A control method for a metal cutting and welding apparatus

Abstract

The application discloses a control method of a metal cutting and welding device, and relates to the technical field of metal processing control. The method obtains initial parameters of a graph neural network proxy model which is pre-trained on multiple historical working conditions in a fusion electric arc energy conservation relationship and a heat conduction partial differential equation physical constraint and a neuron learning mode, collects current working condition small sample data and maps the current working condition small sample data into target domain graph data, performs gradient descent fine-tuning on the model, inputs real-time voltage waveforms, current waveforms and molten pool surface visual features into the fine-tuned model, obtains molten pool internal temperature field and flow field distribution, and further inversely obtains molten depth, penetration state and keyhole shape parameters. The inversion state is taken as a current quantity, an optimal control sequence is solved under physical safety constraints through model predictive control, and a first-step control increment is sent to a device actuator to form a real-time closed loop.

Description

Technical Field

[0001] This invention relates to the field of metal processing control technology, specifically a control method for metal cutting and welding equipment. Background Technology

[0002] Metal cutting and welding are widely used processing technologies in modern manufacturing, involving various equipment forms such as plasma arc cutting, laser cutting, and digital arc welding. In actual production applications, the cutting and welding process is a nonlinear dynamic process with strong coupling of multiple physical fields, including arc physics, fluid heat transfer, metallurgical reaction and mechanical deformation, and accompanied by strong noise. The internal key states of the molten pool, such as the melting depth of the molten pool along the thickness direction of the plate, whether the base material is completely melted through, and the vapor channel morphology formed by the arc or laser under deep melting conditions, have black box characteristics that are difficult to observe directly and quantitatively characterize through conventional means.

[0003] Currently, the control of cutting and welding equipment in industrial settings still largely relies on fixed experience-based process parameter tables developed offline, or on proportional-integral-derivative feedback control methods based on a single variable. This makes it difficult to effectively address issues such as uneven penetration or even internal defects caused by material batch fluctuations, heat accumulation drift, and assembly gap errors in actual processing.

[0004] One of the existing technical solutions that is similar to this invention is a welding and cutting quality control system based on surface visual feature extraction and single-variable proportional-integral-derivative closed-loop feedback. This solution first uses an industrial camera set on the coaxial or off-axis to acquire surface images of the molten pool or the area of ​​the cutting slag in real time. Then, the geometric features such as the width and length of the molten pool surface are extracted as feedback observations through image processing algorithms. Next, the extracted surface features are compared with preset target values ​​to obtain a deviation signal. Then, the proportional-integral-derivative controller calculates the control correction amount according to the deviation signal and a predetermined calculation rule. Finally, the correction amount is sent to the wire feeding mechanism or power module to adjust the wire feeding speed or output current, thereby realizing closed-loop adjustment of the molten pool surface forming features.

[0005] Another relatively close existing technical solution is the deep learning-based control model solution explored by the academic community. This solution is designed for processing scenarios with specific equipment, specific materials and specific plate thicknesses. It collects a large amount of historical processing data and corresponding surface features and cross-sectional annotation results, trains convolutional neural networks or other deep learning models, and directly establishes the mapping relationship between measurable features of the molten pool surface and optimal process parameters. During online inference, the real-time surface features are input into the model to obtain the corresponding control parameter output.

[0006] The aforementioned existing technical solutions have the following three defects;

[0007] First, both the aforementioned proportional-integral-derivative feedback scheme and deep learning control scheme rely solely on measurable features of the molten pool surface as feedback observations, without observing and modeling key states such as molten depth, penetration, and keyholes inside the molten pool. Therefore, their feedback quantities cannot truly represent the core indicators of processing quality, resulting in control results that often only guarantee good surface forming but cannot guarantee internal penetration quality, easily leading to problems where the appearance is acceptable but the internal defects exist.

[0008] Secondly, the training of the aforementioned deep learning control model heavily relies on massive amounts of labeled data collected under specific working conditions, and it does not integrate arc physics and heat transfer mechanisms to form transferable general knowledge during the modeling process. Therefore, the model has low generalization ability. Once the processing equipment is changed, the material type is changed, or the plate thickness is adjusted, the previously trained model becomes invalid, and data must be collected again and retrained. This results in insufficient production line flexibility, high deployment costs, and long response cycles, forming a bottleneck in implementation that is highly dependent on specific equipment and specific working conditions.

[0009] Furthermore, existing digital twin system architectures are mostly deployed in the cloud and run in offline batch processing mode. They do not establish a real-time bidirectional communication interface with the underlying control bus of the equipment. Therefore, their prediction results cannot be converted into equipment execution instructions in real time during the processing. They can only play the role of post-processing visualization monitoring and cannot dynamically correct abnormal trends such as uneven melting penetration during the processing. Summary of the Invention

[0010] The technical problem to be solved by the present invention is that the existing control methods for metal cutting and welding equipment have insufficient generalization ability when dealing with various processing conditions such as material batch fluctuations, plate thickness changes, and equipment model replacements. The control model is highly dependent on the massive amount of labeled data re-collected under specific working conditions to complete retraining and deployment.

[0011] Meanwhile, key quality indicators such as the depth of the melt pool, the penetration of the melt, and the shape of the keyhole during the processing are in a black box state that is invisible and difficult to quantitatively characterize. Existing solutions that rely on high-cost online detection methods are difficult to deploy on a large scale.

[0012] In addition, most existing digital twin systems rely on offline visual monitoring and fail to establish a real-time two-way command interaction channel with the underlying control bus of the equipment. This makes it impossible to form an effective real-time closed-loop control, resulting in quality problems such as uneven melting and defects that occur during the processing not being detected and corrected in a timely manner.

[0013] To address the aforementioned technical problems, this invention provides a control method for metal cutting and welding equipment, comprising: acquiring pre-training initial parameters of a graph neural network surrogate model, wherein the pre-training initial parameters are obtained by training historical processing condition data of various materials, plate thicknesses, and equipment models using a meta-learning approach, under the condition of integrating the physical constraints of the arc energy conservation relationship and the partial differential equation of heat conduction;

[0014] The voltage waveform, current waveform, and visual features of the molten pool surface under the current processing conditions are collected to form a small sample dataset of the target processing conditions, and the small sample dataset of the target processing conditions is mapped to target domain graph data;

[0015] Starting with the pre-trained initial parameters, the graph neural network proxy model is fine-tuned by gradient descent using the target domain graph data to obtain a graph neural network proxy model adapted to the current processing conditions.

[0016] The real-time collected voltage waveforms, current waveforms, and visual features of the molten pool surface are input into the fine-tuned graph neural network surrogate model to obtain the temperature field distribution and flow field distribution inside the molten pool.

[0017] Based on the temperature field distribution and the flow field distribution, the melting depth, melting penetration state and keyhole morphology parameters of the molten pool are obtained by inversion.

[0018] Using the inverted melt depth, melt penetration state, and keyhole morphology parameters as the current state variables, and the finely tuned graph neural network surrogate model as the prediction model, the melt pool state evolution trajectory is predicted in a rolling manner for multiple candidate control input sequences within a preset prediction time domain. The optimal control sequence is obtained by solving the optimization problem.

[0019] The control increment of the first step in the optimal control sequence is sent to the motion controller and power controller of the metal cutting and welding equipment to adjust the corresponding parameters in the wire feeding speed, cutting speed, welding speed and current waveform parameters.

[0020] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0021] 1. This invention combines offline meta-learning pre-training with online few-sample domain adaptive fine-tuning. The graph neural network surrogate model can obtain general process feature extraction capabilities under various historical processing conditions based on the integration of electric arc physics and heat transfer mechanisms. Therefore, after changing equipment, materials, or plate thickness, only a very small amount of new working condition data is needed to complete parameter fine-tuning. Compared with the existing technology that requires the re-collection of massive labeled data and complete model training for specific working conditions, this invention reduces the number of target working condition samples to be collected in the online stage and the number of iteration steps required for gradient descent fine-tuning, thus overcoming the problem of the high dependence of the control model on massive labeled data for specific working conditions in the existing technology.

[0022] 2. This invention uses a graph neural network surrogate model that integrates physical mechanisms and data-driven approaches to predict the distribution of multiple physical fields inside the molten pool. Based on this, it performs online quantitative inversion and reconstruction of key indicators such as melt depth, melt penetration state, and keyhole morphology. This transforms the previously invisible and difficult-to-quantify black box state inside the molten pool into observable and quantifiable feedback information, enabling real-time perception of core processing quality indicators without relying on high-cost online detection equipment.

[0023] 3. This invention combines the prediction results of the graph neural network surrogate model with the model predictive control algorithm, and superimposes deterministic physical boundary constraints characterized by the equipment safety threshold range. Under the premise of meeting industrial-grade safety and reliability requirements, it realizes integrated real-time closed-loop control from the prediction of the internal state of the molten pool to the correction of the action of the underlying actuator of the equipment. It can promptly and dynamically correct abnormal trends of uneven melting penetration that occur during the processing. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the overall system architecture in an embodiment of the present invention.

[0025] Figure 2 This is a flowchart of the offline meta-learning pre-training process in an embodiment of the present invention;

[0026] Figure 3 This is a flowchart of the online adaptive + real-time closed-loop control in an embodiment of the present invention. Detailed Implementation

[0027] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0028] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0029] The control method for metal cutting and welding equipment provided in this embodiment is applicable to various types of metal cutting and welding mainframes such as plasma arc cutting machines, laser cutting machines, and digital arc welding machines. It can achieve rapid generalized control of processing quality and real-time quantitative perception and closed-loop correction of key states inside the molten pool when material batches, plate thickness specifications, or equipment models change.

[0030] I. System Overall Architecture

[0031] The control system in this embodiment includes, from bottom to top, a device execution layer, a perception layer, an edge computing layer, and an offline training layer.

[0032] The equipment execution layer includes a metal cutting and welding host, a wire feeding mechanism, a traveling mechanism, and a power module. The various execution components are physically connected to each other to form the execution body that completes the cutting or welding action.

[0033] The sensing layer includes a high-frequency voltage and current acquisition card and an industrial coaxial vision sensor. The high-frequency voltage and current acquisition card is connected to the power module via a signal line to acquire voltage and current waveforms in real time during the processing.

[0034] The industrial coaxial vision sensor is installed in the molten pool processing area to acquire real-time visual images of the molten pool surface.

[0035] The edge computing layer is implemented by an industrial control computer equipped with an artificial intelligence acceleration chip. It is internally divided into four functional sub-modules: a graph neural network surrogate model inference module, an online domain adaptive fine-tuning module, a molten pool state inversion module, and a model prediction control module. The intermediate calculation results are transmitted sequentially between the sub-modules through data flow, and together they complete all online operations from the input of sensing signals to the output of control commands.

[0036] The offline training layer is deployed in the cloud or on a high-performance computing platform. It includes a multi-condition historical database and a meta-learning pre-training engine, which is used to perform offline pre-training of the graph neural network agent model and distribute the initial parameters after training to the graph neural network agent model inference module of the edge computing layer.

[0037] The edge computing layer and the device execution layer are connected via an industrial control bus to realize a closed-loop path for real-time feedback of device status and issuance of control commands. This enables the control method in this embodiment to complete a complete closed loop of state perception, state inversion and control decision-making within a millisecond time scale.

[0038] II. Meta-learning Pre-training Method for Offline Cross-Condition Neural Network Agent Model

[0039] In this embodiment, a graph neural network proxy model integrating the physical mechanism of electric arc and the heat transfer mechanism is first constructed in the offline stage. Then, the model is pre-trained on historical working condition data of various materials, plate thicknesses and equipment combinations using meta-learning method, so that the model can obtain the ability to extract general process features across working conditions. Specifically, the following processing steps are included.

[0040] First, for various historical processing conditions, namely combinations of different materials, different plate thicknesses and different equipment models, corresponding electric arc physical simulation data and historical processing sensor data are collected. Discrete sampling points on the processing path are used as graph nodes, and the heat transfer and diffusion relationships and electric arc effects between graph nodes are used as graph edges. Based on this, a graph structure dataset is constructed.

[0041] Secondly, the physical mechanisms, such as the arc energy conservation equation and the heat conduction partial differential equation, are transformed into differentiable physical residual terms. These physical residual terms, together with the data-driven prediction loss of the graph neural network surrogate model, constitute a hybrid loss function, achieving the fusion modeling of physical mechanisms and data-driven approaches. The hybrid loss function can be expressed as:

[0042] ;

[0043] ;

[0044] ;

[0045] ;

[0046] in, Indicates the mixed training loss;

[0047] This represents the data-driven prediction loss term;

[0048] Represents the physical residual loss term;

[0049] This is the weighting coefficient between data-driven loss and physical residual loss;

[0050] This represents the number of training samples;

[0051] For the first The true labeled values ​​of each sample;

[0052] For the graph neural network surrogate model for the first Predicted values ​​for each sample;

[0053] Represents the residual operator of the partial differential equation of heat conduction;

[0054] For the first The predicted temperature field corresponding to each sample;

[0055] The density of the material;

[0056] Specific heat capacity of the material;

[0057] For time;

[0058] The thermal conductivity of the material;

[0059] For gradient operators;

[0060] This is the electric arc heat source item.

[0061] Next, each historical processing condition is divided into several sets of meta-training tasks. In each training iteration, several historical processing conditions are randomly sampled to form a meta-task. The data under each meta-task is further divided into a support set and a query set. For each meta-task, the current parameters of the graph neural network surrogate model are virtually updated with a small number of gradient descent steps using its support set data to obtain the task-specific parameters corresponding to the meta-task.

[0062] Then, the prediction loss of the corresponding task-specific parameters is evaluated using the query set data of each meta-task, and the gradient of this loss is calculated with respect to the original parameters of the model. Based on this, the meta-parameters of the model are updated, so that the initial parameters of the model have the ability to converge quickly to new working conditions. The above meta-parameter update process can be expressed as:

[0063] ;

[0064] ;

[0065] in, These are the meta-parameters of the graph neural network surrogate model, i.e., the initial parameters;

[0066] In order to target the Task-specific parameters obtained after virtual updates of each meta-task using the support set;

[0067] The learning rate for the inner loop;

[0068] For the first Individual tasks are parameterized on the support set. The calculated loss;

[0069] The learning rate for the outer loop;

[0070] For the first Each metatask is defined on the query set using task-specific parameters. The calculated loss.

[0071] Repeat the above-described cyclic process of meta-task sampling, task-specific parameter updating, and meta-parameter updating until the meta-parameters converge. Output the initial parameters of the graph neural network surrogate model, which has the ability to extract general process features across materials, plate thicknesses, and equipment, as the pre-trained model parameters for the online deployment phase.

[0072] III. Online Small Sample Domain Adaptive Migration and Real-Time Closed-Loop Control Method

[0073] After completing offline pre-training, this embodiment enters the online deployment stage, which can achieve rapid model adaptation with a very small amount of new working condition data. Specifically, it includes the following processing steps.

[0074] First, under new materials, plate thicknesses, or equipment operating conditions, voltage waveforms, current waveforms, and images of the molten pool surface morphology are acquired in real time using a high-frequency acquisition card and a coaxial vision sensor, forming a small sample target domain dataset. The number of samples is much smaller than the amount of historical training data required in the offline stage.

[0075] Secondly, the collected small sample electrical signal features and visual features are mapped to the graph node attribute space consistent with the offline pre-trained model to construct the target domain graph data.

[0076] Next, the graph neural network surrogate model meta-parameters obtained in the offline phase are loaded as the initial starting point for online fine-tuning. Using small sample data from the target domain, the model parameters are fine-tuned by gradient descent a finite number of times according to the inner loop update rules learned in the pre-training phase, to obtain graph neural network surrogate model parameters adapted to the current working conditions.

[0077] Then, the predicted output of the fine-tuned model is compared with the surface features of the molten pool measured by the visual sensor. If the comparison deviation exceeds the preset threshold, additional small sample collection is triggered and the aforementioned gradient descent fine-tuning steps are repeated.

[0078] If the comparison deviation does not exceed the preset threshold, the decision domain adaptive process is completed, and the system enters the online inference and closed-loop control stage.

[0079] IV. Method for Inverting Critical States Inside the Molten Pool

[0080] After completing the domain adaptation, this embodiment uses a graph neural network surrogate model adapted to the current working conditions to quantitatively invert the invisible state inside the molten pool, specifically including the following processing steps.

[0081] First, the high-frequency voltage waveform, current waveform, and visual features of the molten pool surface, acquired in real time, are used as the observation inputs.

[0082] Secondly, the observation is input into a graph neural network surrogate model that has completed domain adaptation, and forward inference is used to obtain the predicted results of the temperature field distribution and flow field distribution inside the molten pool.

[0083] Next, based on the predicted temperature field distribution and according to the preset isothermal surface determination criteria, the depth of the material melting point isothermal surface along the thickness direction is calculated. This depth is used as the melting depth inversion value, and based on the comparison between the melting depth inversion value and the thickness of the processed plate, it is determined whether the complete melting penetration state has been reached.

[0084] Then, by combining the predicted flow field velocity vector distribution and the preset gas-liquid interface determination conditions, the keyhole morphology characteristic parameters are inverted and reconstructed, specifically including the keyhole depth and keyhole opening diameter.

[0085] Finally, the quantified melt depth inversion value, melt penetration state determination result, and keyhole morphology parameters are output as feedback inputs for subsequent model predictive control, transforming the key indicators inside the molten pool, which were originally in a black box state, into quantitative information that can be directly used by closed-loop control.

[0086] V. Real-time Closed-Loop Parameter Correction Method Based on Model Predictive Control

[0087] After obtaining the inversion state inside the molten pool, this embodiment further utilizes a model predictive control algorithm to achieve millisecond-level real-time closed-loop parameter correction, specifically including the following processing steps.

[0088] First, the current molten pool state obtained by inversion is used as the current state variable of the system. The graph neural network surrogate model that completes domain adaptation is used as the prediction model. In the future finite prediction time domain, the corresponding molten pool state evolution trajectory is predicted in a rolling manner for different candidate control input sequences such as wire feeding speed, cutting speed, welding speed and current waveform parameters.

[0089] Secondly, an optimization objective function is constructed to seek a control sequence that minimizes the deviation between the predicted melting depth and melting penetration state and the target process parameters, while also minimizing the variation amplitude of the control quantity.

[0090] Next, deterministic physical boundary constraints are applied to ensure that control variables such as current, cutting speed, welding speed, and wire feed are always within the equipment's safety threshold range, thus avoiding optimization results from exceeding the equipment's physical feasible domain and reducing the uncontrollable risks brought about by the black box decision-making of the pure data-driven model.

[0091] Then, solve the above constrained optimization problem to obtain the optimal control sequence. Extract only the control increment of the first step in the sequence and send the control increment to the motion controller and power controller of the cutting and welding equipment.

[0092] Finally, in the next sampling period, the state inversion and rolling optimization process is re-executed, and the control quantity is continuously updated in a rolling time domain manner, thereby achieving millisecond-level real-time closed-loop feedback correction. The optimization problem of the above model predictive control can be expressed as:

[0093] ;

[0094] The constraints are:

[0095] ;

[0096] ;

[0097] in, To predict the length of the time domain;

[0098] For the first The control increments include adjustments to wire feed speed, cutting speed, welding speed, and current waveform parameters.

[0099] The first given for the prediction model The predicted values ​​of the molten pool state include the melt depth inversion value, the melt penetration state, and the keyhole morphology parameters;

[0100] For the first Reference values ​​for the target process parameters corresponding to each step;

[0101] and These are the weighted matrices for state deviation and control increment, respectively;

[0102] This represents the state prediction function formed by the graph neural network surrogate model that completes domain adaptation;

[0103] For the first The state of the molten pool at each step;

[0104] For the first Step control input;

[0105] and These are the lower and upper limits of physical safety for controlling input quantities, respectively.

[0106] VI. Other Implementation Methods

[0107] In an alternative embodiment of this invention, the main structure of the graph neural network surrogate model can also be replaced by a temporal-spatial fusion network combining a convolutional neural network and a long short-term memory network. Similarly, residual constraint terms of arc physics and heat transfer mechanism are embedded in its loss function to achieve a physical data fusion modeling effect similar to that of the graph neural network surrogate model. This alternative embodiment only changes the specific network structure of the surrogate model, while the core technical means of offline meta-learning pre-training, online few-sample domain adaptation, and model prediction control closed loop remain unchanged, and the technical objectives of this invention can be achieved in the same way.

[0108] In another alternative implementation of this embodiment, the online adaptive process may not employ gradient-based meta-learning fine-tuning, but instead adopts a metric learning-based nearest neighbor condition retrieval and parameter interpolation method. That is, a mapping library of condition features and corresponding model parameters is established in advance during the offline stage. During online operation, the historical condition most similar to the current small sample features is retrieved from this mapping library, and the model parameters corresponding to the historical condition are weighted and interpolated. The interpolation result is used as the adaptation parameters for the current condition, thereby reducing the computational overhead of the online adaptive process in edge device scenarios with limited computing resources. This alternative implementation can also achieve the real-time closed-loop control objective described in this invention while maintaining the ability to quickly adapt across conditions.

[0109] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. A control method for metal cutting and welding equipment, characterized in that, include: The initial parameters for pre-training of the graph neural network surrogate model are obtained by using meta-learning, which integrates the physical constraints of the electric arc energy conservation relationship and the partial differential equation of heat conduction, and is trained based on historical processing data of various materials, plate thicknesses and equipment models. The voltage waveform, current waveform, and visual features of the molten pool surface under the current processing conditions are collected to form a small sample dataset of the target processing conditions, and the small sample dataset of the target processing conditions is mapped to target domain graph data; Starting with the pre-trained initial parameters, the graph neural network proxy model is fine-tuned by gradient descent using the target domain graph data to obtain a graph neural network proxy model adapted to the current processing conditions. The real-time collected voltage waveforms, current waveforms, and visual features of the molten pool surface are input into the fine-tuned graph neural network surrogate model to obtain the temperature field distribution and flow field distribution inside the molten pool. Based on the temperature field distribution and the flow field distribution, the melting depth, melting penetration state and keyhole morphology parameters of the molten pool are obtained by inversion. Using the inverted melt depth, melt penetration state, and keyhole morphology parameters as the current state variables, and the finely tuned graph neural network surrogate model as the prediction model, the molten pool state evolution trajectory is predicted in a rolling manner for multiple candidate control input sequences within a preset prediction time domain. The optimal control sequence is obtained by solving an optimization problem. The optimization problem aims to minimize the weighted sum of the deviation between the predicted molten pool state and the target process index and the change amplitude of the control quantity, with the control input quantity being within a preset safety threshold range as the constraint condition. The control increment of the first step in the optimal control sequence is sent to the motion controller and power controller of the metal cutting and welding equipment to adjust the corresponding parameters in the wire feeding speed, cutting speed, welding speed and current waveform parameters. In the next sampling period after the control increment is issued, the acquisition of voltage waveform, current waveform and visual features of the molten pool surface, the inversion of the internal state of the molten pool, and the rolling optimization based on the fine-tuned graph neural network surrogate model are re-executed. The control increment issued to the motion controller and the power controller is continuously updated in a rolling time domain manner, forming a millisecond-level real-time closed-loop control that combines graph neural network surrogate model prediction with safety threshold constraint optimization based on physical constraints.

2. The control method according to claim 1, characterized in that, The initial parameters for the pre-training of the graph neural network proxy model are obtained in the following way: discrete sampling points on the processing path are used as graph nodes, and the heat transfer and diffusion relationship and electric arc effect relationship between graph nodes are used as the edges of the graph. A graph structure dataset is constructed based on electric arc physics simulation data and historical processing sensor data of various historical processing conditions. The electric arc energy conservation equation and the heat conduction partial differential equation are transformed into differentiable physical residual terms, and the physical residual terms are combined with the data-driven prediction loss of the graph neural network surrogate model to form a hybrid loss function. Each historical processing condition is divided into multiple meta-training tasks. In each training iteration, multiple historical processing conditions are sampled to form a meta-task, and the data corresponding to each meta-task is divided into a support set and a query set. For each meta-task, gradient descent is used to update the current parameters of the graph neural network proxy model using the support set data of that meta-task, thereby obtaining the task-specific parameters corresponding to that meta-task. The prediction loss of the corresponding task-specific parameters is evaluated using the query set data of each meta-task. The gradient of the prediction loss is calculated with respect to the current parameters of the graph neural network agent model, and the meta-parameters of the graph neural network agent model are updated accordingly. Repeat the process of task sampling, task-specific parameter updating, and meta-parameter updating until the meta-parameters converge. Use the converged meta-parameters as the initial parameters for pre-training.

3. The control method according to claim 1, characterized in that, After obtaining the graph neural network surrogate model adapted to the current processing conditions, the prediction output of the graph neural network surrogate model is compared with the surface features of the molten pool measured by the coaxial vision sensing device. If the comparison deviation exceeds the preset threshold, additional small sample data of the target working condition will be collected and gradient descent fine-tuning will be performed again. If the comparison deviation does not exceed the preset threshold, the decision domain adaptation is completed, and the system enters the online inference and closed-loop control stage.

4. The control method according to claim 1, characterized in that, The inversion of the melt depth of the molten pool includes: calculating the depth of the isothermal surface of the material melting point along the thickness direction in the temperature field distribution according to the preset isothermal surface determination criteria, and using the calculated depth as the melt depth inversion value; Based on the comparison between the inversion value of the melt depth and the thickness of the processed plate, it is determined whether the melt penetration state has been reached.

5. The control method according to claim 1, characterized in that, The keyhole morphology parameters obtained by inversion include: reconstructing the keyhole depth and keyhole opening diameter based on the velocity vector distribution in the flow field distribution and the preset gas-liquid interface determination conditions.

6. The control method according to claim 1, characterized in that, After the control increment is issued, the acquisition of voltage waveform, current waveform and visual features of the molten pool surface, inversion of the internal state of the molten pool and rolling optimization are re-executed in the next sampling cycle, and the control increment issued to the motion controller and the power controller is continuously updated in a rolling time domain manner.

7. The control method according to claim 1, characterized in that, The preset safety threshold range includes the upper and lower limits of equipment safety corresponding to current, cutting speed, welding speed, and wire feed, respectively.

8. A control system for metal cutting and welding equipment, characterized in that, include: The equipment execution module includes a metal cutting and welding host, a wire feeding mechanism, a traveling mechanism, and a power supply module; The sensing module includes a high-frequency voltage and current acquisition device and a coaxial vision sensing device. The high-frequency voltage and current acquisition device is connected to the power module and is used to acquire voltage waveforms and current waveforms. The coaxial vision sensing device is set in the molten pool processing area and is used to acquire visual features of the molten pool surface. The edge computing module includes a graph neural network surrogate model inference unit, an online domain adaptive fine-tuning unit, a molten pool state inversion unit, and a model prediction and control unit. The graph neural network surrogate model inference unit is used to output the internal temperature field distribution and flow field distribution of the molten pool based on the voltage waveform, current waveform, and visual features of the molten pool surface collected by the sensing module. The online domain adaptive fine-tuning unit is used to perform gradient descent fine-tuning on the model parameters in the graph neural network surrogate model inference unit. The molten pool state inversion unit is used to invert the molten pool's melting depth, melting penetration state, and keyhole morphology parameters based on the temperature field distribution and the flow field distribution. The model prediction and control unit is used to continuously solve for the optimal control sequence based on the inverted melting depth, melting penetration state, and keyhole morphology parameters and output the control increment. The offline training module includes a multi-condition historical database and a meta-learning pre-training engine. The meta-learning pre-training engine is used to perform meta-learning pre-training on the graph neural network agent model based on the historical processing condition data in the multi-condition historical database, and to send the initial parameters obtained from the pre-training to the graph neural network agent model inference unit of the edge computing module. The edge computing module is connected to the device execution module via an industrial control bus. It is used to send the control increment output by the model prediction control unit to the motion controller and power controller of the device execution module, and to feed back the processing state change caused by the device execution module executing the control increment to the sensing module.

9. The control system for metal cutting and welding equipment according to claim 8, characterized in that, The graph neural network proxy model in the inference unit of the graph neural network proxy model uses discrete sampling points on the processing path as graph nodes, and the heat transfer and diffusion relationship and electric arc effect relationship between graph nodes as graph edges.

10. The control system for metal cutting and welding equipment according to claim 8, characterized in that, The edge computing module is equipped with an artificial intelligence acceleration chip to support the real-time computation of the graph neural network agent model inference unit, the online domain adaptive fine-tuning unit, the melt pool state inversion unit, and the model prediction and control unit.

Images