Large-diameter pipeline welding process parameter adaptive control method and system

Abstract

The present application belongs to the technical field of intelligent welding manufacturing and automatic control, and provides a large-diameter pipeline welding process parameter self-adaptive control method and system, which has the technical scheme of constructing a multi-modal data set containing bevel geometric characteristics and process parameter time sequence characteristics; obtaining welding quality prediction results based on the multi-modal data set and a trained welding quality prediction model; decoupling the welding parameter control task into an upper global planning layer and a lower real-time control layer, dynamically setting process parameter constraint spaces according to the gravity conditions of different pipeline welding positions in the upper layer, and finely adjusting real-time parameters in the constraint spaces by using a flexible action-evaluation algorithm in the lower layer, introducing an uncertainty quantification mechanism of the welding quality prediction model and welding quality prediction results into the reward function of the algorithm, guiding the agent to safely explore under unknown working conditions, and obtaining an optimized control strategy.

Description

Technical Field

[0001] This invention belongs to the field of intelligent welding manufacturing and automation control technology, and particularly relates to an adaptive control method and system for welding process parameters of large-diameter pipes. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] In the field of automated welding of large-diameter pipes, although precise trajectory motion constitutes the physical basis of automated welding, the key to realizing the intelligence of the welding process lies in elucidating the nonlinear laws behind the complex welding process and enabling the system to dynamically adjust process parameters in response to environmental changes.

[0004] Current automated welding technology still faces significant technical bottlenecks in the practical application of large-diameter pipelines.

[0005] Existing single-layer control structures struggle to resolve cross-scale contradictions within the same parameter system, often leading to welding defects caused by gravity at different angles of the welding torch, lacking adaptive hierarchical constraints for different spatial locations. Predictive models only address the "state awareness" problem; how to make optimal decisions based on prediction results to achieve closed-loop control remains a major challenge. Existing technologies often employ linear PID control strategies or experience-based lookup table methods. However, given the inherent high-dimensional continuous motion space and significant dynamic time-varying characteristics of welding systems, these traditional methods struggle to simultaneously satisfy the dual constraints of real-time response and global optimality when dealing with such strongly coupled nonlinear systems. Circumferential welds in large-diameter pipes are typically long, representing a typical time-varying thermal process. Existing methods often neglect the heat accumulation effect of long-term welding, resulting in drastically different weld depths for the same process parameters in the initial and final arc stages. Existing neural network prediction models typically provide deterministic point predictions, failing to quantify the model's inherent uncertainty. When encountering extreme conditions not covered by training data, the system is prone to blindly and confidently making erroneous decisions, lacking a safe exploration mechanism based on uncertainty to guide the model to adapt to unknown environments. Summary of the Invention

[0006] To address at least one of the technical problems mentioned above, this invention provides an adaptive control method and system for welding process parameters of large-diameter pipes, which achieves accurate modeling of complex welding processes and dynamic optimization of process parameters.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of the present invention provides an adaptive control method for welding process parameters of large-diameter pipes, comprising the following steps: Simultaneously acquire line laser vision information and time-series data of process parameters during the welding process of large-diameter pipes, perform preprocessing, and construct a multimodal dataset containing bevel geometry features and time-series features of process parameters; Welding quality prediction results are obtained based on the multimodal dataset and the trained welding quality prediction model; The welding parameter control task is decoupled into an upper global planning layer and a lower real-time control layer. The upper layer dynamically sets the process parameter constraint space according to the gravity conditions of different welding positions on the pipeline, while the lower layer uses a flexible action-evaluation algorithm to finely adjust the parameters in real time within the constraint space. The uncertainty quantification mechanism of the welding quality prediction model and the welding quality prediction results are introduced into the reward function of the algorithm to guide the agent to conduct safe exploration under unknown working conditions and obtain an optimized control strategy.

[0008] Furthermore, the welding quality prediction model adopts a multi-branch hybrid neural network architecture, including a parallel long short-term memory network and a convolutional neural network. The time series features of the process parameters are extracted based on the long short-term memory network branch, and the spatial correlation features of the bevel geometry are extracted based on the CNN branch. The two features are dynamically coupled through an adaptive fusion layer to obtain a fused high-dimensional feature vector. The fused high-dimensional feature vector is then fed into the subsequent output layer to obtain the quality prediction index.

[0009] Furthermore, during the training of the welding quality prediction model, the hyperparameters of the multi-branch hybrid neural network are globally optimized based on the Grey Wolf optimization algorithm, specifically including: The hyperparameters of the neural network to be optimized are defined as the spatial location vectors of individual gray wolves. ; Using the mean squared error of the welding quality prediction model on the validation set as the fitness function, the population position is iteratively updated by simulating the social hierarchy and encirclement hunting behavior of gray wolf populations, and the global optimal hyperparameter combination is searched. The optimal combination of hyperparameters will be output to construct the final welding quality prediction model.

[0010] Furthermore, the upper layer dynamically sets the process parameter constraint space according to the gravity conditions of different welding positions on the pipeline, including: defining the welding parameter optimization problem as a Markov decision process, and defining the action feasible domain and multi-objective optimization weight of the lower real-time control layer according to the current circumferential angle of the pipeline where the welding torch is located; when the spatial position change exceeds the set threshold or the line laser sensor detects a sudden change in the bevel size, the upper layer is triggered to reset the reference parameters and allowable adjustment range as the constraint boundary of the lower real-time control layer.

[0011] Furthermore, within the allowable adjustment range, the lower real-time control layer outputs specific changes in wire feeding speed, forming an action space. This action space is a vector of incremental adjustment values ​​for process parameters, including changes in welding current, welding voltage, wire feeding speed, welding speed, oscillation amplitude, oscillation frequency, and side-stop time.

[0012] Furthermore, the reward function is: , in, Indicates the state at the next moment. The predicted quality index vector, This represents the preset ideal quality target vector. For boundary penalty terms, It is a penalty for the smoothness of the movement range, to avoid drastic parameter fluctuations; The uncertainty of the model is quantified by using the variance of the predicted output values ​​for the same state through an integrated LSTM network. , , , These are the weighting coefficients.

[0013] Furthermore, when guiding the agent to conduct safe exploration under unknown conditions, the experience is stored in a buffer, data batches are randomly sampled from the buffer, the Critic network is updated to minimize the soft Bellman residual, the Actor network is updated to maximize the Q-value objective of entropy regularization, and the Target network parameters are updated through a soft update mechanism.

[0014] A second aspect of the present invention provides an adaptive control system for welding process parameters of large-diameter pipes, comprising: The dataset construction module is used to synchronously acquire line laser vision information and process parameter time series data during the welding process of large-diameter pipes, and perform preprocessing to construct a multimodal dataset containing bevel geometry features and process parameter time series features. The welding quality prediction module is used to obtain welding quality prediction results based on the multimodal dataset and the trained welding quality prediction model. The adaptive control module decouples the welding parameter control task into an upper global planning layer and a lower real-time control layer. The upper layer dynamically sets the process parameter constraint space based on the gravity conditions of different welding positions on the pipeline, while the lower layer uses a flexible action-evaluation algorithm to finely adjust the parameters in real time within the constraint space. The uncertainty quantification mechanism of the welding quality prediction model and the welding quality prediction results are introduced into the reward function of the algorithm to guide the agent to safely explore under unknown working conditions and obtain an optimized control strategy.

[0015] A third aspect of the present invention provides a computer-readable storage medium.

[0016] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the adaptive control method for welding process parameters of large-diameter pipes as described above.

[0017] A fourth aspect of the present invention provides a computer device.

[0018] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the adaptive control method for welding process parameters of large-diameter pipes as described above.

[0019] Compared with the prior art, the beneficial effects of the present invention are: This invention addresses the dramatic changes in the stress state of the molten pool and the heat accumulation effect caused by long welds in all-position welding of large-diameter pipelines. It dynamically adjusts the control strategy at different positions through a layered constraint mechanism and uses model uncertainty to guide the agent to adapt to unknown working conditions. This effectively eliminates defects such as concave welding in overhead welding and flow in vertical welding, and achieves high-quality, adaptive welding of circumferential welds in all positions of pipelines, significantly improving the consistency of welding quality.

[0020] This invention introduces a line laser sensor to fuse bevel geometry features with time-series process parameters using multimodal features, overcoming the limitations of traditional methods that rely solely on process parameters. A multi-branch hybrid neural network can automatically extract the impact features of geometric mutations on the welding process, significantly improving prediction accuracy under complex conditions. Simultaneously, by combining the Grey Wolf optimization algorithm to adaptively optimize the hyperparameters of the hybrid network, it solves the problem that manual trial and error struggles to find the global optimum in a non-convex search space.

[0021] This invention combines a flexible motion-evaluation-based adaptive control strategy to establish a hierarchical and multi-objective optimization decision-making system. The upper-level planning addresses the process window drift problem under gravity, while the lower-level decision-making is responsible for real-time parameter optimization and adjustment. Combined with a multi-objective reward function, it not only achieves optimal welding quality but also maximizes production efficiency while ensuring process stability.

[0022] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0023] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0024] Figure 1 This is a flowchart of the adaptive control method for welding process parameters of large-diameter pipes provided in an embodiment of the present invention; Figure 2 This is a diagram of the adaptive control system architecture for welding process parameters of large-diameter pipes provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the LSTM unit structure provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the GWO algorithm optimization process provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the SAC algorithm provided in an embodiment of the present invention. Detailed Implementation

[0025] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0026] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0027] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0028] To address the challenges of multivariate strong coupling, nonlinearity, and significant spatial gravity influences in large-diameter pipeline welding, existing control methods struggle to balance multimodal perception capabilities with multi-objective global optimal decision-making under complex conditions. This results in poor weld quality consistency, low welding efficiency, and high uncertainty. This invention provides a welding quality prediction and adaptive control method based on multimodal information fusion and hierarchical reinforcement learning, enabling accurate modeling of complex welding processes and dynamic optimization of process parameters. First, a multi-branch hybrid neural network is constructed based on line laser vision information and process parameters. This network extracts temporal parameter features through a long short-term memory network, extracts bevel geometry features in parallel using a convolutional neural network, and achieves self-extraction of multimodal information features through an adaptive fusion layer. Based on the optimal hyperparameters output by the Grey Wolf optimization algorithm, a high-precision welding quality prediction model is established. Furthermore, this invention establishes a hierarchical and multi-objective optimization decision-making system, decoupling the control task into an upper global planning layer and a lower real-time control layer. The upper layer dynamically sets the process parameter constraint space based on the gravity conditions at different welding positions on the pipeline; the lower layer uses a flexible motion-evaluation algorithm to finely adjust parameters in real-time within the constraint space. This invention also introduces a model uncertainty quantification mechanism as an exploration reward to guide the agent to adapt to unknown working conditions. This method effectively solves problems such as poor quality consistency, parameter adjustment lag, and heat accumulation effects in all-position welding of large-diameter pipelines, achieving intelligent closed-loop control of the welding process.

[0029] Example 1 like Figure 1 and Figure 2 As shown, this embodiment provides an adaptive control method for welding process parameters of large-diameter pipes, including the following steps: Step 1: Simultaneously acquire line laser vision information and time-series data of process parameters during the welding process of large-diameter pipes, and preprocess them to construct a multimodal dataset containing bevel geometry features and time-series features of process parameters; In order to achieve comprehensive perception of complex welding environments, visual information from a line laser profile sensor is received while acquiring welding process parameters. It should be noted that the large-diameter pipelines mentioned in this invention specifically refer to industrial conveying pipelines with a diameter of not less than 600 mm.

[0030] Furthermore, the bevel width is obtained by performing spatial dimension filtering on the visual information data. Misalignment and centering deviation ; The specific process parameters for welding include welding current. ,Voltage Wire feeding speed Welding speed Swing amplitude oscillation frequency and stop time ; The processed multi-source heterogeneous data are aligned according to timestamps to construct a standard multimodal dataset containing both geometric and temporal features, denoted as . It is divided into training and validation sets for subsequent model training and evaluation.

[0031] Step 2: Obtain welding quality prediction results based on a multimodal dataset containing bevel geometry features and temporal features of process parameters and a trained welding quality prediction model; In this embodiment, the welding quality prediction model adopts a multi-branch hybrid neural network architecture, including a parallel Long Short-Term Memory (LSTM) network and a Convolutional Neural Network (CNN). It not only extracts the temporal dependence of a single modality, but also automatically learns the coupling weights between geometric features and process parameters through a fully connected fusion layer.

[0032] Specifically, the steps include the following: Step 201: Extract time series features of process parameters based on LSTM branching; The core of an LSTM network lies in its cell state, which serves as the information transmission channel across all time steps. To control the retention and forgetting of information within this channel, such as... Figure 3 As shown, the LSTM unit is configured to include three key gating structures: a forget gate, an input gate, and an output gate.

[0033] The forget gate is configured to determine which information to discard from the cell state at the previous time step; this module receives the output from the previous time step. and the input vector at the current time step .

[0034] In this embodiment, the input vector It includes seven dimensions of process parameters: wire feed speed Welding speed Swing amplitude oscillation speed Duration of stay Welding current and welding voltage The forget gate outputs a value through the Sigmoid function. The calculation formula is as follows: (1), in, This is the weight matrix. For bias terms, This is the Sigmoid activation function.

[0035] The input gate is configured to determine the extent to which new information from the current time step updates the cell state. The structure contains two parallel layers: a sigmoid layer to determine the update rate. And a Tanh layer for creating new candidate vectors. Its mathematical expression is as follows: (2), (3), in, Here is the weight matrix of the input gate. This is the weight matrix for the candidate cell states. For the bias term of the input gate, This is a bias term for the candidate cell state.

[0036] After calculation using equations (2) and (3), the system updates the old cell state. To generate the cell state at the current moment Specifically, by... and Element-wise multiplication is used to discard invalid historical heat information, and then the data is accumulated. and The product of: .

[0037] The output gate is based on the updated cell state. Determines the final output value at the current moment. This output value is directly mapped to quality indicators such as melt penetration and defect rate. The calculation process is as follows: (4), (5), in, The output value of the output gate at the current moment. This is the weight matrix of the output gate. This is the bias term for the output gate.

[0038] The output vector obtained by the output gate operation It contains characteristic information of welding process parameters in the time dimension.

[0039] (6), As shown in equation (6), after completing the calculations for all time steps, the LSTM network will perform the calculations for the last time step. Temporal feature vectors The overall output of the current branch The vectors are fed into the fusion layer and fused with the feature vectors output by the CNN branch.

[0040] Step 202: Extract spatial correlation features of bevel geometry based on CNN branches; The raw bevel contour data acquired by the visual sensor is preprocessed and transformed into a sequence of feature vectors containing information such as bevel width, misalignment, and alignment deviation. The CNN captures the local spatial relationships of the bevel morphology by sliding multiple convolutional kernels across the feature space. The operation process of the convolutional layer is as follows: (7), in, Indicates the first Layer Each feature map at location The output, Indicates the first The layer output is located at the index The element value, For activation function, and These are the kernel width, weights, and bias, respectively.

[0041] To extract deep nonlinear features, pooling layers are followed by convolutional layers to reduce feature dimensionality and improve the robustness of the model: (8), in, Representing the The first feature vector output after pooling is the th layer. Each element. This represents the input features of this layer, i.e., the local geometric features extracted by the convolutional layer in the previous layer. To pool the window width, Step size, This represents the local sampling index within the pooling window.

[0042] The CNN output vector, after spatial feature extraction, is mapped through a fully connected layer to the same dimension as the LSTM main branch features to obtain the spatial geometric feature vector. The calculation process is as follows: (9), in, This is the weight matrix of the fully connected layer. The feature vectors are processed by the pooling layer. This is the bias term for the fully connected layer.

[0043] Step 203: Dynamically couple the two feature paths through an adaptive fusion layer to obtain a fused high-dimensional feature vector; After obtaining the time-series feature vector from the LSTM branch After the spatial feature vectors of the CNN branch The system utilizes a fusion layer to dynamically couple the two feature streams. The coupling process introduces adaptive weight coefficients, calculated using the following formula: (10) in, This represents the time-series feature vector of process parameters extracted by the LSTM branch. This represents the spatial geometric feature vector extracted by the auxiliary branch. and These are the coupling weight parameters that the model learns automatically during training. This structure can effectively integrate the influence of bevel shape changes on welding quality and achieve complementary cross-modal information.

[0044] Step 204: The fused high-dimensional feature vector is fed into the subsequent output layer to obtain the quality prediction index, thereby achieving accurate prediction of indicators such as weld penetration and forming quality.

[0045] Step 205: The hyperparameters of the welding quality prediction model are optimized using the Grey Wolf optimization algorithm, and the model is trained using the optimized hyperparameters. For example... Figure 4 As shown, during the training of the welding quality prediction model, the hyperparameters of the multi-branch hybrid neural network are globally optimized based on the Grey Wolf Optimizer (GWO) algorithm, specifically including: The hyperparameters of the neural network to be optimized are defined as the spatial location vectors of individual gray wolves. In this embodiment, the hyperparameter combination sets the optimization dimension. ,in Represents the number of neurons in the hidden layer of the network. Represents the initial learning rate. Represents the number of hidden layers. Represents the dropout ratio. This represents the width of the convolution kernel.

[0046] Using the mean squared error of the welding quality prediction model on the validation set as the fitness function, the population position is iteratively updated by simulating the social hierarchy and encirclement hunting behavior of a gray wolf population. The globally optimal hyperparameter combination is then searched, and the output optimal hyperparameter combination is used to construct the final high-precision welding quality prediction model. Specifically, Through fitness function The quality of the parameter set is evaluated by a function defined as the mean squared error (MSE) of the prediction model on the validation set. (11), in, This is actual welding quality data. For parameter-based The hybrid network predicts the output. For the first A sample of actual welding quality data. According to the first The predicted output calculated from a sample of actual data.

[0047] The GWO algorithm updates its parameters by simulating the encirclement and hunting behavior of gray wolves. The mathematical model of the encirclement mechanism is as follows: (12) (13) in, This is the current optimal parameter position. and The coefficient vector that controls the step size and direction. The difference between the current optimal parameters and other parameters. For the first The current position of the candidate parameters at the next iteration. For the first The current optimal parameter position at the next iteration. For candidate parameters in The new position at any given moment.

[0048] Step 3: Decouple the welding parameter control task into an upper global planning layer and a lower real-time control layer. The upper layer dynamically sets the process parameter constraint space according to the gravity conditions of different welding positions on the pipeline, and the lower layer uses the SoftActor-Critic (SAC) algorithm to perform real-time fine adjustment of parameters within the constraint space.

[0049] The global planning layer, as the upper layer, is responsible for establishing a reinforcement learning environment based on the high-precision prediction model obtained for the gravity condition differences at different welding positions of large-diameter pipelines. It also models the welding parameter optimization and adjustment problem as a Markov decision process and dynamically sets the constraint space and target benchmark of process parameters according to the weight of efficiency and quality. Specifically, the welding parameter optimization problem is defined as a Markov Decision Process (MDP), and the feasible domain of the lower-level controller and the weight allocation of multi-objective optimization are defined based on the current circumferential angle of the welding torch. Triggering is initiated when the spatial position change exceeds a set threshold or the laser sensor detects a sudden change in the bevel size, and its output is the reference parameter for the current sector. and allowable adjustment range .

[0050] Based on the determined reference parameters MDP state space Defined as ,in Indicates welding current, Indicates welding voltage, Indicates wire feeding speed, Indicates welding speed, Indicates the amplitude of swing, Indicates the oscillation frequency, Indicates the stopping time. This is the quality metric predicted by the model at the previous time step, used to guide the exploration.

[0051] The real-time control layer, acting as the lower layer, incorporates the SAC deep reinforcement learning algorithm to construct an adaptive control policy, enabling real-time fine-tuning of parameters within a given search space. This algorithm employs a maximum entropy-based reinforcement learning framework, whose objective function simultaneously maximizes the expected cumulative reward and policy entropy to balance exploration and exploitation. Through an Actor-Critic network structure, the algorithm utilizes reparameterization techniques to generate continuous actions and adjusts welding process parameters in real-time based on feedback from the quality prediction model, forming a closed-loop control. The decision system constructs a multi-objective reward function encompassing welding quality, production efficiency, and process stability, guiding the agent towards the ideal quality target. A model uncertainty quantification mechanism is introduced, using the variance of the prediction model as an exploration reward, guiding the agent to safely explore under unknown or highly uncertain conditions, rather than blindly exploiting resources.

[0052] Within the allowable adjustment range Internally, the SAC algorithm is used to output incremental actions. This enables rapid suppression of abnormal situations. This layer receives constraints from the upper layer and outputs specific changes in the wire feeding speed, thus forming the motion space. Action space Defined as an incremental adjustment vector of seven process parameters ,in Indicates the change in welding current, Indicates the change in welding voltage, Indicates the change in wire feed speed, Indicates the change in welding speed, Indicates the change in swing amplitude, Indicates the change in oscillation frequency. This indicates the change in stopping time.

[0053] reward function It is constructed as a multi-objective weighted form, and uncertainty is introduced for guidance: (14) in, Indicates the state at the next moment. The predicted quality index vector, This represents the preset ideal quality target vector. For boundary penalty terms, It is a penalty for the smoothness of the movement range, to avoid drastic parameter fluctuations; The uncertainty of the model is quantified by using the variance of the predicted output values ​​for the same state through an integrated LSTM network. , , , These are the weighting coefficients.

[0054] The SAC algorithm employs a maximum entropy framework, and its objective function is... Aimed at maximizing cumulative reward and policy entropy: (15) in, Represents the mathematical expectation. Represents the state-action edge distribution. Indicates an immediate reward. Indicates temperature parameter, Represents the entropy of the strategy.

[0055] By encoding the spatial position in the model, the unique characteristics of the molten pool's variation with angle due to gravity during all-position welding of large-diameter pipes are addressed, thereby avoiding the significant heat accumulation effect caused by long weld seams. Compared to general welding methods, this method can effectively suppress problems such as the concavity of the molten pool in the overhead welding position and the flow in the vertical welding position, and adaptively compensate for the penetration depth deviation caused by heat accumulation.

[0056] Step 4: The agent undergoes iterative training in a simulated environment to dynamically optimize the control strategy; The upper layer establishes a Markov decision process based on the welding environment and determines the action boundary according to the sector where the welding torch is currently located. The lower-level intelligent agent observes the current state and outputs incremental actions for process parameters.

[0057] During the iteration process, the reward value is calculated according to Equation (14). The uncertainty reward is fed back to the agent by calculating the variance of the model prediction value in the environment, guiding it to conduct safe exploration under unknown working conditions.

[0058] Each iteration generates the action at the current moment. With rewards Based on the current state and the state in the next moment The experience that constitutes this iteration The experience is stored in a buffer, data batches are randomly sampled from the buffer, the Critic network is updated to minimize the soft Bellman residual, the Actor network is updated to maximize the Q-value objective of entropy regularization, and the parameters of the Target network are updated through a soft update mechanism.

[0059] like Figure 5 As shown, the agent comprises a Critic network for estimating the soft Q value, an Actor network for generating actions, and a Target network for providing a stable standard for the learning of the Critic network.

[0060] The Actor network employs a reparameterization technique, by adjusting Gaussian noise... Perform transformation to generate actions To achieve gradient backpropagation: (16) Where tanh represents the hyperbolic tangent activation function. Represents the mean vector. This represents the standard deviation vector.

[0061] The agent continuously updates the weight parameters of the Critic and Actor networks by calculating the loss function.

[0062] The loss function used by the Critic network is: (17) (18) in, For the Critic network in parameters The loss value below, For experience replay buffer, These are the internal weight parameters of the Critic network. This represents the predicted evaluation value output by the Critic network for the current state and action combination. The target evaluation value includes both immediate rewards and future expectations. As a discount factor, The evaluation criteria value output by the Target network. Indicates temperature parameter, Let be the logarithmic probability of the action policy.

[0063] The loss function used by the Actor network is: (19) in, For Actor networks in parameters The loss value below, These are the internal weight parameters of the Actor network. For experience replay buffer, Indicates temperature parameter, Let be the logarithmic probability of the action policy. The Actor network adjusts its actions based on the newly generated increments from the current state. This is the quality score given by the Critic network for newly generated actions.

[0064] Through continuous iterative calculations, the agent gradually grasps the influence relationship between process parameters and welding quality of large-diameter pipelines, dynamically optimizes control strategies, and ultimately outputs precise parameter increments in a stable manner, thereby achieving closed-loop control.

[0065] Step 5: Apply the trained hierarchical strategy to the construction of circumferential welds in large-diameter pipelines to achieve adaptive control of welding quality in all positions.

[0066] The trained hierarchical strategy was applied to the construction of circumferential welds in large-diameter pipelines. Based on real-time sensing of bevel geometry fluctuations and spatial position changes, the upper layer dynamically adjusted the process window, while the lower layer adjusted the wire feed speed, welding speed, and oscillation parameters in real time. Through the synergy of multimodal perception and hierarchical decision-making, defects such as penetration deviation caused by heat accumulation and molten pool flow due to gravity were effectively eliminated, achieving consistent control of welding quality across all positions.

[0067] Example 2 This embodiment provides an adaptive control system for welding process parameters of large-diameter pipes, including: The dataset construction module is used to synchronously acquire line laser vision information and process parameter time series data during the welding process of large-diameter pipes, and perform preprocessing to construct a multimodal dataset containing bevel geometry features and process parameter time series features. The welding quality prediction module is used to obtain welding quality prediction results based on the multimodal dataset and the trained welding quality prediction model. The adaptive control module decouples the welding parameter control task into an upper global planning layer and a lower real-time control layer. The upper layer dynamically sets the process parameter constraint space based on the gravity conditions of different welding positions on the pipeline, while the lower layer uses a flexible action-evaluation algorithm to finely adjust the parameters in real time within the constraint space. The uncertainty quantification mechanism of the welding quality prediction model and the welding quality prediction results are introduced into the reward function of the algorithm to guide the agent to safely explore under unknown working conditions and obtain an optimized control strategy.

[0068] It should be noted that the specific implementation of the adaptive control system for welding process parameters of large-diameter pipes in this embodiment of the invention is similar to the specific implementation of the adaptive control method for welding process parameters of large-diameter pipes in this embodiment of the invention. For details, please refer to the description in the method section. To reduce redundancy, it will not be repeated here.

[0069] Example 3 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the adaptive control method for welding process parameters of large-diameter pipes as described above.

[0070] Example 4 This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the adaptive control method for welding process parameters of large-diameter pipes as described above.

[0071] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0072] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. 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, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0073] 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.

[0074] 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.

[0075] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0076] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An adaptive control method for welding process parameters of large-diameter pipes, characterized in that, Includes the following steps: Simultaneously acquire line laser vision information and time-series data of process parameters during the welding process of large-diameter pipes, perform preprocessing, and construct a multimodal dataset containing bevel geometry features and time-series features of process parameters; Welding quality prediction results are obtained based on the multimodal dataset and the trained welding quality prediction model; The welding parameter control task is decoupled into an upper global planning layer and a lower real-time control layer. The upper layer dynamically sets the process parameter constraint space according to the gravity conditions of different welding positions on the pipeline. The lower layer uses a flexible action-evaluation algorithm to finely adjust the parameters in real time within the constraint space. The uncertainty quantification mechanism of the welding quality prediction model and the welding quality prediction results are introduced into the reward function of the algorithm to guide the agent to conduct safe exploration under unknown working conditions and obtain an optimized control strategy. The upper layer dynamically sets the process parameter constraint space according to the gravity conditions of different welding positions on the pipeline, including: defining the welding parameter optimization problem as a Markov decision process, and defining the action feasible domain and multi-objective optimization weight of the lower real-time control layer according to the current circumferential angle of the pipeline where the welding torch is located; when the spatial position change exceeds the set threshold or the line laser sensor detects a sudden change in the bevel size, the upper layer is triggered to reset the reference parameters and allowable adjustment range as the constraint boundary of the lower real-time control layer. The reward function is: ， in, Indicates the state at the next moment. The predicted quality index vector, This represents the preset ideal quality target vector. For boundary penalty terms, It is a penalty for the smoothness of the movement range, to avoid drastic parameter fluctuations; To quantify the uncertainty of a model by utilizing the variance of the predicted output values ​​for the same state using an integrated LSTM network, , , , These are the weighting coefficients.

2. The adaptive control method for welding process parameters of large-diameter pipes as described in claim 1, characterized in that, The welding quality prediction model adopts a multi-branch hybrid neural network architecture, including a parallel long short-term memory network and a convolutional neural network. The time series features of the process parameters are extracted based on the long short-term memory network branch, and the spatial correlation features of the bevel geometry are extracted based on the CNN branch. The two features are dynamically coupled through an adaptive fusion layer to obtain a fused high-dimensional feature vector. The fused high-dimensional feature vector is then fed into the subsequent output layer to obtain the quality prediction index.

3. The adaptive control method for welding process parameters of large-diameter pipes as described in claim 1, characterized in that, When training the welding quality prediction model, the hyperparameters of the multi-branch hybrid neural network are globally optimized based on the Grey Wolf optimization algorithm, specifically including: The hyperparameters of the neural network to be optimized are defined as the spatial location vectors of individual gray wolves. ; Using the mean squared error of the welding quality prediction model on the validation set as the fitness function, the population position is iteratively updated by simulating the social hierarchy and encirclement hunting behavior of gray wolf populations, and the global optimal hyperparameter combination is searched. The optimal combination of hyperparameters will be output to construct the final welding quality prediction model.

4. The adaptive control method for welding process parameters of large-diameter pipes as described in claim 1, characterized in that, Within the allowable adjustment range, the lower real-time control layer outputs specific changes in wire feeding speed, forming an action space. This action space is a vector of incremental adjustment values ​​for process parameters, including changes in welding current, welding voltage, wire feeding speed, welding speed, oscillation amplitude, oscillation frequency, and side-stop time.

5. The adaptive control method for welding process parameters of large-diameter pipes as described in claim 1, characterized in that, When guiding the agent to conduct safe exploration under unknown conditions, the experience is stored in a buffer, data batches are randomly sampled from the buffer, the Critic network is updated to minimize the soft Bellman residual, the Actor network is updated to maximize the Q-value objective of entropy regularization, and the Target network parameters are updated through a soft update mechanism.

6. An adaptive control system for welding process parameters of large-diameter pipes, characterized in that, The method for adaptive control of welding process parameters for large-diameter pipes according to any one of claims 1-5 includes: The dataset construction module is used to synchronously acquire line laser vision information and process parameter time series data during the welding process of large-diameter pipes, and perform preprocessing to construct a multimodal dataset containing bevel geometry features and process parameter time series features. The welding quality prediction module is used to obtain welding quality prediction results based on the multimodal dataset and the trained welding quality prediction model. The adaptive control module decouples the welding parameter control task into an upper global planning layer and a lower real-time control layer. The upper layer dynamically sets the process parameter constraint space based on the gravity conditions of different welding positions on the pipeline, while the lower layer uses a flexible action-evaluation algorithm to finely adjust the parameters in real time within the constraint space. The uncertainty quantification mechanism of the welding quality prediction model and the welding quality prediction results are introduced into the reward function of the algorithm to guide the agent to safely explore under unknown working conditions and obtain an optimized control strategy.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the adaptive control method for welding process parameters of large-diameter pipes as described in any one of claims 1-5.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the adaptive control method for welding process parameters of large-diameter pipes as described in any one of claims 1-5.

Images