Self-adaptive induction-electric arc hybrid welding method and system

Abstract

The invention provides a self-adaptive induction-electric arc hybrid welding method and system, belongs to the technical field of automatic welding, and aims to solve the problems of non-uniform preheating of a complex joint and poor welding quality caused by a fixed form of a heating device. The method comprises the steps that three-dimensional geometric information of a to-be-welded area is sensed in real time through a three-dimensional visual sensor; a pre-training decision model based on reinforcement learning is used, geometric information is used as state input, and a control instruction is calculated in real time; dynamically controlling the power of a modular array formed by a plurality of induction heating modules according to the instruction to form a customized heating area matched with the local geometrical characteristics, and controlling an electric arc welding gun to perform welding at the same time; the system comprises a three-dimensional visual sensor, a modular induction heating array, an electric arc welding gun and a process planning controller which are required for executing the method. Accurate self-adaptive heating of complex welding seams can be achieved, and the welding quality stability and the automation level are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of automated welding technology, and in particular to an adaptive induction-arc hybrid welding method and system. Background Technology

[0002] Induction-arc hybrid welding technology is an advanced welding process that improves welding efficiency and weld quality by preheating the base material using an induction coil in front of an electric arc. In existing technologies, inductors with fixed shapes are typically used, such as ring or U-shaped coils made of rigid copper tubing. These inductors are mainly designed for joints with simple and regular geometries, such as butt welds on flat plates or circumferential welds on straight pipes. During welding operations, a welding robot carries this fixed-shape inductor and moves along a pre-programmed trajectory to heat the weld area.

[0003] However, in fields such as engineering machinery, shipbuilding, and steel structures, there are numerous complex three-dimensional joints, including T-joints, corner joints, and variable cross-section components. Existing technologies exhibit significant shortcomings when dealing with such complex configurations. First, rigid and fixed-shape sensors cannot dynamically conform to the geometric contours of complex joints. As the welding head moves along the weld seam, the distance and coupling posture between the sensor and various points on the workpiece surface change drastically, resulting in severely uneven energy distribution during induction heating and an inability to form an ideal and uniform preheating temperature field in the welding area. Second, due to inconsistent preheating effects, the stability of the subsequent arc and the flow behavior of the molten pool fluctuate violently, easily leading to welding defects such as incomplete fusion and incomplete penetration. Furthermore, fixed process parameters cannot adapt to tolerances generated during workpiece manufacturing and assembly, resulting in poor weld quality consistency. Even when using sensing technologies such as 3D vision to acquire weld seam geometry, the lack of a matching, flexibly adjustable heating actuator still fails to fundamentally solve the problem of precise, adaptive heating of complex contours, leading to low automation levels in the entire system and severely limiting its application scenarios. Summary of the Invention

[0004] The purpose of this application is to provide an adaptive induction-arc hybrid welding method and system to solve the technical problems of uneven preheating, poor welding quality and low automation caused by the fixed shape of the heating device in the prior art, which cannot adapt to the geometric changes of complex three-dimensional joints.

[0005] To achieve the above objectives, this application provides an adaptive induction-arc hybrid welding system, including a welding head and a process planning controller. The welding head includes: a three-dimensional vision sensor for real-time acquisition of three-dimensional geometric information of the area to be welded; a modular induction heating array composed of multiple independently controllable induction heating modules for induction heating of the area to be welded; and an arc welding torch. The process planning controller includes a reinforcement learning-based decision module, adapted to: receive real-time three-dimensional geometric information acquired by the three-dimensional vision sensor as state input; and output control commands in real-time, through a pre-trained neural network decision model, for controlling the output power of the multiple induction heating modules and the welding parameters of the arc welding torch respectively. The process planning controller establishes a communication connection with the three-dimensional vision sensor, the modular induction heating array, and the arc welding torch in the welding head to receive the three-dimensional geometric information and control the multiple induction heating modules and the arc welding torch according to the control commands.

[0006] To achieve the above objectives, this application also provides an adaptive induction-arc hybrid welding method, characterized by the following steps: real-time perception, continuously scanning and acquiring the three-dimensional geometric information of the area to be welded using a three-dimensional vision sensor; intelligent decision-making, using the three-dimensional geometric information as state input, and utilizing a pre-trained reinforcement learning-based neural network decision model to calculate control commands in real time, the commands including power values ​​for controlling multiple independently controllable induction heating modules and welding parameters for controlling an arc welding torch; and coordinated execution, using a motion mechanism to drive the multiple induction heating modules and the arc welding torch to move along the welding path, and dynamically controlling the power of the multiple induction heating modules according to the control commands to form a customized heating area matching the local three-dimensional geometric features, while simultaneously controlling the arc welding torch to perform welding.

[0007] Optionally, the induction heating module is an electromagnetic lens module, which contains an induction coil and a soft magnetic ferrite core that has the function of converging and focusing magnetic fields.

[0008] Furthermore, the geometry of the soft magnetic ferrite core is parabolic or lenticular.

[0009] Furthermore, the electromagnetic lens module is a pluggable module.

[0010] Optionally, the system further includes multiple current or temperature sensors, each associated with one of the electromagnetic lens modules, which are communicatively connected to the process planning controller. The process planning controller is further adapted to: monitor the current or temperature of the electromagnetic lens modules using the sensors to determine if they have failed, and, if any electromagnetic lens module fails, adjust the power of other modules according to a preset power redistribution algorithm for compensation.

[0011] Optionally, the pre-trained neural network decision model is obtained through reinforcement learning in a digital twin simulation environment. Further, the digital twin simulation environment is a multiphysics coupled simulation model of electromagnetic field, temperature field, and flow field.

[0012] Optionally, the control command further includes fine-tuning commands for controlling the welding motion speed and attitude, and the fine-tuning commands are executed in the cooperative execution step.

[0013] Compared with existing technologies, this application has the following beneficial effects: First, by employing a modular electromagnetic lens array, a customized heating area that precisely matches the geometric contour of complex joints can be generated in real time according to software programming, realizing "software-defined heating" and fundamentally solving the problem of uneven preheating, exhibiting extremely high heating adaptability and precision. Second, through a reinforcement learning-based decision-making system, the complex process planning process is transformed into machine autonomous learning and real-time decision-making. It can autonomously discover and execute optimal process parameters based on real-time perceived weld changes, significantly improving the stability and consistency of welding quality and effectively reducing defects such as incomplete fusion and burn-through. Furthermore, the pluggable design of the electromagnetic lens module makes the replacement of faulty units simple and quick, greatly reducing maintenance costs and downtime, and improving the reliability and maintainability of the system. Finally, this invention enables induction-arc hybrid welding technology to be truly applied to actual industrial products with complex geometries and large assembly tolerances, providing key technical support for high-quality automated welding in heavy equipment, shipbuilding, steel structures, and other fields, and powerfully promoting the automation process of high-end equipment manufacturing. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 A schematic diagram of the overall architecture of an adaptive induction-arc hybrid welding system provided in this application embodiment; Figure 2This is a schematic diagram of the structure of the welding head provided in an embodiment of this application; Figure 3 This is a cross-sectional view of the electromagnetic lens module provided in an embodiment of this application; Figure 4 A schematic flowchart of an adaptive induction-arc hybrid welding method provided in an embodiment of this application; Figure 5 A functional block diagram of the process planning controller provided in the embodiments of this application; Figure 6 The signaling interaction timing diagram of the main components in the adaptive induction-arc hybrid welding method provided in the embodiments of this application is shown.

[0016] Key reference numerals: 1-Industrial robot; 2-Welding head; 3-Process planning controller; 4-Induction heating power supply; 5-Welding power supply; 6-Workpiece; 21-3D vision sensor; 22-Modular electromagnetic lens array; 23-Arc welding torch; 24-Array substrate; 221-Induction coil; 222-Field focusing magnetic core; 223-Shell; 224-Interface; 31-Data acquisition module; 32-Decision module; 33-Neural network model; 34-Digital twin environment; S101-Step of robot moving welding head; S102-Step of scanning and acquiring 3D geometric information; S103-Step of receiving information and using it as status input; S104-Step of calculating optimal action command; S105-Step of issuing command; S106-Step of determining whether welding is finished. Detailed Implementation

[0017] To better understand the technical solutions of this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be noted that the specific embodiments described herein are only some embodiments of this application, not all of them, and are only used to explain this application, not to limit this application.

[0018] Example 1

[0019] This embodiment provides an adaptive induction-arc hybrid welding system and its working method, which aims to perform high-quality automated welding of T-joints with variable gap characteristics.

[0020] Please see Figure 1This illustration shows the overall architecture of an adaptive induction-arc hybrid welding system according to one embodiment of this application. The system mainly includes a six-axis industrial robot 1, a welding head 2 mounted on the end effector of the industrial robot 1, a process planning controller 3 as the core of the system control, a multi-channel induction heating power supply 4 providing energy for induction heating, and a welding power supply 5 providing energy for arc welding. During system operation, the industrial robot 1, according to instructions issued by the process planning controller 3, moves the welding head 2 along the complex weld path of the workpiece 6 to be welded. In this embodiment, the workpiece 6 is specifically a T-joint made of 25mm thick Q345B steel plate, which, after assembly, has a non-uniform gap ranging from 1mm to 3.5mm.

[0021] Please see Figure 2 The internal structure of welding head 2 is shown in detail. Welding head 2 adopts a highly integrated design. At its front end, a laser line scanner, serving as a 3D vision sensor 21, is installed to scan and acquire the 3D geometric information of the area to be welded in real time during the welding process. Following the 3D vision sensor 21 is a modular electromagnetic lens array 22, which is the core actuator for achieving adaptive heating. Specifically, this array consists of 100 square electromagnetic lens modules, each 15mm x 15mm in size, arranged in a 10x10 checkerboard pattern and fixed on an array substrate 24. At the center of the modular electromagnetic lens array 22, a gas-shielded welding torch, serving as an arc welding torch 23, is integrated for metal filler welding in the area after induction preheating. It can be understood that the relative positions of the 3D vision sensor 21, the modular electromagnetic lens array 22, and the arc welding torch 23 on welding head 2 have been precisely calibrated to ensure the spatial coordination of the three actions: sensing, preheating, and welding.

[0022] Please see Figure 3The diagram shows a cross-sectional view of the internal structure of a single electromagnetic lens module. Each module constitutes an independent, precisely controllable heating unit. Structurally, each module includes an induction coil 221 for generating a high-frequency alternating magnetic field. This coil is preferably a planar helical coil to accommodate the module's flattened design. Behind the induction coil 221, a field-focusing magnetic core 222 made of a high-permeability soft magnetic ferrite material is positioned. The geometry of this field-focusing magnetic core 222 is specially designed as a parabolic shape, acting similarly to an optical lens converging light. It can converge and focus the divergent magnetic field lines generated by the induction coil 221, allowing the magnetic field energy to be highly concentrated and projected onto a specific small area of ​​the workpiece 6, thereby achieving high-precision and high-efficiency heating. The entire module is protected by a high-temperature resistant, insulating shell 223. At the base of the module, a standardized interface 224 is provided, which integrates mechanical fixing structures and electrical connection terminals. This design makes each electromagnetic lens module a pluggable module, which can be quickly removed from the array substrate 24 and replaced when any module fails, thereby greatly improving the maintainability and reliability of the system.

[0023] The following is combined with Figure 4 and Figure 5 The working method and control logic of this embodiment will be described in detail. Figure 4 The overall flow of the method in this application is shown. Figure 5 This demonstrates the functional modules within the process planning controller 3. The entire welding process constitutes a continuous closed-loop control cycle.

[0024] In the initial stage of the process, namely in step S101 where the robot moves the welding head, the process planning controller 3 controls the industrial robot 1 to move the welding head 2 along the weld direction of the T-joint at a preset base speed (e.g., 30 cm / min).

[0025] During the movement, the system enters step S102, which involves scanning to acquire three-dimensional geometric information. The three-dimensional vision sensor 21 (i.e., a laser line scanner) located at the front end of the welding head 2 continuously projects a laser line forward onto the area to be welded at a high frequency (e.g., 100 Hz). Its internal camera captures the deformation of the laser line on the surface of the workpiece 6, and the three-dimensional point cloud data of the weld area is calculated in real time using the principle of triangulation. This data accurately describes the local geometric contour of the weld, including the root position of the T-joint, the bevel angle, and the crucial assembly gap size.

[0026] Subsequently, in step S103, where information is received and used as status input, the aforementioned 3D point cloud data is sent to the process planning controller 3. The data acquisition module 31 inside the controller 3 filters, denoises, and extracts features from the raw point cloud data, converting it into structured status information, such as a gap value of 1.5mm at the current position and a bevel width of 12mm. This status information is then transmitted to the decision module 32 as the "status input" for the current moment.

[0027] The next step is step S104, which calculates the optimal action command and is the core of the intelligent decision-making in this application. A pre-trained reinforcement learning-based neural network model 33 is deployed within the decision module 32. Upon receiving the current state input, this model performs a forward propagation calculation and outputs a set of optimal "action commands" within a very short time (e.g., within 10 milliseconds). This set of action commands is a multi-dimensional vector containing precise control parameters for each actuator in the system. Specifically, it includes: a 100-dimensional power vector, corresponding to the target output power values ​​(range 0-500W) of the 100 electromagnetic lens modules in the modular electromagnetic lens array 22; and a set of arc welding parameters, such as welding current, arc voltage, and wire feed speed.

[0028] In step S105, the process planning controller 3 parses the calculated action commands and sends them to the multi-channel induction heating power supply 4 and the welding power supply 5 via a high-speed communication bus (e.g., EtherCAT). The induction heating power supply 4 can independently control the current amplitude and frequency supplied to each electromagnetic lens module according to the commands, thereby achieving precise adjustment of the output power of each module; the welding power supply 5 sets the corresponding arc parameters according to the commands.

[0029] During the collaborative execution phase, when the welding head 2 reaches a position with a gap of 1.5mm, the instruction output by the decision module 32 may be as follows: activate three rows of 30 electromagnetic lens modules directly opposite the weld root and the bevel surfaces on both sides; allocate a higher power of 350W to the 10 modules in the row directly opposite the weld root to ensure sufficient preheating of the root area; and allocate a slightly lower power of 250W to the two rows of 20 modules in the bevel areas on both sides. Through this differentiated power allocation, a uniform preheating temperature field with a temperature of approximately 200°C, precisely matching the T-joint profile, is formed on the surface of the workpiece 6. Accordingly, the instruction may set the parameters of the arc welding torch 23 to 280A current and 30V voltage, which is a set of stable welding parameters suitable for this gap and preheating state.

[0030] As the welding head 2 continues to advance, when the 3D vision sensor 21 scans a location where the gap widens to 3.5mm, the new state input will cause the decision module 32 to adjust its output strategy in real time. At this point, the new action command might be: to prevent the bevel edge from overheating or even burning through due to the increased gap, the command would slightly reduce the output power of the module corresponding to the bevel edge area; simultaneously, to ensure the increased gap is completely filled and sufficient weld penetration is achieved, the command would increase the arc current to 300A and possibly increase the wire feed speed accordingly. As a preferred implementation, the control command could also include fine-tuning instructions for the welding motion speed. For example, in areas where the gap widens, to ensure sufficient weld penetration, the command might reduce the movement speed of the industrial robot 1 from 30 cm / min to 25 cm / min.

[0031] Finally, in step S106, which determines whether welding has ended, the system checks whether the weld end point has been reached. If not, the process returns to step S101 and continues the closed-loop control of "real-time perception - intelligent decision-making - collaborative execution" to achieve point-by-point, real-time, and adaptive welding of the entire weld. If the end point has been reached, the process ends.

[0032] After welding, samples were taken from the T-joint weld and subjected to metallographic analysis. Experimental results showed that the weld penetration and width remained highly consistent throughout the entire weld, whether in a narrow 1.5mm gap area or a wide 3.5mm gap area. Furthermore, the internal structure was dense, without welding defects such as incomplete fusion, incomplete penetration, slag inclusions, or porosity. This fully demonstrates that the solution provided in this embodiment can effectively overcome the challenges posed by complex joint geometry variations, ensuring the stability and consistency of weld quality.

[0033] As an optional implementation, the system in this embodiment can also achieve fault self-diagnosis and compensation functions. Specifically, a current sensor can be integrated into the power supply circuit of each electromagnetic lens module, or a temperature sensor can be attached to the module housing 223, and these sensors can be connected to the process planning controller 3. The controller 3 can determine the operating status of each module by monitoring the current or temperature of each module in real time. For example, when the current of a module suddenly drops to zero or the temperature rises abnormally, the controller 3 can determine that the module has failed. At this time, the controller 3 will immediately start a preset power redistribution algorithm, such as increasing the output power of several neighboring normally operating modules according to a certain weight, to compensate for the local heating loss caused by the failed module, thereby ensuring the stability of the overall preheating temperature field and further enhancing the robustness of the system.

[0034] Example 2

[0035] This embodiment is a variation of Embodiment 1, the main difference being the structural design of the modular electromagnetic lens array 22. To achieve better heating performance in applications requiring extremely smooth heat distribution, this embodiment employs a honeycomb-shaped array structure.

[0036] Specifically, in this embodiment, the units constituting the modular electromagnetic lens array 22 are no longer square modules, but multiple regular hexagonal electromagnetic lens modules. These hexagonal modules are tightly embedded in the array substrate 24 in a honeycomb structure. Compared with the checkerboard arrangement in Embodiment 1, the honeycomb arrangement has a higher space fill rate, resulting in smaller gaps between modules and almost seamless splicing.

[0037] The other components of the system, such as the industrial robot 1, the three-dimensional vision sensor 21 on the welding head 2, the arc welding gun 23, and the process planning controller 3, have the same structure and function as in Example 1.

[0038] During operation, as the welding head 2 moves along the weld seam, the process planning controller 3 activates and controls the corresponding hexagonal modules based on the weld seam contour perceived in real time by the 3D vision sensor 21. Because the modules are more densely packed and the heating units are less discrete, the controller can perform more precise and continuous power control. For example, when welding a T-joint, for the transition rounded corner area from the weld root to the sidewall bevel, the controller can apply a smooth, gradual power distribution to a row of hexagonal modules, thereby achieving a smooth heat transition in this transition area and effectively avoiding localized thermal stress concentration that may occur due to sudden changes in heating power.

[0039] The modular electromagnetic lens array 22, arranged in a honeycomb pattern, can generate customized heating zones with more complex shapes and smoother power distribution profiles. This superior thermal field distribution not only helps stabilize the molten pool and improve weld formation, but also further reduces residual stress and deformation generated during welding, which is of great significance for improving the fatigue life and service reliability of welded joints.

[0040] Example 3

[0041] This embodiment details the offline training process of the core neural network model 33 in the process planning controller 3 described in Embodiments 1 and 2. This process is a prerequisite for realizing intelligent decision-making in welding processes and specifically demonstrates the application of "digital twin" technology.

[0042] The training process takes place on an offline computing server equipped with a high-performance graphics processor, and the entire process can be divided into three key steps: First, a digital twin environment is built. A high-fidelity digital twin environment 34 is constructed using professional multiphysics simulation software (such as COMSOL Multiphysics or ANSYS). This environment is a virtual model capable of accurately simulating the real welding process, and it is completely consistent with the physical system described in Example 1 (including the geometry of the welding head 2, the electromagnetic characteristics of the electromagnetic lens module, the material properties of the workpiece 6, etc.) and the welding physical process. The core of this model lies in its multiphysics coupling characteristics, which can simultaneously solve for three interacting physical fields: electromagnetic field, temperature field, and flow field. Specifically, the model can simulate: the eddy current distribution and corresponding Joule heating generated inside the workpiece 6 when a given high-frequency current passes through the induction coil 221 (electromagnetic field-temperature field coupling); how this heat diffuses inside the workpiece through heat conduction to form a preheating temperature field (heat transfer); and the process of metal melting, flowing, and solidifying to form a weld under the action of an electric arc (computational fluid dynamics). This digital twin environment 34 constitutes a virtual "trial and error" platform.

[0043] Secondly, the reinforcement learning agent is defined and trained. An advanced reinforcement learning algorithm, namely the proximal policy optimization algorithm, is used to train a deep neural network as the decision-making agent (i.e., the final deployed neural network model 33). The state space is defined as the environmental information that the agent can "observe" at each decision moment. In this embodiment, the state consists of two parts: one part is the weld contour point cloud data acquired by the virtual 3D vision sensor, which is processed to form a vector describing the geometric features; the other part is the workpiece surface temperature distribution map acquired by the virtual temperature sensor, forming a matrix describing the current thermal state. The action space is defined as the control operations that the agent can perform. In this embodiment, the action is a continuous multi-dimensional vector, including the power values ​​of 100 electromagnetic lens modules (each value is continuously adjustable between 0-500W), and 3 arc parameters (welding current, arc voltage, and wire feed speed). The reward function is the key to guiding the learning direction of the agent. This embodiment designs a reward function with the following form: in, and These are the weld depth and weld width obtained from simulation in the digital twin environment. and It refers to the desired depth and width of the weld. It is the volume of defective areas such as incomplete fusion or incomplete penetration detected in the simulation. These are the weighting coefficients. The goal of this reward function is to guide the agent to learn an action policy that makes the simulated weld geometry as close as possible to the target value, while minimizing the occurrence of defects.

[0044] Training was initiated on a server, where the agent explored a vast amount of "virtual welding" within thousands of independent digital twin environments running in parallel, executing over 20 million decision steps. In each step, the agent observed the state, output an action, and the environment simulated the action, returning a reward value. Through a proximal policy optimization algorithm that continuously updated the neural network weights based on the obtained rewards, the agent gradually learned the complex nonlinear mapping between any given weld state (geometry + temperature) and the optimal welding action (power + arc parameters).

[0045] Finally, the model is deployed. After sufficient training and convergence, the weights of the neural network are fixed, forming a high-performance decision model 33. This fixed model file is exported from the training server and deployed to the industrial PC in the actual welding system, namely the decision module 32 of the process planning controller 3.

[0046] The training process described in this embodiment enables the development of an intelligent decision-making core capable of rapidly and accurately calculating optimal control commands within 10 milliseconds based on real-time sensor input. This frees the entire welding system from reliance on manual programming and tedious process experiments, achieving complete autonomy in welding process decision-making and demonstrating an extremely high level of intelligence.

[0047] Example 4

[0048] This embodiment demonstrates the versatility and adaptability of the proposed solution under different sensing methods and more complex application scenarios.

[0049] In terms of system structure, the main difference between this embodiment and Embodiment 1 lies in the type of the 3D vision sensor 21 and the geometry of the workpiece 6 to be welded. In this embodiment, the 3D vision sensor 21 is replaced by a structured light 3D camera. Unlike a laser line scanner that can only acquire a single line outline, the structured light 3D camera can capture and reconstruct a 3D point cloud of an area (e.g., 50mm x 50mm) in front of it at once by projecting coded structured light patterns (such as stripes or grids), thereby providing richer and more comprehensive geometric information.

[0050] The welding object is a butt weld of a variable cross-section S-shaped pipe fitting commonly found in marine piping systems. The complexity of this workpiece is reflected in two aspects: first, the weld path is a complex three-dimensional spatial curve with a constantly changing radius of curvature; second, the wall thickness of the pipe fitting also changes along the weld path, for example, smoothly transitioning from 5mm to 8mm.

[0051] Let's combine them again below. Figure 4 and Figure 6 The working process of this embodiment will be described. Figure 6The real-time interactions between the components are clearly illustrated in a time sequence diagram. Industrial robot 1 drives welding head 2 to move along the centerline path of the S-shaped weld. During the movement, the structured light 3D camera periodically collects the area point cloud data of the weld area in front and sends this "sensor data" to the process planning controller.

[0052] Upon receiving the data, the process planning controller is activated. Its data acquisition module 31 can not only identify the shape and gap of the bevel from the point cloud, but also further calculate the local radius of curvature of the current weld segment and the changing trend of the pipe wall thickness. This rich geometric information together constitutes the state input of the decision module 32.

[0053] Based on these complex inputs, the neural network model 33 within the decision module 32 calculates the optimal action commands in real time. These commands enable higher-level adaptive control: when the system reaches a point with a large curvature in an S-shaped bend, the model outputs commands to appropriately reduce the movement speed of the industrial robot 1 to ensure a smooth transition of the welding torch posture and a stable molten pool; on the other hand, it adjusts the activation area and power distribution of the modular electromagnetic lens array 22 to precisely match the shape of the heating zone with the curved weld. When the system senses that the pipe wall ahead will transition from a 5mm thin-walled section to an 8mm thick-walled section, the model responds in advance. Its output commands increase the output power of the electromagnetic lens module corresponding to the thick-walled area to ensure that the thicker base material can be fully preheated to the target temperature before the arc arrives. Simultaneously, the commands smoothly increase the welding current and wire feed speed to match the increased wall thickness, ensuring a uniform transition of penetration depth in the thick-thin transition zone, thereby avoiding incomplete penetration defects.

[0054] These complex control commands are then sent to the controllers of the induction power supply, welding power supply, and industrial robot 1 to coordinate the execution of heating, welding, and movement. The entire process is repeated cyclically, enabling high-quality automated welding of complex spatial curve welds with varying cross-sections and curvatures.

[0055] This embodiment demonstrates that the technical solution proposed in this application has strong versatility, is compatible with different three-dimensional vision sensing technologies, and can effectively cope with various complex geometric configurations other than T-joints, providing strong technical support for the application of this technology in high-end equipment fields such as shipbuilding, aerospace, and nuclear power.

[0056] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method of adaptive induction-arc hybrid welding, characterized in that, Includes the following steps: Real-time sensing: Continuously scans and acquires the three-dimensional geometric information of the area to be welded through a three-dimensional vision sensor; Intelligent decision-making: Using the three-dimensional geometric information as state input, a pre-trained neural network decision-making model based on reinforcement learning is used to calculate control commands in real time. The commands include power values ​​for controlling multiple independently controllable induction heating modules and welding parameters for controlling an arc welding torch. Collaborative execution: The motion mechanism drives the multiple induction heating modules and the arc welding gun to move along the welding path, and dynamically controls the power of the multiple induction heating modules according to the control command to form a customized heating area that matches the local three-dimensional geometric features, while simultaneously controlling the arc welding gun to perform welding.

2. The method of claim 1, wherein, In the intelligent decision-making step, the pre-trained neural network decision-making model is obtained through reinforcement learning in a digital twin simulation environment.

3. The method of claim 1 or claim 2, wherein, The control commands also include fine-tuning commands for controlling the welding motion speed and posture, and the fine-tuning commands are executed in the cooperative execution step.

4. An adaptive induction-arc hybrid welding system, comprising a welding head and a process planning controller, characterized in that, The welding head includes: a three-dimensional vision sensor for acquiring three-dimensional geometric information of the area to be welded in real time; a modular induction heating array, consisting of multiple independently controllable induction heating modules, for induction heating of the area to be welded; and an arc welding torch. The process planning controller includes a reinforcement learning-based decision module, which is adapted to: receive real-time three-dimensional geometric information acquired by the three-dimensional vision sensor as a state input; and output control commands in real time through a pre-trained neural network decision model for controlling the output power of the plurality of induction heating modules and the welding parameters of the arc welding gun, respectively. The process planning controller establishes a communication connection with the three-dimensional vision sensor, the modular induction heating array, and the arc welding gun in the welding head to receive the three-dimensional geometric information and control the multiple induction heating modules and the arc welding gun according to the control instructions.

5. The system of claim 4, wherein, The induction heating module is an electromagnetic lens module, which contains an induction coil and a soft magnetic ferrite core that has the function of converging and focusing magnetic fields.

6. The system of claim 5, wherein, The soft magnetic ferrite core has a parabolic or lenticular geometry.

7. The system of claim 5, wherein, The electromagnetic lens module is a pluggable module.

8. The system of claim 7, wherein, The system also includes: multiple current or temperature sensors respectively associated with the electromagnetic lens module, the sensors being communicatively connected to the process planning controller; The process planning controller is also adapted to: monitor the current or temperature of the electromagnetic lens module through the sensor to determine whether it has failed, and when any of the electromagnetic lens modules fails, adjust the power of other modules to compensate according to a preset power redistribution algorithm.

9. The system of claim 4, wherein, The pre-trained neural network decision model was obtained through reinforcement learning in a digital twin simulation environment.

10. The system of claim 9, wherein, The digital twin simulation environment is a multi-physics coupled simulation model of electromagnetic field, temperature field, and flow field.

Images