A dynamic parameter adaptive TIG welding automation method with low deformation of thin-walled workpieces

Abstract

The application provides a thin-walled workpiece low-deformation dynamic parameter self-adaptive TIG welding automation method, including the following steps: S1, a multi-element heterogeneous sensing fusion system is constructed, a molten pool visual sensor, an arc sensing module and a laser positioning module are integrated; S2, a deformation prediction digital twin model driven by data and mechanism is established, and finite element thermal-mechanical coupling analysis and machine learning algorithms are fused; S3, a deep reinforcement learning and self-adaptive PID compound controller is designed, and the pulse frequency, the welding speed and the heat input welding parameters are dynamically adjusted according to the real-time sensing information and the deformation prediction result; S4, a welding heat sink and a flexible compression auxiliary means are integrated, and the residual stress and the out-of-plane deformation are inhibited in real time in the welding process; the application improves the welding forming consistency and the production efficiency, and reduces the overall process landing cost, and is suitable for the high-precision automatic welding needs of various thin-walled components in the fields of aerospace, precision equipment and the like.

Description

Technical Field

[0001] This invention belongs to the field of thin-walled workpiece splicing and manufacturing technology, specifically a dynamic parameter adaptive TIG welding automation method for thin-walled workpieces with low deformation. Background Technology

[0002] Thin-walled components, with their lightweight and high specific strength characteristics, are increasingly used in aerospace, precision instruments, and new energy storage equipment. TIG welding, as a high-quality welding process with low heat input, is the mainstream technical solution for splicing and manufacturing such thin-walled workpieces.

[0003] However, most existing solutions use offline-fixed welding parameters and passive rigid clamping fixtures, which cannot adapt to the real-time fluctuations in the thermo-mechanical state of thin-walled workpieces during welding. On the one hand, the prediction of pure mechanism finite element deformation models has strong lag, making it difficult to predict welding deformation in advance. Conventional open-loop parameter control modes cannot cope with random disturbances such as workpiece clamping deviations and bevel gap fluctuations during the welding process. On the other hand, single passive deformation suppression methods cannot synchronously match real-time heat input changes, which can easily lead to problems such as residual stress concentration, local out-of-plane depressions, or excessive bulging. In actual production, the post-weld straightening process for thin-walled workpieces accounts for as much as 30%, which seriously restricts the welding forming accuracy and production efficiency of thin-walled components.

[0004] To address the problems raised in the background art, those skilled in the art have proposed a dynamic parameter adaptive TIG welding automation method for low deformation of thin-walled workpieces. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an automated TIG welding method with low deformation of thin-walled workpieces based on dynamic parameters. This method solves the problem that most existing technical solutions employ offline-fixed welding parameters and passive rigid clamping fixtures, which cannot adapt to the real-time fluctuations in the thermo-mechanical state of thin-walled workpieces during welding. On the one hand, purely mechanistic finite element deformation models have strong predictive lag, making it difficult to anticipate welding deformation. Conventional open-loop parameter control modes cannot cope with random disturbances such as workpiece clamping deviations and bevel gap fluctuations during welding. On the other hand, single passive deformation suppression methods cannot synchronously match real-time heat input changes, easily leading to problems such as residual stress concentration, local out-of-plane depressions, or excessive bulging. In actual production, the post-weld straightening process for thin-walled workpieces accounts for more than 30%, severely restricting the welding forming accuracy and production efficiency of thin-walled components.

[0006] An automated method for low-deformation dynamic parameter adaptive TIG welding of thin-walled workpieces includes the following steps:

[0007] S1. Construct a multi-source heterogeneous sensing fusion system, integrating a molten pool vision sensor, an arc sensing module, and a laser positioning module to collect multi-source heterogeneous information during the welding process of thin-walled workpieces, and achieve full-domain high-fidelity perception of the welding status.

[0008] S2. Establish a digital twin model for deformation prediction driven by a combination of data and mechanism, and integrate finite element thermo-mechanical coupling analysis with machine learning algorithms to achieve online advanced prediction of welding deformation and pre-optimization of process parameters.

[0009] S3. Design a deep reinforcement learning and adaptive PID composite controller. Based on real-time sensing information and deformation prediction results, dynamically adjust key welding parameters such as pulse frequency, welding speed and heat input to form a closed-loop adaptive control architecture of "sensing-prediction-decision-execution".

[0010] S4 integrates welding heat sink and flexible clamping auxiliary methods to suppress residual stress and out-of-plane deformation in real time during the welding process, and realize the whole process collaborative control of welding deformation of thin-walled workpieces.

[0011] Preferably, the specific implementation of the multi-heterogeneous sensing fusion system in step S1 is as follows:

[0012] S1.1, The molten pool vision sensor uses a high-speed CMOS camera with a frame rate of... Acquire molten pool morphology images at ≥2000fps and extract molten pool length. ,width and area Geometric features;

[0013] S1.2, Arc sensing module collects arc voltage in real time. With welding current Signal, calculate instantaneous arc power and arc impedance ;

[0014] S1.3 The laser positioning module uses the structured light triangulation principle to obtain the three-dimensional contour of the weld and extract the weld deviation. Gaps with bevel ;

[0015] S1.4. An improved adaptive weighted multi-source information fusion algorithm is used to perform spatiotemporal alignment and fusion of the above-mentioned multi-source heterogeneous data. The multi-source information fusion vector is defined as:

[0016] ;

[0017] in, For the first Sensor-like devices at all times The original feature vector, For normalized mapping functions, The adaptive weight coefficients are dynamically calculated using the following attention mechanism:

[0018] ;

[0019] ;

[0020] in, For the first Hidden layer representation of sensor-like features v For trainable attention parameters, The temperature coefficient is used; when the signal-to-noise ratio of a certain type of sensor signal is lower than a preset threshold, its corresponding weight automatically decays to near zero, ensuring the robustness of the fusion result.

[0021] Preferably, the specific construction method of the data and mechanism hybrid-driven deformation prediction digital twin model in step S2 is as follows:

[0022] S2.1 Establish a thermo-mechanical coupled finite element model for welding thin-walled workpieces. A double ellipsoidal heat source model is used to describe the heat source input, and the heat source power density distribution is as follows:

[0023] ;

[0024] in, For effective heat input, a, b, and c are the three semi-axial parameters of the ellipsoidal heat source, respectively, and are related to the welding speed. and sheet thickness Related:

[0025] ;

[0026] in, Thermal conductivity, Characteristic temperature difference, , , This is an empirical correction factor;

[0027] S2.2. Based on finite element simulation, multiple sets of temperature and deformation field datasets under various working conditions are generated. The Physical Information Neural Network (PINN) is used for data augmentation and accelerated solution. Its loss function is defined as:

[0028] ;

[0029] ;

[0030] in, For the residual loss of the heat conduction equation, For the fitting loss of finite element simulation data, For boundary condition constraint loss, , , For balancing weighting coefficients;

[0031] S2.3 Construct a data-mechanism hybrid prediction model, embedding the finite element analytical solution as a physical prior into a Long Short-Term Memory (LSTM) network to predict time steps. Deformation amount:

[0032] ;

[0033] in, The deformation components are calculated in real time for the finite element model. This is the nonlinear compensation amount output by the residual correction network.

[0034] Preferably, the specific design method of the deep reinforcement learning and adaptive PID composite controller in step S3 is as follows:

[0035] S3.1 Model welding parameter control as a Markov Decision Process (MDP) and define the state space. ,in, This represents the current deformation error;

[0036] S3.2, The motion space is defined as the welding parameter adjustment vector. These correspond to the pulse frequency increment, welding speed increment, and heat input increment, respectively.

[0037] S3.3. Train the deep reinforcement learning policy network using the Proximal Policy Optimization (PPO) algorithm. Its objective function is:

[0038] ;

[0039] in, For generalized advantage estimation, ϵ is the pruning parameter;

[0040] S3.4 The adaptive PID controller receives the coarse-tuning instructions from deep reinforcement learning and then performs fine-tuning. Its control law is as follows:

[0041] ;

[0042] The PID parameters are self-tuned in real time based on the nonlinear characteristics of the welding process.

[0043] , ;

[0044] in, This is a stability index for the welding process, obtained by normalizing the standard deviation of the arc signal. , , , This is the adaptive gain coefficient.

[0045] Preferably, the specific implementation of the heat sink during welding in step S4 is as follows:

[0046] S4.1 A pre-set cooling channel is installed on the back side of the thin-walled workpiece during welding. Cooling medium circulates within the channel. The distance between the cooling channel and the weld toe is... Keep it within the range of 3-8mm;

[0047] S4.2 Based on the finite element pre-analysis results, a cooling channel layout optimization model is established with the goal of minimizing out-of-plane deformation:

[0048] ;

[0049] ;

[0050] in, For the out-of-plane displacement of each monitoring point, Given a vector of cooling channel locations and flow parameters, constraints are set to ensure that cooling does not lead to overcooling and cracking.

[0051] S4.3 The cooling medium flow rate is dynamically adjusted based on real-time heat input:

[0052] ;

[0053] in, As the baseline flow rate, This is the flow regulation coefficient. This is the reference thermal input power.

[0054] Preferably, the specific implementation of the flexible pressing in step S4 is as follows:

[0055] S4.4 Multiple flexible clamping devices are arranged on both sides of the welding area of ​​the thin-walled workpiece, and the pressure at each clamping point is... Independently adjustable;

[0056] S4.5. Establish a coupling relationship model between clamping force and welding deformation, and use the response surface methodology to establish a proxy model for the clamping force parameters and the maximum deformation:

[0057] ;

[0058] in, For each clamping point, the force vector is... These are polynomial basis functions;

[0059] S4.6. Real-time optimization of pressure at each clamping point based on online gradient descent method:

[0060] ;

[0061] in, To optimize the step size of the clamping force, the gradient is obtained by analytical differentiation of the surrogate model, while constraining the pressure at each point. To avoid crushing thin-walled workpieces.

[0062] Preferably, the deformation prediction digital twin model in step S2 further includes an online incremental learning mechanism, specifically:

[0063] When the actual welding deformation monitoring value Compared with the predicted value Deviation exceeds tolerance threshold At that time, the model is triggered to update online:

[0064] ;

[0065] ;

[0066] in, For online learning rate, To prevent catastrophic forgetting in the model, regularization coefficients are used. Approximate calculation is performed using backpropagation with finite samples; incrementally learned samples are stored in the experience replay pool. The pool capacity is A priority sampling strategy is adopted:

[0067] ;

[0068] in, For the first The prediction error for each sample, where ω is the priority exponent and ϵ is a small constant.

[0069] Preferably, in step S3, the deep reinforcement learning policy network adopts a multi-objective reward function design to balance deformation control accuracy and welding efficiency.

[0070] ;

[0071] in, For the squared term of deformation error, This is a penalty term for changes in heat input. This is a penalty for incremental welding time. This is a welding safety constraint indication function, which takes the value -1 when the arc parameters exceed the safe range and 0 otherwise. , , , These are multi-objective weighting coefficients;

[0072] Simultaneously, a course learning strategy is introduced, with a relatively high tolerance for deformation set in the early stages of training. With each training round attenuation:

[0073] ;

[0074] in, To determine the course decay rate, the strategy network is gradually guided from coarse regulation to fine regulation.

[0075] Preferably, the method also includes an online construction and dynamic updating mechanism for the safety domain of welding process parameter constraints:

[0076] Based on real-time acquired arc characteristics, a safe operating domain for the welding process is constructed. :

[0077] ;

[0078] The constraint functions are defined as follows:

[0079] ,

[0080] ;

[0081] Deep reinforcement learning action outputs are corrected by safe region projection:

[0082] ;

[0083] Ensure that all control commands are always within the safe operating range to prevent defects such as burn-through and lack of fusion.

[0084] Preferably, the "perception-prediction-decision-execution" closed-loop architecture adopts a hierarchical control strategy:

[0085] The upper layer is the digital twin prediction layer, based on the prediction cycle. =50-200ms runtime deformation advance prediction, output pre-optimized parameter set ;

[0086] The middle layer is the deep reinforcement learning decision layer, based on the decision cycle. =20-50ms Parameters are adjusted based on real-time status, and control commands are output. ;

[0087] The lower layer is the adaptive PID execution layer, based on the execution cycle. Fine-tuning of parameters and servo tracking are performed within 5-10ms.

[0088] The three layers communicate via status information. With motion information The asynchronous message passing mechanism is coupled, and its overall closed-loop transfer function is approximately as follows:

[0089] ;

[0090] in, For the adaptive PID controller transfer function, For the equivalent transfer function of deep reinforcement learning strategies, For the delayed link in digital twin prediction, It is a multi-source sensing feedback loop; the collaborative work of each layer keeps the total system response delay within 100ms, meeting the real-time control requirements under the rapid thermal response characteristics of thin-walled workpieces.

[0091] Compared with the prior art, the present invention has the following beneficial effects:

[0092] This low-deformation dynamic parameter adaptive TIG welding automation technology for thin-walled workpieces breaks through the technical bottleneck of traditional TIG welding, which relies on offline solidification of process parameters and passive shape control by rigid tooling. By constructing a multi-source heterogeneous high-fidelity sensing system, a digital twin prediction model driven by a hybrid of data and mechanism, and a millisecond-level closed-loop control architecture combining deep reinforcement learning and adaptive PID, coupled with thermo-mechanical collaborative shape control methods of heat sink and dynamic flexible clamping, the total system response delay is controlled within 100ms. This not only solves common defects in traditional TIG welding such as tungsten inclusion, porosity, and deformation deviation, but also significantly reduces the risk of instability and deformation caused by clamping deviation and working condition fluctuations during the welding of thin-walled workpieces. It can achieve high-precision, low-deformation welding of thin-walled components without relying on high-cost, high-precision special equipment, effectively reducing the proportion of post-weld straightening processes. While improving the consistency of welding formation and production efficiency, it also reduces the overall process implementation cost, making it suitable for the high-precision automated welding needs of various thin-walled components in aerospace, precision equipment and other fields. Attached Figure Description

[0093] Figure 1 The flowchart shows the dynamic parameter adaptive TIG welding automation method for low deformation of thin-walled workpieces according to the present invention.

[0094] Figure 2 This is a framework diagram of the dynamic parameter adaptive TIG welding automation system for low deformation of thin-walled workpieces according to the present invention. Detailed Implementation

[0095] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.

[0096] Example 1: This invention provides an automated method for low-deformation dynamic parameter adaptive TIG welding of thin-walled workpieces, such as... Figure 1As shown, it includes the following steps:

[0097] S1. Construct a multi-source heterogeneous sensing fusion system, integrating a molten pool vision sensor, an arc sensing module, and a laser positioning module to collect multi-source heterogeneous information during the welding process of thin-walled workpieces, and achieve full-domain high-fidelity perception of the welding status.

[0098] S2. Establish a digital twin model for deformation prediction driven by a combination of data and mechanism, and integrate finite element thermo-mechanical coupling analysis with machine learning algorithms to achieve online advanced prediction of welding deformation and pre-optimization of process parameters.

[0099] S3. Design a deep reinforcement learning and adaptive PID composite controller. Based on real-time sensing information and deformation prediction results, dynamically adjust key welding parameters such as pulse frequency, welding speed and heat input to form a closed-loop adaptive control architecture of "sensing-prediction-decision-execution".

[0100] S4 integrates welding heat sink and flexible clamping auxiliary methods to suppress residual stress and out-of-plane deformation in real time during the welding process, and realize the whole process collaborative control of welding deformation of thin-walled workpieces.

[0101] As shown above, this method achieves high-fidelity perception of the welding state across the entire domain by constructing a multi-heterogeneous sensing fusion system. It is combined with a deformation prediction digital twin model driven by a hybrid data and mechanism approach to complete online advance prediction of welding deformation. Relying on deep reinforcement learning and an adaptive PID composite controller, a closed-loop adaptive control architecture of "perception-prediction-decision-execution" with a total response delay of less than 100ms is built. This architecture can adapt to the real-time fluctuations in the thermo-mechanical state during the welding process of thin-walled workpieces and effectively cope with random disturbances such as workpiece clamping deviation and bevel gap fluctuation. Furthermore, by combining the collaborative auxiliary means of welding heat sink and flexible clamping, it synchronously matches the real-time heat input and suppresses residual stress and out-of-plane deformation in real time. This completely eliminates the limitations of traditional offline solidification welding parameters and passive rigid clamping fixtures, significantly reduces the proportion of post-weld straightening processes for thin-walled workpieces, and significantly improves the welding forming accuracy and production efficiency of thin-walled components.

[0102] Specifically, the implementation method of the multi-heterogeneous sensing fusion system described in step S1 is as follows:

[0103] S1.1, The molten pool vision sensor uses a high-speed CMOS camera with a frame rate of... Acquire molten pool morphology images at ≥2000fps and extract molten pool length. ,width and area Geometric features;

[0104] S1.2, Arc sensing module collects arc voltage in real time. With welding current Signal, calculate instantaneous arc power and arc impedance ;

[0105] S1.3 The laser positioning module uses the structured light triangulation principle to obtain the three-dimensional contour of the weld and extract the weld deviation. Gaps with bevel ;

[0106] S1.4. An improved adaptive weighted multi-source information fusion algorithm is used to perform spatiotemporal alignment and fusion of the above-mentioned multi-source heterogeneous data. The multi-source information fusion vector is defined as:

[0107] ;

[0108] in, For the first Sensor-like devices at all times The original feature vector, For normalized mapping functions, The adaptive weight coefficients are dynamically calculated using the following attention mechanism:

[0109] ;

[0110] ;

[0111] in, For the first Hidden layer representation of sensor-like features v For trainable attention parameters, The temperature coefficient is used; when the signal-to-noise ratio of a certain type of sensor signal is lower than a preset threshold, its corresponding weight automatically decays to near zero, ensuring the robustness of the fusion result.

[0112] As can be seen from the above, this multi-dimensional heterogeneous sensor fusion system completes multi-dimensional welding information acquisition through the complementary advantages of three types of sensors: the molten pool vision sensor accurately captures the dynamically changing molten pool morphology with a high frame rate of no less than 2000fps using a high-speed CMOS camera, stably outputting core geometric features such as the length, width, and area of ​​the molten pool, and reproducing the real-time dynamic behavior of the molten pool; the arc sensing module continuously and synchronously acquires arc voltage and welding current signals, calculates instantaneous arc power and arc impedance in real time, and accurately feeds back the real-time working status of the arc; the laser positioning module accurately obtains the three-dimensional contour of the weld seam based on the mature structured light triangulation principle, and reliably extracts weld seam deviations. By measuring the weld bevel gap and accurately identifying spatial morphological deviations, an improved adaptive weighted multi-source information fusion algorithm with an attention mechanism is used to complete the spatiotemporal alignment of all heterogeneous multi-source data. The attention mechanism can dynamically calculate the adaptive weights of each sensor. When the signal-to-noise ratio of a certain type of sensor signal is lower than a preset threshold, its corresponding weight will automatically decay to near zero, eliminating invalid and distorted sensor data interference from the root. This ultimately achieves full-domain high-fidelity perception of the entire welding process, providing a robust and highly confident raw data foundation for subsequent deformation prediction models and closed-loop control systems, and avoiding deviations in subsequent decisions due to sensor data distortion.

[0113] Specifically, the construction method of the data and mechanism hybrid-driven deformation prediction digital twin model in step S2 is as follows:

[0114] S2.1 Establish a thermo-mechanical coupled finite element model for welding thin-walled workpieces. A double ellipsoidal heat source model is used to describe the heat source input, and the heat source power density distribution is as follows:

[0115] ;

[0116] in, For effective heat input, a, b, and c are the three semi-axial parameters of the ellipsoidal heat source, respectively, and are related to the welding speed. and sheet thickness Related:

[0117] ;

[0118] in, Thermal conductivity, Characteristic temperature difference, , , This is an empirical correction factor;

[0119] S2.2. Based on finite element simulation, multiple sets of temperature and deformation field datasets under various working conditions are generated. The Physical Information Neural Network (PINN) is used for data augmentation and accelerated solution. Its loss function is defined as:

[0120] ;

[0121] ;

[0122] in, For the residual loss of the heat conduction equation, For the fitting loss of finite element simulation data, For boundary condition constraint loss, , , For balancing weighting coefficients;

[0123] S2.3 Construct a data-mechanism hybrid prediction model, embedding the finite element analytical solution as a physical prior into a Long Short-Term Memory (LSTM) network to predict time steps. Deformation amount:

[0124] ;

[0125] in, The deformation components are calculated in real time for the finite element model. This is the nonlinear compensation amount output by the residual correction network.

[0126] As shown above, this data-mechanism hybrid-driven deformation prediction digital twin model first establishes a thermo-mechanical coupled finite element model adapted to the characteristics of thin-walled workpieces. It uses the industry-standard double-ellipsoidal heat source model to accurately describe the asymmetric three-dimensional energy distribution of the arc heat source, directly linking the three semi-axis parameters of the ellipsoidal heat source to welding speed and plate thickness. This not only restores the dynamic changes of the heat source in real welding scenarios but also avoids the simulation accuracy deviations caused by traditional fixed heat source parameters. Next, based on a large dataset of temperature and deformation fields generated from multi-condition finite element simulations, a Physical Information Neural Network (PINN) is introduced for data augmentation and accelerated solving. The network is trained using a multi-objective loss function composed of residual loss from the heat conduction equation, fitting loss from finite element simulation data, and boundary condition constraint loss. This significantly reduces the computational cost of traditional pure finite element simulations. While reducing computational time, the system avoids extrapolation errors that do not conform to the physical laws of welding by constraining the network through physical equations, achieving high-precision and rapid solutions for the welding thermo-mechanical evolution process. Finally, the finite element analytical solution is embedded as a physical prior into a Long Short-Term Memory (LSTM) network to construct a hybrid prediction model. This model not only ensures the physical rationality of the prediction results by leveraging the mechanistic analysis results of the finite element model, but also corrects the nonlinear errors in the welding process that are difficult for pure mechanistic models to capture through the nonlinear compensation output of the residual correction network. This accurately predicts the welding deformation at the next moment, completely solving the pain points of strong prediction lag in traditional pure mechanistic finite element models and poor generalization of pure data-driven models. It enables online advance prediction of welding deformation and pre-optimization of process parameters, providing high-precision deformation prediction support for the subsequent closed-loop control system and effectively reducing the risk of excessive deformation during the welding of thin-walled workpieces.

[0127] Specifically, the design method of the deep reinforcement learning and adaptive PID composite controller in step S3 is as follows:

[0128] S3.1 Model welding parameter control as a Markov Decision Process (MDP) and define the state space. ,in, This represents the current deformation error;

[0129] S3.2, The motion space is defined as the welding parameter adjustment vector. These correspond to the pulse frequency increment, welding speed increment, and heat input increment, respectively.

[0130] S3.3. Train the deep reinforcement learning policy network using the Proximal Policy Optimization (PPO) algorithm. Its objective function is:

[0131] ;

[0132] in, For generalized advantage estimation, ϵ is the pruning parameter;

[0133] S3.4 The adaptive PID controller receives the coarse-tuning instructions from deep reinforcement learning and then performs fine-tuning. Its control law is as follows:

[0134] ;

[0135] The PID parameters are self-tuned in real time based on the nonlinear characteristics of the welding process.

[0136] , ;

[0137] in, This is a stability index for the welding process, obtained by normalizing the standard deviation of the arc signal. , , , This is the adaptive gain coefficient.

[0138] As shown above, this deep reinforcement learning and adaptive PID composite controller models the welding parameter control process as a Markov decision process, incorporating multi-source sensing features, advanced deformation prediction, and real-time deformation error into the state space. This allows the control decision to simultaneously consider the current welding state and subsequent deformation trends. The corresponding action space directly and accurately covers the incremental adjustment of three core welding parameters: pulse frequency, welding speed, and heat input. It can perform targeted control of welding heat input and process posture as needed. The controller uses the industry-mature Proximal Policy Optimization (PPO) algorithm for policy network training. Relying on the objective function under pruning constraints, it effectively balances training stability and policy exploration space. This avoids the policy oscillation problem that easily occurs in traditional deep reinforcement learning training and can converge quickly under complex nonlinear welding conditions. The system employs a highly adaptable coarse-tuning strategy to output the globally optimal parameter control direction. Subsequently, an adaptive PID controller takes over the coarse-tuning instructions to complete the final high-precision fine-tuning. The PID's proportional, integral, and derivative control parameters can be dynamically self-tuned in real time, completely following the welding stability index obtained by normalizing the arc signal. This perfectly adapts to the nonlinear characteristics of thin-walled welding processes that continuously change with the working conditions. It solves the pain points of traditional pure PID control in the face of highly nonlinear welding scenarios, such as parameter fixation, large overshoot, and slow dynamic response. Furthermore, relying on a composite control architecture, it balances global optimal decision-making capabilities with millisecond-level high-precision execution tracking capabilities. This provides a highly reliable and fast-response control core support for the entire "perception-prediction-decision-execution" closed-loop system, ensuring the deformation control accuracy and process stability of thin-walled workpiece welding.

[0139] Specifically, the specific implementation of the heat sink during welding in step S4 is as follows:

[0140] S4.1 A pre-set cooling channel is installed on the back side of the thin-walled workpiece during welding. Cooling medium circulates within the channel. The distance between the cooling channel and the weld toe is... Keep it within the range of 3-8mm;

[0141] S4.2 Based on the finite element pre-analysis results, a cooling channel layout optimization model is established with the goal of minimizing out-of-plane deformation:

[0142] ;

[0143] ;

[0144] in, For the out-of-plane displacement of each monitoring point, Given a vector of cooling channel locations and flow parameters, constraints are set to ensure that cooling does not lead to overcooling and cracking.

[0145] S4.3 The cooling medium flow rate is dynamically adjusted based on real-time heat input:

[0146] ;

[0147] in, As the baseline flow rate, This is the flow regulation coefficient. This is the reference thermal input power.

[0148] As can be seen from the above, this on-the-fly heat sink implementation method pre-sets a cooling channel on the back side of the thin-walled workpiece with a weld toe spacing strictly controlled at 3-8mm. This can form a strong cooling zone immediately behind the heat-affected zone of the weld without interfering with the normal operation of the welding arc on the front side. This effectively removes excess heat accumulated during the welding process, significantly reducing the local high-temperature dwell time and the range of high-temperature areas on the workpiece. The precise arrangement of the channel positions is achieved through a cooling channel layout optimization model aimed at minimizing out-of-plane deformation. At the same time, constraints are set to strictly avoid the risk of welding cold cracks caused by over-cooling. Finally, the flow rate of the cooling medium is dynamically adjusted based on the real-time heat input signal, so that the cooling intensity can be perfectly matched with the fluctuation of the real-time welding heat input. This avoids the problem of mismatch between cooling capacity and heat input in the traditional fixed flow cooling mode. It can effectively suppress the accumulation of residual stress and out-of-plane deformation during the welding process of thin-walled workpieces. Together with the closed-loop parameter adaptive control system mentioned above, it forms a two-way synergistic matching of heat input and heat output, further controlling the welding deformation of thin-walled workpieces to an extremely low level and significantly reducing the workload of subsequent post-weld correction.

[0149] Specifically, the flexible pressing method in step S4 is implemented as follows:

[0150] S4.4 Multiple flexible clamping devices are arranged on both sides of the welding area of ​​the thin-walled workpiece, and the pressure at each clamping point is... Independently adjustable;

[0151] S4.5. Establish a coupling relationship model between clamping force and welding deformation, and use the response surface methodology to establish a proxy model for the clamping force parameters and the maximum deformation:

[0152] ;

[0153] in, For each clamping point, the force vector is... These are polynomial basis functions;

[0154] S4.6. Real-time optimization of pressure at each clamping point based on online gradient descent method:

[0155] ;

[0156] in, To optimize the step size of the clamping force, the gradient is obtained by analytical differentiation of the surrogate model, while constraining the pressure at each point. To avoid crushing thin-walled workpieces.

[0157] As shown above, this flexible clamping device arranges multiple independently adjustable clamping points on both sides of the welding area. A proxy model of the clamping force parameters and maximum deformation is constructed using the response surface methodology, accurately mapping the complex nonlinear coupling relationship between the clamping force at multiple points and the welding deformation of the thin-walled workpiece. Real-time dynamic optimization of the clamping force at each point is performed using an online gradient descent algorithm. The gradient direction can be quickly obtained by analytical differentiation directly through the proxy model. Under the premise of strictly limiting the upper and lower limits of the pressure at each point and avoiding local crushing damage to the thin-walled workpiece, the distribution and magnitude of the clamping force can adapt to the constantly changing thermal stress and deformation trend during the welding process. Precise flexible constraints are applied to areas prone to instability and warping during welding, directly suppressing out-of-plane instability deformation of the thin-walled component from a mechanical perspective. This synergizes with the thermal regulation effect of the welding heat sink, further enhancing the deformation suppression effect throughout the entire process. It can adapt to the clamping requirements of thin-walled workpieces of different specifications and completely overcomes the shortcomings of traditional rigid integral clamping fixtures, such as poor adaptability and easy local stress concentration, significantly improving the dimensional and positional stability of thin-walled workpieces during welding.

[0158] Specifically, the deformation prediction digital twin model in step S2 also includes an online incremental learning mechanism, specifically:

[0159] When the actual welding deformation monitoring value Compared with the predicted value Deviation exceeds tolerance threshold At that time, the model is triggered to update online:

[0160] ;

[0161] ;

[0162] in, For online learning rate, To prevent catastrophic forgetting in the model, regularization coefficients are used. Approximate calculation is performed using backpropagation with finite samples; incrementally learned samples are stored in the experience replay pool. The pool capacity is A priority sampling strategy is adopted:

[0163] ;

[0164] in, For the first The prediction error for each sample, where ω is the priority exponent and ϵ is a small constant.

[0165] As shown above, the online incremental learning mechanism of this deformation prediction digital twin model can automatically trigger model updates when the deviation between the actual welding deformation monitoring value and the model prediction value exceeds a preset tolerance threshold. By introducing a loss function with regularization constraints to carry out small-sample gradient iterative optimization, the model parameters are fine-tuned using only a small number of newly added measured samples. At the same time, the parameter update magnitude is constrained by the regularization term, which fundamentally avoids the catastrophic forgetting problem when the model adapts to new working conditions, ensuring that the prediction accuracy of the mature working conditions already trained will not decline significantly. The fixed-capacity experience replay pool adopts a priority sampling strategy, prioritizing the selection of high-value samples with larger prediction errors for subsequent iterative training, which greatly improves the utilization efficiency of new samples. It can not only adapt the model to complex on-site disturbances such as clamping deviations of different batches of workpieces and material property fluctuations without relying on massive labeled data, but also continuously improve the prediction accuracy and generalization ability of the model in the long-term industrial application process. This ensures that the digital twin model always fits the real welding working conditions on site, and continuously outputs high-confidence advanced deformation prediction results for the subsequent closed-loop control system.

[0166] Specifically, in step S3, the deep reinforcement learning policy network adopts a multi-objective reward function design to balance deformation control accuracy and welding efficiency.

[0167] ;

[0168] in, For the squared term of deformation error, This is a penalty term for changes in heat input. This is a penalty for incremental welding time. This is a welding safety constraint indication function, which takes the value -1 when the arc parameters exceed the safe range, and 0 otherwise. , , , These are multi-objective weighting coefficients;

[0169] Simultaneously, a course learning strategy is introduced, with a relatively high tolerance for deformation set in the early stages of training. With each training round attenuation:

[0170] ;

[0171] in, To determine the course decay rate, the strategy network is gradually guided from coarse regulation to fine regulation.

[0172] As shown above, the deep reinforcement learning strategy network employs a multi-objective reward function design. It prioritizes the welding positional accuracy of thin-walled workpieces through the squared deformation error term, while a heat input change penalty term suppresses drastic parameter fluctuations to prevent process instability. The welding time increment penalty term guides the strategy towards higher welding efficiency. Furthermore, a safety constraint indicator function strictly limits the arc parameters within the process safety range. Ultimately, by adjusting the multi-objective weight coefficients, it can flexibly adapt to the differentiated requirements for deformation control, welding efficiency, and process stability under different working conditions. The accompanying course learning strategy sets a large deformation tolerance in the early training stages, then gradually narrows the accuracy requirements exponentially with each training round. This progressively guides the strategy network from coarse-grained working condition adaptation and control to micron-level high-precision fine-tuning. This significantly reduces the training difficulty in complex multi-objective scenarios, effectively avoiding the problem of strategy convergence difficulties caused by excessively high accuracy requirements in the early training stages. It also ensures that the final trained control strategy simultaneously considers low deformation control of thin-walled workpieces, welding production efficiency, and overall process safety in actual production, further enhancing the comprehensive implementation performance of the entire closed-loop adaptive system.

[0173] Example 2: This example is basically the same as the previous example, except that the method also includes an online construction and dynamic update mechanism for the safety domain of welding process parameter constraints:

[0174] Based on real-time acquired arc characteristics, a safe operating domain for the welding process is constructed. :

[0175] ;

[0176] The constraint functions are defined as follows:

[0177] ,

[0178] ;

[0179] Deep reinforcement learning action outputs are corrected by projection over the safe region:

[0180] ;

[0181] Ensure that all control commands are always within the safe operating range to prevent defects such as burn-through and lack of fusion.

[0182] As can be seen from the above, the online construction and dynamic update mechanism of the safety domain of the welding process parameter constraint can dynamically generate a multi-dimensional safety operation boundary covering the upper limit of arc current, upper limit of arc voltage, unit volume heat input threshold, and peak temperature critical value based on the real-time collected arc characteristics. It includes all dangerous parameter ranges that are prone to defects such as burn-through, lack of fusion, and coarse grains within the constraint range. At the same time, the original actions output by deep reinforcement learning are verified and bounded in real time through the safety domain projection correction link, ensuring that all issued welding parameter control commands are always within the dynamically updated safety operation domain. This not only solves the pain point that the traditional fixed process window cannot adapt to the dynamic changes in working conditions during thin-walled welding, but also builds a safety protection barrier for the entire adaptive closed-loop control system from the underlying architecture level. It completely avoids welding defects caused by intelligent algorithms outputting parameters that exceed the limits under extreme working conditions, allowing the low deformation adaptive welding process of thin-walled workpieces to maintain extremely high process stability and safety under all working conditions.

[0183] Example 3: This example is basically the same as the previous example, except that the "perception-prediction-decision-execution" closed-loop architecture adopts a hierarchical control strategy:

[0184] The upper layer is the digital twin prediction layer, based on the prediction cycle. =50-200ms runtime deformation advance prediction, output pre-optimized parameter set ;

[0185] The middle layer is the deep reinforcement learning decision layer, based on the decision cycle. =20-50ms Parameters are adjusted based on real-time status, and control commands are output. ;

[0186] The lower layer is the adaptive PID execution layer, based on the execution cycle. Fine-tuning of parameters and servo tracking are performed within 5-10ms.

[0187] The three layers communicate via status information. With motion information The asynchronous message passing mechanism is coupled, and its overall closed-loop transfer function is approximately as follows:

[0188] ;

[0189] in, For the adaptive PID controller transfer function, For the equivalent transfer function of deep reinforcement learning strategies, For the delayed link in digital twin prediction, It is a multi-source sensing feedback loop; the collaborative work of each layer keeps the total system response delay within 100ms, meeting the real-time control requirements under the rapid thermal response characteristics of thin-walled workpieces.

[0190] As can be seen from the above, the hierarchical control strategy adopted by this "perception-prediction-decision-execution" closed-loop architecture achieves precise division of labor and efficient collaboration among the three layers by setting differentiated operating cycles: the upper-layer digital twin prediction layer completes the advanced prediction of welding deformation and pre-optimization of process parameters with a prediction cycle of 50-200ms, locking in a reasonable parameter control range for the system in advance, and fundamentally avoiding the risk of process instability caused by large parameter jumps; the middle-layer deep reinforcement learning decision layer completes the global optimal parameter dynamic adjustment based on the real-time welding status with a decision cycle of 20-50ms, and outputs precise control commands adapted to the current working conditions; the lower-layer adaptive PID execution layer completes the global optimal parameter dynamic adjustment based on the real-time welding status with a decision cycle of 5- With a high execution cycle of 10ms, the system completes millisecond-level fine-tuning of parameters and servo tracking, ensuring the rapid execution of control commands. The three layers rely on an asynchronous message passing mechanism to achieve efficient coupling of state and action information, ultimately strictly controlling the overall closed-loop response delay of the system to within 100ms. This perfectly adapts to the characteristics of fast thermal response and strong nonlinear deformation evolution in the welding process of thin-walled workpieces. It not only solves the industry pain point that a single controller cannot simultaneously take into account global optimization capabilities and millisecond-level response speed, but also significantly improves the dynamic control accuracy and working condition adaptability of the entire adaptive welding system, providing a highly reliable closed-loop control architecture to support high-precision, low-deformation welding of thin-walled workpieces.

[0191] Application example: TIG welding verification of thin-walled aluminum alloy tank shells

[0192] To verify the practical effect of the method proposed in this invention, a typical thin-walled workpiece—a certain type of aerospace aluminum alloy tank shell (material: 6061-T6, wall thickness 1.2 mm, diameter φ300 mm, length 400 mm)—was selected for butt welding TIG welding tests. The workpiece used a V-groove with a 60° bevel angle and a 0.5 mm blunt edge. The assembly gap fluctuated randomly within the range of 0.2–0.8 mm. ER4043 welding wire with a diameter of 1.0 mm was used as the welding base material. The test was conducted on a six-axis robot platform equipped with a digital TIG welding power supply (maximum output current 300 A). The system integrated a high-speed CMOS camera (2000 fps), an arc voltage / current acquisition module (sampling rate 10 kHz), and a laser structured light positioning sensor (resolution 0.02 mm). The cooling channel used during welding was a copper water-cooled block (channel distance 5 mm from weld toe). The flexible clamping device consisted of eight independently servo-controlled cylinders (pressure range 0–200 N).

[0193] Three comparative experiments were set up (each group was repeated 5 times and the mean was taken):

[0194] Control group 1: Traditional offline curing parameters (pulse frequency 60 Hz, welding speed 120 mm / min, heat input reference value) + rigid integral clamping fixture, without sensor feedback and parameter self-adaptation.

[0195] Control group 2: The method of this invention uses a multi-sensor fusion + digital twin prediction + reinforcement learning and PID composite controller (i.e. closed-loop adaptive control), but does not enable the welding heat sink and flexible clamping auxiliary modules.

[0196] Experimental group: The complete method of this invention (closed-loop adaptive control + heat sink during welding + flexible clamping online optimization).

[0197] All tests were conducted on the same batch of workpieces under the same ambient temperature. During the welding process, the arc voltage, current, molten pool image, weld tracking deviation, and temperature field on the back of the workpiece were recorded in real time (infrared thermal imager). After welding, a three-dimensional optical scanner (accuracy 0.02 mm) was used to measure the out-of-plane deformation distribution, and samples were taken for X-ray flaw detection and metallographic observation. The main verification indicators included: maximum out-of-plane deformation (taking the maximum absolute value of displacement from 10 equidistant sections along the weld length), angular deformation (height difference at 20 mm on both sides of the weld center), welding time (time to complete the entire circumferential weld), and weld formation qualification rate (the percentage of weld lengths with no surface cracks, lack of fusion, porosity, or other defects and a weld reinforcement height ≤ 1.0 mm, according to GJB 1718A-2005 standard).

[0198] The verification results are shown in the table below:

[0199] Control group 1 1.87 ± 0.23 2.14 ± 0.31 186 ± 12 72.3 45 min Control group 2 0.92 ± 0.11 1.02 ± 0.15 165 ± 8 88.6 15 min experimental group 0.41 ± 0.06 0.48 ± 0.07 152 ± 6 96.2 5 min

[0200] As shown in the table above, control group 1, due to its fixed parameters and rigid clamping, could not adapt to the fluctuations in assembly gaps and the effects of heat accumulation. The maximum out-of-plane deformation was close to 1.9 mm, and the angular deformation exceeded 2°, requiring extensive post-weld straightening, resulting in a pass rate of only 72.3%. While control group 2 reduced the deformation to 0.92 mm and the angular deformation to 1.02° through closed-loop adaptive control, improving welding efficiency by approximately 11%, it lacked synergistic suppression through auxiliary means. The out-of-plane deformation still exceeded the typical requirement of ≤0.5 mm for aerospace thin-walled structural components, and there were minor warpings caused by residual stress in some areas. In contrast, the test group using the complete solution of this invention achieved a maximum out-of-plane deformation of only 0.41 mm and an angular deformation of less than 0.5°, meeting the requirements for high-precision assembly. The welding time was reduced by 18.3% compared to control group 1, the weld formation pass rate increased to 96.2%, and the post-weld straightening time was drastically reduced to 5 min / piece, significantly lowering subsequent manufacturing costs. Furthermore, during the welding process, when the groove gap suddenly increased to 0.7 mm, the digital twin model predicted the deformation increase trend 1.2 seconds in advance. The reinforcement learning controller immediately increased the pulse frequency (from 60 Hz to 78 Hz) and slightly reduced the welding speed (from 120 mm / min to 108 mm / min). At the same time, the welding cooling flow rate dynamically increased by 12%, and the pressure of the flexible clamping point was redistributed online, successfully suppressing the deformation surge caused by the transient disturbance. There were no burn-through or non-fusion defects throughout the process, verifying the robustness of the method of the present invention to random disturbances and the effectiveness of the whole process collaborative control.

[0201] A dynamic parameter adaptive TIG welding automation system for low deformation of thin-walled workpieces, such as... Figure 2 As shown, the dynamic parameter adaptive TIG welding automation method suitable for low deformation of the above-mentioned thin-walled workpieces includes:

[0202] The multi-source heterogeneous sensing fusion system integrates a molten pool vision sensor, an arc sensing module, and a laser positioning module to collect multi-source heterogeneous information during the welding process of thin-walled workpieces, achieving full-domain high-fidelity perception of the welding status.

[0203] A data- and mechanism-driven deformation prediction digital twin model is used to integrate finite element thermo-mechanical coupling analysis with machine learning algorithms to achieve online advanced prediction of welding deformation and pre-optimization of process parameters.

[0204] A deep reinforcement learning and adaptive PID composite controller is used to dynamically adjust key welding parameters such as pulse frequency, welding speed and heat input based on real-time sensing information and deformation prediction results, forming a closed-loop adaptive control architecture of "sensing-prediction-decision-execution".

[0205] The welding heat sink and flexible clamping auxiliary module is used to suppress residual stress and out-of-plane deformation in real time during the welding process, so as to realize the whole-process coordinated control of welding deformation of thin-walled workpieces.

[0206] As can be seen from the above, the automated system adapted to the aforementioned low-deformation adaptive TIG welding method achieves high-fidelity perception of the entire welding process through a multi-dimensional heterogeneous system integrating three types of sensing units. It completes deformation prediction and parameter pre-optimization based on a digital twin model driven by a hybrid data mechanism. Combined with a DRL and adaptive PID composite controller, it constructs a closed-loop control link with millisecond-level response. Furthermore, it combines a welding heat sink and a flexible clamping module to achieve thermal-mechanical synergistic deformation suppression throughout the entire process. It can complete the entire link response within 100ms, adapting to the rapid thermal response characteristics of thin-walled workpieces, significantly reducing residual stress and out-of-plane deformation after welding, obtaining high-quality welds with high dimensional accuracy and good consistency, significantly reducing the proportion of subsequent straightening processes, and improving the automated welding production efficiency and process stability of thin-walled components.

[0207] This application provides an electronic device applicable to the aforementioned automated TIG welding method for low deformation of thin-walled workpieces using dynamic parameters, comprising:

[0208] Memory, used to store computer programs and data;

[0209] A processor is used to run system programs.

[0210] This application provides a computer storage medium applicable to the aforementioned automated TIG welding method for low deformation of thin-walled workpieces using dynamic parameters, and implements hierarchical confidentiality management of the system and data in accordance with confidentiality management requirements.

[0211] Those skilled in the art will understand that embodiments of this application can be provided as a system or a computer program product. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0212] This application is described with reference to flowchart illustrations and / or block diagrams of devices (systems) and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.

[0213] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0214] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0215] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0216] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0217] Computer-readable media include both permanent and non-permanent, removable and non-removable media, which can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0218] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, article, or apparatus that includes that element.

[0219] The embodiments of the present invention are given for the purposes of illustration and description. Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. An automated method for low-deformation dynamic parameter adaptive TIG welding of thin-walled workpieces, characterized in that, Includes the following steps: S1. Construct a multi-source heterogeneous sensing fusion system, integrating a molten pool vision sensor, an arc sensing module, and a laser positioning module to collect multi-source heterogeneous information during the welding process of thin-walled workpieces. S2. Establish a deformation prediction digital twin model driven by a combination of data and mechanism, integrating finite element thermo-mechanical coupling analysis with machine learning algorithms; S3. Design a deep reinforcement learning and adaptive PID composite controller. Based on real-time sensing information and deformation prediction results, dynamically adjust the pulse frequency, welding speed and heat input welding parameters to form a closed-loop adaptive control architecture of "sensing-prediction-decision-execution". S4 integrates welding heat sink and flexible clamping auxiliary means to suppress residual stress and out-of-plane deformation in real time during the welding process.

2. The automated method for low-deformation dynamic parameter adaptive TIG welding of thin-walled workpieces as described in claim 1, characterized in that: The specific implementation method of the multi-heterogeneous sensing fusion system described in step S1 is as follows: S1.1, The molten pool vision sensor uses a high-speed CMOS camera with a frame rate of... Acquire molten pool morphology images at ≥2000fps and extract molten pool length. ,width and area Geometric features; S1.2, Arc sensing module collects arc voltage in real time. With welding current Signal, calculate instantaneous arc power and arc impedance ; S1.3 The laser positioning module uses the structured light triangulation principle to obtain the three-dimensional contour of the weld and extract the weld deviation. Gaps with bevel ; S1.

4. An improved adaptive weighted multi-source information fusion algorithm is used to perform spatiotemporal alignment and fusion of the above-mentioned multi-source heterogeneous data. The multi-source information fusion vector is defined as: ； in, For the first Sensor-like devices at all times The original feature vector, For normalized mapping functions, The adaptive weight coefficients are dynamically calculated using the following attention mechanism: ； ； in, For the first Hidden layer representation of sensor-like features v For trainable attention parameters, The temperature coefficient is used; when the signal-to-noise ratio of a certain type of sensor signal is lower than a preset threshold, its corresponding weight automatically decays to near zero.

3. The automated method for low-deformation dynamic parameter adaptive TIG welding of thin-walled workpieces as described in claim 1, characterized in that: The specific construction method of the data and mechanism hybrid-driven deformation prediction digital twin model in step S2 is as follows: S2.1 Establish a thermo-mechanical coupled finite element model for welding thin-walled workpieces. A double ellipsoidal heat source model is used to describe the heat source input, and the heat source power density distribution is as follows: ； in, For effective heat input, a, b, and c are the three semi-axial parameters of the ellipsoidal heat source, respectively, and are related to the welding speed. and sheet thickness Related: ； in, Thermal conductivity, Characteristic temperature difference, , , This is an empirical correction factor; S2.

2. Based on finite element simulation, multiple sets of temperature and deformation field datasets under various working conditions are generated. The Physical Information Neural Network (PINN) is used for data augmentation and accelerated solution. Its loss function is defined as: ； ； in, For the residual loss of the heat conduction equation, For the fitting loss of finite element simulation data, For boundary condition constraint loss, , , For balancing weighting coefficients; S2.3 Construct a data-mechanism hybrid prediction model, embedding the finite element analytical solution as a physical prior into a Long Short-Term Memory (LSTM) network to predict time steps. Deformation amount: ； in, The deformation components are calculated in real time for the finite element model. This is the nonlinear compensation amount output by the residual correction network.

4. The automated method for low-deformation dynamic parameter adaptive TIG welding of thin-walled workpieces as described in claim 1, characterized in that: The specific design method of the deep reinforcement learning and adaptive PID composite controller in step S3 is as follows: S3.1 Model welding parameter control as a Markov Decision Process (MDP) and define the state space. ,in, This represents the current deformation error; S3.2, The motion space is defined as the welding parameter adjustment vector. These correspond to the pulse frequency increment, welding speed increment, and heat input increment, respectively. S3.

3. Train the deep reinforcement learning policy network using the Proximal Policy Optimization (PPO) algorithm. Its objective function is: ； in, For generalized advantage estimation, ϵ is the pruning parameter; S3.4 The adaptive PID controller receives the coarse-tuning instructions from deep reinforcement learning and then performs fine-tuning. Its control law is as follows: ； The PID parameters are self-tuned in real time based on the nonlinear characteristics of the welding process. ， ； in, This is a stability index for the welding process, obtained by normalizing the standard deviation of the arc signal. , , , This is the adaptive gain coefficient.

5. The automated method for low-deformation dynamic parameter adaptive TIG welding of thin-walled workpieces as described in claim 1, characterized in that: The specific implementation method of the heat sink during welding in step S4 is as follows: S4.1 A pre-set cooling channel is installed on the back side of the thin-walled workpiece during welding. Cooling medium circulates within the channel. The distance between the cooling channel and the weld toe is... Keep it within the range of 3-8mm; S4.2 Based on the finite element pre-analysis results, a cooling channel layout optimization model is established with the goal of minimizing out-of-plane deformation: ； ； in, For the out-of-plane displacement of each monitoring point, Given a vector of cooling channel locations and flow parameters, constraints are set to ensure that cooling does not lead to overcooling and cracking. S4.3 The cooling medium flow rate is dynamically adjusted based on real-time heat input: ； in, As the baseline flow rate, This is the flow regulation coefficient. This is the reference thermal input power.

6. The automated method for low-deformation dynamic parameter adaptive TIG welding of thin-walled workpieces as described in claim 1, characterized in that: The specific implementation method of the flexible pressing in step S4 is as follows: S4.4 Multiple flexible clamping devices are arranged on both sides of the welding area of ​​the thin-walled workpiece, and the pressure at each clamping point is... Independently adjustable; S4.

5. Establish a coupling relationship model between clamping force and welding deformation, and use the response surface methodology to establish a proxy model for the clamping force parameters and the maximum deformation: ； in, For each clamping point, the force vector is... These are polynomial basis functions; S4.

6. Real-time optimization of pressure at each clamping point based on online gradient descent method: ; in, To optimize the step size of the clamping force, the gradient is obtained by analytical differentiation of the surrogate model, while constraining the pressure at each point. To avoid crushing thin-walled workpieces.

7. The automated method for low-deformation dynamic parameter adaptive TIG welding of thin-walled workpieces as described in claim 1, characterized in that: The deformation prediction digital twin model in step S2 also includes an online incremental learning mechanism, specifically: When the actual welding deformation monitoring value Compared with the predicted value Deviation exceeds tolerance threshold At that time, the model is triggered to update online: ; ; in, For online learning rate, To prevent catastrophic forgetting in the model, regularization coefficients are used. Approximate calculation is performed using backpropagation with finite samples; incrementally learned samples are stored in the experience replay pool. The pool capacity is A priority sampling strategy is adopted: ; in, For the first The prediction error for each sample, where ω is the priority exponent and ϵ is a small constant.

8. The automated method for low-deformation dynamic parameter adaptive TIG welding of thin-walled workpieces as described in claim 1, characterized in that: In step S3, the deep reinforcement learning policy network employs a multi-objective reward function design to balance deformation control accuracy and welding efficiency. ; in, For the squared term of deformation error, This is a penalty term for changes in heat input. This is a penalty for incremental welding time. This is a welding safety constraint indication function, which takes the value -1 when the arc parameters exceed the safe range and 0 otherwise. , , , These are multi-objective weighting coefficients; Simultaneously, a course learning strategy is introduced, with a relatively high tolerance for deformation set in the early stages of training. With each training round attenuation: ; in, To determine the course decay rate, the strategy network is gradually guided from coarse regulation to fine regulation.

9. The automated method for low-deformation dynamic parameter adaptive TIG welding of thin-walled workpieces as described in claim 1, characterized in that: This method also includes an online construction and dynamic update mechanism for the safety domain of welding process parameter constraints: Based on real-time acquired arc characteristics, a safe operating domain for the welding process is constructed. : ; The constraint functions are defined as follows: , ; Deep reinforcement learning action outputs are corrected by safe region projection: ; Ensure that all control commands are always within the safe operating range.

10. The automated method for low-deformation dynamic parameter adaptive TIG welding of thin-walled workpieces as described in claim 1, characterized in that: The "perception-prediction-decision-execution" closed-loop architecture adopts a hierarchical control strategy: The upper layer is the digital twin prediction layer, based on the prediction cycle. =50-200ms runtime deformation advance prediction, output pre-optimized parameter set ; The middle layer is the deep reinforcement learning decision layer, based on the decision cycle. =20-50ms Parameters are adjusted based on real-time status, and control commands are output. ; The lower layer is the adaptive PID execution layer, based on the execution cycle. Fine-tuning of parameters and servo tracking are performed within 5-10ms. The three layers communicate via status information. With motion information The asynchronous message passing mechanism is coupled, and its overall closed-loop transfer function is approximately as follows: ; in, For the adaptive PID controller transfer function, For the equivalent transfer function of deep reinforcement learning strategies, For the delayed link in digital twin prediction, This is a multi-source sensing feedback loop; The collaborative work of each layer keeps the total system response delay within 100ms, meeting the real-time control requirements under the rapid thermal response characteristics of thin-walled workpieces.

Images