A physical-data joint driving welding temperature field digital twin method

Abstract

This invention discloses a physics-data jointly driven digital twin method for welding temperature fields. The specific method involves: constructing a physics-data fusion dataset containing information related to heat source parameters based on heat transfer mechanisms and experimental data; constructing and training a real-time prediction model for heat source parameters based on a liquid neural network, and simultaneously constructing and training a KAN welding temperature field prediction model incorporating heat transfer constraints; building a lightweight adaptive adjustment network to adapt to new data conditions; collecting and preprocessing real-time welding data; predicting real-time heat source parameters using the LNN model; calculating the heat flux density, and inputting the heat flux density into the KAN model adjusted by the lightweight network to achieve temperature field prediction. This invention, through physics-data joint driving, integrates the temporal dynamic adaptive advantages of LNN with the strong nonlinear fitting capability of KAN, balancing prediction accuracy, real-time performance, and interpretability. The lightweight network significantly improves the efficiency of adapting to new data.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and welding technology, and specifically relates to a physical-data jointly driven digital twin method for welding temperature field. Background Technology

[0002] Accurate real-time prediction of the temperature field during welding is a core prerequisite for ensuring weld quality and achieving high-precision control of the welding process. By understanding the distribution and variation patterns of the temperature field, thermal stress, deformation, and residual stress can be effectively controlled, thereby improving the integrity and mechanical properties of the welded structure.

[0003] Existing methods for predicting welding temperature fields have significant limitations: traditional monitoring methods such as thermocouples and infrared thermal imaging can only acquire surface temperature and cannot reflect the internal temperature distribution; numerical simulation methods such as the finite element method can achieve three-dimensional temperature field analysis, but they are computationally complex and time-consuming, making it difficult to meet real-time requirements, and they require remodeling for new working conditions, resulting in insufficient flexibility; among existing data-driven models, time-series networks such as LSTM can process time-series data, but their dynamic adaptive ability is limited, they rely on a large amount of labeled data, have weak generalization ability, lack physical mechanism constraints, and have poor interpretability; KAN networks, as a new type of neural network with strong nonlinear fitting ability and certain interpretability, have the core advantage of accurately fitting nonlinear relationships through the coefficients of spline activation functions, but when directly applied, they face the problems of low efficiency and high computational cost of full retraining under new working conditions, and existing technologies lack a solution to combine dynamic time-series prediction models with KAN to achieve real-time prediction and dynamic adaptation of the entire heat source parameter-temperature field chain, which cannot meet the flexible needs of multi-working-condition manufacturing scenarios. Summary of the Invention

[0004] 1. The technical problem to be solved:

[0005] Existing technologies suffer from several problems, including difficulty in achieving both accuracy and real-time performance in welding temperature field prediction, insufficient dynamic adaptability of heat source parameter prediction, low efficiency in adapting to new working conditions, and inadequate dynamic updating capability of temperature field prediction.

[0006] 2. Technical Solution:

[0007] To address the above problems, this invention provides a physical-data jointly driven digital twin method for welding temperature fields, comprising the following steps:

[0008] Step S01: Construct a physics-data fusion dataset, which includes finite element simulation data based on heat transfer mechanism and real data collected from experiments; each data sample includes welding parameters, molten pool surface image, infrared image, workpiece node coordinates, material thermal property parameters, heat flux density, and corresponding temperature value.

[0009] Step S02: Construct a real-time prediction model for heat source parameters based on liquid neural networks (LNNs). With minimizing the prediction error of heat source parameters as the objective function, train the model using the heat source parameter data in the dataset from step S01 to obtain the real-time prediction model for heat source parameters.

[0010] Step S03: Construct a welding temperature field prediction model based on KAN network. With minimizing the prediction error under the constraint of heat transfer mechanism as the objective function, train the model using the dataset from step S01 to obtain the welding temperature field prediction model.

[0011] Step S04: Construct a lightweight adaptive adjustment network, which is used to extract the difference features between the new working condition data and the dataset, and to locally adjust the parameters of the KAN model trained in step S03.

[0012] Step S05: Collect real-time data of the welding process through sensors, infrared cameras and industrial cameras. The real-time data includes welding current, arc voltage and surface image of the molten pool. The actual molten pool surface width parameter is extracted through image processing.

[0013] Step S06: Preprocess the real-time monitoring data during the welding process; input the preprocessed real-time monitoring data into the heat source parameter prediction model to obtain the predicted real-time heat source parameters; calculate the real-time heat flux density using the heat source model; input the real-time heat flux density into the temperature field prediction model and calculate the error; input the error into a lightweight adaptive adjustment network to adjust the temperature field prediction model to the optimized KAN model; use the optimized KAN model to calculate the accurate temperature values ​​of each unit node of the welded workpiece.

[0014] 3. Beneficial effects:

[0015] (1) This invention introduces a liquid neural network (LNN) to realize real-time prediction of heat source parameters. Its dynamic neuron connection and adaptive activation mechanism can accurately capture the temporal dynamic characteristics of the welding process. Compared with the traditional LSTM network, it improves the real-time performance and dynamic adaptability of heat source parameter prediction, and provides a high-precision heat source input basis for temperature field prediction.

[0016] (2) A phased training strategy is adopted. In the first phase, the KAN model is trained with a joint loss function to take into account both the data fitting accuracy and the physical laws of heat transfer, so as to avoid the problem that the model “fits the data well but violates the common sense of physics”. In the second phase, a lightweight network is trained with pure data loss to focus on the correction of prediction error under new working conditions. There is no need to retrain all parameters of KAN, which greatly improves the adaptation efficiency.

[0017] (3) The lightweight adaptive adjustment network fine-tunes the core spline coefficients of KAN to accurately adapt to the temperature field changes under new working conditions. Combined with the phased training logic, it further enhances the flexibility and practicality of the method under multiple working conditions, taking into account both physical consistency and dynamic adaptability.

[0018] (4) The data acquisition and preprocessing process is standardized. The configuration of sensors and cameras and the image processing algorithm ensure the accuracy and reliability of real-time data, providing high-quality data support for LNN heat source parameter prediction and KAN temperature field prediction. Attached Figure Description

[0019] Figure 1 This is the overall flowchart of this method.

[0020] Figure 2 This is a schematic diagram of the temperature field during the welding process solved using the finite element method.

[0021] Figure 3 Comparison of the predicted and actual values ​​of the semi-axis length in the melt width direction of the heat source parameter prediction model.

[0022] Figure 4 This is a schematic diagram of the liquid neural network used in the heat source parameter prediction model.

[0023] Figure 5 This is a schematic diagram of the KAN temperature field prediction model. Detailed Implementation

[0024] The present invention will now be described in detail with reference to the accompanying drawings.

[0025] like Figure 1 As shown, a physics-data jointly driven digital twin method for welding temperature fields includes the following steps:

[0026] Step S01: Construct a physics-data fusion dataset, which includes finite element simulation data based on heat transfer mechanism and real data collected from experiments; each data sample includes welding parameters, molten pool surface image, infrared image, workpiece node coordinates, material thermal property parameters, heat flux density, and corresponding temperature value.

[0027] Step S02: Construct a real-time prediction model for heat source parameters based on a liquid neural network. With minimizing the prediction error of heat source parameters as the objective function, train the model using the heat source parameter data in the dataset from step S01 to obtain the real-time prediction model for heat source parameters.

[0028] Step S03: Construct a welding temperature field prediction model based on KAN (Kolmogorov-Arnold Network). With minimizing the prediction error under the constraint of heat transfer mechanism as the objective function, train the model using the dataset from step S01 to obtain the welding temperature field prediction model.

[0029] Step S04: Construct a lightweight adaptive adjustment network, which is used to extract the difference features between the new working condition data and the dataset, and to locally adjust the parameters of the KAN model trained in step S03.

[0030] Step S05: Collect real-time data of the welding process through sensors, infrared cameras and industrial cameras. The real-time data includes welding current, arc voltage and surface image of the molten pool. The actual molten pool surface width parameter is extracted through image processing.

[0031] Step S06: Preprocess the real-time monitoring data during the welding process; input the preprocessed real-time monitoring data into the heat source parameter prediction model to obtain the predicted real-time heat source parameters; calculate the real-time heat flux density using the heat source model; input the real-time heat flux density into the temperature field prediction model and calculate the error; input the error into a lightweight adaptive adjustment network to adjust the temperature field prediction model to the optimized KAN model; use the optimized KAN model to calculate the accurate temperature values ​​of each unit node of the welded workpiece.

[0032] In step S01, the method for constructing the physical-data fusion dataset includes the following steps:

[0033] Step S11: Use finite element software to establish a three-dimensional geometric model of the welded workpiece, define the geometric dimensions and divide the structured mesh, and obtain the three-dimensional coordinate values ​​of each element node.

[0034] Step S12: Input the material thermal properties of the workpiece to be welded, including thermal conductivity, density, specific heat, and latent heat. The thermal conductivity and latent heat are set in temperature segments to conform to the actual heat transfer law.

[0035] Step S13: Set welding parameters and heat source parameters, create a double ellipsoid moving heat source function, and calculate the heat flux density of each unit node at different times.

[0036] The welding parameters include welding current, arc voltage, welding speed, and heat source parameters; the heat source parameters include the length of the front half-axis of the welding direction, the length of the rear half-axis of the welding direction, the length of the half-axis of the weld width direction, and the length of the half-axis of the weld depth direction; the material thermal properties include thermal conductivity, density, specific heat, and latent heat, and the thermal conductivity and latent heat vary with temperature; the finite element simulation data are obtained by establishing a geometric model of the welded workpiece, setting boundary conditions, and solving the double ellipsoidal moving heat source function.

[0037] The power density distribution of the first half of the double ellipsoidal moving heat source function is shown in the following equation:

[0038]

[0039] The power density distribution in the latter half of the double-ellipsoidal moving heat source is as follows:

[0040]

[0041] In the above two formulas, It is the distance from the center of the heat source in the welding direction. It is the distance from the center of the heat source in the direction of the melt width. It is the distance from the center of the heat source in the direction of melting depth. It is the length of the first half of the axis in the welding direction. It is the length of the rear half-axis in the welding direction. It is the length of the semi-axis in the direction of weld width. It is the length of the semi-axis in the direction of melting depth. It is the intensity of the volumetric heat source. and The energy ratio of the front and rear parts. + =2.

[0042] Step S14: Define the initial temperature and thermal convection boundary conditions, solve the transient heat conduction equation, and obtain the simulated temperature values ​​of each node.

[0043] Step S15: Collect real data under different working conditions through experiments, including welding current, arc voltage, molten pool surface image, infrared image, heat source parameters and corresponding temperature measurement values.

[0044] Step S16: Perform spatiotemporal alignment between simulated data and real data, and use an outlier detection algorithm to remove outlier data. The simulated molten pool surface width in the heat source parameter dataset is obtained by calculating the distance between unit nodes at the left and right edges of the welded workpiece geometric model with the temperature value being the metal melting point temperature, thus forming a physical-data fusion dataset.

[0045] Step S02 specifically includes the following steps:

[0046] Step S21: The input layer receives time-series feature data, specifically the welding current, arc voltage, and weld pool surface width sequence at times [tn,t] before the current time t. The data is input into the model after standardization.

[0047] In one embodiment, n is 40.

[0048] Step S22: The dynamic hidden layer adopts the core structure of a liquid neural network, which includes a dynamic neuron pool and an adaptive connection weight matrix. The connection strength between neurons can be adjusted in real time according to the changes in the input time series data. The activation function adopts a dynamic nonlinear activation mechanism, which can accurately capture the dynamic correlation between the heat source parameters and the input time series data during the welding process.

[0049] In one embodiment, the number of hidden layer neurons is set to 32, balancing computational efficiency and prediction accuracy.

[0050] Step S23: The output layer consists of 4 output nodes, which correspond to the predicted values ​​of the heat source parameters (first half-axis length of the welding direction, second half-axis length of the welding direction, half-axis length of the weld width direction, and half-axis length of the weld depth direction), and are output using a linear activation function;

[0051] Step S24: Using the mean squared error between the predicted heat source parameters and the actual heat source parameters in the dataset as the loss function, the adaptive gradient descent algorithm is used to minimize the loss, and the dynamic connection weights and neuron activation parameters of the model are iteratively updated.

[0052] In one embodiment, the number of iterations is set to 50 rounds, the batch size is set to 32, and a real-time prediction model for heat source parameters is obtained after training.

[0053] Step S03 specifically includes the following steps:

[0054] S31: The input layer receives six dimensions of features: the six dimensions are: the three-dimensional coordinates of the welding workpiece unit node, the time step, the instantaneous heat flux density, and the temperature at the previous moment.

[0055] S32: The KAN hidden structure consists of two hidden layers, using spline basis functions as activation functions. The layers are fully connected to achieve feature transfer and incorporate the constraints of the heat transfer partial differential equation, ensuring that the model prediction results conform to physical laws.

[0056] In one embodiment, the width from the input layer to the hidden layer is 6 and 128 neurons.

[0057] S33: A single output node in the output layer, outputting the real-time temperature value of the corresponding unit node.

[0058] S34: Using the mean square error between the predicted temperature value and the temperature value in the dataset as the base loss, the residual term of the heat transfer partial differential equation is added as the physical constraint loss to construct a joint loss function. The Adam optimizer is used to minimize the joint loss, and the model parameters are iteratively updated to obtain the trained basic KAN prediction model.

[0059] In step S04, the training objective of the lightweight adaptive adjustment network is to minimize the temperature prediction error of the KAN model under the new operating conditions. The gradient descent algorithm is used to optimize the network parameters, and the training data consists of a small sample of real data under the new operating conditions. The specific method includes the following steps:

[0060] Step S41: A lightweight spline modulation structure is adopted, which includes a residual encoder and a spline parameter generator, and the total number of parameters is controlled within 10% of the basic KAN model.

[0061] Step S42: The residual encoder inputs the prediction residual of the basic KAN model and extracts the error distribution features through the fully connected layer; the residual is the absolute error sequence between the current predicted temperature and the actual temperature.

[0062] Step S43: The core function of the spline parameter generator is to convert error features into modulation parameters of spline control points. These parameters can dynamically correct the prediction output of the basic KAN model by changing the morphological features of the spline function, including two adjustment methods: control point position offset and scale scaling. Its underlying logic is to precisely adjust the coefficients of KAN to optimize the prediction behavior of the model.

[0063] Step S44: Use small sample data (5%-10% of the base dataset) under the new operating conditions to train the spline modulator end-to-end, and use the mean square error of temperature prediction of the modulated model as the loss function to optimize the modulator parameters.

[0064] The specific method of step S05 includes the following steps:

[0065] Step S51: Collect welding current and arc voltage in real time using an arc sensor to ensure that the data covers the base and peak phases of the welding pulse.

[0066] In one embodiment, the sampling frequency is 100Hz.

[0067] Step S52: Capture visible light images of the molten pool surface using a high dynamic range (HDR) industrial camera. The lens is equipped with a bandpass filter with a center wavelength of 685±2nm and a full width at half maximum (FWHM) of 10±2nm to suppress welding arc interference. The image format is Mono8 with a resolution of 300×300.

[0068] In one embodiment, the sampling frequency is 100Hz.

[0069] Step S53: Acquire infrared images of the welding area using an infrared camera with a wavelength range of 7.5-13μm, a temperature measurement range of -40℃-1500℃, and a sampling frequency of 100Hz. This image is used to extract the true surface temperature value of the welding area, providing a basis for subsequent error calculation.

[0070] Step S54: Process the visible light image of the molten pool surface: Correct the image using the checkerboard method to achieve accurate matching between pixels and actual spatial positions; extract the region of interest (DOI); enhance image contrast using the Retinex algorithm to highlight the molten pool boundary; obtain continuous and complete molten pool edges using Canny edge detection and morphological operations (dilation, erosion); fit the left and right boundaries using the least squares method, and calculate the actual molten pool surface width based on pixel distance.

[0071] Step S55: Preprocess the infrared image: remove noise by Gaussian filtering, and convert the pixel grayscale values ​​into actual temperature values ​​by combining the infrared camera calibration parameters. Select the real temperature data of the molten pool area as the benchmark for error calculation.

[0072] The specific method of step S06 includes the following steps:

[0073] Step S61: During real-time welding, the pre-processed real-time time series data (welding current, arc voltage, and molten pool surface width at time [t-40,t]) is input into the LNN heat source parameter prediction model trained in step S02, and the predicted value of the heat source parameters at the current time is output in real time.

[0074] Step S62: Substitute the real-time heat source parameters predicted by LNN into the double ellipsoidal moving heat source function, and calculate the real-time heat flux density of each unit node of the welded workpiece by combining the real-time welding current and arc voltage.

[0075] Step S63: Input the instantaneous heat flux density, unit node coordinates, time step, and previous temperature into the basic KAN model trained in step S03, and output the initial predicted temperature value of each unit node in real time.

[0076] Step S64: Simultaneously calculate the absolute error sequence between the initial predicted temperature value and the infrared measured temperature value to form real-time prediction residual data; input the residual data into the lightweight adaptive adjustment network trained in step S04 in real time, the network extracts error features in real time through the feature extraction module, and outputs the spline coefficient correction amount of the KAN model in real time through the parameter adjustment module.

[0077] Step S65: Immediately input the real-time generated correction value into the basic KAN model, dynamically adjust the coefficients of the hidden layer spline activation function, and realize the real-time optimization of the basic KAN model; the optimized KAN model outputs the accurate temperature value of each unit node;

[0078] In one embodiment, the system includes a data acquisition module, a data preprocessing module, an LNN heat source parameter prediction module, a KAN temperature field prediction module, and a lightweight adjustment module. The data acquisition module includes an arc sensor, an infrared camera, and an industrial camera. The LNN heat source parameter prediction module, the KAN temperature field prediction module, and the lightweight adjustment module are implemented through software programs. Specific Implementation

[0079] This embodiment uses pulsed gas shielded tungsten inert gas (TIG) welding of 304L stainless steel plates as the application scenario. In terms of hardware configuration, the welding experimental platform adopts a pulsed gas shielded TIG welding device with a tungsten electrode diameter of 2.4mm, using argon as the shielding gas at a flow rate of 6L / min, a welding pulse frequency of 50Hz, and a duty cycle of 50%. The workpiece being welded is a 304L stainless steel plate with dimensions of 40mm × 30mm × 2mm, fixed to a stepper motor-driven motion platform. The data acquisition equipment includes an arc sensor, a high dynamic range (HDR) industrial camera, and an infrared camera. The arc sensor collects welding current and arc voltage in real time at a sampling frequency of 100Hz. The industrial camera lens is equipped with a bandpass filter to capture images of the molten pool surface at a resolution of 300×300 and a sampling frequency of 100Hz. The infrared camera collects infrared images of the welding area to obtain the actual temperature value. The data processing equipment is equipped with a multi-core processor and a GPU to ensure the real-time performance of model inference and data processing.

[0080] In terms of software environment, Abaqus was used for finite element simulation to build the geometric model of the welded workpiece and solve the temperature field, corresponding to the simulation results in Figure 2. The deep learning framework used Python and PyTorch, along with tools such as OpenCV and NumPy, to achieve model training and data preprocessing.

[0081] During the construction of the physical-data fusion dataset, as shown in the attached document... Figure 2 As shown, attached Figure 2 This is a schematic diagram of the welding temperature field solved by the finite element method. Figure 2The image illustrates the workpiece mesh generation and temperature distribution characteristics. The red area represents the high-temperature zone of the molten pool, and the blue area represents the ambient temperature zone, visually presenting the temperature field distribution pattern of the simulated data. Specifically, a three-dimensional geometric model of the workpiece is established using Abaqus, material thermophysical parameters are input, initial temperature and thermal convection boundary conditions are set, a double ellipsoidal moving heat source function subroutine is created, and the transient temperature field under multiple sets of different welding parameters is solved to obtain simulated data. At the same time, welding current, arc voltage, molten pool surface image, infrared image, and temperature measurement values ​​under different working conditions are collected through welding experiments to form real data. The two types of data are spatiotemporally aligned, outliers are removed, and the simulated molten pool surface width is calculated, finally forming a physical-data fusion dataset, which is divided into training set, validation set, and test set according to the proportion.

[0082] During model training, the LNN heat source parameter prediction model is shown in the attached figure. Figure 4 As shown, the model includes an input layer, a dynamic hidden layer, and an output layer. The input layer receives time-series data, the dynamic hidden layer captures time-series features, and the output layer outputs four heat source parameters. The model is trained using a dataset, with mean squared error as the loss function, and the parameters are optimized using a gradient descent algorithm. After training, the model can accurately predict heat source parameters.

[0083] As attached Figure 3 As shown, the predicted values ​​highly match the actual values, verifying the model's accuracy. The KAN temperature field prediction model is attached. Figure 5 As shown, the input layer receives 6-dimensional features, the hidden layer incorporates heat transfer mechanism constraints, and the lightweight adaptive adjustment network optimizes the KAN parameters through error feedback. A joint loss function is constructed to balance data fitting accuracy and physical laws. After training, the basic KAN model is obtained. The lightweight adaptive adjustment network is then trained with small sample data under new working conditions. By extracting prediction error features, the KAN model parameter correction is output, enabling the model to quickly adapt to new working conditions.

[0084] Real-time temperature field prediction and performance verification are performed according to the process shown in Figure 1. After collecting and preprocessing real-time data, the data is input into the LNN model to obtain heat source parameters. The heat flux density is calculated by substituting the parameters into the double ellipsoidal moving heat source function. The initial predicted temperature is then input into the basic KAN model. After adjustment by a lightweight network, the accurate temperature value is output. The entire process meets the real-time requirements. The optimized model has high prediction accuracy, and the temperature prediction error is controlled within the industrial allowable range. When adapting to multiple operating conditions, there is no need for full retraining. It can be quickly adapted by adjusting the lightweight network, which significantly improves efficiency.

[0085] The operating environment for this example is shown in the table below.

[0086] Table 1: Operating Environment .

[0087] In summary, this invention, through physical-data joint driving, integrates the core advantages of LNN and KAN, and combines a lightweight adaptive adjustment mechanism, effectively solving the shortcomings of existing technologies and providing a generalized solution for high-precision control of the welding process. Although the present invention has been disclosed above with preferred embodiments, they are not intended to limit the present invention. Any person skilled in the art can make various changes or modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be defined by the scope of protection of the claims of this application.

Claims

1. A physics-data co-driven welding temperature field digital twin method, characterized in that: Includes the following steps: Step S01: Construct a physics-data fusion dataset, which includes finite element simulation data based on heat transfer mechanism and real data collected from experiments; each data sample includes welding parameters, molten pool surface image, infrared image, workpiece node coordinates, material thermal property parameters, heat flux density, and corresponding temperature value; Step S02: Construct a real-time prediction model for heat source parameters based on a liquid neural network. With minimizing the prediction error of heat source parameters as the objective function, train the model using the heat source parameter data in the dataset from step S01 to obtain the real-time prediction model for heat source parameters. Step S03: Construct a welding temperature field prediction model based on KAN network. With minimizing the prediction error under the constraint of heat transfer mechanism as the objective function, train the model using the dataset from step S01 to obtain the welding temperature field prediction model. Step S04: Construct a lightweight adaptive adjustment network, which is used to extract the difference features between the new working condition data and the dataset, and to locally adjust the parameters of the KAN model trained in step S03; Step S05: Collect real-time data of the welding process through sensors, infrared cameras and industrial cameras. The real-time data includes welding current, arc voltage and molten pool surface image. The actual molten pool surface width parameter is extracted through image processing. Step S06: Preprocess the real-time monitoring data during the welding process; input the preprocessed real-time monitoring data into the heat source parameter prediction model to obtain the predicted real-time heat source parameters; calculate the real-time heat flux density using the heat source model; The real-time heat flux density is input into the temperature field prediction model and the error is calculated. The error is then input into a lightweight adaptive adjustment network to adjust the temperature field prediction model to the optimized KAN model. The optimized KAN model is then used to calculate the accurate temperature values ​​of each unit node of the welded workpiece.

2. The physics-data co-driven weld temperature field digital twin method of claim 1, wherein: In step S01, the method for constructing the physical-data fusion dataset includes the following steps: Step S11: Use finite element software to establish a three-dimensional geometric model of the welded workpiece, define the geometric dimensions and divide the structured mesh, and obtain the three-dimensional coordinate values ​​of each element node; Step S12: Input the material thermal properties of the workpiece to be welded, including thermal conductivity, density, specific heat, and latent heat. The thermal conductivity and latent heat are set according to temperature segments to conform to the actual heat transfer law. Step S13: Set welding parameters and heat source parameters, create a double ellipsoid moving heat source function, and calculate the heat flux density of each unit node at different times; Step S14: Define the initial temperature and thermal convection boundary conditions, solve the transient heat conduction equation, and obtain the simulated temperature values ​​of each node; Step S15: Collect real data under different working conditions through experiments, including welding current, arc voltage, molten pool surface image, infrared image, heat source parameters and corresponding temperature measurement values; Step S16: Perform spatiotemporal alignment between simulated data and real data, and use an outlier detection algorithm to remove outlier data. The simulated molten pool surface width in the heat source parameter dataset is obtained by calculating the distance between unit nodes at the left and right edges of the welded workpiece geometric model with the temperature value being the metal melting point temperature, thus forming a physical-data fusion dataset.

3. The physics-data co-driven weld temperature field digital twin method of claim 2, wherein: The welding parameters include welding current, arc voltage, welding speed, and heat source parameters; the heat source parameters include the length of the front half-axis of the welding direction, the length of the rear half-axis of the welding direction, the length of the half-axis of the weld width direction, and the length of the half-axis of the weld depth direction; the material thermal properties include thermal conductivity, density, specific heat, and latent heat, and the thermal conductivity and latent heat vary with temperature; the finite element simulation data are obtained by establishing a geometric model of the welded workpiece, setting boundary conditions, and solving the double ellipsoidal moving heat source function.

4. The physics-data co-driven weld temperature field digital twin method of claim 3, wherein: The power density distribution of the first half of the double ellipsoidal moving heat source function is shown in the following equation: The power density distribution in the latter half of the double-ellipsoidal moving heat source is as follows: In the above two equations, is the distance from the center of the heat source in the welding direction, is the distance from the center of the heat source in the fusion width direction, is the distance from the center of the heat source in the fusion depth direction, is the front half axis length in the welding direction, is the rear half axis length in the welding direction, is the half axis length in the fusion width direction, is the half axis length in the fusion depth direction, is the volumetric heat source intensity, and is the front-to-rear energy ratio, + = 2.

5. The physics-data co-driven weld temperature field digital twin method of claim 4, wherein: In step S02, the method for constructing a real-time prediction model of heat source parameters based on a liquid neural network includes the following steps: Step S21: The input layer receives time-series feature data, specifically the welding current, arc voltage, and weld pool surface width sequence at times [tn,t] before the current time t. The data is input into the model after standardization. Step S22: The dynamic hidden layer adopts the core structure of a liquid neural network, which includes a dynamic neuron pool and an adaptive connection weight matrix. The connection strength between neurons is adjusted in real time according to the changes in the input time series data. The activation function adopts a dynamic nonlinear activation mechanism, which can accurately capture the dynamic correlation between the heat source parameters and the input time series data during the welding process. Step S23: The output layer has 4 output nodes, each corresponding to the predicted value of the heat source parameters, and is output using a linear activation function; Step S24: Using the mean squared error between the predicted heat source parameters and the actual heat source parameters in the dataset as the loss function, the adaptive gradient descent algorithm is used to minimize the loss. The dynamic connection weights and neuron activation parameters of the model are updated iteratively. The number of iterations and the batch size are customized. After training, a real-time prediction model for heat source parameters is obtained.

6. The physics-data co-driven weld temperature field digital twin method of claim 5, wherein: In step S21, n is set to 40; in step S22, the number of hidden layer neurons is set to 32; in step S24, the number of iterations is set to 50 rounds and the batch size is set to 32.

7. The physics-data co-driven weld temperature field digital twin method of claim 6, wherein: In step S03, the method for constructing a welding temperature field prediction model based on a KAN network includes the following steps: Step S31: The input layer receives six dimensions of features: the six dimensions are: the three-dimensional coordinates of the welding workpiece unit node, the time step, the instantaneous heat flux density, and the temperature at the previous moment; Step S32: The KAN hidden structure consists of two hidden layers, using spline basis functions as activation functions. The feature transfer between layers is achieved through a fully connected approach, incorporating the constraints of the heat transfer partial differential equation. Step S33: The single output node of the output layer outputs the real-time temperature value of the corresponding unit node; Step S34: Using the mean square error between the predicted temperature value and the temperature value in the dataset as the base loss, add the residual term of the heat transfer partial differential equation as the physical constraint loss, construct a joint loss function, use the Adam optimizer to minimize the joint loss, iteratively update the model parameters, and obtain the trained basic KAN prediction model.

8. The physics-data co-driven weld temperature field digital twin method of claim 7, wherein: The specific method for step S04 is as follows: Step S41: A lightweight spline modulation structure is adopted, which includes a residual encoder and a spline parameter generator, and the total number of parameters is controlled within 10% of the basic KAN model. Step S42: The residual encoder inputs the prediction residual of the basic KAN model and extracts the error distribution features through a fully connected layer; the prediction residual is the absolute error sequence between the current predicted temperature and the actual temperature. Step S43: The core function of the spline parameter generator is to convert error features into modulation parameters of spline control points. The modulation parameters dynamically correct the prediction output of the basic KAN model by changing the morphological features of the spline function, including two adjustment methods: control point position offset and scale scaling. Step S44: Use small sample data under the new working condition to train the spline modulator end-to-end, and use the mean square error of temperature prediction of the modulated model as the loss function to optimize the modulator parameters. The small sample data is 5%-10% of the data in the base dataset.

9. The physics-data co-driven weld temperature field digital twin method of claim 8, wherein: The specific method of step S05 includes the following steps: Step S51: The welding current and arc voltage are collected in real time by an arc sensor at a sampling frequency of 100Hz to ensure that the data covers the base value and peak value phase of the welding pulse; Step S52: Capture visible light images of the molten pool surface using a high dynamic range (HDR) industrial camera. The lens is equipped with a bandpass filter with a center wavelength of 685±2nm and a full width at half maximum (FWHM) of 10±2nm to suppress welding arc interference. The image format is Mono8 with a resolution of 300×300. Step S53: Acquire infrared images of the welding area using an infrared camera with a wavelength range of 7.5-13μm, a temperature measurement range of -40℃-1500℃, and a sampling frequency of 100Hz. This image is used to extract the true surface temperature value of the welding area, providing a basis for subsequent error calculation. Step S54: Process the visible light image of the molten pool surface: Correct the image using the checkerboard method to achieve accurate matching between pixels and actual spatial positions; extract the region of interest; enhance image contrast using the Retinex algorithm to highlight the molten pool boundary; obtain continuous and complete molten pool edges using Canny edge detection and morphological operations; fit the left and right boundaries using the least squares method, and calculate the actual molten pool surface width based on pixel distance. Step S55: Preprocess the infrared image: remove noise by Gaussian filtering, and convert the pixel grayscale values ​​into actual temperature values ​​by combining the infrared camera calibration parameters. Select the real temperature data of the molten pool area as the benchmark for error calculation.

10. The physics-data co-driven weld temperature field digital twin method of claim 9, wherein: The specific method of step S06 includes the following steps: Step S61: During real-time welding, the pre-processed real-time time series data (welding current, arc voltage, and molten pool surface width at time [t-40,t]) is input into the LNN heat source parameter prediction model trained in step S02, and the predicted value of the heat source parameters at the current time is output in real time. Step S62: Substitute the real-time heat source parameters predicted by LNN into the double ellipsoidal moving heat source function, and calculate the real-time heat flux density of each unit node of the welded workpiece by combining the real-time welding current and arc voltage. Step S63: Input the instantaneous heat flux density, unit node coordinates, time step, and previous temperature into the basic KAN model trained in step S03, and output the initial predicted temperature value of each unit node in real time. Step S64: Simultaneously calculate the absolute error sequence between the initial predicted temperature value and the infrared measured temperature value to form real-time prediction residual data; input the residual data into the lightweight adaptive adjustment network trained in step S04 in real time, the network extracts error features in real time through the feature extraction module, and outputs the spline coefficient correction amount of the KAN model in real time through the parameter adjustment module. Step S65: Immediately input the real-time generated correction value into the basic KAN model, dynamically adjust the coefficients of the hidden layer spline activation function, and realize the real-time optimization of the basic KAN model; the optimized KAN model outputs the accurate temperature value of each unit node.

Images