A method for rapid optimization of dissimilar metal welding process parameters based on transfer learning

Abstract

This invention relates to the interdisciplinary field of artificial intelligence and welding process optimization, specifically disclosing a rapid optimization method for dissimilar metal welding process parameters based on transfer learning. This method aims to address the core problems of long optimization cycles and high costs in dissimilar metal welding due to scarce experimental data and reliance on trial and error in process development. Its core steps include: constructing a welding process knowledge base integrating multi-source data; generating a dynamic prediction model adapted to the target task using hierarchical transfer learning; establishing a digital twin integrating this model as a virtual optimization environment; applying multi-objective reinforcement learning in the twin for rapid optimization of process parameters; and finally achieving online dynamic control of the process strategy and system generalization across materials. This invention significantly reduces data dependence and achieves rapid and high-quality process optimization under small sample conditions.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and welding process optimization, and specifically discloses a method for rapid optimization of welding process parameters of dissimilar metals based on transfer learning. Background Technology

[0002] Currently, existing technical solutions in this field can be mainly divided into two categories: one is the traditional optimization method based on a large number of physical experiments and classical statistical models; the other is the data-driven method based on deep learning models to construct the mapping relationship between process parameters and joint performance and to perform optimization. For example, the published patent CN114833439A, "A method for welding high-melting-point dissimilar metals by pre-placed T-shaped full barrier layer," represents a typical traditional process solution, which improves weldability through a specific physical barrier structure. Another category, such as some published neural network-based welding parameter prediction methods, attempts to establish a static mapping model between process parameters and quality indicators.

[0003] The core flaw of traditional methods like these lies in their over-reliance on physical experiments and expert experience for specific material combinations for parameter tuning and process design. This approach is essentially a fixed process scheme based on trial and error, failing to build a quantifiable process-performance prediction model. Consequently, it lacks adaptability to dynamic changes in the welding process and cannot quickly adapt to new welding tasks using existing knowledge data. This results in long process development cycles, high costs, and difficulty in achieving automatic collaborative optimization of multiple objective parameters.

[0004] While existing data-driven optimization methods have achieved parameter prediction to some extent, they still have fundamental limitations. These methods typically require the collection and labeling of a large amount of high-quality training data for each new combination of dissimilar metals or operating conditions, meaning model training starts entirely from scratch. This approach cannot effectively reuse historically accumulated or simulation-generated welding knowledge and fails to address the "small sample" cold start problem commonly encountered in industrial practice. Consequently, its optimization efficiency and engineering practicality are severely constrained, making it difficult to meet the needs of rapid process development. Summary of the Invention

[0005] The purpose of this invention is to provide a rapid optimization method for dissimilar metal welding process parameters based on transfer learning. Addressing the core shortcomings of existing dissimilar metal welding process optimization methods, such as heavy reliance on specific experiments, inability to utilize existing knowledge, and the limitations of data-driven methods due to small sample data limitations, this invention proposes a rapid optimization method based on transfer learning. This invention constructs a welding process knowledge base integrating multi-source data and utilizes hierarchical transfer learning technology to rapidly adapt general welding knowledge into a dedicated prediction model for the target task. This allows for the establishment of accurate process parameter-quality mapping relationships with minimal target experimental data. Subsequently, a digital twin is created based on this model, and multi-objective reinforcement learning is used in a virtual environment for efficient optimization to generate the optimal process strategy. Ultimately, this achieves online dynamic control and rapid cross-material generalization of the actual welding process, forming a complete closed loop from knowledge transfer and virtual optimization to online application. This fundamentally solves the technical problems of long development cycles and high costs associated with traditional methods, as well as the strong data dependence of existing intelligent methods.

[0006] The objective of this invention can be achieved through the following technical solutions: A rapid optimization method for dissimilar metal welding process parameters based on transfer learning includes the following steps: S1: Utilize multiphysics simulation data and historical welding experimental data to construct a welding process knowledge base, forming a transfer learning knowledge base that includes general welding physical laws and process mapping relationships; S2: Based on the aforementioned transfer learning knowledge base, perform hierarchical transfer learning and model fine-tuning for the target dissimilar metal welding task to generate a dynamic prediction model for welding quality and parameters that is adapted to the target task. S3: Using the aforementioned dynamic prediction model for welding quality and parameters, create a virtual welding optimization environment and establish a digital twin capable of simulating multi-objective interaction and dynamic process response; S4: In the digital twin, a multi-objective reinforcement learning method is applied to quickly optimize the candidate process parameter set and output the optimal welding process strategy set that satisfies the balance of conflicting performance indicators. S5: Using the aforementioned optimal welding process strategy set, guide the online dynamic parameter control and meta-learning adaptation of the actual dissimilar metal welding process, and finally obtain an intelligent welding optimization system with cross-material generalization capability.

[0007] Preferably, S1 specifically refers to: Multi-source heterogeneous data from different material combinations, welding methods, and quality standards are integrated; through data cleaning, normalization, and feature engineering, feature sets strongly correlated with welding thermal cycles, weld pool dynamics, and joint performance are extracted; finally, a structured and queryable welding process knowledge graph is constructed as the carrier of the transfer learning knowledge base.

[0008] Preferably, in S2, hierarchical transfer learning and model fine-tuning specifically involve: The common data in the transfer learning knowledge base are used to pre-train the deep neural network base model to learn the common rules of welding. A domain-adaptive algorithm is employed to minimize the feature distribution differences between small sample experimental data in the target domain and data in the source domain. The model is fine-tuned using target domain data to quickly adapt to the process-performance mapping relationship of specific dissimilar metal combinations.

[0009] Preferably, the dynamic prediction model for welding quality and parameters is a hybrid model that integrates attention mechanisms and long short-term memory networks. The long short-term memory network layer is used to process the input welding current, voltage and welding speed parameter sequences in time step order to extract their time dynamic features and long-term dependencies. The attention mechanism layer is connected after the long short-term memory network layer. It calculates the attention score of each time step feature through a trainable feedforward neural network and assigns differentiated weights to the features of different time steps accordingly, focusing on the process stages that have a key impact on the final quality indicators. The final output of the model is the weighted sum of the hidden states and their corresponding weights at all time steps. After being mapped by a fully connected layer, it simultaneously generates predicted values ​​for multiple quality indicators such as weld penetration, tensile strength, and forming coefficient. The model parameters obtained through transfer learning in S2 are used as the initial state for training the hybrid model, and the mean square error between the predicted and true values ​​is used as the loss function for supervised fine-tuning optimization.

[0010] Preferably, in step S3, establishing a digital twin includes: The prediction model is encapsulated into a callable simulation agent model; a multiphysics calculation engine is integrated to simulate the evolution of temperature and stress fields; and a multi-objective interactive evaluation module is designed to quantify the trade-off between multiple conflicting objectives such as weld formation quality, mechanical properties, and thermal damage in real time.

[0011] Preferably, in step S4, the application of the multi-objective reinforcement learning method specifically involves: Welding process parameters are used as the action space of the agent, and real-time welding status and multi-objective evaluation results are used as the state space and reward signal. A deep deterministic policy gradient algorithm based on Pareto front improvement is used to explore and learn in the digital twin to iteratively generate the optimal welding process policy set.

[0012] Preferably, the optimal welding process strategy set includes a Pareto optimal solution set for given welding conditions and quality requirements, and a recommendation strategy generator that performs online weight allocation based on actual production preferences.

[0013] Preferably, the online dynamic parameter control in S5 specifically includes: The actual welding process is monitored online using visual, spectral, and acoustic emission sensors; the monitoring data is input into the prediction model for real-time status diagnosis; when the predicted indicators deviate from expectations, the meta-learning adapter is triggered to quickly fine-tune the model based on a small amount of new data, and the strategy set is called to generate adjusted process parameters for closed-loop control.

[0014] Preferably, the meta-learning adapter adopts a model-independent meta-learning algorithm framework, the goal of which is to enable the prediction model to quickly adapt to new dissimilar metal welding tasks by meta-training on a variety of welding tasks covered by the knowledge base, with only a single gradient update.

[0015] The beneficial effects of this invention are: This invention effectively overcomes the inherent challenge of scarce high-quality experimental data in dissimilar metal welding by constructing a welding process knowledge base and applying hierarchical transfer learning. It can rapidly transfer and adapt well-validated general welding principles and knowledge contained in simulation data to new, specific welding tasks. This allows for the establishment of a high-precision dynamic prediction model of process parameters and quality with minimal experimental samples in the target domain, significantly reducing reliance on expensive and time-consuming physical experiments and achieving a fundamental shift in process development from "data scarcity-driven" to "knowledge reuse-driven."

[0016] This invention achieves efficient and in-depth collaborative optimization of welding process parameters by creating a digital twin of an integrated prediction model and using multi-objective reinforcement learning in this virtual environment. This method replaces traditional physical trial-and-error and single-objective optimization, enabling the exploration of massive parameter combinations in a digital space at near-zero cost. It simultaneously balances multiple conflicting quality objectives such as weld formation and mechanical properties, automatically finding the Pareto optimal strategy set that satisfies specific preferences, thereby significantly improving the efficiency, breadth, and intelligence of the optimization process.

[0017] The intelligent optimization system constructed in this invention possesses online dynamic control and meta-learning adaptation capabilities, forming a closed loop of continuous self-improvement. The system can dynamically adjust parameters based on real-time feedback from the actual welding process and quickly adapt to new material combinations or changes in working conditions through a meta-learning mechanism. Simultaneously, new data and strategies generated online can be fed back into the knowledge base, enabling the system's prediction and optimization capabilities to continuously evolve during use, ultimately becoming an adaptive intelligent solution with excellent cross-material generalization capabilities and engineering practicality. Attached Figure Description

[0018] Figure 1This is a flowchart illustrating a rapid optimization method for dissimilar metal welding process parameters based on transfer learning, according to the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] Example: Figure 1 As shown, a rapid optimization method for dissimilar metal welding process parameters based on transfer learning includes the following steps: S1: Utilize multiphysics simulation data and historical welding experimental data to construct a welding process knowledge base, forming a transfer learning knowledge base that includes general welding physical laws and process mapping relationships; S2: Based on the aforementioned transfer learning knowledge base, perform hierarchical transfer learning and model fine-tuning for the target dissimilar metal welding task to generate a dynamic prediction model for welding quality and parameters that is adapted to the target task. S3: Using the aforementioned dynamic prediction model for welding quality and parameters, create a virtual welding optimization environment and establish a digital twin capable of simulating multi-objective interaction and dynamic process response; S4: In the digital twin, a multi-objective reinforcement learning method is applied to quickly optimize the candidate process parameter set and output the optimal welding process strategy set that satisfies the balance of conflicting performance indicators. S5: Using the aforementioned optimal welding process strategy set, guide the online dynamic parameter control and meta-learning adaptation of the actual dissimilar metal welding process, and finally obtain an intelligent welding optimization system with cross-material generalization capability.

[0021] In this embodiment, a welding process knowledge base is constructed using multiphysics simulation data and historical welding experiment data, forming a transfer learning knowledge base that includes general welding physical laws and process mapping relationships. The specific implementation method is as follows: This step aims to systematically integrate and process multi-source heterogeneous data, transforming it into a structured knowledge carrier. The specific implementation includes the following core components: The integration and standardization of multi-source heterogeneous data is performed. This multi-source data mainly encompasses two categories: firstly, multiphysics simulation data generated through finite element analysis, computational fluid dynamics, and other methods; and secondly, historical welding experimental data collected from historical production records, scientific literature, and laboratory accumulation. These data exhibit significant differences in format, units, and precision. The integration process involves establishing a unified data pool, cleaning the original data to remove outliers and missing items, and performing normalization to eliminate the influence of units. This process can be formally represented as integrating datasets from n different source domains. ,in, Represents the dataset from the i-th source. This is the integrated total dataset. Each data sample can be represented as a high-dimensional vector. ,in Representative process parameters, Representative physical field characteristics of the process This represents the final performance metric.

[0022] Performing in-depth feature engineering to extract a feature set that is highly explanatory of weld quality is a key step in elevating raw data into valuable knowledge. Feature engineering focuses on constructing derived features that are strongly correlated with the welding thermal cycle, weld pool dynamics, and joint performance, based on the principles of welding physics and metallurgy. For temperature field data, the core is to analyze its thermal cycle curves and extract features such as peak temperatures. High temperature dwell time The time to cool from 800°C to 500°C, and the instantaneous cooling rate Key features include cooling rate. As a core factor determining microstructure and performance, it can be defined and calculated using the derivative of the temperature-time curve within a specific temperature range: In practical data processing, estimations are often based on simplified models. Referring to the Rosencell line heat source model, at a given temperature T, the relationship between the cooling rate and welding parameters can be approximated as: in, Thermal conductivity, The initial temperature, Where Q is the thermal efficiency factor and Q is the heat input power. For welding speed. This formula reveals the relationship between cooling rate and line energy. The physical law is inversely proportional. Simultaneously, the molten pool length can be extracted from the molten pool simulation data. Aspect Ratio Isomorphic features; principal components of the residual stress tensor can be extracted from stress field data. , , These refined features together constitute a high-dimensional feature space F that comprehensively reflects the physical essence of the welding process and its correlation with the final performance.

[0023] Based on the processed feature set, a structured and queryable welding process knowledge graph is constructed, serving as the final carrier of the transfer learning knowledge base. The knowledge graph is organized using a triplet format of "entity-relationship-attribute". Core entity types include: material entities, process method entities, process parameter entities, physical process entities, and performance index entities. Relationships define the causal or influence connections between entities. The construction process includes defining a schema layer using domain rules and extracting instances from the data using data mining techniques. The graph structure of the knowledge graph can be formally defined as: Where V is a finite set of entity nodes, Let A be the set of edges representing relations, where A is the set of attributes associated with nodes and edges. For example, a specific piece of knowledge can be represented as a triple. Its attributes May include factors affecting strength coefficient The resulting knowledge graph is a semantic network containing knowledge in the welding field. Through graph query and reasoning capabilities, it provides a solid and operational structured knowledge foundation for effective knowledge transfer from the source domain to the target domain. This knowledge base will continue to evolve, with new knowledge generated during subsequent optimization processes being assimilated into the database, enabling the system to enrich and evolve itself.

[0024] In this embodiment, relying on the aforementioned transfer learning knowledge base, hierarchical transfer learning and model fine-tuning are performed for the target dissimilar metal welding task, generating a dynamic prediction model for welding quality and parameters adapted to the target task. The specific implementation method is as follows: In this step, hierarchical transfer learning and model fine-tuning are the core intelligent transformation processes for achieving accurate predictions from general welding knowledge to specific tasks. This process aims to overcome the challenge of scarce experimental data for target dissimilar metal welding. Through a three-level progressive strategy—pre-training, domain adaptation, and fine-tuning—it efficiently and robustly transfers the rich prior knowledge contained in the source domain to the target domain, thereby constructing a high-precision dynamic prediction model.

[0025] The first stage involves model pre-training based on general knowledge. The goal of this stage is to enable the deep neural network base model to learn the common physical laws and statistical characteristics of the welding process from large-scale source domain data. Specifically, from the welding process knowledge graph, all triplet data related to the process-performance mapping are extracted to construct the source domain training sample set. ,in This is a sequence of welding parameters arranged in time steps. This represents the corresponding multidimensional quality index vector. The base model employs a hybrid architecture that integrates an attention mechanism and a long short-term memory network. The pre-training task is defined as minimizing the model's prediction error on the source domain data, and its loss function is in the form of mean squared error: By optimizing this loss function, the model parameters It is initialized to a state with the ability to understand general welding rules, laying the foundation for subsequent migration.

[0026] The second level is feature distribution alignment based on domain adaptation. Although the pre-trained model has learned general knowledge, its data distribution differs from the distribution of the limited experimental data from the new target task, and direct application will lead to performance degradation. This embodiment employs a domain adaptation algorithm to minimize the distribution difference between the source and target domains in the feature space. Let the target domain's small sample dataset be... ,and Extract the feature output of a certain intermediate layer from the hybrid model, and calculate the feature distribution of the source domain respectively. and target domain feature distribution The difference in mean. Using the maximum mean difference as a measure of distribution distance, its empirical estimate is: Where h is the feature vector extracted by the model. The function is mapped to the reproducing kernel Hilbert space H, typically implemented using a Gaussian kernel, etc. Loss is predicted through joint optimization. Alignment loss with distribution The feature representations learned by the model become less sensitive to changes in the domain, thereby improving the model's generalization potential in the target domain.

[0027] The third stage involves fine-tuning the model based on a small sample from the target domain. After distribution alignment, the model has initially adapted to the data characteristics of the target domain. To further accurately adapt to the process-performance mapping details of specific dissimilar metal combinations, this step uses all target domain data. Supervised fine-tuning of the model is then performed. A small learning rate is used in this stage to prevent catastrophic forgetting of pre-trained knowledge. The objective function for fine-tuning is a simple target domain prediction loss. ,in These are the model parameters after pre-training and domain adaptation. Through fine-tuning, the model ultimately combines the general rules learned in the source domain with a small number of precise samples specific to the target domain, generating a dynamic prediction model for welding quality and parameters that is fully adapted to the current task. This model can not only make high-precision predictions of quality indicators under new parameter combinations, but its internal attention weights can also reveal the key impacts of different process stages on the final quality, providing in-depth insights for process analysis and optimization, thereby achieving rapid and reliable modeling under small sample conditions.

[0028] In this embodiment, a virtual welding optimization environment is created using the aforementioned welding quality and parameter dynamic prediction model, establishing a digital twin capable of simulating multi-objective interactions and dynamic process responses. The specific implementation method is as follows: This step aims to construct a high-fidelity, computable virtual welding optimization environment, i.e., a digital twin. This twin is not simply a 3D visualization model, but a dynamic simulation system deeply integrated with data, models, and algorithms. Its core function is to replace expensive and time-consuming physical trial and error, providing an unlimited, low-cost, and observable "testing ground" for subsequent intelligent optimization algorithms. The key to its implementation lies in the organic integration of the intelligent prediction model obtained in the previous steps with the physical law simulation engine, and the establishment of a multi-objective quantitative evaluation system.

[0029] The foundation of a digital twin is the construction of a simulation proxy model, which uses the welding quality and parameter dynamic prediction model trained in step S2. ,in The fine-tuned optimal parameters are then encapsulated in an engineering manner. Specifically, the neural network model is deployed as a microservice with standard input / output interfaces, enabling it to accept any given sequence of welding process parameters. , Let be the parameter vector at time t, and output the corresponding multidimensional quality index prediction values ​​in real time. The significant advantage of this proxy model lies in its extremely high computational efficiency. A single forward propagation can be completed within milliseconds. Compared to traditional multiphysics numerical simulations that take minutes or even hours, it makes it possible to rapidly evaluate massive parameter schemes in a virtual environment. It is a core component supporting real-time interaction of optimization algorithms such as reinforcement learning.

[0030] To enhance the physical consistency and process evolution interpretation capabilities of the twins, a high-fidelity multiphysics computational engine needs to be integrated. This engine, based on the finite element method or finite volume method, solves the partial differential equations governing the key physical fields controlling the welding process. Temperature field The evolution is governed by a nonlinear heat conduction equation: ,in, For material density, For specific heat capacity, Thermal conductivity, For laser heat source items, This represents the losses from thermal convection and radiation. Based on the thermal field calculations, the quasi-static equilibrium equations are solved through thermo-mechanical coupling analysis to obtain the stress-strain field: ,in, Let b be the Cauchy stress tensor and b be the volume force. In this embodiment, the macroscopic process effects predicted by the surrogate model are used as boundary conditions or source terms to guide high-precision physical field simulation through a co-simulation interface. At the same time, the fine-grained results of the physical field simulation calculation are fed back to verify and calibrate the surrogate model, forming a hybrid modeling paradigm in which data-driven and physical models mutually verify and enhance each other.

[0031] The decision-oriented function of the twin is implemented by a multi-objective interactive evaluation module. The core task of this module is to quantify in real time the achievement status of multiple conflicting welding objectives caused by any set of process parameters. A set of M objective functions that need to be optimized simultaneously are defined. . This represents the negative deviation of the weld penetration from the expected value. Represents the width of the heat-affected zone. This represents the tensile strength of the joint. To achieve a comprehensive balance within the same dimension, each objective needs to be normalized, and weights can be dynamically assigned by the user or the upper-level algorithm. The evaluation module outputs a comprehensive trade-off evaluation value V and the position of the current solution in the objective space: in .

[0032] Even more advanced is the inclusion of an online Pareto front estimator within the module. It continuously records and analyzes all solution points generated during the virtual optimization process. It also dynamically updates the currently known non-dominated solution set, providing rich gradient information and convergence direction for the optimization algorithm, thereby guiding the search to the optimal process parameter region that can balance various performance aspects.

[0033] The digital twin created in this step is a composite system that integrates rapid proxy models, high-fidelity physical simulation, and intelligent multi-objective evaluation. It possesses the speed advantage of data-driven models, the reliability guaranteed by physical laws, and provides direction for optimization through multi-objective quantitative evaluation, thus laying a solid environmental foundation for the next step of efficient and accurate process parameter optimization in virtual space.

[0034] In this embodiment, a multi-objective reinforcement learning method is applied to rapidly optimize the candidate process parameter set in the digital twin, outputting an optimal welding process strategy set that satisfies the balance of conflicting performance indicators. The specific implementation method is as follows: This step aims to address the core challenge of balancing conflicting performance indicators in dissimilar metal welding, which is difficult to achieve through human experience or single-objective optimization. The implementation involves establishing a multi-objective reinforcement learning agent within a digital twin constructed using S3. Through autonomous interaction and trial and error with the virtual environment, this agent automatically explores and discovers process parameter strategies that optimally balance various quality indicators. This method formalizes the process optimization problem as a sequential decision-making process: in each optimization iteration and virtual welding experiment, the agent observes the current welding task status, selects a set of process parameters, and issues them to the digital twin for execution. The twin simulates the welding process under these parameters and calculates multiple performance indicators. Finally, the agent obtains a multi-dimensional reward signal based on the achievement of these indicators and updates its decision-making strategy accordingly, pursuing the maximization of long-term cumulative rewards.

[0035] Specifically, the state space S is defined as the set of all observable information related to the welding process and the target, including the physical properties of the target material, such as thermal conductivity. Specific heat capacity The data includes joint geometry and real-time features extracted from twin physics calculations. The motion space A corresponds to the welding process parameters to be optimized, such as laser power P, scanning speed v, and defocusing amount. These parameters are typically normalized to a continuous interval. The core design lies in the multi-objective reward function. At time t, when the agent takes an action... Then, the multi-objective evaluation module of the digital twin will output a vector containing M normalized objective values. The reward vector is defined as the improvement in the objective achieved by the action. In this way, the feedback received by the agent at each step is a vector. Each component corresponds to an immediate reward for an optimization objective. The ultimate goal of the agent is to learn a policy. This allows the trajectory it guides to obtain the maximum cumulative vectorized reward. Since vectors are usually not directly comparable, this introduces a core challenge in multi-objective optimization.

[0036] To address this challenge, this embodiment employs a deep deterministic policy gradient algorithm based on an improvement of the Pareto front. This algorithm is a multi-objective extension of the classic deep deterministic policy gradient framework. Its core innovation lies in the fact that the update direction of the policy network is not determined by the gradient of a single scalar reward, but is guided by a dynamically maintained set of non-dominated solutions. The agent maintains a replay buffer B, which stores a large number of state transition tuples. During each update, the algorithm samples a batch of data from the buffer and uses a multi-objective evaluation network to estimate the long-term cumulative vectorized reward for each state-action pair under the current policy. ,in These are network parameters. Strategy. The update objective is no longer to maximize a single Q-value, but rather to drive the resulting action to approach the currently estimated Pareto frontier. A key update criterion is based on Pareto dominance: for an action... If its corresponding expected cumulative return vector Dominate another action If the reward vector is such that the strategy should be directed towards... The direction is updated. This can be achieved by solving a constrained optimization problem whose objective function can be formalized as: ,in, It is a reference point related to a threshold of the current Pareto front on target m. This design encourages strategies to explore actions that can improve the weakest target or advance the Pareto front overall.

[0037] After sufficient learning and iteration, the agent will output an optimal welding process strategy set, which is precisely the set of Pareto optimal strategies discovered and saved by the algorithm throughout the exploration process. This strategy set consists of two parts: one is the Pareto optimal solution set. Each of these strategies Both correspond to a process parameter scheme that cannot be comprehensively surpassed by other strategies in the target space; the second is a recommendation strategy generator. In actual production, operators or upper-level production management systems may have dynamically changing preferences for various quality indicators. This generator allows users to input a real-time preference weight vector. Then, in the Pareto optimal solution set, a solution that maximizes the weighted utility is quickly calculated and recommended. Optimal strategy Therefore, this invention not only automatically discovers a wide range of high-quality process parameters, but also provides a flexible and interpretable human-machine interface, enabling the optimization results to be directly adapted to complex actual production needs.

[0038] In this embodiment, the optimal welding process strategy set is used to guide the online dynamic parameter control and meta-learning adaptation of the actual dissimilar metal welding process, ultimately obtaining an intelligent welding optimization system with cross-material generalization capabilities. The specific implementation method is as follows: This step marks the transition of the entire optimization method from the virtual world to the physical world, forming a complete closed loop of "perception-decision-execution-evolution". The core task of this step is to safely and accurately apply the theoretically optimal process strategy set output by S4 to actual dissimilar metal welding equipment. Furthermore, it aims to enable the system to possess real-time self-adjustment and rapid adaptive capabilities when facing fluctuations in real-world operating conditions or new material tasks, ultimately refining it into an intelligent welding optimization system with cross-material generalization capabilities. Its implementation relies on a sophisticated online sensor network, a real-time diagnostic and decision engine embedded in edge computing units, and a meta-learning core capable of rapid adaptation to small sample sizes.

[0039] Online sensing and real-time status diagnosis are the starting point for closed-loop control. In actual welding processes, a multimodal sensor network provides comprehensive synchronous monitoring: high-speed vision sensors capture images of the molten pool morphology, keyhole dynamics, and weld formation; spectral sensors analyze the spectral characteristics of the plasma plume to infer the molten pool temperature and elemental loss; and acoustic emission sensors collect elastic stress wave signals during welding to monitor defects such as microcrack formation. These high-dimensional, high-speed time-series data streams... The data is transmitted in real time to the edge computing unit. The system then invokes the deployed dynamic prediction model for welding quality and parameters. However, the input at this time is not a preset parameter, but a deep feature vector extracted in real time from the sensor data that is isomorphic to the model during training. , This is a feature extraction network. Based on these real-time features, the model can online infer the estimated values ​​of key quality indicators under the current welding condition. The system continues to use this predicted value. Compared with the target expectation of the current welding task A comparison is performed, and if the deviation of any indicator exceeds the preset safety threshold... Upon this time, a state deviation is determined to have occurred, triggering the control mechanism. This deviation can be quantified as: in, The norm is a weighted average based on the importance of each indicator. This is the triggering condition.

[0040] Meta-learning for rapid adaptation and policy replanning is key to handling unknown disturbances. Once regulation is triggered, traditional fine-tuning methods cannot meet real-time requirements due to the need for large amounts of data and iteration time. This embodiment employs a "meta-learning adapter." This adapter has been meta-trained on numerous differential welding tasks covered by the S1 knowledge base. Its goal is not to learn how to solve a specific task, but rather to learn "how to quickly learn" to solve new tasks. It adopts a model-independent meta-learning algorithm framework. Assuming a master prediction model... The parameters are The meta-learner learns an optimal initial state of parameters. And an effective parameter update rule. When faced with a small amount of new data collected in real time. At that time, the adapter is not trained from scratch, but based on Perform one or more gradient descent iterations to quickly generate model parameters adapted to the current new operating conditions. : ,in, To adapt the learning rate, this process can be completed within hundreds of milliseconds, achieving "second-level online calibration" of the model.

[0041] Closed-loop dynamic parameter control and system evolution are the final implementation steps. This utilizes a rapidly adapted new model. The system reassesses the current state. Based on the latest state diagnosis results, the decision engine queries or invokes the optimal welding process strategy set generated by S4. This strategy set is essentially a mapping library from states to optimal actions. The system then... And the possible real-time production preference weights w, retrieve the optimal process parameter adjustment amount from the strategy set or calculate it instantaneously. And adjust the parameters The data is sent to the welding equipment actuator to complete one closed-loop control cycle. Simultaneously, the effective data generated during this online control process... The data is automatically labeled and stored. Once a certain batch of this new data has accumulated, the system will initiate a background asynchronous evolution process: using this data as a new "small sample task" not only to update the main prediction model, but more importantly, to retrain the meta-learning adapter itself, thereby optimizing its initial parameters. The updated rules enable the system to adapt faster and more accurately when facing similar problems in the future.

[0042] Through a cycle of "real-time monitoring - rapid diagnosis - meta-learning adaptation - closed-loop control - continuous evolution", the intelligent welding optimization system finally obtained by this method can not only accurately implement the optimal strategy obtained by virtual space optimization, but also cope with uncertainties such as material fluctuations, assembly errors, and environmental interference in actual production. Furthermore, by rapidly learning from each new task, it can precipitate its experience into the system's inherent generalization ability, truly realizing the leap from solving "one" problem to solving "a class" of problems.

[0043] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, 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 process, method, article, or apparatus.

[0044] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for rapid optimization of dissimilar metal welding process parameters based on transfer learning, characterized in that, Includes the following steps: S1: Utilize multiphysics simulation data and historical welding experimental data to construct a welding process knowledge base, forming a transfer learning knowledge base that includes general welding physical laws and process mapping relationships; S2: Based on the aforementioned transfer learning knowledge base, perform hierarchical transfer learning and model fine-tuning for the target dissimilar metal welding task to generate a dynamic prediction model for welding quality and parameters that is adapted to the target task. S3: Using the aforementioned dynamic prediction model for welding quality and parameters, create a virtual welding optimization environment and establish a digital twin capable of simulating multi-objective interaction and dynamic process response; S4: In the digital twin, a multi-objective reinforcement learning method is applied to quickly optimize the candidate process parameter set and output the optimal welding process strategy set that satisfies the balance of conflicting performance indicators. S5: Using the aforementioned optimal welding process strategy set, guide the online dynamic parameter control and meta-learning adaptation of the actual dissimilar metal welding process, and finally obtain an intelligent welding optimization system with cross-material generalization capability.

2. The method for rapid optimization of dissimilar metal welding process parameters based on transfer learning according to claim 1, characterized in that, Specifically, S1 is: Multi-source heterogeneous data from different material combinations, welding methods, and quality standards are integrated; through data cleaning, normalization, and feature engineering, feature sets strongly correlated with welding thermal cycles, weld pool dynamics, and joint performance are extracted; finally, a structured and queryable welding process knowledge graph is constructed as the carrier of the transfer learning knowledge base.

3. The method for rapid optimization of dissimilar metal welding process parameters based on transfer learning according to claim 1, characterized in that, In S2, hierarchical transfer learning and model fine-tuning specifically refer to: The common data in the transfer learning knowledge base are used to pre-train the deep neural network base model to learn the common rules of welding. A domain-adaptive algorithm is employed to minimize the feature distribution differences between small sample experimental data in the target domain and data in the source domain. The model is fine-tuned using target domain data to quickly adapt to the process-performance mapping relationship of specific dissimilar metal combinations.

4. The method for rapid optimization of dissimilar metal welding process parameters based on transfer learning according to claim 3, characterized in that, The dynamic prediction model for welding quality and parameters is a hybrid model that integrates attention mechanisms and long short-term memory networks. The long short-term memory network layer is used to process the input welding current, voltage and welding speed parameter sequences in time step order to extract their time dynamic features and long-term dependencies. The attention mechanism layer is connected after the long short-term memory network layer. It calculates the attention score of each time step feature through a trainable feedforward neural network and assigns differentiated weights to the features of different time steps accordingly, focusing on the process stages that have a key impact on the final quality indicators. The final output of the model is the weighted sum of the hidden states and their corresponding weights at all time steps. After being mapped by a fully connected layer, it generates predicted values ​​for multiple quality indicators such as weld penetration, tensile strength, and forming coefficient. The model parameters obtained through transfer learning are used as the initial state for training the hybrid model, and the mean square error between the predicted and true values ​​is used as the loss function for supervised fine-tuning optimization.

5. The method for rapid optimization of dissimilar metal welding process parameters based on transfer learning according to claim 1, characterized in that, In step S3, establishing a digital twin includes: The prediction model is encapsulated into a callable simulation agent model; a multiphysics calculation engine is integrated to simulate the evolution of temperature and stress fields; and a multi-objective interactive evaluation module is designed to quantify the trade-off between multiple conflicting objectives such as weld formation quality, mechanical properties, and thermal damage in real time.

6. The method for rapid optimization of dissimilar metal welding process parameters based on transfer learning according to claim 1, characterized in that, In S4, the application of the multi-objective reinforcement learning method specifically involves: Welding process parameters are used as the action space of the agent, and real-time welding status and multi-objective evaluation results are used as the state space and reward signal. A deep deterministic policy gradient algorithm based on Pareto front improvement is used to explore and learn in the digital twin to iteratively generate the optimal welding process policy set.

7. The method for rapid optimization of dissimilar metal welding process parameters based on transfer learning according to claim 6, characterized in that, The optimal welding process strategy set includes a Pareto optimal solution set for given welding conditions and quality requirements, and a recommendation strategy generator that performs online weight allocation based on actual production preferences.

8. The method for rapid optimization of dissimilar metal welding process parameters based on transfer learning according to claim 1, characterized in that, The online dynamic parameter control in S5 specifically includes: The actual welding process is monitored online using visual, spectral, and acoustic emission sensors; the monitoring data is input into the prediction model for real-time status diagnosis; when the predicted indicators deviate from expectations, the meta-learning adapter is triggered to quickly fine-tune the model based on a small amount of new data, and the strategy set is called to generate adjusted process parameters for closed-loop control.

9. A method for rapid optimization of dissimilar metal welding process parameters based on transfer learning according to claim 8, characterized in that, The meta-learning adapter adopts a model-independent meta-learning algorithm framework. Its goal is to enable the prediction model to quickly adapt to new dissimilar metal welding tasks with only a single gradient update by meta-training on a variety of welding tasks covered by the knowledge base.

Images