A welding stress-strain rapid prediction control method and system based on an LSTM network

Abstract

The application discloses the field of welding stress, and discloses a kind of welding stress strain fast prediction control method and system based on LSTM network, the method is first based on original welding simulation and actual welding process in high fidelity data, using laser power, welding speed, defocusing amount and spot diameter and other welding process parameters and welding workpiece and other parameters as input, corresponding welding stress and strain as the output of prediction to train LSTM agent model.The application can be implemented in the butt joint state of various welding, can adapt to a variety of states, the accuracy of cross-condition identification exceeds 80%, and can meet the requirements of normal work.The application realizes the rapid calculation of welding stress and strain based on LSTM network training model, the detection of welding process for convenient backtracking adjustment, and dynamic adjustment of welding process based on the advantages of model to realize closed-loop optimization of welding process.

Description

Technical Field

[0001] This invention belongs to, but is not limited to, the field of welding stress, and particularly relates to a method and system for rapid prediction and control of welding stress and strain based on LSTM networks. Background Technology

[0002] Laser welding technology, characterized by its high welding speed, high specific energy, and high welding efficiency, is widely used in fields such as rail transportation, shipbuilding, and marine manufacturing. However, due to the nature of laser welding, the process involves rapid heating and cooling, resulting in a steep temperature gradient in the weld and its surrounding area, leading to uneven thermal expansion and contraction. Under the constraints of the base material and fixtures, these thermal cycles induce welding stress and strain, manifesting as welding deformation and residual stress. Even with a relatively small heat-affected zone, deformation can still be significant for thin-walled parts, long welds, and large components, potentially directly impacting assembly accuracy, performance, and defect susceptibility (e.g., misalignment, poor assembly, or cracking risk). Therefore, monitoring and controlling welding deformation is crucial for ensuring laser welding quality and manufacturing consistency.

[0003] Traditional laser welding stress and strain measurements often rely on specialized instruments after welding, and depend on empirical curves or manual statistical calculations based on empirical data. This method is characterized by high cost, low efficiency and feasibility, and lack of effective control. For example, patent CN 118551626 A uses small-angle X-ray scattering technology to measure the lattice distortion degree of the weld joint micro-region to assess welding stress in order to improve measurement efficiency, but this operation is complex. With the development of computational simulation technology, finite element simulation using the FAE method is an important method for calculating welding stress and strain. Through simulation calculation, stress and strain during the welding process can be calculated and predicted. Simulation using software such as Ansys often requires precise control of key physical information such as the model and heat source. This is characterized by cumbersome model control and high computational cost. To further improve the calculation efficiency of welding stress and strain, patent CN116882252 A uses files based on the inherent deformation library to interpolate data in the welding model to achieve rapid calculation of the mechanical properties of arc-welded structural components. As the size of the structural unit increases, or by using the method in patent CN 109926767 A, which employs equivalent heat sources or equivalent joints, the calculation model can be simplified, thereby improving the rapid calculation method for stress and strain in large-scale welding. Alternatively, a method similar to that used in patent CN116629067 A, which compares and analyzes the maximum stress and maximum strain values, can be employed to quickly lock the convergence region, ensuring accurate and efficient convergence of components, especially large components, during the welding model process, thus improving the efficiency of welding simulation.

[0004] In summary, current direct calculations of welding deformation and stress often rely on experience or manual methods, resulting in low measurement efficiency and complex measurement operations. Traditional computational simulations, on the other hand, rely heavily on model quality and component dimensions for accuracy and efficiency, leading to high computational costs, insufficient generalization ability of the welding model, and the need for repeated calculations. Furthermore, adjusting the model to accelerate computation often results in a loss of accuracy. Using internally stored databases requires large amounts of data, and simple fitting methods cannot effectively improve computational accuracy or effectively control welding deformation and stress during the welding process.

[0005] Long Short-Term Memory (LSTM) networks are recurrent neural networks (RNNs) specifically designed for modeling sequential data and learning long-term temporal dependencies. Unlike feedforward neural networks that treat each sample as an independent entity, LSTMs process input data sequentially and maintain an internal cell state that acts as a memory of historical information. At each time step, LSTMs use a set of gating mechanisms (typically including forget gates, input gates, and output gates) to regulate the retention, updating, and output of information in the cell state. By controlling the flow of information through these gating mechanisms, LSTMs mitigate the vanishing gradient problem common in traditional RNNs, enabling them to stably learn dependencies over long time spans. LSTM networks have unique advantages in predicting relevant temporal signals, particularly suitable for predicting and fitting time-series data on stress and strain in laser welding. Summary of the Invention

[0006] To address the shortcomings and improvement needs of existing technologies, this invention provides a rapid prediction and control method for welding stress and strain based on LSTM networks.

[0007] The specific implementation method of this invention is as follows: a fast prediction and control method for welding stress and strain based on LSTM network, the method comprising:

[0008] Step 1: Organize the multi-source input information related to welding conditions into a unified time sequence;

[0009] Step 2: Based on the welding path and the geometric relationship of the workpiece, determine the set of key locations in the welding area for mechanical prediction;

[0010] Step 3: After associating the multi-source input information with the key location set, input the information into a pre-trained long short-term memory network to obtain the stress and strain prediction results of the key locations in the future time period.

[0011] Step 4: Based on the degree of deviation between the predicted results and historical welding mechanical characteristics, determine the quality trend of the welding process;

[0012] Step 5: Based on the quality trend, select either to perform the working mode of only performing prediction output, or to perform the working mode of adjusting the welding mechanical state based on the prediction results.

[0013] Step 6: Store the input information, prediction results, quality judgment results, and control behaviors during the welding process in a unified manner.

[0014] Furthermore, the multi-source input information includes at least welding process parameters, welding path parameters, and workpiece structure parameters, and the multi-source input information constitutes a time sequence input according to the welding time order.

[0015] Furthermore, the set of key locations includes multiple spatial locations in the weld area, heat-affected zone, and substrate area, used to characterize the stress and strain evolution characteristics of different mechanical response areas during the welding process.

[0016] Another objective of this invention is to provide a time-series prediction method for welding stress and strain, wherein the time-series prediction method for welding stress and strain includes:

[0017] By standardizing welding condition parameters, path parameters, and constraint parameters, and introducing derived features reflecting the relationship between welding heat input and mechanical constraints into the normalized features,

[0018] A structured feature sequence is constructed for input to a long short-term memory network, enabling the long short-term memory network to simultaneously learn the time dependence and physical constraint relationships in the welding process, thereby outputting stress and strain prediction results at multiple locations.

[0019] Furthermore, the derived features include at least heat input features characterizing the intensity of welding heat input, and constraint features characterizing the strength of clamping constraints.

[0020] The constraint features are determined by the magnitude of the clamping force and the spatial relationship between the clamping position and the predicted position.

[0021] Furthermore, the structured feature sequence is constructed in the same way as the model training phase during the prediction phase, and is managed uniformly through versioning to avoid input distribution shift.

[0022] Another objective of this invention is to provide a fast prediction and control system for welding stress and strain based on an LSTM network, the system comprising:

[0023] The data organization module is used to perform time-series processing on multi-source input information related to the welding process;

[0024] The critical location identification module is used to determine the critical locations for mechanical prediction based on the welding path and workpiece structure.

[0025] The prediction module is used to input the multi-source input information and the key locations into a long short-term memory network to obtain the time-series prediction results of stress and strain;

[0026] The decision-making module is used to determine the welding quality trend based on the prediction results and to choose whether to trigger control actions.

[0027] The execution and recording module is used to execute the corresponding welding control operations and store the data of the entire process.

[0028] Furthermore, the decision-making module is configured to support at least two operating states, including a prediction state that only outputs prediction results, and a control state that regulates the welding process based on the prediction results.

[0029] Furthermore, the prediction module employs a long short-term memory network that includes at least one memory layer for extracting welding time-dependent features, and at least one output layer for outputting multi-position stress-strain prediction results.

[0030] Furthermore, the execution and recording module is configured to associate and store welding input information, prediction results, quality judgment results, and control behaviors to form a traceable welding process data record.

[0031] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0032] First, it significantly reduces computation time and costs. While maintaining basic computational accuracy to meet industrial needs, computation time is reduced by 98%. As shown in the figure below, in a preliminary calculation (with laser welding power of 15 kW, welding speed of 20 mm / s, workpiece being welded as S700MC high-strength steel, workpiece thickness of 20 mm, workpiece length of 2200*1000*20, butt weld), using Ansys would take approximately 12 hours. Calculations using experimental instruments would require 4 days. This invention, based on the LSTM algorithm, achieves a minimum computational accuracy of 86% (compared to ANSYS), calculating stress and strain at a point in 67 seconds. This significantly reduces computation time and production costs.

[0033] This invention can be applied to various welding butt joint states (in the detection mode, the model learns the features of relevant data), can adapt to multiple states, and has an accuracy rate of over 80% for cross-working conditions, which can meet the requirements of normal work.

[0034] The control mode of this invention can significantly reduce welding stress and strain, with a minimum reduction of 80%. In the control mode, the welding force is controlled by an LSTM model, thereby achieving precise regulation of the welding force.

[0035] Secondly, as supplementary evidence of the inventive step of the claims of this invention, it is also reflected in the following important aspects:

[0036] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0037] The expected benefits after the technical solution of this invention is transformed are as follows:

[0038] There are two main methods for transforming the invention: its fast calculation mode can be directly applied to the welding calculation process, or the relevant network model can be embedded into SolidWorks or ANSYS as a plug-in module for fast pre-calculation during simulation. This has significant commercial value and can generate substantial expected revenue.

[0039] Meanwhile, the monitoring mode of this invention can be directly applied to the production process of relevant welding production lines, enabling parameter traceability and overall intelligentization of the welding production line. Furthermore, its intelligent control mode can be directly applied to intelligent welding production lines, significantly reducing welding scrap rates, lowering R&D and production costs, possessing significant commercial value, and potentially yielding substantial expected returns.

[0040] (2) The technical solution of this invention fills a technical gap in the industry both domestically and internationally:

[0041] This invention fills a gap in the intelligent prediction and advanced control of stress and strain in laser welding, both domestically and internationally. Existing technologies in the industry often rely on optimizing welding paths, exploring based on relevant experience, or post-weld adjustments. Calculations typically use specialized ANSYS simulation software, which is time-consuming, costly, and lacks real-time monitoring and control technologies. This invention, based on the characteristics of LSTM neural networks and addressing the highly time-series nature of welding process data, enables real-time stress control to ensure welding quality. It fills the gap in the industry's inability to control welding deformation and stress in real time. It has significant commercial value and is expected to yield good returns.

[0042] (3) The technical solution of the present invention solves a technical problem that people have long wanted to solve but have never been able to solve successfully:

[0043] This invention solves a long-standing technical problem that has hindered the rapid, high-fidelity calculation of stress and strain in welding. Previously, rapid calculations often required adjusting mesh parameters or optimizing the model, sacrificing significant simulation accuracy and increasing operational complexity, demanding highly skilled technicians. This invention, based on an LSTM neural network, establishes a high-fidelity model for rapidly predicting welding stress and strain. In basic data comparison and analysis, the prediction accuracy for welding stress and strain exceeds 80%, resolving a long-standing and unsolved technical challenge.

[0044] Meanwhile, this invention also solves the long-standing problem of unsuccessfully controlling the stress and strain of welding in real time. Based on the strong temporal characteristics of welding physical signals, this invention uses an LSTM network to predict relevant data in advance, providing a basis for real-time control of welding. Furthermore, in the control mode, the method provided by this invention achieves good results in controlling the stress and strain of welding, solving the problem of previously unsuccessful or ineffective stress and strain control in welding.

[0045] (4) The technical solution of the present invention overcomes technical bias:

[0046] This invention overcomes the technical bias of whether welding stress and strain can be precisely controlled in advance. Previously, welding stress and strain were often controlled using traditional welding processes or by over-positioning and clamping with fixtures, often resulting in over-adjustment and untimely control after welding. This invention, based on an LSTM network, achieves rapid prediction and control of welding stress and strain, overcoming the limitations of previous methods that could not predict and adjust welding stress with high fidelity or achieve precise control, ultimately achieving better results. Attached Figure Description

[0047] Figure 1 A flowchart of a rapid prediction and control method for welding stress and strain based on LSTM network is provided for an embodiment of the present invention;

[0048] Figure 2 A comparison diagram of the predicted force and the actual force of the model in a real working scenario provided by an embodiment of the present invention;

[0049] Figure 3 A detailed flowchart of a rapid prediction and control method for welding stress and strain based on LSTM network provided in an embodiment of the present invention;

[0050] Figure 4 Specific operation interface diagrams (fast calculation mode) provided for embodiments of the present invention;

[0051] Figure 5 This is an operation interface diagram (detection mode) provided for an embodiment of the present invention;

[0052] Figure 6 This is an operation interface diagram (advanced control mode) provided for an embodiment of the present invention.

[0053] Figure 7 A block diagram of a welding stress-strain fast prediction and control system based on an LSTM network is provided for an embodiment of the present invention.

[0054] Figure 8This is a diagram of the interface for detecting welding stress in detection mode, provided as an embodiment of the present invention.

[0055] Figure 9 This is a diagram of an interface for controlling a simple welding process in control mode, provided as an embodiment of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0057] like Figure 1 As shown, this embodiment of the invention provides a fast prediction and control method for welding stress and strain based on LSTM networks, including:

[0058] S101, performs unified timing organization of multi-source input information related to welding conditions;

[0059] S102, Based on the geometric relationship between the welding path and the workpiece, determine the set of key locations in the welding area for mechanical prediction;

[0060] S103, after the multi-source input information is matched with the key location set, it is input into a pre-trained long short-term memory network to obtain the stress and strain prediction results of the key locations in the future time period.

[0061] S104, Based on the degree of deviation between the predicted results and historical welding mechanical characteristics, determine the quality trend of the welding process;

[0062] S105, based on the quality trend, select to execute either the working mode that only performs prediction output or the working mode that adjusts the welding mechanical state based on the prediction results.

[0063] S106, which uniformly stores the input information, prediction results, quality judgment results and control behaviors during the welding process.

[0064] The embodiments of the present invention provide multi-source input information including at least welding process parameters, welding path parameters, and workpiece structure parameters, wherein the multi-source input information constitutes a time sequence input in the order of welding time.

[0065] The embodiments of the present invention provide a set of key locations including multiple spatial locations in the weld area, heat-affected zone and substrate area, which are used to characterize the stress and strain evolution characteristics of different mechanical response areas during the welding process.

[0066] This invention provides a time-series prediction method for welding stress and strain, the method comprising:

[0067] By standardizing welding condition parameters, path parameters, and constraint parameters, and introducing derived features reflecting the relationship between welding heat input and mechanical constraints into the normalized features,

[0068] A structured feature sequence is constructed for input to a long short-term memory network, enabling the long short-term memory network to simultaneously learn the time dependence and physical constraint relationships in the welding process, thereby outputting stress and strain prediction results at multiple locations.

[0069] The embodiments of the present invention provide derived features including at least a heat input feature for characterizing the intensity of welding heat input, and a constraint feature for characterizing the strength of clamping constraint.

[0070] The constraint features are determined by the magnitude of the clamping force and the spatial relationship between the clamping position and the predicted position.

[0071] The present invention provides a structured feature sequence that is consistent with the feature construction method in the prediction stage and the model training stage, and is managed uniformly through versioning to avoid input distribution shift.

[0072] like Figure 2 This is a comparison chart between the model's predicted force and the actual force in a real-world working scenario. In terms of prediction accuracy, the prediction accuracy exceeds 80%.

[0073] like Figure 3 As shown, this embodiment of the invention provides a fast prediction and control method for welding stress and strain based on LSTM networks, which includes the following detailed components:

[0074] (1): Preparatory work, mainly to prepare for the system to start working;

[0075] (2): The purpose of the input parameter acquisition stage is to input the preconditions;

[0076] (3): After the above preprocessing, key points will be displayed. Based on the welding path and geometry, the system identifies and displays the key locations that need to be predicted for stress and strain.

[0077] (4): After completing the preliminary preparations, input the relevant data into the trained LSTM network and make accurate predictions on the stress and strain of the welded workpiece during the experiment.

[0078] (5): In order to meet the requirements of different experimental conditions and working scenarios, after basic calculation, the system has the following three working modes according to the specific working scenario: fast calculation mode, detection mode and advanced control mode; in order to adapt to laser welding of various docking conditions, based on the data characteristics of the previous relevant welding conditions, new data characteristics are predicted and the basic working mode is performed; the characteristics of the stress and strain related data of the welding stored in the previous period are statistically analyzed for the current data characteristics. If the data at the relevant moment exceeds the threshold, there is a relatively high possibility of welding failure. At the same time, based on this characteristic, the welding characteristics of various docking states are identified and controlled, thereby achieving better welding.

[0079] (6): After the above process is completed, the relevant data will be saved to generate a work log for subsequent query.

[0080] like Figure 3 As shown, (2) specifically includes:

[0081] The prerequisites are mainly divided into three categories: welding process parameters, welding path parameters, and workpiece geometric parameters. Among them, the welding process parameters mainly include important parameters such as laser welding power, spot energy distribution, defocusing amount, and bevel shape. These parameters collectively determine the linear energy coefficient. Size;

[0082]

[0083] here V is the linear energy coefficient, P is the power of laser welding (in W), and V is the welding speed (in mm / s).

[0084] At the same time, it should be noted that the linear energy coefficient varies depending on the different laser welding process parameters mentioned above. It will also change with these process parameters. The subsequent model has embedded the correspondence between different process parameters and line energy absorption coefficient to ensure the accuracy of welding deformation prediction and results. A large amount of finite element and experimental data has confirmed that the line energy index is strongly correlated with the magnitude of residual stress.

[0085] The second type of parameter is the welding path parameter, which mainly includes the geometric classification of the welding path (linear, circular, or complex), the discretized distribution of the path, and the coordinates of the start and end points of the welding path and their relative positions in the welded workpiece. Extensive experimental and simulation data have shown a strong correlation between welding path parameters and welding stress and strain; the welding path affects the deformation force and amount of deformation. Path discretization captures the temporal characteristics of the welding operation. Since welding is essentially a gradual process of different regions being heated at different times, the spatial sequence of the path directly affects the overlapping and interaction of thermal cycles; especially in multi-pass welding operations, the final stress distribution exhibits significant path dependence.

[0086] The third type is the input of workpiece geometric parameters, which mainly includes the workpiece size, joint type, material type, and clamping configuration. The clamping configuration mainly includes the contact type of the fixture, the position of the fixture's action point, and the magnitude of the clamping force. The clamping configuration can be used to guide the model in rapid calculation and inspection in the early stages. At the same time, in the control mode, the model can predict the welding stress in advance, and then use the fixture to make targeted adjustments to the welding stress and strain to achieve better welding and ensure the quality of welding.

[0087] The physical meaning of each point in (3) is as follows:

[0088] Weld centerline points: Located along the centerline of the fusion zone, they are discretely distributed along the path; these points typically exhibit the highest tensile residual stress. Heat-affected zone boundary points: Perpendicular to the weld centerline, approximately 5-8 mm away; the heat-affected zone is the transition area between the weld metal and the unaffected substrate, whose material properties change due to heat exposure. Substrate reference points: Located 15-20 mm from the weld centerline in the unaffected area; these points typically exhibit compressive stress to balance the tensile stress between the weld and the heat-affected zone.

[0089] After basic processing, the input preprocessing and feature construction follow. First, the range and physical consistency of the input are verified. Range verification is to verify whether each parameter combination is within the range that ensures good welding quality. The specific range is determined by the laser welding parameters, the workpiece to be welded, and the aforementioned parameters. For physically impossible parameter combinations, the completeness of the parameters is also checked. All necessary parameters must be provided. Missing key parameters will lead to prediction failure and will display a clear error message indicating the required parameters.

[0090] Next, feature normalization is performed, with each input parameter undergoing min-max normalization. Normalization balances the influence of each parameter, enabling the network to learn the true relative importance of each feature through training. The normalization formula is shown below:

[0091]

[0092] After normalization, the system calculates derived features incorporating physical information and encapsulates important parameter relationships; its main parameters are as follows:

[0093] Heat input: Its importance has been discussed above, specifically regarding the coefficient of linear energy. Targeted corrections were made to improve the accuracy of predicting residual stress in welding;

[0094] Clamping stiffness coefficient: For each key point, the system calculates the constraint strength index based on the distance and magnitude of all clamping forces; the main calculation method is as follows:

[0095]

[0096] Where Fi is the clamping force of the i-th fixture, and di is the distance from the fixture to the evaluation point; this index reflects the mechanical boundary conditions: the high clamping force near the weld restricts thermal expansion, generates greater compressive plastic strain during heating, and thus forms higher tensile residual stress after cooling.

[0097] Thermal severity parameter: An index that comprehensively characterizes heating intensity and cooling rate, calculated from heat input and material thickness; this parameter helps the model distinguish between scenarios with similar heat inputs but different thermal cycle characteristics;

[0098] After performing the above preprocessing data calculations and encapsulation, all normalized features are assembled into a structured input array. For a typical single-point static prediction scenario, the input array dimension is (1, 1, n_features), where n_features typically varies between 8 and 15 depending on the number of fixtures and optional parameters. The dimension of the dynamic test scenario will increase to include the time series.

[0099] The feature order in the input vector is standardized to match the training configuration: [power, energy distribution, defocus, thermal input, clamp 1 distance, clamp 1 force, clamp 2 distance, clamp 2 force, ..., thickness, ...].

[0100] The preprocessing stage is crucial to prediction accuracy; the normalization parameters and feature engineering process must be completely consistent with those used during model training; any discrepancies will lead to distribution shifts, resulting in prediction errors even if the base LSTM model is accurate; therefore, the preprocessing module and normalization constants are version-controlled and deployed together with the LSTM model as an integrated system.

[0101] Specifically, (4) includes:

[0102] First, the model is loaded and initialized, instantiating the complete network structure, including all LSTM layers, Dropout layers, fully connected layers, and output layers, along with specific configurations. Next, the relevant weight parameters are loaded, loading the training weight matrix W and bias vector b for all layers from the saved model file—a total of approximately 140,000-200,000 parameters. These parameters encode the learned mapping from welding parameters to stress-strain distribution, representing the knowledge accumulated from hundreds or thousands of training samples. Then, the evaluation device is selected: for basic computation and detection, CPU is used; for advanced tuning, GPU is used for accelerated computation. In evaluation mode, the network provides deterministic predictions—the same input always produces the same output.

[0103] At the specific network structure level, the network adopts a five-layer architecture: two LSTM layers, containing 128 and 64 memory units respectively, for temporal feature extraction; two fully connected layers, with 128 and 64 units and ReLU activation, for nonlinear feature transformation; and a linear output layer with 32 units, generating multi-point stress-strain predictions; the total number of parameters is approximately 150,000. The LSTM layers achieve selective information transmission through a gating mechanism: the forget gate controls the retention of historical information, the input gate controls the reception of new information, and the output gate controls the output of information. Cell states serve as long-term memory carriers, enabling the network to capture the temperature evolution characteristics across time steps in the welding thermal cycle. The first LSTM layer, with 128 units, extracts primary temporal features, and the second LSTM layer, with 64 units, performs hierarchical feature refinement, with only the hidden state at the final moment being passed to subsequent fully connected layers. By combining physical information with LSTM, welding calculations can be performed for different laser welding conditions, and support is also provided for mode selection.

[0104] like Figure 4 As shown, the fast calculation mode in (5) specifically includes:

[0105] This mode is suitable for rapid testing of stress and strain before welding. Previous experiments and simulations often lacked reference data, requiring trial welding or coarsening of the simulation mesh parameters to reduce the number of experiments or save simulation time before formal welding or simulation. This required a high level of experience from operators. The testing mode provides guidance before experiments and simulations by outputting welding stress and strain under relevant working conditions, allowing operators to make a preliminary judgment on the feasibility of welding. Furthermore, by adjusting welding process parameters and clamping parameters, the relevant process window can be quickly found, improving welding efficiency and reducing welding costs.

[0106] like Figure 5 As shown, the detection mode specifically includes: This mode is applicable to the quality assurance stage of the welding process; During the welding process, the stress and deformation during welding can be detected in real time by relevant fixtures equipped with force sensors. The relevant detection information will be transmitted to the system and compared with the data calculated by the system; Based on the 3σ principle of normal distribution, if the relevant stress and strain exceed the range, the quality of the welding can be judged, and the relevant workpiece needs to be re-inspected to ensure the quality of the welding; At the same time, if the relevant data meets the calculation requirements, the system will also record the relevant situation to provide guidance for the next production process and facilitate the assembly of the workpieces in the later stages;

[0107] like Figure 6 As shown, the advanced control mode specifically includes the following: The preparation and calculation process of the advanced control mode are the same as those of the detection mode. In particular, during the welding process, based on the data currently obtained on the fixture, an advanced prediction is made for the data in the next 60 milliseconds. The prediction result is compared with the result calculated in the model to determine the quality trend of the welding system. If the welding quality meets the requirements, no operation will be performed. If the requirements are not met, the system will guide the fixture to adjust the clamping force based on the predicted data and the calculated data, thereby ensuring good welding quality.

[0108] The specific information to be stored in (6) mainly includes input information, welding process parameters, welding path parameters, workpiece parameters, welding start and end times, dynamic comparison data between calculated results and actual results, detection mode and advanced control mode, laser welding quality judgment, advanced control intervention time period, and analysis data and field data of welding stress and strain at key points of the whole.

[0109] Meanwhile, to facilitate future system maintenance and upgrades, in addition to the built-in model, if new experimental or simulation data becomes available later, supplementary training data (including experimental and simulation data) can be added back to the welding stress-strain rapid calculation and prediction model. This supplements the model's relevant knowledge base, further ensuring the quality of welding and the quality and efficiency of stress-strain prediction.

[0110] In summary, this method and system for rapid calculation and prediction of stress and strain in laser welding, trained on a prediction model using limited data from existing experiments and high-precision simulations, significantly reduces simulation time and experimental costs while maintaining basic prediction accuracy. Furthermore, given the highly time-series nature of welding data, this invention specifically utilizes an LSTM model to predict welding data, enabling the identification of stress and strain trends and allowing for targeted control of the welding process, thereby ensuring welding quality.

[0111] like Figure 7 This invention provides a fast prediction and control system for welding stress and strain based on an LSTM network, the system comprising:

[0112] The data organization module is used to perform time-series processing on multi-source input information related to the welding process;

[0113] The critical location identification module is used to determine the critical locations for mechanical prediction based on the welding path and workpiece structure.

[0114] The prediction module is used to input the multi-source input information and the key locations into a long short-term memory network to obtain the time-series prediction results of stress and strain;

[0115] The decision-making module is used to determine the welding quality trend based on the prediction results and to choose whether to trigger control actions.

[0116] The execution and recording module is used to execute the corresponding welding control operations and store the data of the entire process.

[0117] The present invention provides a decision module configured to support at least two working states, including a prediction state that only outputs prediction results, and a control state that regulates the welding process based on the prediction results.

[0118] The long short-term memory network used in the prediction module provided in this embodiment of the invention includes at least one memory layer for extracting welding time-dependent features, and at least one output layer for outputting multi-position stress-strain prediction results.

[0119] The present invention provides an execution and recording module configured to associate and store welding input information, prediction results, quality judgment results and control behaviors to form a traceable welding process data record.

[0120] The welding stress-strain rapid prediction and control system based on LSTM network provided in this invention achieves forward-looking judgment and adaptive control of welding quality by temporal modeling and mechanical state prediction of multi-source welding information during the welding operation. During system operation, the data organization module first uniformly accesses and preprocesses the multi-source input information collected in real time during the welding process. This multi-source input information includes, but is not limited to, welding current, voltage, welding speed, welding torch pose, welding path parameters, workpiece material properties, and environmental parameters. The data organization module synchronizes, aligns, and temporally processes the above information according to the welding timeline, constructing time-series data that meets the input requirements of long short-term memory networks, thereby providing stable and continuous feature inputs for subsequent predictions.

[0121] After data organization, the key location identification module determines representative key locations along the weld line and in its neighborhood, based on preset welding path information and workpiece geometry model, which are representative of the evolution of welding stress and strain. By spatially mapping the welding path to the workpiece structure, the system can automatically identify areas prone to stress concentration or deformation sensitivity, giving subsequent mechanical predictions a clear spatial orientation and avoiding redundant calculations of the entire location, thereby improving prediction efficiency.

[0122] The prediction module inputs the time-series processed multi-source input information along with the key location information into a Long Short-Term Memory (LSTM) network. The LSM network models the time-dependent characteristics of the welding process through its memory layer, effectively capturing the evolution of welding heat input, heat conduction, and mechanical response over time. The output layer simultaneously provides predictions of stress and strain at multiple key locations. These predictions are output in time-series format, enabling rapid prediction of the changing trends of the mechanical state during welding.

[0123] The decision-making module determines the welding quality trend based on the stress-strain time series results output by the prediction module. In the prediction state, the decision-making module only analyzes and outputs the prediction results for process monitoring or process evaluation; in the control state, the decision-making module further compares the prediction results with preset quality criteria, and generates corresponding control decision instructions when the prediction results indicate the presence of welding defect risks or quality deterioration trends.

[0124] The execution and recording module performs corresponding welding control operations according to the instructions of the decision-making module, such as adjusting welding current, welding speed or heat input parameters. At the same time, it links and stores welding input information, prediction results, quality judgment results and actual control behaviors to form a complete and traceable welding process data record, providing data support for subsequent quality analysis, model optimization and process improvement.

[0125] Example 1: Rapid Prediction of Laser Butt Welding

[0126] In this embodiment, as Figure 2 As shown, taking thick stainless steel laser butt welding as an example, welding power, welding speed, defocusing amount, welding path geometry information, and workpiece size information are collected before welding begins. The system automatically identifies the weld centerline and key predicted positions near it based on the welding path, and constructs the above parameters into an input sequence according to the welding time sequence. After inputting into a trained Long Short-Term Memory (LSTM) network, without any control intervention, the system only outputs the stress and strain prediction results at each key position during the welding process for pre-welding process feasibility assessment. This embodiment verifies the system's ability to rapidly obtain welding mechanical trends under experimental and simulation-free conditions.

[0127] Example 2: Cross-condition prediction under multi-pass welding path conditions

[0128] In this embodiment, as Figure 6 As shown, for welding conditions involving multiple welding passes and multiple butt joints, the system introduces discretized spatial sequence information of the welding path during the input stage, enabling continuous input across different weld passes in the time dimension. The system simultaneously covers the intersection areas of each weld pass when identifying key locations. The Long Short-Term Memory (LSTM) network utilizes its temporal memory capability to capture the influence of the previous welding thermal cycle on the stress evolution of the subsequent weld pass, achieving stress-strain prediction across weld passes. This embodiment demonstrates the overall modeling capability of this technology for path-dependent welding mechanical behavior.

[0129] In this embodiment, during the welding process, sensing units on the fixture collect welding deformation and stress information in real time, and compare it synchronously with the system's prediction results. The system performs statistical analysis on the deviation between the predicted and detected values. When the deviation exceeds a preset threshold, it determines that there is an abnormal risk in the welding quality and records the corresponding time period and spatial location. This embodiment demonstrates the technical effect of online welding quality assessment achieved through the collaborative work of the prediction model and real-time detection.

[0130] Example 3: Example of proactive regulation based on prediction results

[0131] like Figure 6In this embodiment, the system predicts stress and strain in advance for future time periods during the welding process. When the prediction results indicate an out-of-limit trend at a critical location, the system issues a control command before welding is completed to adjust the clamping force, thereby changing the subsequent thermal deformation path. After the adjustment is completed, the system continuously updates the prediction results and evaluates the control effect, achieving closed-loop operation. This embodiment demonstrates an active intervention mechanism where welding control is directly driven by prediction results.

[0132] Example 4: Adaptive Examples with Different Materials and Clamping Configurations

[0133] like Figure 8 In this embodiment, different materials and clamping arrangements are used, adjusting only the input parameters without modifying the model structure. During the prediction phase, the system incorporates clamping constraint strength-related features, enabling the model to automatically distinguish welding stress evolution characteristics under different boundary conditions. This embodiment demonstrates that this technology is not dependent on fixed working conditions but possesses adaptive capabilities to changes in materials and constraint conditions.

[0134] Example 6: Welding Process Data Accumulation and Continuous Model Optimization

[0135] like Figure 6 In this embodiment, the system uniformly stores the input parameters, prediction results, detection results, and control behaviors for each welding process, forming a continuous welding dataset. This dataset can then be used to retrain or fine-tune the model, thereby gradually improving prediction accuracy and control performance. This embodiment illustrates that this technology is not only applicable to single welding tasks but can also achieve long-term performance improvements through data accumulation.

[0136] Evidence related to the technical effects obtained by the embodiments of the present invention.

[0137] In a fast calculation mode, such as Figure 4 As shown, this is the interface in the fast calculation mode. With the relevant training model, the stress prediction time is only 0.03s, while the calculation time of the same ANSYS model is at least 2 hours and 7200s. The method provided by this invention has excellent computational performance.

[0138] In the inspection mode, the inspection is based on the stress of the weld, such as... Figure 8 This invention addresses the stress and strain handling process during welding in practical applications. The method proposed in this invention specifically monitors the entire welding process, facilitating subsequent traceability. It also simultaneously observes the trends in stress and strain during welding, allowing for deviation analysis and subsequent parameter adjustments.

[0139] In control mode, such as Figure 9As shown, by controlling the simple welding process, the welding stress was specifically controlled, and the relevant welding stress and strain were reduced to a minimum within the sensor resolution range.

[0140] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0141] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A fast prediction and control method for welding stress and strain based on LSTM network, characterized in that, Step 1: Organize the multi-source input information related to welding conditions into a unified time sequence; Step 2: Based on the welding path and the geometric relationship of the workpiece, determine the set of key locations in the welding area for mechanical prediction; Step 3: After associating the multi-source input information with the key location set, input the information into a pre-trained long short-term memory network to obtain the stress and strain prediction results of the key locations in the future time period. Step 4: Based on the degree of deviation between the predicted results and historical welding mechanical characteristics, determine the quality trend of the welding process; Step 5: Based on the quality trend, select either to perform the working mode of only performing prediction output, or to perform the working mode of adjusting the welding mechanical state based on the prediction results. Step 6: Store the input information, prediction results, quality judgment results, and control behaviors during the welding process in a unified manner.

2. The method according to claim 1, characterized in that, The multi-source input information includes at least welding process parameters, welding path parameters, and workpiece structure parameters, and the multi-source input information is arranged in the order of welding time to form a time sequence input.

3. The method according to claim 1, characterized in that, The set of key locations includes multiple spatial locations in the weld area, heat-affected zone, and substrate area, used to characterize the stress and strain evolution characteristics of different mechanical response areas during the welding process.

4. A time-series prediction method for welding stress and strain, implementing the fast prediction and control method for welding stress and strain based on LSTM networks as described in any one of claims 1-3, characterized in that, The time-series prediction method for welding stress and strain shown includes: By standardizing welding condition parameters, path parameters, and constraint parameters, and introducing derived features reflecting the relationship between welding heat input and mechanical constraints into the normalized features, A structured feature sequence is constructed for input to a long short-term memory network, enabling the long short-term memory network to simultaneously learn the time dependence and physical constraint relationships in the welding process, thereby outputting stress and strain prediction results at multiple locations.

5. The method according to claim 4, characterized in that, The derived features include at least heat input features characterizing the intensity of welding heat input, and constraint features characterizing the strength of clamping constraints. The constraint features are determined by the magnitude of the clamping force and the spatial relationship between the clamping position and the predicted position.

6. The method according to claim 4, characterized in that, The structured feature sequences are constructed in the same way during the prediction phase as during the model training phase, and are managed uniformly through versioning to avoid input distribution shifts.

7. A welding stress-strain rapid prediction and control system based on an LSTM network, implementing the welding stress-strain rapid prediction and control method based on any one of claims 1-3, characterized in that, The system includes: The data organization module is used to perform time-series processing on multi-source input information related to the welding process; The critical location identification module is used to determine the critical locations for mechanical prediction based on the welding path and workpiece structure. The prediction module is used to input the multi-source input information and the key locations into a long short-term memory network to obtain the time-series prediction results of stress and strain; The decision-making module is used to determine the welding quality trend based on the prediction results and to choose whether to trigger control actions. The execution and recording module is used to execute the corresponding welding control operations and store the data of the entire process.

8. The system according to claim 7, characterized in that, The decision module is configured to support at least two operating states, including a prediction state that only outputs prediction results, and a control state that regulates the welding process based on the prediction results.

9. The system according to claim 7, characterized in that, The prediction module employs a long short-term memory network, which includes at least one memory layer for extracting welding time-dependent features, and at least one output layer for outputting multi-position stress-strain prediction results.

10. The system according to claim 7, characterized in that, The execution and recording module is configured to associate and store welding input information, prediction results, quality judgment results, and control behaviors to form a traceable welding process data record.

Images