Electric arc additive manufacturing temperature field and stress field real-time reconstruction method

Abstract

The invention discloses an electric arc additive manufacturing temperature field and stress field real-time reconstruction method, which comprises the following steps of: 1, constructing a sequential coupling finite element model in which heat conduction analysis and thermoelastic-plastic mechanical analysis are combined on the basis of process parameters of an electric arc additive manufacturing process and three-dimensional geometry of parts, generating a multi-physical field coupling data set among heat flux density, temperature and stress; step 2, constructing and training a neural network model based on a multi-physics field coupling data set, and realizing mapping from heat flux density to temperature and stress; 3, dynamically optimizing heat source model parameters in combination with real-time monitoring data and process parameters in the electric arc additive manufacturing process; and 4, based on the optimized heat source model parameters and the trained neural network model, real-time reconstruction of a temperature field and a stress field in the electric arc additive manufacturing process is achieved. According to the invention, the requirements of rapid sensing and intelligent regulation and control of temperature and stress states on an additive manufacturing site are met.

Description

Technical Field

[0001] The present invention relates to the technical field of additive manufacturing, and in particular to a method for real-time reconstruction of temperature and stress fields in arc additive manufacturing. Background Art

[0002] Arc additive manufacturing (AM) is an additive manufacturing method that uses an arc heat source to deposit metal wire layer by layer. It boasts significant advantages, including high forming efficiency, high material utilization, and low cost, and is widely used in aerospace, marine engineering, and rail transportation. During the AM process, uneven heat input can lead to significant temperature gradients, which can further induce residual stress and deformation within parts, severely impacting structural performance and forming quality. Therefore, accurately determining the temperature and stress field distribution is crucial to avoiding defects such as mechanical property degradation, component deformation, cracks, and fractures.

[0003] Currently, traditional methods rely heavily on numerical simulations of thermal-mechanical coupling. However, these simulations are computationally complex and time-consuming, making it difficult to meet the real-time and high-efficiency requirements of actual manufacturing processes. With the development of sensing technology, infrared temperature measurement, multispectral imaging, and other methods can achieve real-time observation of temperature fields, but they still cannot obtain stress information simultaneously, and data utilization is limited. Furthermore, current finite element and physical modeling methods typically fix heat source parameters, ignoring the dynamic nature of heat source parameters as they change with working conditions during the actual process, making it difficult to accurately reflect the actual heat input state.

[0004] In recent years, artificial intelligence methods such as deep learning have demonstrated tremendous potential in complex physical field modeling and data-driven prediction. By combining neural networks with finite element simulation results, efficient mapping between physical parameters and field variables can be achieved, with good generalization and real-time performance. Therefore, there is an urgent need for a comprehensive modeling approach that integrates physical models, real-time monitoring, and deep learning. This approach can quickly and accurately reconstruct temperature and stress fields during arc additive manufacturing, providing critical support for manufacturing process optimization and quality control. Summary of the Invention

[0005] This invention aims to address key technical issues such as low efficiency, insufficient accuracy, and difficulty in real-time response in temperature and stress field reconstruction during arc additive manufacturing. It provides a method for real-time reconstruction of temperature and stress fields during arc additive manufacturing. This method should exhibit high response speed and prediction accuracy, and be adaptable to dynamic changes in the manufacturing process, meeting the demands for rapid sensing and intelligent control of temperature and stress states in additive manufacturing sites.

[0006] The technical solutions for achieving the purpose of the present invention are:

[0007] A method for real-time reconstruction of temperature and stress fields in arc additive manufacturing, comprising:

[0008] Step 1: Based on the process parameters of the arc additive manufacturing process and the three-dimensional geometry of the part, a sequential coupled finite element model combining heat conduction analysis and thermo-elasto-plastic analysis is constructed to generate a multi-physics coupling data set between heat flux density, temperature, and stress.

[0009] Step 2: Build and train a neural network model based on the multi-physics coupling dataset to map heat flux to temperature and stress.

[0010] Step 3: Dynamically optimize the heat source model parameters by combining real-time monitoring data and process parameters during the arc additive manufacturing process;

[0011] Step 4: Based on the optimized heat source model parameters and the trained neural network model, real-time reconstruction of the temperature field and stress field during the arc additive manufacturing process is achieved.

[0012] Compared with the prior art, the present invention has the following significant advantages:

[0013] The present invention proposes a method for real-time reconstruction of temperature and stress fields in the arc additive manufacturing process. By integrating finite element multi-physics modeling, deep learning network training, and multi-spectral sensor data feedback, it achieves rapid and accurate prediction of complex thermal-mechanical behaviors in the manufacturing process. Compared with traditional methods that rely on thermal-mechanical coupling simulation, this method replaces the physical solution process with high computational complexity with a neural network, significantly improving the prediction efficiency and meeting the demand for real-time perception of temperature and stress states at the manufacturing site. By constructing a multi-physics coupling data set of heat flux density, temperature, and stress, and using three-dimensional convolution and Fourier neural operators to construct a neural network architecture, effective modeling of the spatiotemporal evolution process is achieved, giving the model good generalization ability and accuracy.

[0014] Furthermore, the present invention introduces a multispectral camera to capture the surface temperature field of the melt pool, and combines it with the L-BFGS-B optimization algorithm to dynamically correct the heat source model parameters, ensuring that the reconstruction results closely match the actual heat input state. This resolves the issue of traditional fixed heat source models not being consistent with process reality and enhances the system's adaptability to manufacturing fluctuations. This method fully integrates the theoretical foundations of physical modeling, the expressive power of deep learning, and the data support of sensing technology, significantly improving the quality and reliability of temperature and stress information reconstruction during arc additive manufacturing. This provides a solid technical foundation for online monitoring, process control, and defect warning of the manufacturing process, and has significant engineering application value and promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is the overall flow chart of the present invention.

[0016] Figure 2 Schematic diagram of arc additive manufacturing according to the present invention.

[0017] Figure 3 This is a temperature field distribution diagram in the arc additive manufacturing of the present invention.

[0018] Figure 4 This is a diagram of the stress field distribution in arc additive manufacturing according to the present invention. DETAILED DESCRIPTION

[0019] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0020] Combine Figures 1-4 The present invention provides a method for real-time reconstruction of temperature and stress fields in arc additive manufacturing, which is characterized by comprising the following steps:

[0021] Step 1: Based on the geometry and process conditions of the arc additive manufacturing process, a sequential coupled finite element model combining heat conduction analysis and thermo-elasto-plastic analysis is constructed to generate a multi-physics coupling dataset between heat flux density, temperature, and stress.

[0022] Step 2: Build and train a neural network model based on the multi-physics coupling dataset to achieve efficient mapping of heat flux density to temperature and stress;

[0023] Step 3: Dynamically optimize the heat source model parameters by combining real-time monitoring data and process parameters during the arc additive manufacturing process;

[0024] Step 4: Based on the optimized heat source model parameters and the trained neural network model, real-time reconstruction of the temperature field and stress field during the arc additive manufacturing process is achieved.

[0025] Furthermore, step one specifically includes:

[0026] Furthermore, the "sequentially coupled finite element model" in step one refers to a modeling approach that solves the temperature and stress fields step by step. This model first calculates the temperature field distribution using the transient heat conduction equation. It then uses the temperature field results as load input to solve the thermoplastic constitutive equations and obtain stress field information. Unlike strongly coupled models, the sequential coupling approach effectively reduces computational complexity while maintaining high simulation accuracy, making it suitable for modeling the thermal-mechanical field evolution during arc additive manufacturing.

[0027] The construction of the model includes the following steps:

[0028] (1) Establish a finite element model including the substrate and multilayer deposition path based on the three-dimensional geometric information of the arc additive manufacturing part; set the material's thermophysical and mechanical properties that vary with temperature;

[0029] (2) The arc heat source is modeled using a double ellipsoid heat source model, and the heat source area is divided into the front half and the back half, and the heat flux density q at the three-dimensional space coordinate point (x, y, z) is defined respectively. f (x,y,z) and q r (x,y,z), the specific formula is as follows:

[0030]

[0031]

[0032] where q f (x,y,z) and q r (x, y, z) are the heat flux densities at the spatial points (x, y, z) in the front and back half of the heat source, respectively. When x in space satisfies x≤x0 (the center of the heat source), it belongs to the front half of the heat source, otherwise it belongs to the back half; η is the welding thermal efficiency; V is the arc voltage; I is the welding current; b is the semi-axis length in the weld width direction; a f is the front half axis length in the welding direction, a f =b;a r is the rear half axis length in the welding direction, a r =2b; c is the semi-axis length in the direction of penetration; x is the distance from the center of the heat source in the welding direction; y is the distance from the center of the heat source in the direction of weld width; z is the distance from the center of the heat source in the direction of penetration; f1 and f2 are the distribution coefficients of the total energy of the heat source between the front half and the back half (i.e., energy proportional factors), f1+f2=2.

[0033] (3) Define the heat source movement trajectory according to the process path and set the time step and total duration of the simulation calculation;

[0034] (4) Apply heat conduction boundary conditions, convection and radiation boundary conditions, and initial temperature conditions;

[0035] (4) In the framework of thermo-mechanical sequential coupling, the transient heat conduction equation is first solved to obtain the time series temperature field, and then the temperature field is input as the load into the thermo-elasto-plastic constitutive equation to obtain the stress field distribution that evolves synchronously with the deposition process;

[0036] (5) Repeat the above (1-4) process under the typical double ellipsoid heat source model parameters (welding thermal efficiency, semi-axis length in the direction of weld width, semi-axis length in the direction of weld depth) and process parameter combination (current, voltage, speed) to obtain the temperature field and stress field (to obtain the equivalent stress value) results covering various working conditions.

[0037] Furthermore, the multi-physics coupling dataset between heat flux, temperature, and stress in step 1 is defined as follows: for each set of simulation conditions, during the entire additive manufacturing time history, the heat flux input value, temperature value, and equivalent stress value are collected for all nodes (discrete points in the finite element model) i in the finite element model at each time step t; the data (heat flux input value, temperature value, and equivalent stress value) of all time steps t and nodes i are uniformly coded according to the "time step-space node number" as the index to construct the multi-physics coupling dataset:

[0038]

[0039] Where N is the total number of nodes in the finite element model; T is the total time step number of the thermal-mechanical coupling simulation, q i (t) is the heat flux density input value of the i-th node at time step t; T i (t) is the temperature value of the i-th node at time step t; is the equivalent stress value of the i-th node at time step t;

[0040] The multi-physics field coupling data set is normalized and used to train a neural network model to achieve a nonlinear mapping relationship from heat flux density input values ​​to temperature values ​​and equivalent stress values.

[0041] Furthermore, step 2 specifically includes:

[0042] The heat flux input data of the entire additive manufacturing process is organized into a four-dimensional tensor Q:

[0043] Q∈R X×Y×Z×T (4)

[0044] Where R represents the real number domain, X, Y, Z are the total number of nodes of the three-dimensional additive structure under finite element discretization; the elements in the tensor are denoted as Q x,y,z,t , represents the heat flux input value at the position (x,y,z) at time step t.

[0045] The tensor Q covers the entire part geometry at all time steps, and for the spatial regions that have not yet been additively manufactured, the heat flux density value at all time steps is set to 0;

[0046] For time steps that have not yet occurred (i.e., time steps after the current prediction moment), the heat flux density values ​​of all spatial positions in the corresponding tensor are also uniformly assigned to 0 to ensure that the input tensor has the same structure during the training and inference stages and can represent the historical heat input state before any time;

[0047] The tensor Q is used as the input of the neural network model, and after spatiotemporal feature extraction and mapping operations, the three-dimensional temperature field corresponding to the current time step t is output. and equivalent stress field

[0048] Furthermore, the neural network in step 2 is a spatiotemporal joint regression network based on a three-dimensional convolutional network and a Fourier neural operator. Its structure includes: using multiple stacked three-dimensional convolutional layers on the input tensor Q, combining the normalization layer (BatchNorm3D) and the activation function (ReLU) to extract spatiotemporal fusion features, and obtaining the intermediate feature tensor:

[0049] F1∈R X×Y×Z×D (5)

[0050] Where D is the number of intermediate feature channels;

[0051] The intermediate feature tensor F1 is input into the Fourier neural operator (FNO) network, and a three-dimensional fast Fourier transform is applied to each intermediate feature channel. Some low-frequency modes are selected in the frequency domain for weighted mapping, and then restored to the spatial domain through inverse Fourier transform to obtain the enhanced spatial feature tensor F2∈R X×Y×Z×D ; Then F2 is input into two parallel output branches, each of which is composed of several convolutional layers or fully connected layers, respectively regressing the temperature field at the current time step and equivalent stress field

[0052] Furthermore, the neural network model training in step 2 includes: dividing the sample data set (multi-physics field coupling data set) into a training set and a validation set, which are used for model parameter update and performance evaluation respectively; using a supervised learning method during the training process, the heat flux density tensor Q is input into the neural network model, and the three-dimensional temperature field regression result of each spatial node at the current time step is obtained through forward propagation. and the equivalent stress field regression results And construct a joint loss function based on the mean square error (MSE) between it and the true label:

[0053]

[0054] Where N represents the number of all spatial nodes in the finite element model at the current time step t; i represents the spatial position index, that is, the number of each spatial node; T i (t), is the true label value obtained by finite element simulation; Output results for the neural network.

[0055] Adam is used as the optimization algorithm, and the learning rate, batch size, and number of iterations are set to update the network parameters (all learnable parameters in the neural network). During the training process, the minimum loss on the validation set is used as the selection criterion, and the corresponding network weights (the set of all learnable parameters in the neural network structure) are saved as the final inference model parameters.

[0056] Furthermore, step three specifically includes:

[0057] Get the surface temperature field T of the molten pool collected in real time by the multispectral camera at the current time step exp (x,y,z);

[0058] Under the heat source parameter set θ of the current time step (welding thermal efficiency η, semi-axis length b in the direction of weld width, semi-axis length c in the direction of weld depth), the double ellipsoid heat source model is used to calculate the heat flux density input tensor Q, and the corresponding three-dimensional temperature field is obtained through the neural network. and equivalent stress field

[0059] Construct the temperature field output by the neural network at the current time step The objective function between the measured temperature field and the actual temperature field;

[0060] The key parameters in the heat source model (welding thermal efficiency η, semi-axis length b in the direction of weld width, and semi-axis length c in the direction of weld depth) are used as optimization variables, and reasonable boundary ranges are set. The L-BFGS-B optimization algorithm is used to optimize the objective function. Minimize and solve to get the optimal parameter set θ * .

[0061] Furthermore, in step three, the temperature field of the molten pool surface is obtained by colorimetric temperature measurement using a multispectral camera, specifically including: collecting molten pool images through a multispectral camera system arranged above the additive manufacturing area, the camera is equipped with two photosensitive chips at the same time, which respond to the visible light band and the near-infrared band respectively, and synchronously collect the radiation intensity of the molten pool surface at two different wavelengths; based on the principle of colorimetric temperature measurement, using Planck's radiation law and Wien's approximation formula, a temperature-grayscale ratio relationship model is established at two selected wavelengths λ1 and λ2; before actual temperature measurement, experimental calibration is carried out using a high-temperature metal sample with known temperature characteristics, and a calibration curve is established by fitting the grayscale ratio of the red light channel and the near-infrared channel of the multispectral image at different temperatures for subsequent temperature inversion; in order to reduce the colorimetric temperature measurement error, a blue light filter and a three-band filter with specific transmission spectrum characteristics are installed at the front end of the multispectral camera to suppress glare and dust interference; at the same time, parameters such as the lens aperture and exposure time are adjusted to ensure the image signal-to-noise ratio and improve the accuracy of temperature measurement.

[0062] Furthermore, the heat source model parameter optimization in step 3 is performed using the L-BFGS-B algorithm, which specifically includes: setting the heat source parameter set to be optimized θ = [η, b, c], and specifying upper and lower limits for each parameter; converting the temperature field output by the neural network at the current time step into Compare with the measured temperature field and construct the residual objective function:

[0063]

[0064] Where j represents the node number on the melt pool surface; M represents the total number of mesh nodes; Predict the temperature for the neural network generated by the current heat source parameter set θ; is the measured temperature at the corresponding position.

[0065] The L-BFGS-B algorithm is used for iterative optimization to calculate the objective function gradient, update the parameters based on the approximate Hessian matrix, and keep the parameter value within the preset boundary range in each step of the search; when the gradient norm is less than the set threshold or the number of iterations reaches the upper limit, the optimization is terminated and the optimal heat source model parameter set θ is output. * .

[0066] Furthermore, step four specifically includes:

[0067] Based on the optimal parameter set θ obtained in step 3 * Update the double ellipsoid heat source model and calculate the heat flux density q(x, y, z, t) at the current time step t based on the actual arc voltage V and welding current I collected at the current time step;

[0068] Update the neural network input tensor Q using the heat flux density q(x,y,z,t) at the current time step;

[0069] The updated input tensor Q is input into the neural network model that has completed training and loaded the final weights, and the temperature field at the current time step t is regressed through forward inference. and equivalent stress field Realize real-time reconstruction of temperature and stress fields during arc additive manufacturing.

[0070] Example:

[0071] The flow chart of this method is a real-time reconstruction method of temperature field and stress field in arc additive manufacturing process. Figure 1 As shown; the schematic diagram of arc additive manufacturing involved in the embodiment is shown in the attached Figure 2 The temperature field distribution in arc additive manufacturing is shown in the attached Figure 3 The stress field distribution in arc additive manufacturing is shown in the attached Figure 4 The specific steps of this embodiment are:

[0072] During the implementation process, a 3D thermo-mechanical coupled finite element model was first constructed using ANSYS, using 316L stainless steel as the target material, based on the desired workpiece geometry and additive manufacturing path strategy. The constructed geometry consisted of a 300mm × 100mm × 5mm substrate with 10 weld beads deposited layer by layer along the Z-axis. Each layer was 200mm long, 1.5mm thick, and 5mm wide. SOLID226 elements were used for modeling, with a 0.5mm edge length. The model supported thermo-mechanical coupled solutions and volume activation to support layered loading.

[0073] In the model's material property definition, parameters such as thermal conductivity, specific heat capacity, density, Young's modulus, and yield strength are imported from the Materials Manual and expressed as multi-segment piecewise linear interpolation functions to express their temperature-dependent relationships. The material model uses temperature-dependent elastic-plastic behavior and assumes no phase changes during the manufacturing process.

[0074] To simulate the arc heat source, a double ellipsoid heat flux model was used and loaded into the ANSYS thermal analysis module through the APDL custom function. The initial parameters were set to thermal efficiency η = 0.8, and the front and rear semi-axes of the ellipsoid were a f =3mm,a r =6mm, width semi-axis b = 3mm, depth semi-axis c = 2mm, energy distribution coefficient f1 = 1.2, f2 = 0.8. The heat source motion trajectory and time are loaded through the *DEFINE_TRANSIENT path in APDL, and the welding speed is set to 9mm / s, the average wire feed speed is 150mm / s, the average arc current is 150A, and the average arc voltage is 20V. These layers are deposited by moving back and forth, and the idle time between layers is 20s. The iterative solution is performed with a time step of 0.2s. In terms of boundary conditions, the bottom of the substrate is fixed, and the remaining surfaces are set as convection boundaries (convection coefficient 20W / (m 2 ·K), ambient temperature 298K), and a radiation boundary (emissivity 0.3) is applied, and the initial overall temperature is set to room temperature 298K.

[0075] Within the framework of sequential thermal-mechanical coupling, a transient heat conduction analysis is first performed, generating a temperature field for all nodes at each time step. This temperature field is then fed into the structural analysis module as a load, calculating the equivalent stress at the corresponding time step. The dataset uses a 1-second time step interval, a total of 420 seconds, and 136,741 nodes. All simulation outputs are structured into a set of data pairs:

[0076]

[0077] where q i(t) is the heat flux density input value of the i-th node at time t (calculated by the heat source model), T i (t) and are the node temperature and equivalent stress respectively.

[0078] In the neural network modeling stage, a neural network developed based on the PyTorch2.7 framework is used, and the structure combines a three-dimensional convolutional network and a Fourier neural network. The input data is encoded into a shape of Q∈R 401×11×31×420 The four-dimensional tensor covers the entire three-dimensional space and the entire time step. The heat flux density of the nodes in the area not scanned by the heat source is assigned to 0 to ensure the uniformity of the tensor structure.

[0079] The 3D convolutional network structure includes a convolution kernel of 3×3×3, ReLU activation, BatchNorm, and 128 output intermediate feature channels. The Fourier neural operator network performs a 3D fast Fourier transform on each feature channel, selects the first 10 low-frequency modes for mapping, and then inversely transforms them, which enhances the network's ability to model non-local features in physical processes. The last two output branches are composed of a combination of convolution and fully connected layers, respectively predicting the temperature field. and equivalent stress field

[0080] The network loss function is set as:

[0081]

[0082] The training set uses 80% of the total simulation data, and the remaining 20% ​​is the validation set. The Adam optimizer is used with a learning rate of 1e-4, a batch size of 16, and 1000 training rounds.

[0083] In the actual additive manufacturing experiment, a multispectral camera (Datasheet_AD-130GE) was used to collect images of the molten pool surface. It was equipped with two photosensitive chips sensitive to visible light and near-infrared wavelengths. The fitting curve between the grayscale ratio of the two channels and the actual temperature was calibrated using high-temperature samples.

[0084] In the real-time monitoring stage, the image is collected and the surface temperature distribution T is obtained by inversion through the colorimetric temperature measurement algorithm. exp (x, y, z), and compare it with the predicted value of the neural network under the current heat source parameter θ By contrast, construct the objective function:

[0085]

[0086] Where θ is the heat source parameter set [η, b, c], which is iteratively optimized by the L-BFGS-B algorithm in the interval η∈[0.7,0.9], b∈[2.5,3.5], c∈[1.8,2.5], with one update per second.

[0087] The final optimized heat source parameter θ * It is used to update the input tensor Q and fed into the neural network in real time for temperature and stress field prediction.

[0088] This embodiment obtains high-fidelity temperature and stress field simulation data through a thermal-mechanical coupling finite element model built on the ANSYS platform, ensuring the physical consistency and coverage of the neural network training data. In terms of neural network structure, the combination of three-dimensional convolution and Fourier neural operators realizes deep modeling and feature extraction of complex spatiotemporal heat flow inputs, significantly improving the accuracy and generalization ability of predictions. At the same time, a multispectral camera system is introduced to realize non-contact, high-resolution monitoring of the surface temperature of the molten pool, and by optimizing the heat source model parameters in real time, the deviation between the simulation model and the actual working conditions is effectively bridged, so that the model can dynamically adapt to the fluctuations of heat input during the manufacturing process. This method has significant engineering value and promotion potential, and can provide key support for online quality monitoring, defect warning and closed-loop control in the arc additive manufacturing process.

[0089] Although the present invention has been disclosed above in terms of preferred embodiments, they are not intended to limit the present invention. Anyone 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 based on the scope of protection defined by the claims of this application.

Claims

1. A method for real-time reconstruction of temperature and stress fields in arc additive manufacturing, characterized in that: include: Step 1: Based on the process parameters of the arc additive manufacturing process and the three-dimensional geometry of the part, a sequential coupled finite element model combining heat conduction analysis and thermo-elasto-plastic analysis is constructed to generate a multi-physics coupling data set between heat flux density, temperature, and stress. Step 2: Build and train a neural network model based on the multi-physics coupling dataset to map heat flux to temperature and stress. Step 3: Dynamically optimize the heat source model parameters by combining real-time monitoring data and process parameters during the arc additive manufacturing process; Step 4: Based on the optimized heat source model parameters and the trained neural network model, real-time reconstruction of the temperature field and stress field during the arc additive manufacturing process is achieved.

2. The method for real-time reconstruction of temperature and stress fields in arc additive manufacturing according to claim 1, characterized in that: The establishment of the sequential coupled finite element model in step 1 includes: (1) Establish a finite element model including the substrate and multilayer deposition path based on the three-dimensional geometric information of the arc additive manufacturing part; set the material's thermophysical and mechanical properties that vary with temperature; (2) The arc heat source is modeled using a double ellipsoid heat source model. The heat source area is divided into the front half and the back half, and the heat flux density at the three-dimensional space coordinate point (x, y, z) is defined as follows: where q f (x,y,z) and q r (x, y, z) are the heat flux densities at the points (x, y, z) in the front and back half of the heat source, respectively; η is the welding thermal efficiency; V is the arc voltage; I is the welding current; b is the semi-axis length in the weld width direction; a f is the front half axis length in the welding direction; a r is the length of the rear semi-axis in the welding direction; c is the length of the semi-axis in the penetration direction; x is the distance from the center of the heat source in the welding direction; y is the distance from the center of the heat source in the weld width direction; z is the distance from the center of the heat source in the penetration depth direction; f1 and f2 are the distribution coefficients of the total energy of the heat source between the front and rear half zones; (3) Define the heat source movement trajectory and set the time step and total duration of the simulation calculation; (4) Apply heat conduction boundary conditions, convection and radiation boundary conditions, and initial temperature conditions; (4) First, the transient heat conduction equation is solved to obtain the time series temperature field, and then the temperature field is input as the load into the thermo-elasto-plastic constitutive equation to obtain the stress field distribution that evolves synchronously with the deposition process; (5) Repeat the above (1-4) process under the combination of typical double ellipsoid heat source model parameters and process parameters to obtain temperature field and stress field results covering various working conditions.

3. The method for real-time reconstruction of temperature and stress fields in arc additive manufacturing according to claim 1, characterized in that: The generated multiphysics coupling data set is: Where N is the total number of nodes in the finite element model; T is the total time step number of the thermal-mechanical coupling simulation, q i (t) is the heat flux density input value of the i-th node at time step t; T i (t) is the temperature value of the i-th node at time step t; is the equivalent stress value of the i-th node at time step t.

4. The method for real-time reconstruction of temperature and stress fields in arc additive manufacturing according to claim 1, characterized in that: Step 2 specifically includes: The heat flux input data of the entire additive manufacturing process is organized into a four-dimensional tensor Q: Q∈R X×Y×Z×T Where R represents the real number domain, X, Y, Z are the total number of nodes of the three-dimensional additive structure under finite element discretization, and T is the total time step number of thermal-mechanical coupling simulation; The tensor Q is used as the input of the neural network model, and after spatiotemporal feature extraction and mapping operations, the three-dimensional temperature field corresponding to the current time step t is output. and equivalent stress field 5. The method for real-time reconstruction of temperature and stress fields in arc additive manufacturing according to claim 4, characterized in that: The neural network in step 2 is a spatiotemporal joint regression network based on a three-dimensional convolutional network and a Fourier neural operator. Its structure includes: using multiple stacked three-dimensional convolutional layers on the input tensor Q, combining normalization layers and activation functions to extract spatiotemporal fusion features, and obtaining an intermediate feature tensor: F1∈R X ×Y×Z×D Where D is the number of intermediate feature channels; The intermediate feature tensor F1 is input into the Fourier neural operator network. By applying a three-dimensional fast Fourier transform to each intermediate feature channel, some low-frequency modes are selected in the frequency domain for weighted mapping, and then restored to the spatial domain through inverse Fourier transform to obtain the enhanced spatial feature tensor F2∈R X×Y×Z×D ; Then input F2 into two parallel output branches to regress the temperature field at the current time step respectively and equivalent stress field 6. The method for real-time reconstruction of temperature and stress fields in arc additive manufacturing according to claim 5, characterized in that: The input tensor Q is input to the neural network model, and the temperature regression result at the current time step is obtained through forward propagation and the equivalent stress field regression results And construct a joint loss function based on the mean square error (MSE) between it and the true label: Where N represents the number of all spatial nodes in the finite element model at the current time step t; i represents the spatial position index, that is, the number of each spatial node; T i (t), is the true label value obtained by finite element simulation; Output results for the neural network.

7. The method for real-time reconstruction of temperature and stress fields in arc additive manufacturing according to claim 1, characterized in that: Step three includes: Get the molten pool surface temperature field T collected in real time at the current time step exp (x,y,z); According to the heat source model parameters of the current time step, the heat source model parameter set θ is set, and the heat flux density input tensor Q is calculated using the double ellipsoid heat source model. The corresponding three-dimensional temperature field is obtained through the neural network. and equivalent stress field Construct the temperature field output by the neural network at the current time step The objective function between the measured temperature field and the actual temperature field; The heat source model parameters are used as optimization variables, the boundary range is set, and the L-BFGS-B optimization algorithm is used to optimize the objective function. Minimize and solve to obtain the optimal heat source model parameter set θ * .

8. The method for real-time reconstruction of temperature and stress fields in arc additive manufacturing according to claim 7, characterized in that: The heat source model parameter optimization is performed using the L-BFGS-B algorithm, which specifically includes: setting the heat source parameter set to be optimized θ = [η, b, c], and specifying the upper and lower limits for each parameter; converting the temperature field output by the neural network at the current time step into Compare with the measured temperature field and construct the residual objective function: Where j represents the node number on the melt pool surface; M represents the total number of mesh nodes; Predict the temperature for the neural network generated by the current heat source parameter set θ; is the measured temperature at the corresponding position, η, b, and c are the welding thermal efficiency, the semi-axis length in the direction of weld width, and the semi-axis length in the direction of weld depth, respectively; The L-BFGS-B algorithm is used for iterative optimization to calculate the objective function gradient, update the parameters based on the approximate Hessian matrix, and keep the parameter value within the preset boundary range in each step of the search; when the gradient norm is less than the set threshold or the number of iterations reaches the upper limit, the optimization is terminated and the optimal parameter set θ is output. * .

9. The method for real-time reconstruction of temperature and stress fields in arc additive manufacturing according to claim 7, characterized in that: The surface temperature field of the molten pool is obtained by colorimetric temperature measurement using a multispectral camera, specifically including: collecting molten pool images through a multispectral camera system arranged above the additive manufacturing area, the camera is simultaneously equipped with two photosensitive chips, which respond to the visible light band and the near-infrared band respectively, and synchronously collect the radiation intensity of the molten pool surface at two different wavelengths; establishing a temperature-grayscale ratio relationship model at two selected wavelengths; before actual temperature measurement, experimental calibration is carried out using a metal sample with known temperature characteristics, and a calibration curve is established by fitting the grayscale ratio of the red light channel and the near-infrared channel of the multispectral image at different temperatures for subsequent temperature inversion.

10. The method for real-time reconstruction of temperature and stress fields in arc additive manufacturing according to claim 1, characterized in that: Based on the optimal heat source model parameter set obtained in step 3, the double ellipsoid heat source model is updated, and the heat flux density at the current time step is calculated based on the actual arc voltage and welding current collected at the current time step; Update the neural network input tensor using the heat flux density at the current time step; The updated input tensor is input into the neural network model that has completed training and loaded with the final weights. The temperature field and equivalent stress field at the current time step are regressed through forward reasoning to achieve real-time reconstruction of the temperature field and stress field in the arc additive manufacturing process.

Images