Steel bridge deck slab welding deformation intelligent prediction method based on physical-virtual double-domain cooperation and CNN-LSTM fusion modeling

Abstract

The invention discloses an intelligent prediction method for welding deformation of a steel bridge deck based on physical-virtual double-domain collaboration and CNN-LSTM fusion modeling, relates to the technical field of welding process optimization and intelligent manufacturing, and aims to solve the problem that traditional welding deformation control mainly depends on empirical formula derivation and finite element thermoelastic-plastic analysis. However, the former is difficult to accurately reflect a nonlinear relation of a complex welding condition, and the latter has the problems of high computing resource consumption, low modeling efficiency and the like, and particularly, the prediction precision and the real-time performance are difficult to meet intelligent manufacturing requirements in the face of a multi-welding-seam coupling effect and a large-scale structure. The invention provides an intelligent prediction method for welding deformation of a steel bridge deck based on physical-virtual double-domain cooperation and CNN-LSTM fusion modeling, and the method comprises the steps: deeply fusing the dynamic mapping capability of the physical-virtual double-domain cooperation and the spatial-temporal feature modeling advantages of a CNN-LSTM fusion model, and synchronously integrating an NSGA-II algorithm to construct a process optimization closed-loop system; and high-precision prediction and process dynamic optimization of the welding deformation of the steel bridge deck are realized.

Description

Technical Field

[0001] This invention relates to the fields of welding process optimization and intelligent manufacturing technology, specifically to an intelligent prediction method for welding deformation of steel bridge decks based on physical-virtual dual-domain collaboration, CNN-LSTM fusion modeling, and NSGA-II multi-objective optimization algorithm. It is particularly suitable for welding U-ribs of orthotropic steel bridge decks with plate thickness of 12-40mm, and can achieve synergistic optimization of deformation control and manufacturing efficiency. Background Technology

[0002] In recent years, with the rapid development of long-span steel bridges, orthotropic steel bridge decks have become a core structural form in modern bridge engineering due to their excellent mechanical properties and ease of construction. However, during the welding and manufacturing process of U-ribs in steel bridge decks, the non-uniform thermal stress field generated by local high-temperature thermal cycling leads to complex residual deformation in the welded area. This deformation not only affects the geometric accuracy and assembly quality of the components but also causes stress concentration, significantly reducing the fatigue life and load-bearing capacity of the structure. Traditional welding deformation control mainly relies on empirical formula derivation and finite element thermo-elastic-plastic analysis. However, the former is difficult to accurately reflect the nonlinear relationship of complex welding conditions, while the latter suffers from high computational resource consumption and low modeling efficiency. Especially when facing the coupling effect of multiple welds and large-scale structures, the prediction accuracy and real-time performance are difficult to meet the needs of intelligent manufacturing.

[0003] Current applications of digital twin technology in the welding field are mostly one-way data mapping from the physical domain to the virtual domain, lacking real-time feedback correction from the virtual domain to the physical domain, resulting in insufficient synergy between the two domains. At the same time, existing deep learning models either rely solely on CNN to extract spatial features or solely on LSTM to capture temporal features, making it difficult to take into account both the 'spatial distribution complexity' and 'temporal evolution continuity' of welding deformation, thus limiting prediction accuracy.

[0004] Digital twin technology provides a new approach for the full-element digital reproduction of the welding process by constructing a real-time interactive mapping between physical and virtual spaces. Deep learning technology exhibits unique advantages in handling high-dimensional nonlinear problems, particularly the ability of convolutional neural networks to extract spatial features and the ability of long short-term memory networks to model temporal dynamics, providing technical support for the in-depth mining of multi-source heterogeneous welding data. The NSGA-II algorithm, as an efficient multi-objective optimization tool, can find Pareto optimal solutions among conflicting objectives, achieving multi-objective balance without pre-setting weights. Therefore, this invention deeply integrates physical-virtual dual-domain collaboration, CNN-LSTM fusion modeling, and the NSGA-II algorithm to propose an intelligent prediction and process optimization method for welding deformation of steel bridge decks. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides an intelligent prediction method for steel bridge deck welding deformation based on physical-virtual dual-domain collaboration and CNN-LSTM fusion modeling, specifically solving the problems of low accuracy, poor efficiency, and difficulty in balancing multiple objectives in traditional methods.

[0006] The main idea of ​​the technical solution adopted in this invention is as follows: The "physical-virtual dual-domain data channel" is a two-way real-time communication and data interaction mechanism connecting the welding physical site (physical domain) and the digital twin simulation system (virtual domain). Its technical implementation relies on industrial Ethernet transmission and edge computing modules. By constructing a multi-source data fusion system, it integrates multi-dimensional data such as welding current, temperature field, and three-dimensional deformation. An innovative CNN-LSTM finite element hybrid model is established, utilizing convolutional networks to process 1280×1024 pixel thermal images, LSTM to capture 100Hz temporal features, and coupling with elastoplastic finite element physical field calculations. Multimodal feature fusion is achieved through an attention mechanism. The system is equipped with a dynamic twin iteration mechanism, updating the material constitutive model every 5 seconds to correct the elastic modulus or thermal expansion coefficient at that temperature. Combined with transfer learning, it adapts to plate thickness conditions of 12–40 mm, achieving a prediction deviation control accuracy of ±0.3 mm. A closed-loop process optimization system based on the NSGA-II algorithm was developed concurrently. Through non-dominated sorting, crowding calculation, and elite retention strategies, it finds the Pareto optimal solution between the two conflicting objectives of "minimizing welding deformation" and "maximizing welding efficiency". It can generate a recommended welding sequence scheme within 30 seconds and dynamically adjust the heat input. It improves the computational efficiency by 80 times compared with the traditional finite element method. In actual engineering applications, it reduces welding deformation by 37%, providing an intelligent solution for the manufacturing of complex steel structures.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A smart prediction method for welding deformation of steel bridge deck based on physical-virtual dual-domain collaboration and CNN-LSTM fusion modeling includes the following steps: Welding parameters during the steel bridge deck welding process are collected and recorded in real time. A three-dimensional thermo-elastic-plastic finite element model of the steel bridge deck was established, and real-time data injection and dynamic parameter calibration were achieved through a physical-virtual dual-domain data channel. A neural network integrating CNN-LSTM was constructed for feature extraction, and joint training was performed using virtual welding data generated by the digital twin model and actual measurement data. The pre-processed new welding parameters are input into the model to predict deformation trends. Multi-objective optimization is performed on the prediction results.

[0008] Furthermore, the above technical solutions include welding parameters such as temperature field distribution, micro-strain data, and molten pool image characteristics.

[0009] The above technical solutions further enable the following: welding parameters acquired in real time from the physical domain are injected into the virtual domain finite element model to calibrate the Gaussian heat source function parameters and material thermal property parameters; the deviation between the virtual domain simulation results and the measured data is transmitted back to the physical domain to drive the sensor control unit to dynamically adjust the sampling frequency, calibrate the sensor accuracy, or optimize the monitoring point layout.

[0010] Furthermore, through the above technical solutions, the neural network extracts the spatial features of the temperature field collected in the physical domain through CNN, and captures the temporal deformation trend of virtual domain simulation and physical domain measurement through LSTM.

[0011] Based on the above technical solutions, the prediction output formula is further as follows: in: This is the weight matrix of the output gate; This is the bias vector for the output gate; For the predicted deformation value, This represents the hidden state of the LSTM at time step t.

[0012] Based on the above technical solutions, the specific steps for multi-objective process optimization of the prediction results include: implementing dynamic correction and incremental training based on multi-dimensional error assessment; Pareto optimal solutions are generated based on the NSGA-II algorithm, and pre-deformation compensation schemes are generated by combining the prediction results and updating the parameters of the digital twin model. By optimizing the iterative parameters using Bayesian methods and combining them with the pre-deformation compensation scheme of the NSGA-II algorithm, the process continues until a predetermined termination condition is met. The termination condition is: weighted validation set... or number of iterations .

[0013] Furthermore, based on the above technical solutions, dynamic correction is performed according to spatial consistency, and its expression is: in, The actual deformation field matrix has dimensions m×n; To predict the deformation field matrix, the dimension is... same; The covariance of matrices A and B is calculated using the following formula: ,in , The mean; This represents the standard deviation of the measured deformation field. To predict the standard deviation of the deformation field.

[0014] Based on the above technical solution, the weight update expression for incremental training is further as follows: in, Let be the weight matrix of the LSTM network at time t; The learning rate; For loss function Weights The gradient; This is a newly added experimental dataset.

[0015] Based on the above technical solutions, the expression for pre-deformation compensation is further as follows: in, This is the amount of pre-deformation compensation; For the finite element simulation model function, input compensation amount Output predicted deformation ; The amount of welding deformation predicted by the LSTM model; It is an L2 norm.

[0016] A smart prediction system for welding deformation of steel bridge decks based on physical-virtual dual-domain collaborative modeling and CNN-LSTM fusion modeling includes: The data acquisition module collects and records welding parameters in real time during the steel bridge deck welding process; The collaborative modeling module establishes a three-dimensional thermo-elastic-plastic finite element model of the steel bridge deck, and realizes real-time data injection and dynamic parameter calibration through a physical-virtual dual-domain data channel; The prediction module constructs a neural network that integrates CNN-LSTM, and performs joint training with virtual welding data generated by the digital twin model and measured data. The preprocessed welding parameters are then input into the model to predict deformation trends. The control module implements dynamic correction and incremental training based on multi-dimensional error assessment. It achieves multi-objective optimization by minimizing welding deformation and maximizing welding efficiency through the NSGA-II algorithm, generates a recommended welding sequence scheme and dynamically adjusts the heat input, generates a pre-deformation compensation scheme based on the prediction results, updates the parameters of the digital twin model, and iterates the parameters through Bayesian optimization until the set termination conditions are met.

[0017] The beneficial effects of this invention are: 1. The "Physical-Virtual Dual-Domain Data Channel" is a two-way real-time communication and data interaction mechanism connecting the welding physical site (physical domain) and the digital twin simulation system (virtual domain). Its technical implementation relies on industrial Ethernet transmission and edge computing modules. By constructing a multi-source data fusion system, it integrates multi-dimensional data such as welding current, temperature field, and three-dimensional deformation. It innovatively establishes a CNN-LSTM finite element hybrid model, utilizing convolutional networks to process 1280×1024 pixel thermal images and LSTM to capture 100Hz temporal features, coupled with elastoplastic finite element physical field calculations. Multimodal feature fusion is achieved through an attention mechanism. The system is equipped with a dynamic twin iteration mechanism, updating the material constitutive model every 5 seconds. Combined with transfer learning, it adapts to plate thickness conditions ranging from 12 to 40 mm, achieving a control accuracy of ±0.3 mm in prediction deviation. Simultaneously develop a process optimization closed-loop system based on the NSGA-II algorithm, which can generate a recommended welding sequence scheme within 30 seconds and dynamically adjust the heat input. It improves the calculation efficiency by 80 times compared with the traditional finite element method. In actual engineering applications, it reduces welding deformation by 37%, providing an intelligent solution for the manufacturing of complex steel structures.

[0018] 2. This invention achieves high-precision prediction and dynamic optimization of steel bridge deck welding deformation by deeply integrating the dynamic mapping capabilities of physical-virtual dual-domain collaboration with the spatiotemporal feature modeling advantages of the CNN-LSTM fusion model. This saves significant costs and reduces the analysis cycle to the hour level. Its dual-domain collaborative digital twin combined with the fusion neural network model can simultaneously capture multi-dimensional deformation features, directly guiding process parameter adjustments, pre-deformation compensation design, and real-time construction monitoring, increasing the welding qualification rate to over 98%. Simultaneously, it supports transfer learning to adapt to different steel types and structural forms, significantly improving the level of intelligent construction in bridge engineering. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the system of the present invention; Figure 2 This is a schematic diagram of a thermo-elastic-plastic finite element model. Figure 3 This is a schematic diagram of the distribution steps of the present invention; Figure 4 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] The inventors found that traditional welding deformation control mainly relies on empirical formula derivation and finite element thermo-elastic-plastic analysis. However, the former is difficult to accurately reflect the nonlinear relationship of complex welding conditions, while the latter has problems such as high computational resource consumption and low modeling efficiency. Especially when facing the coupling effect of multiple weld seams and large-scale structures, the prediction accuracy and real-time performance are difficult to meet the needs of intelligent manufacturing.

[0022] Based on the above findings, this application proposes an intelligent prediction method for steel bridge deck welding deformation based on physical-virtual dual-domain collaboration and CNN-LSTM fusion modeling. By deeply integrating the dynamic mapping capabilities of physical-virtual dual-domain collaboration with the spatiotemporal feature modeling advantages of the CNN-LSTM fusion model, high-precision prediction and dynamic process optimization of steel bridge deck welding deformation are achieved, saving costs and compressing the analysis cycle to the hour level. Its model, combining a dual-domain collaborative digital twin with a fusion neural network, can simultaneously capture multi-dimensional deformation features, directly guiding process parameter adjustments, pre-deformation compensation design, and real-time construction monitoring, increasing the welding qualification rate to over 98%. It also supports transfer learning to adapt to different steel types and structural forms, significantly improving the level of intelligent construction in bridge engineering.

[0023] Example 1 See Figure 1 - Figure 4 This application discloses an intelligent prediction method for welding deformation of steel bridge decks based on physical-virtual dual-domain collaboration and CNN-LSTM fusion modeling, including the following steps: S1. Welding parameters during the steel bridge deck welding process are collected and recorded in real time using sensors; S2. Establish a three-dimensional thermo-elastic-plastic finite element model of the steel bridge deck, and realize real-time data injection and dynamic parameter calibration through a physical-virtual dual-domain data channel; S3. Construct a hybrid neural network that integrates CNN-LSTM, extract features, and perform joint training with virtual deformation data generated by the digital twin model and actual test data; S4. Real-time acquisition of welding parameters to be predicted, and input of them into a hybrid neural network for deformation trend prediction; S5. Perform multi-objective optimization on the prediction results.

[0024] The specific steps for multi-objective process optimization of the prediction results in S5 include: S51. Implement dynamic correction and incremental training based on multi-dimensional error assessment; S52. Generate Pareto optimal solution based on NSGA-II algorithm, generate pre-deformation compensation scheme based on prediction results, and update digital twin model parameters; S53. Optimize the iterative parameters using Bayesian optimization, combined with the pre-deformation compensation scheme of the NSGA-II algorithm, until the set termination condition is met. The termination condition is: weighted validation set. or number of iterations . Example 2

[0025] S1. Record relevant data in real time during the welding process. Deploy a high-resolution infrared thermal imager around the weld to acquire temperature field images of 1280×1024 pixels and monitor temperature changes. Install fiber Bragg grating (FBG) strain sensors at key locations to accurately monitor minute strain changes. At the same time, use a high-speed multispectral camera system to acquire images of the molten pool with welding deformation.

[0026] Welding parameters during the steel bridge deck welding process are collected and recorded in real time. Then, preprocessing is performed to eliminate noise in the original data, unify the data format, and extract key features to ensure that the data can be adapted to the input requirements of the CNN-LSTM deep learning model.

[0027] The temperature field exhibits standard three-dimensional nonlinear unsteady-state heat conduction. The three-dimensional heat conduction governing equation for the welding process is: Basic form: Unfolding format: in, For material density ; Specific heat capacity of the material k is the thermal conductivity of the material. t is the heat transfer time (s); T is the temperature field distribution function; To solve for the intensity of the internal heat source in the region (W / m) 3 ); in the formula Both are functions of temperature T.

[0028] S2. Based on the collected welding parameters, a three-dimensional thermo-elastic-plastic finite element model of the steel bridge deck is established, and the Gaussian heat source function parameters are dynamically adjusted according to the real-time welding parameters. This model is used to simulate the temperature field changes during welding. Real-time data collected by sensors is injected into the model through a physical-virtual dual-domain data channel. At the same time, the simulation deviation of the model is fed back to the physical domain sensor calibration stage to dynamically adjust the Gaussian heat source function parameters and boundary conditions, forming a real-time mapping of the virtual welding process.

[0029] Data collected in the physical domain is transmitted to the virtual domain in real time via industrial Ethernet. The model correction instructions in the virtual domain are fed back to the sensor control unit in the physical domain through the edge computing module, so that the data interaction delay between the two domains is ≤100ms. The core functions of the physical and virtual dual-domain data channel include: (1) dynamically injecting the temperature field distribution, micro-strain data and molten pool image features collected in real time in the physical domain into the finite element model of the virtual domain for calibration of Gaussian heat source function parameters and material thermal property parameters; (2) transmitting the deviation between the simulation results in the virtual domain and the measured data (such as temperature field distribution error and strain prediction deviation) in reverse to the physical domain, driving the sensor control unit to dynamically adjust the sampling frequency, calibrate the sensor accuracy or optimize the monitoring point layout, forming a closed-loop calibration mechanism for dual-domain collaboration.

[0030] The parameters of the Gaussian heat source function include the effective heating radius R of the heat source and the maximum heat flux density at the center of the heat source. The thermal properties of materials include density ρ and specific heat capacity as they change with temperature. And thermal conductivity k.

[0031] The established virtual model temperature field numerical simulation adopts a Gaussian distributed heat convection density model, and the following equation is in the form of a Gaussian heat source function, representing the heat flux density at any point from the heating center: In the formula: The maximum heat flux density at the center of the heating range; The effective heating radius of the heat source; Let P be the distance from any point P to the heating center O.

[0032] Based on the law of conservation of energy and Fourier's law, the heat conduction equation for the welding process can be expressed as: In the formula: Let be the spatiotemporal temperature field function (°C); The density of the material varies with temperature; For specific heat capacity, nonlinear temperature dependence; Thermal conductivity; anisotropic at high temperatures; The power density of the welding heat source; The convective heat transfer coefficient; Surface emissivity; This is a Stefan-Boltzmann constant; The ambient temperature.

[0033] S3. Finite element analysis is performed in Perfect-Welding software based on a finite element model. Deformation features are constructed using a digital twin model. The digital twin model simulates the deformation distribution under different welding parameters, and representative points in key areas are extracted as prediction targets. Based on this, deformation-related features are constructed as input to a deep learning model.

[0034] The design incorporates a deep learning model, specifically a hybrid neural network model that integrates CNN (Convolutional Neural Network) and LSTM (Long Short-Term Memory). Specifically, CNN extracts spatial features of the molten pool image and temperature field distribution from the physical domain, such as the distribution of high-temperature regions and gradient changes. LSTM captures the temporal deformation trends at a sampling frequency of 100Hz in both virtual domain simulation and physical domain measurements. This data is then jointly trained using virtual welding data generated by the digital twin model and the measured data from S1 to enhance the model's reliability. Virtual welding data refers to the simulated data generated in the digital twin model, mimicking the entire welding process under different combinations of welding parameters, and whose dimensions match those of the measured data in the physical domain.

[0035] The training iterations were 200, with a learning rate of 0.001. Keypoint locations were encoded using a sine function injected into the model to enhance spatial correlation modeling capabilities.

[0036] S4. Real-time acquisition of welding parameters and sensor data to be processed. After preprocessing in step 1, the data is input into a hybrid neural network, which outputs the deformation prediction trend in the short term. The prediction results are then fed back into the digital twin model. Model quantization and hardware acceleration technologies are used to ensure the real-time performance of each prediction.

[0037] Below is the time series formula for a hybrid neural network, using the ReLU function and a CNN feature extraction layer: In the formula: The weight matrix of the convolution kernel; The bias vector of the convolutional layer; Input data; This is a feature map for a CNN.

[0038] LSTM timing modeling: In the formula: Let t be the hidden state of the LSTM at time step t; The cell state at time step t represents the LSTM.

[0039] The calculations for the forget gate, input gate, and output gate are as follows: Final prediction output: In the formula: , and These are the weight matrices for the forget gate, input gate, and output gate, respectively. , and These are the bias vectors for each gate; It is the Sigmoid activation function. ; The predicted deformation value has the same dimensions as the target output.

[0040] CNN uses three convolutional layers (with kernel sizes of 3×3, 5×5, and 3×3) to extract spatial features of the temperature field, while LSTM uses two hidden layers (64 neurons per layer) to capture temporal deformation trends. The two are fused into feature vectors through an attention mechanism.

[0041] S51. Evaluate the prediction error from multiple dimensions such as numerical value, spatial consistency and temporal correlation, and perform dynamic correction. Use the latest measured data for incremental training and update the weights of the LSTM network in real time. Error assessment and dynamic correction, including numerical, spatial consistency, and temporal correlation.

[0042] Numerical error (weighted MAE): Where N is the total number of samples (the number of welding deformation measurement points); The weight of the i-th measurement point (set according to the importance of the welding area, such as higher weight for critical areas); This represents the actual deformation value at the i-th measurement point (from actual sensor measurement data). Predict the deformation value (LSTM output) for the model at the i-th measurement point.

[0043] Spatial consistency (spatial correlation coefficient): in, The actual deformation field matrix has dimensions m×n (e.g., the measured deformation values ​​of the mesh nodes on the surface of the weldment). To predict the deformation field matrix, the dimension is... same; The covariance of matrices A and B is calculated using the following formula: ,in , The mean; This represents the standard deviation of the measured deformation field. To predict the standard deviation of the deformation field.

[0044] Incremental training weight update (online gradient descent): in, Let be the weight matrix of the LSTM network at time t; The learning rate (a hyperparameter that controls the step size of parameter updates, such as 0.001); For loss function Weights The gradient; This is a newly added measured dataset (containing the latest welding deformation data).

[0045] The core objective of incremental training is to enable the model to quickly learn the deformation patterns of new working conditions without retraining the entire model when acquiring new real-world datasets. This is achieved through "local weight updates," ensuring that the prediction accuracy remains within the control threshold of ±0.3mm. This formula is the mathematical core of incremental training, based on the "online gradient descent" mechanism. The model does not load all data at once but receives new data as a "data stream," calculating the gradient of the loss function only with the latest data and gradually adjusting the weights.

[0046] The newly added experimental dataset comes from physical domain sensors and requires preprocessing identical to the initial training data, including processing of temporal and spatial data to avoid model input dimension mismatch. The preprocessed D_new is then input into the current model to calculate the loss function. For the current weight gradient ;Calculate the new weights at time t+1 according to the weight update formula. After the weights are updated, the model accuracy needs to be verified through the "physical-virtual dual-domain data channel." The updated model is then used to predict the "virtual deformed data" generated by the virtual domain digital twin model. The deviation between the virtual simulation results and the model prediction results is compared. If the deviation is less than a preset threshold, the weight update is valid; if it exceeds the threshold, the learning rate needs to be adjusted in reverse. This continues until the dual-domain data consistency requirement is met.

[0047] S52. Generate Pareto optimal solution based on NSGA-II algorithm, generate pre-deformation compensation scheme based on prediction results, and update digital twin model parameters.

[0048] The multi-objective process optimization based on the prediction results using the NSGA-II algorithm includes the following steps: Determine the optimization variables and objectives: The optimization variables include welding current, welding voltage, welding speed, and welding sequence; the optimization objectives are: Objective 1: Welding deformation ≤ ±0.3mm (based on CNN-LSTM model prediction results), Objective 2: Welding efficiency ≥ 350mm / min (based on equipment capacity and process requirements); Population initialization and fitness calculation: Initialize 50 combinations of process parameters as the initial population. Each combination is input into the CNN-LSTM model to calculate the deformation. Combined with the equipment parameters, the efficiency is calculated and used as the fitness value. Non-dominated sorting: Select non-dominated individuals in the population (no other individual is better at both objectives), divide them into different sorting levels, and prioritize retaining Pareto optimal individuals in the first level; Crowding degree calculation: For individuals in the same sorting layer, calculate the sparsity of the surrounding solutions, retain individuals with high crowding degree, and ensure the diversity of solutions (adapting to different plate thicknesses of 12 to 40 mm). Elite preservation and genetic operations: merge the parent and offspring populations, select the best individuals through sorting and crowding, generate a new population through selection, crossover and mutation, and output the Pareto optimal solution set after 50 generations. Solution selection and execution: Based on the actual working conditions of the physical domain (such as current plate thickness and structural stress requirements), the target solution is selected from the optimal set, welding sequence recommendation instructions are generated, and the heat input is dynamically adjusted. The actual welding data is fed back to the digital twin model to update parameters and achieve closed-loop optimization.

[0049] S6 involves optimization algorithms, such as PID control parameter adjustment or genetic algorithms. Pre-deformation compensation may require an inverse model, for example, deriving the compensation amount from the predicted deformation. Updating material constitutive parameters can use finite element analysis or parameter estimation methods, such as least squares or Kalman filtering. Database storage and continued model training may involve data augmentation or transfer learning.

[0050] Inverse model of compensation amount: in, This is the amount of pre-deformation compensation (such as the amount of adjustment for the initial shape of the weldment). For the finite element simulation model function, input compensation amount Output predicted deformation ; The amount of welding deformation predicted by the LSTM model; It is the L2 norm (the sum of squared errors).

[0051] Dynamic adjustment of welding parameters (PID control): in, To control the output (such as the adjustment amount of welding current or welding torch movement speed); , The deformation error at time t is the difference between the measured value and the predicted value. , and These are the proportional, integral, and derivative coefficients, respectively (which need to be adjusted according to the welding process). This is the cumulative error (total historical error). This is the rate of change of error (the difference between the current error and the previous error).

[0052] Constitutive parameter gradient estimation: in, These are the constitutive parameters of the material; Update the step size for the parameters (similar to the learning rate, such as 0.01). This represents the error in finite element simulation. The gradient of the error with respect to the material parameters (calculated by the adjoint method or automatic differentiation).

[0053] S53. Evaluate whether the deviation between the measured deformation value and the predicted value is within the allowable range. Search for the optimal network parameters through Bayesian optimization, with the optimization objective being the weighted average squared error on the validation set. Combine this with the pre-deformation compensation scheme of the NSGA-II algorithm. If the set stopping condition is met, the iteration ends; otherwise, it continues. Finally, the optimized parameters are imported into the industrial control platform to achieve automatic control.

[0054] To determine if the error is within acceptable limits, Bayesian optimization is used to find the optimal parameters. The objective function is the weighted mean square error (MSE) of the validation set. The core of Bayesian optimization is a Gaussian process and a data acquisition function, such as EI. The formula for the weighted MSE requires assigning different weights to different samples. An early stopping mechanism may be implemented for the number of iterations, and the updated parameters are imported into the control platform.

[0055] Weighted Validation Set (MSE): in, The number of samples in the validation set; The weight of the j-th sample (e.g., the sample weight of the high-stress region). ); To verify the actual deformation value of the j-th sample in the set; Let be the model's predicted value for the j-th sample.

[0056] Gaussian process surrogate model: in, Hyperparameter vectors (such as the number of layers, number of neurons, and learning rate of the LSTM); It is the mean function (usually set to a constant or zero); This is the kernel function.

[0057] Expected Improvement (EI) Acquisition Function: in, Objective function (i.e., validation set weighted MSE); Current optimal objective function value (minimum MSE among the hyperparameters already tried); Expected value (calculated based on the posterior distribution of the Gaussian process).

[0058] Termination conditions and deployment: Error threshold; This represents the current iteration number; Maximum allowed number of iterations; The optimal combination of hyperparameters (the parameters that minimize the MSE of the validation set).

[0059] Application Cases In the construction project of a large cross-sea steel bridge, the orthotropic steel bridge deck uses Q355D steel, and the welding process of U-ribs and transverse diaphragms is complex.

[0060] The construction team first used the thermo-elastic-plastic finite element method and SYSWELD software to construct a three-dimensional model of the local U-rib-diaphragm, accurately simulating the welding temperature field, residual stress, and deformation. Simultaneously, various sensors were deployed at the welding site, such as an infrared thermal imager acquiring high-resolution temperature field images around the weld every second, a fiber Bragg grating strain sensor monitoring micro-strain changes at a frequency of 100Hz, and a high-speed multispectral camera system capturing images of the molten pool, collecting a large amount of welding process data.

[0061] After preprocessing such as filtering and denoising, the data is input into a neural network model that combines CNN and LSTM. CNN is responsible for extracting spatial features from the melt pool image and temperature field distribution, while LSTM focuses on capturing the temporal trend of deformation.

[0062] After being trained on a large amount of virtual and measured data, the model can predict deformation trends in real time based on newly input welding parameters.

[0063] For example, during a U-rib welding process, the model predicted in advance that the deformation of a key part would exceed the allowable range in the next 5 seconds. The team then used the NSGA-II algorithm to perform multi-objective optimization on parameters such as welding current, voltage, speed and sequence, and combined Bayesian optimization to iteratively update the model parameters to generate a pre-deformation compensation scheme. Ultimately, the actual welding deformation was effectively controlled within ±0.3mm, and the welding qualification rate reached more than 98%, which greatly ensured the welding quality and construction accuracy of the steel bridge deck.

[0064] This invention achieves high-precision prediction and dynamic optimization of welding deformation in steel bridge decks by deeply integrating the dynamic mapping capabilities of physical-virtual dual-domain collaboration with the spatiotemporal feature modeling advantages of the CNN-LSTM fusion model. This saves significant costs and reduces the analysis cycle to the hour level. Its dual-domain collaborative digital twin combined with the fusion neural network model can simultaneously capture multi-dimensional deformation features, directly guiding process parameter adjustments, pre-deformation compensation design, and real-time construction monitoring. This increases the welding qualification rate to over 98%, while also supporting transfer learning to adapt to different steel types and structural forms, significantly improving the level of intelligent construction in bridge engineering.

[0065] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A smart prediction method for welding deformation of steel bridge deck based on physical-virtual dual-domain collaboration and CNN-LSTM fusion modeling, characterized in that, Includes the following steps: Welding parameters during the steel bridge deck welding process are collected and recorded in real time using sensors. A three-dimensional thermo-elastic-plastic finite element model of the steel bridge deck was established, and real-time data injection and dynamic parameter calibration were achieved through a physical-virtual dual-domain data channel. A hybrid neural network integrating CNN-LSTM is constructed for feature extraction, and joint training is performed using virtual deformation data generated by the digital twin model and actual test data. The welding parameters to be predicted are collected in real time and input into a hybrid neural network to predict deformation trends. Multi-objective optimization is performed on the prediction results.

2. The intelligent prediction method for welding deformation of steel bridge deck based on physical-virtual dual-domain collaboration and CNN-LSTM fusion modeling as described in claim 1, characterized in that: Welding parameters include temperature field distribution, micro-strain data, and molten pool image characteristics.

3. The intelligent prediction method for welding deformation of steel bridge deck based on physical-virtual dual-domain collaboration and CNN-LSTM fusion modeling as described in claim 2, characterized in that: Welding parameters acquired in real time from the physical domain are injected into the finite element model of the virtual domain to calibrate the Gaussian heat source function parameters and material thermal property parameters. The deviation between the virtual domain simulation results and the measured data is transmitted back to the physical domain, driving the sensor control unit to dynamically adjust the sampling frequency, calibrate the sensor accuracy, or optimize the monitoring point layout.

4. The intelligent prediction method for welding deformation of steel bridge deck based on physical-virtual dual-domain collaboration and CNN-LSTM fusion modeling as described in claim 3, characterized in that: In a hybrid neural network, CNN extracts the molten pool image and temperature field spatial features acquired in the physical domain, while LSTM captures the temporal deformation trend of virtual domain simulation and physical domain measurement.

5. The intelligent prediction method for welding deformation of steel bridge deck based on physical-virtual dual-domain collaboration and CNN-LSTM fusion modeling as described in claim 4, characterized in that, The formula for predicting the output is: in: This is the weight matrix of the output gate; This is the bias vector for the output gate; For the predicted deformation value, This represents the hidden state of the LSTM at time step t.

6. The intelligent prediction method for welding deformation of steel bridge deck based on physical-virtual dual-domain collaboration and CNN-LSTM fusion modeling as described in claim 5, is characterized in that, The specific steps for multi-objective process optimization based on the prediction results include: Dynamic correction and incremental training are implemented based on multi-dimensional error assessment; Pareto optimal solutions are generated based on the NSGA-II algorithm, and pre-deformation compensation schemes are generated by combining the prediction results and updating the parameters of the digital twin model. By optimizing the iterative parameters using Bayesian methods and combining them with the pre-deformation compensation scheme of the NSGA-II algorithm, the process continues until a predetermined termination condition is met. The termination condition is: weighted validation set... or number of iterations .

7. The intelligent prediction method for welding deformation of steel bridge deck based on physical-virtual dual-domain collaboration and CNN-LSTM fusion modeling as described in claim 6, characterized in that: Dynamic correction based on spatial consistency is expressed as follows: in, The actual deformation field matrix has dimensions m×n; To predict the deformation field matrix, the dimension is... same; The covariance of matrices A and B is calculated using the following formula: ,in , The mean; This represents the standard deviation of the measured deformation field. To predict the standard deviation of the deformation field.

8. The intelligent prediction method for welding deformation of steel bridge deck based on physical-virtual dual-domain collaboration and CNN-LSTM fusion modeling as described in claim 7, characterized in that: The weight update expression for incremental training is: in, Let be the weight matrix of the LSTM network at time t; The learning rate; loss function Weights The gradient; This is a newly added experimental dataset.

9. The intelligent prediction method for welding deformation of steel bridge deck based on physical-virtual dual-domain collaboration and CNN-LSTM fusion modeling as described in claim 8, characterized in that: The expression for pre-deformation compensation is: in, This is the amount of pre-deformation compensation; For the finite element model function, input compensation amount Output predicted deformation ; The amount of welding deformation predicted by the finite element model; It is an L2 norm.

10. A smart prediction system for welding deformation of steel bridge decks based on physical-virtual dual-domain collaboration and CNN-LSTM fusion modeling, characterized in that, include: The data acquisition module collects and records welding parameters in real time during the steel bridge deck welding process; The collaborative modeling module establishes a three-dimensional thermo-elastic-plastic finite element model of the steel bridge deck, and realizes real-time data injection and dynamic parameter calibration through a physical-virtual dual-domain data channel; The prediction module constructs a neural network that integrates CNN-LSTM, and performs joint training with virtual welding data generated by the digital twin model and measured data. The preprocessed welding parameters are then input into the model to predict deformation trends. The control module uses the NSGA-II algorithm to achieve multi-objective optimization of minimizing welding deformation and maximizing welding efficiency, generates a recommended welding sequence, and dynamically adjusts the heat input to form a pre-deformation compensation scheme.

Images