first The position of the welding head in the workpiece coordinate system within each control cycle. Indicates the first Spatial coordinates of each weld condition node This indicates the sequence number of the weld status node closest to the current position of the weld joint. This represents the Euclidean norm.

4. The method as described in claim 3, characterized in that, In step S3, mapping the real-time welding data to the current weld status node and its neighboring weld status nodes specifically includes: The real-time welding data includes molten pool images, keyhole images, temperature fields, attitude data, laser energy field parameters, and local vacuum levels; the real-time welding data is then processed by feature extraction to form a real-time feature vector. According to the mapping weight Mapped to the The state node of the weld seam is obtained; The weld condition node at the [number]th weld [location] Node characteristics of each control cycle It can be expressed by the formula: in, Indicates the first The weld condition node at the [number]th weld [location] Node characteristics of each control cycle Represents the normalization function. This represents vector concatenation. Indicates the first Initial pre-welding characteristics of each weld joint state node Indicates the first Real-time feature vector within each control cycle Mapped to the Mapping weights for each weld state node, This indicates the sequence number of the weld status node receiving the real-time feature map. Represents the set of neighboring nodes The index of any node used for normalized summation, Indicates the current weld state node The set of neighboring nodes centered on the center This indicates the radius affected by real-time feature mapping. Indicates the first Spatial coordinates of each weld condition node This represents the spatial coordinates of the weld state node that is closest to the current position of the weld joint. Represents the spatial coordinates of any neighboring node.

5. The method as described in claim 4, characterized in that, In step S3, the feature confidence level is determined based on splash interference, reflection interference, image blur, and instantaneous feature jump degree between adjacent frames, and is expressed by the formula: in, Indicates the first The weld condition node at the [number]th weld [location] Characteristic confidence level of each control cycle; Indicates the degree of splash interference; Indicates the degree of reflectivity and interference; Indicates image blur; Indicates the instantaneous jump degree of features between adjacent frames; , , and These represent the confidence attenuation coefficients for splash interference, reflection interference, image blur, and instantaneous jump in features between adjacent frames, respectively.

6. The method as described in claim 5, characterized in that, In step S4, the adjacency matrix is ​​composed of edge weights. The matrix formed by the edge weights Indicates the first Nodes within each control cycle To the node The physical coupling strength during information propagation; the edge weights Based on the spatial adjacency relationship of the weld, the thermal conduction coupling relationship, the attitude and gravity projection relationship, the laser energy input relationship, and the vacuum environment constraint relationship, the following conditions are met: in, Represents a node With nodes An indicator of whether there is a physical coupling edge between them; and Representing nodes respectively With nodes In the Characteristic confidence level of each control cycle; Represents a node With nodes The equivalent distance of heat conduction between them; Indicates the influence factor of attitude gravity projection; Indicates the laser energy coupling factor; Indicates the factors affecting the vacuum environment; , , and These represent the distance attenuation coefficient, attitude influence coefficient, energy coupling coefficient, and vacuum environment influence coefficient, respectively; the vacuum environment influence factor is based on the local vacuum level. Target vacuum degree and vacuum degree change rate Calculated; The edge features are expressed by the following formula: in, Representing edge features, This indicates transpose.

7. The method as described in claim 6, characterized in that, In step S5, the multiple physical constraints include energy balance constraints, heat conduction constraints, molten pool / keyhole stability constraints, attitude gravity projection constraints, and vacuum environment constraints. The physical balance residuals include energy balance residuals, heat conduction residuals, molten pool / keyhole stability residuals, attitude gravity projection residuals, and vacuum environment residuals, satisfying: in, Indicates the first Each node at time... The physical equilibrium residual; Indicates the first Each node at time... The energy balance residual; Indicates the first Each node at time... Thermal conductivity residual; Indicates the first Each node at time... Molten pool / keyhole stability residual; Indicates the first Each node at time... The attitude gravity projection residual; Indicates the first Each node at time... Vacuum environment residuals; , , , and These represent the weighting coefficients of the corresponding residuals; the vacuum environment residuals specifically refer to the degree to which the current vacuum environment influencing factors deviate from the stable target range. The physical balance residual between nodes is based on the nodes. Physical equilibrium residuals, nodes Physical equilibrium residuals and edge weights Determined, satisfies: in, Indicates time node To the node The physical balance residuals between nodes corresponding to the information propagation process. and Representing nodes respectively With nodes The physical equilibrium residual.

8. The method as described in claim 7, characterized in that, Step S6 specifically includes: S61. In any propagation layer of the physically informed gating graph attention network, the node features in the node feature matrix are mapped to hidden node features, and the node feature-related scores are determined by combining the edge features. In the In the layer propagation layer, the node feature-related scores satisfy: in, Indicates the first Layer nodes To the node Node feature-related scores during information dissemination; Represents the attention parameter vector; Represents the hidden feature transformation matrix of a node; Represents the edge feature transformation matrix; and They represent the first Layer nodes With nodes In the Hidden features of nodes in each control cycle; Represents a node To the node Edge features during information propagation; S62. Based on the node feature-related scores and the physical balance residuals between nodes, calculate the attention weight, expressed by the formula: Based on the node feature-related scores, edge features, and node hiding features, the gated propagation coefficient is calculated and expressed by the formula: in, Indicates attention weight; This represents the residual penalty coefficient; Represents a node The set of adjacent nodes, Indicates the index of any node in the set of adjacent nodes; Indicates the first Layer propagation nodes To the node Gating propagation coefficient during information transmission; Represents a nonlinear activation function; Represents the gate weight matrix; Represents the gate bias vector; This represents the residual suppression coefficient; S63. Update nodes based on the attention weights and gating propagation coefficients. The hidden features of a node in the next layer are expressed by the following formula: in, Indicates the first The node at the th In the first control cycle Hidden features of layers; Indicates the first Layer feature transformation matrix; S64. Input the node hidden features updated by the last propagation layer into the fully connected regression layer, and obtain the predicted welding state value at the next moment through linear transformation and nonlinear activation function mapping.

9. The method as described in claim 8, characterized in that, In step S7, the control adjustment amount is determined based on the predicted welding state value at the next moment, the preset welding quality target, the current control amount, the quality deviation weight matrix, the control smoothing weight matrix, the predicted physical balance residual, the physical balance residual penalty weight, the lower limit of the control amount, and the upper limit of the control amount, satisfying the following: And satisfy the constraints: in, Indicates the first Control adjustment amount per control cycle; This represents the candidate control adjustment amount during the constraint optimization process; Indicates the current control quantity; This indicates the target welding state at the next moment corresponding to the preset welding quality target; This indicates the predicted welding status at the next moment; This represents the quality deviation weight matrix; This represents the control smoothing weight matrix; This represents the predicted physical equilibrium residual; Indicates the penalty weight for physical equilibrium residuals; Indicates the lower limit of the control quantity; This indicates the upper limit of the control quantity.

10. A vacuum laser welding all-attitude adaptive control system, characterized in that, The method described in any one of claims 1-9 includes the following modules: Pre-welding data acquisition module: used to acquire basic data of the workpiece to be welded and the welding robot, and extract three-dimensional trajectory data of the weld; the three-dimensional trajectory data of the weld includes spatial position data of the weld and local geometric state data; Weld joint node generation module: used to discretize the three-dimensional trajectory data of the weld joint according to a preset spatial step size to generate multiple weld joint state nodes; wherein, each weld joint state node is bound to the corresponding weld joint spatial position data and local geometric state data; Feature extraction and confidence calibration module: used to determine the current weld state node based on the current position of the welding head during vacuum laser welding; collect real-time welding data and map the real-time welding data to the current weld state node and its neighboring weld state nodes to form a node feature matrix and feature confidence; Physical coupling construction module: used to physically couple the weld state node, attitude node, laser energy field node and material environment node, and generate adjacency matrix and edge features based on weld spatial adjacency relationship, heat conduction coupling relationship, attitude gravity projection relationship, laser energy input relationship and vacuum environment constraint relationship; Physical balance residual calculation module: used to calculate physical balance residual and inter-node physical balance residual based on the node feature matrix, adjacency matrix, edge features and multiple physical constraints; Physically informed gating graph attention network prediction module: In the physically informed gating graph attention network, it calculates attention weights and gating propagation coefficients based on the node feature matrix, edge features and physical balance residuals between nodes, and updates nodes based on the attention weights and gating propagation coefficients to obtain the predicted welding state value at the next moment. Scene-specific control output module: It is used to generate control adjustment amount based on the predicted value of welding state at the next moment and preset control parameters, and adjust the laser power, scanning speed, focal position, spot shape, attitude compensation amount and local vacuum degree according to the control adjustment amount, so as to realize the full attitude adaptive control of vacuum laser welding.

Images