A method and system for real-time compensation of skeleton welding deformation based on dynamic feedback

Abstract

This invention relates to the field of mechanical manufacturing and welding automation technology, and discloses a method and system for real-time compensation of welding deformation of a skeleton based on dynamic feedback. The method includes acquiring three-dimensional geometric model data of the skeleton, material thickness distribution data, and welding path sequence data; processing this data to obtain an initial heat distribution map and a thermal deformation lag time series; extracting the superposition parameters of sequential thermal effects from the sequence, inputting these parameters into a preset neural network prediction model to obtain a predicted value of asynchronous deformation trend; if the predicted value exceeds a preset trend threshold, acquiring an adjustment signal to obtain a real-time compensation vector; updating the welding path sequence data based on the real-time compensation vector to obtain an optimized thermal deformation lag time series, which is then input into the neural network prediction model for iterative verification; if the superposition parameters of sequential thermal effects converge, obtaining the final predicted deformation distribution map; and generating a welding process control command sequence based on the distribution map. This method can solve the problem of low geometric accuracy in welded components.

Description

Technical Field

[0001] This invention relates to the field of mechanical manufacturing and welding automation technology, and in particular to a method and system for real-time compensation of skeleton welding deformation based on dynamic feedback. Background Technology

[0002] Currently, skeleton structure welding is a core process in the manufacturing of large-scale mechanical equipment, engineering components, and aerospace precision parts. It directly determines the overall geometric accuracy, structural strength, and service life of the product, and its deformation control level has become a key indicator of the advancement of welding technology. As high-end equipment develops towards larger size, lighter weight, and higher precision, the complexity of skeleton structures is constantly increasing, the material thickness distribution is becoming more uneven, and the difficulty of welding path planning is significantly increasing. This places more stringent requirements on the precision control of metal cutting and welding equipment such as automatic and semi-automatic electric arc and plasma arc welding machines.

[0003] In existing technologies, welding deformation control mainly relies on two types of solutions. One is pre-weld process parameter optimization, which uses simulation to preset parameters such as welding current, voltage, and speed to try to reduce deformation at the source. The other is post-weld mechanical correction, which corrects the deformation that has occurred after welding through pressure correction, flame correction, and other methods. While these methods can play a certain role in welding simple structures, they have significant limitations in welding complex skeleton structures, especially in meeting the continuous operation and high-cycle production requirements of metal cutting and welding equipment manufacturing processes such as automatic and semi-automatic electric arc welding machines and plasma arc welding machines. The expansion and contraction of materials caused by welding heat input are not instantaneous, but have significant hysteresis propagation characteristics. The process of heat transfer, accumulation, and release from the weld point to the surrounding area often takes several seconds or even longer. Existing feedforward control schemes only perform welding based on preset parameters and cannot respond to real-time thermal deformation dynamics during the welding process; simple feedback adjustment does not consider the time delay characteristics of thermal deformation, resulting in compensation actions always being a step too late, making it difficult to offset the cumulative deviation caused by the lag, ultimately causing the deformation of critical parts of the structure to exceed the allowable range. This problem is particularly prominent in continuous welding operations of metal cutting and welding equipment such as automatic and semi-automatic electric arc and plasma arc welding machines.

[0004] Existing technologies suffer from low geometric accuracy in welded components. Summary of the Invention

[0005] This invention provides a method and system for real-time compensation of welding deformation of skeleton based on dynamic feedback, in order to solve the problem of low geometric accuracy of welded components.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a method for real-time compensation of skeleton welding deformation based on dynamic feedback, comprising: The three-dimensional geometric model data and material thickness distribution data of the skeleton structure are obtained, and the three-dimensional geometric model data and the material thickness distribution data are processed by the finite element simulation method to obtain the initial thermal distribution mapping. Obtain welding path sequence data, and interpolate the welding path sequence data according to the initial heat distribution mapping to obtain the thermal deformation hysteresis time series. The parameters of the superposition of sequential thermal effects are extracted from the thermal deformation hysteresis time series, and the parameters of the superposition of sequential thermal effects are input into a preset neural network prediction model to obtain the predicted value of asynchronous deformation trend. Obtain geometric shape influence data; if the predicted value of asynchronous deformation trend exceeds a preset trend threshold, obtain an adjustment signal from the neural network prediction model; and fuse the geometric shape influence data according to the adjustment signal to obtain a real-time offset compensation vector. Based on the real-time offset compensation vector, the welding path sequence data is updated, and the heat distribution mapping is recalculated using the preset simulation method to obtain the optimized thermal deformation hysteresis time series. The optimized thermal deformation lag time series is input into the neural network prediction model for iterative verification. If the superposition parameters of the sequential thermal effects converge, the final predicted deformation distribution map is obtained. Based on the final predicted deformation distribution map, a welding process control command sequence is generated, and the control command sequence is transmitted to the execution system to obtain the geometric accuracy index of the skeleton structure welding.

[0007] Secondly, the present invention provides a real-time compensation system for skeleton welding deformation based on dynamic feedback, comprising: The initial thermal distribution mapping module is used to acquire the three-dimensional geometric model data and material thickness distribution data of the skeleton structure, and process the three-dimensional geometric model data and the material thickness distribution data through the finite element simulation method to obtain the initial thermal distribution mapping. The thermal deformation hysteresis time series module is used to acquire welding path sequence data, and interpolate the welding path sequence data according to the initial heat distribution mapping to obtain the thermal deformation hysteresis time series. The asynchronous deformation prediction module is used to extract the superposition parameters of sequential thermal effects from the thermal deformation lag time series, input the superposition parameters of sequential thermal effects into a preset neural network prediction model, and obtain the asynchronous deformation trend prediction value. The real-time offset compensation vector module is used to acquire geometric shape influence data. If the asynchronous deformation trend prediction value exceeds the preset trend threshold, an adjustment signal is obtained from the neural network prediction model. The geometric shape influence data is fused according to the adjustment signal to obtain the real-time offset compensation vector. The optimized lag time series module is used to update the welding path sequence data according to the real-time offset compensation vector, and recalculate the heat distribution mapping using the finite element simulation method to obtain the optimized thermal deformation lag time series. The final deformation prediction module is used to input the optimized thermal deformation lag time series into the neural network prediction model for iterative verification. If the superposition parameters of the sequential thermal effects converge, the final predicted deformation distribution map is obtained. The control command generation and execution module is used to generate a welding process control command sequence based on the final predicted deformation distribution map, transmit the control command sequence to the execution system, and obtain the geometric accuracy index of the skeleton structure welding.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention constructs a thermal deformation lag time series containing the duration of heat accumulation and release through finite element simulation, and accurately quantifies the thermal response delay law of the skeleton structure by combining phase lag parameters, thus solving the core pain point that the existing technology cannot capture the thermal deformation time delay characteristics. This invention first constructs a three-dimensional mesh model containing thickness gradient information based on the three-dimensional geometric model data and material thickness distribution data of the skeleton structure. Through unsteady heat conduction calculation and heat flow vector analysis, an initial heat distribution mapping that can reflect the thermal resistance difference in different regions is obtained. Then, combined with the welding path sequence data, the heat flow density vector flow trajectory is tracked, and the delay data from heat input to deformation response in each region is accurately analyzed to form a complete thermal deformation lag time series. This design deeply couples the geometry of the skeleton structure, the material thickness distribution and the thermal response delay, and can clearly identify the deformation lag differences in different regions caused by thickness gradient and local stiffness differences. This allows the compensation action to accurately match the actual deformation evolution rhythm, avoids the cumulative deviation caused by the late compensation in the existing technology, and significantly improves the welding accuracy and stability of complex skeleton structures in the manufacturing of metal cutting and welding equipment such as automatic and semi-automatic electric arc and plasma arc welding machines.

[0009] (2) This invention extracts the superposition parameters of sequential thermal effects through convolution operations and uses a long short-term memory network model to deeply capture the asynchronous deformation trends of different regions, achieving accurate prediction and targeted offsetting of complex deformations such as twisting and undulation, breaking through the limitation of existing technologies that are difficult to quantify the superposition effect of thermal effects. This invention does not simply ignore the superposition effect of sequential welds, but generates a superposition weight matrix of thermal effects by performing convolution operations on the lag time series of thermal deformation, accurately quantifying the cumulative contribution intensity of thermal input at different times at each node in space; then, combined with the node thermal conduction attenuation values, it constructs a feature vector group containing the contribution of thermal effects at multiple time scales, which, after being input into the long short-term memory network model, can effectively capture the long-range dependence of the welding thermal process and output the local thermal strain evolution sequence of each region. Through gradient calculation and nonlinear regression analysis of this sequence, the order of deformation initiation, development speed and final amplitude differences of different regions can be accurately predicted. The real-time offset compensation vector generated based on the prediction results further integrates the geometric shape influence data and the superposition attenuation law of thermal effects, which can specifically offset the superposition effect of thermal effects caused by continuous operation of multiple weld points.

[0010] (3) This invention constructs a dynamic closed-loop control logic to dynamically update the welding path and heat distribution mapping in real time by offsetting the compensation vector. Iterative verification ensures the convergence of the heat effect superposition parameters, significantly improving the intelligence level and adaptability of the welding process. It perfectly adapts to the continuous operation and high-cycle production requirements of metal cutting and welding equipment manufacturing such as automatic and semi-automatic electric arc and plasma arc welding machines. After generating the real-time offsetting compensation vector, this invention does not perform the compensation action all at once. Instead, it uses a topology reconstruction algorithm to generate a corrected welding path sequence containing compensation information, recalculates the heat distribution mapping matrix, and couples it to the thermo-elastic-plastic constitutive model to form a dynamic deformation field. By comparing the cumulative deviation gradient between the dynamic deformation field and the theoretical design model, the thermal deformation lag time series is optimized and then input into the neural network prediction model for iterative verification until the heat effect superposition parameters converge. Finally, a high-precision predicted deformation distribution map is output. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the real-time compensation method for skeleton welding deformation based on dynamic feedback provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the real-time compensation system for skeleton welding deformation based on dynamic feedback provided in the second embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Reference Figure 1 The first embodiment of the present invention provides a real-time compensation method for skeleton welding deformation based on dynamic feedback, including the following steps: S11, Obtain the three-dimensional geometric model data and material thickness distribution data of the skeleton structure, and process the three-dimensional geometric model data and the material thickness distribution data through the finite element simulation method to obtain the initial thermal distribution mapping; S12, Obtain welding path sequence data, and interpolate the welding path sequence data according to the initial heat distribution mapping to obtain the thermal deformation hysteresis time series; S13, extract the superposition parameters of sequential thermal effects from the thermal deformation hysteresis time series, input the superposition parameters of sequential thermal effects into a preset neural network prediction model, and obtain the predicted value of asynchronous deformation trend. S14, acquire geometric shape influence data; if the asynchronous deformation trend prediction value exceeds the preset trend threshold, acquire the adjustment signal from the neural network prediction model, and fuse the geometric shape influence data according to the adjustment signal to obtain a real-time offset compensation vector. S15, based on the real-time offset compensation vector, update the welding path sequence data, and recalculate the heat distribution mapping using the finite element simulation method to obtain the optimized thermal deformation hysteresis time sequence; S16, The optimized thermal deformation lag time series is input into the neural network prediction model for iterative verification. If the superposition parameters of the sequential thermal effects converge, the final predicted deformation distribution map is obtained. S17. Based on the final predicted deformation distribution map, a welding process control instruction sequence is generated, and the control instruction sequence is transmitted to the execution system to obtain the geometric accuracy index of the skeleton structure welding.

[0014] In step S11, the three-dimensional geometric model data and material thickness distribution data of the skeleton structure are acquired. The three-dimensional geometric model data and the material thickness distribution data are processed by the finite element simulation method to obtain the initial thermal distribution mapping, including: Obtain the three-dimensional geometric model data and material thickness distribution data of the skeleton structure, and construct a three-dimensional mesh model of the skeleton containing thickness gradient information based on the three-dimensional geometric model data and the material thickness distribution data. Unsteady-state heat conduction calculations were performed using the aforementioned skeleton 3D mesh model to obtain transient response temperature field data. Perform heat flux vector analysis on the transient response temperature field data to obtain the initial heat distribution mapping; The phase lag parameter is calculated based on the initial heat distribution mapping, and the delay time corresponding to the phase lag parameter is determined to complete the delay characteristic analysis of the skeleton structure.

[0015] In one implementation, this embodiment constructs a technical solution from data acquisition to latency characteristic quantification. During the data acquisition phase, a multi-source data fusion strategy is employed to ensure the integrity and accuracy of the basic data. The three-dimensional geometric model data of the skeleton structure is acquired using a high-precision three-dimensional laser scanner with a scanning resolution set to 0.02 mm, comprehensively covering all external surfaces and key internal cavities of the skeleton. After the scanned data is exported in STL format, it undergoes noise reduction, hole filling, and topology optimization processing to completely eliminate defects such as burrs and missing data generated during the scanning process. Material thickness distribution data is acquired through a dual-channel approach. Visible areas on the external surface are measured point-by-point using an ultrasonic thickness gauge, with the measurement interval set to 5 to 10 mm based on the structural complexity to ensure the spatial continuity of the thickness data. For invisible internal areas, three-dimensional reconstruction is performed using skeleton structure design drawings and CT scan images. Gray-scale thresholding and edge extraction are performed on the CT slice images to accurately reconstruct the internal thickness distribution. Finally, a one-to-one correspondence mapping relationship is established between the processed three-dimensional geometric model data and the material thickness distribution.

[0016] It is important to note that the construction of the 3D mesh model of the skeleton prioritizes preserving thickness gradient information to achieve a balance between computational accuracy and efficiency. Based on the correlated 3D geometric model and thickness data, an adaptive meshing algorithm is employed. Mesh refinement is applied to regions with drastic thickness variations, with mesh cell sizes controlled between 0.5 and 2 mm. For regions with uniform thickness, the mesh size is appropriately increased, with mesh cell sizes set between 2 and 5 mm, forming a non-uniform structured mesh that incorporates thickness gradient information. Eight-node hexahedral cells are selected for the mesh, as this type of cell offers advantages in terms of numerical stability and high computational accuracy in heat conduction calculations. After meshing, a mesh quality check tool is used to rigorously verify indicators such as cell twist and aspect ratio, ensuring that the twist of all cells is less than 15 degrees and the aspect ratio is controlled within 1:5 to avoid computational errors caused by poor mesh quality. Simultaneously, thickness attribute labels are embedded in the mesh model, ensuring that each mesh cell contains corresponding material thickness information, providing support for subsequent thermal resistance calculations and phase hysteresis analysis.

[0017] In another implementation, the finite element thermal analysis model and unsteady-state heat conduction calculations strictly adhere to the laws of thermophysics to ensure the accuracy of the temperature field data. First, based on the characteristics of the skeleton material, corresponding thermophysical parameters, including thermal conductivity, specific heat capacity, and density, are retrieved from the material library. All parameters are set as functions of temperature to accurately reflect the material's thermal properties as a function of temperature. Then, boundary conditions conforming to actual welding conditions are applied. A moving heat source is used to load the welding area, and a double-elliptical heat source model is selected, with its major axis, minor axis, and energy distribution coefficient determined experimentally based on welding process parameters. Convective heat transfer boundary conditions are applied to the non-welding areas, with the convective heat transfer coefficient set to 15 to 30 Kelvin per square meter based on the ambient temperature and structural surface condition. Adiabatic boundary conditions are applied to the structural symmetry planes to prevent heat transfer along these planes. Unsteady-state heat conduction calculations employ an implicit time-integration scheme, with the time step dynamically adjusted based on the thermal response rate. During the welding phase, where heat changes rapidly, the time step is set to 0.01 seconds to ensure accurate capture of transient temperature changes. In the cooling phase, where heat diffuses slowly, the time step is automatically adjusted to 0.1 to 0.5 seconds to balance computational accuracy and efficiency. During the calculation, the temperature change curves at each node are monitored in real time. When the rate of temperature change between adjacent time steps is less than 0.5 degrees Celsius per second, the region is considered to have reached a thermally stable state. Finally, transient response temperature field data across the entire time domain and space are output.

[0018] It should be noted that the heat flux vector analysis and thermal resistance calculation employ refined processing methods to accurately construct the initial heat distribution map. Heat flux vectors are extracted from the transient response temperature field data to obtain the heat flux direction and density at each node. The heat flux vectors are then superimposed on the temperature field for visualization, clearly showing the propagation path and accumulation areas of heat within the skeletal structure. Thermal resistance calculations utilize a unit-level refined calculation approach, precisely calculating the thickness, thermal conductivity, and temperature gradient of each grid unit. The equivalent thermal resistance of each region is then obtained by summarizing the node coupling relationships. Based on the heat flux vector distribution and equivalent thermal resistance calculation results, an interpolation algorithm is used to construct a three-dimensional initial heat distribution map. This map, using grid nodes as carriers, contains multi-dimensional information such as temperature, heat flux density, and thermal resistance, comprehensively reflecting the heat distribution state and heat transfer characteristics of the skeletal structure.

[0019] Specifically, signal analysis methods were used to calculate the phase lag parameter and determine the delay time, quantifying the delay characteristics of the skeleton structure. Key monitoring points on the skeleton structure were selected, including nodes along the welding path, nodes with abrupt thickness changes, and stress concentration nodes. The temperature response time series and heat flow excitation signal time series of each monitoring point were extracted. A Fast Fourier Transform (FFT) was used to transform the two time series to the frequency domain, and the phase difference between the temperature response signal and the heat flow excitation signal at each frequency component was calculated, which is the phase lag parameter corresponding to that frequency. By analyzing the variation of the phase lag parameter with frequency, the phase lag value at the dominant frequency was determined, and the corresponding delay time was then calculated. Statistical analysis of the delay times of all key monitoring points was performed, and a delay time distribution cloud map was drawn, clearly showing the differences in thermal response delay in different regions of the skeleton structure, thus completing the delay characteristic analysis.

[0020] It is worth noting that, based on the principle of frequency domain analysis, the phase lag value, in radians, is converted into a delay time, in seconds. After determining the dominant frequency, the delay time is calculated by dividing the phase lag value by the product of 2π and the dominant frequency. The dominant frequency is determined by analyzing the frequency components with the most concentrated energy in the spectrum of the heat flow excitation signal, or by selecting the low-frequency components that contribute the most to the thermal deformation of the structure. Through this conversion, the abstract phase relationship can be quantified into a specific time lag, providing an accurate time reference for the subsequent construction of the thermal deformation lag time series.

[0021] In step S12, welding path sequence data is acquired, and the welding path sequence data is interpolated according to the initial heat distribution mapping to obtain a thermal deformation hysteresis time series, including: Obtain welding path sequence data, process the welding path sequence data using spatial interpolation, and generate a sequence of weld point coordinates with timestamps; Nonlinear transient thermal structure coupling calculations are performed based on the timestamped weld point coordinate sequence to obtain transient temperature field evolution data; By tracing the heat flux density vector flow trajectory through the transient temperature field evolution data, a set of heat input propagation process paths can be obtained; Deformation response delay data is calculated along the set of heat input propagation paths. The deformation response delay data is then analyzed, and the heat accumulation time and heat release time are calculated based on the deformation response delay data to obtain the thermal deformation hysteresis time series.

[0022] In one implementation, this embodiment employs a multi-source collaborative acquisition strategy for the welding path sequence data acquisition stage to ensure data integrity and accuracy. Pre-set welding path planning data is exported through the welding robot control system, including the three-dimensional coordinates of the starting point, turning point, and ending point of each weld, as well as preset welding speed parameters. Simultaneously, the welding process is captured in real-time by a high-definition vision camera, and the actual welding torch trajectory is extracted using image recognition algorithms to correct the preset path data, eliminating path deviations caused by mechanical errors. For complex curved surface skeleton structures, a laser tracker is additionally used to acquire welding torch position information in real-time, with a sampling frequency set to 100Hz to ensure the capture of subtle changes in the welding torch movement. Finally, this data is integrated to form complete welding path sequence data containing the preset path and actual trajectory corrections.

[0023] It should be noted that the spatial interpolation process focuses on generating a high-precision, timestamped sequence of weld point coordinates. A cubic spline interpolation algorithm is used to interpolate the discrete welding path sequence data, dynamically adjusting the interpolation interval based on the welding speed to ensure uniform distances between adjacent interpolation points. This expands the discrete path nodes into a continuous sequence of weld point coordinates, with each coordinate point corresponding to a unique timestamp, accurately recording the spatial position of the welding torch at different times. After interpolation, the coordinate sequence undergoes smoothing filtering to remove high-frequency noise interference, ensuring the continuity and stability of the sequence and providing an accurate spatiotemporal reference for subsequent thermal-structural coupling calculations.

[0024] In another implementation, the nonlinear transient thermal-structural coupling calculation follows the physical laws of thermo-mechanical coupling to obtain real transient temperature field evolution data. The time-stamped weld point coordinate sequence is imported into the software, and a nonlinear transient thermal-structural coupling model is established by combining the material properties of the skeleton structure (such as the thermal conductivity, specific heat capacity, and elastic modulus of high-temperature alloys as a function of temperature). A moving double-elliptical heat source model is used for the welding heat source. The energy distribution coefficient of the heat source is calibrated according to the welding process parameters (welding current, voltage, duty cycle) to ensure that the heat source energy input is consistent with the actual welding conditions. During the calculation, the latent heat of phase change option is enabled to consider the heat absorption and release during the material melting and solidification process; simultaneously, the large deformation option is enabled to adapt to the significant deformation of the structure during welding. An adaptive adjustment strategy is adopted for the time step. During the welding start-up and arc termination phases with drastic changes in heat input, the time step is set to 0.001s to ensure the rapid fluctuations in the temperature field are captured; during the welding stabilization phase, the time step is adjusted to 0.01s to balance calculation accuracy and efficiency. Through multiphysics coupling calculation, the system outputs full-time transient temperature field evolution data from the start of welding to the end of cooling, including the temperature value of each grid node at different times.

[0025] It should be noted that the heat flux density vector flow trajectory tracking stage aims to accurately delineate the heat input propagation path. The vector tracking tool in the processing module analyzes the transient temperature field evolution data time-by-time, extracting the heat flux density vector at each node, including the heat flow direction and intensity. The heat flux density vectors at different times are superimposed with the temperature field for visualization, clearly presenting the dynamic process of heat propagation from the weld point to the surrounding area. For special parts such as hollows and reinforcing ribs in complex skeletal structures, the focus is on tracking the branching and convergence of heat flow, identifying the main channel and branch paths of heat flow propagation. For example, in the welding of the combustion chamber casing of an aero-engine, heat diffuses laterally from the main channel of the weld to the base material on both sides, with heat flow convergence occurring near the reinforcing ribs, while in thin-walled areas, the heat flow propagation speed is faster and the direction is more concentrated. Based on the tracking results, paths with heat flux density greater than a preset threshold are selected, forming a set of heat input propagation process paths. Each path includes the starting point, ending point, propagation direction, and heat flow intensity variation pattern.

[0026] It is important to note that the core of the deformation response delay data calculation and analysis lies in quantifying the duration of heat accumulation and release. Along the set of heat input propagation paths, temperature-time curves and deformation-time curves of key nodes on each path are extracted. By comparing the peak times of the two curves, the delay time of the deformation response relative to the heat input is determined. For example, if a key node on a certain path reaches its temperature peak 3.2 seconds after the heat input, while the deformation peak occurs 1.5 seconds after the temperature peak, the deformation response delay time of that node is 1.5 seconds. Further analysis of the deformation response delay data defines the heat accumulation time as the time from when the node temperature begins to rise significantly (exceeding room temperature by 50°C) to when the node begins to undergo plastic deformation; and the heat release time as the time from when the heat source leaves the node's region to when the node deformation stabilizes (deformation change is less than 0.001 mm for three consecutive time steps). The heat accumulation time and heat release time are calculated separately for different paths. For example, the heat accumulation time of the main weld channel is about 4.5s and the heat release time is about 12s; the heat accumulation time of the branch path near the stiffener is about 3s and the heat release time is about 15s. The deformation response delay time, heat accumulation time, and heat release time of all paths are integrated and arranged in timestamp order to form a complete thermal deformation hysteresis time series.

[0027] In step S13, the superposition parameters of sequential thermal effects are extracted from the thermal deformation hysteresis time series, and the superposition parameters of sequential thermal effects are input into a preset neural network prediction model to obtain the predicted value of asynchronous deformation trend, including: The thermal deformation hysteresis time series is convolved to obtain the thermal effect superposition weight matrix; Calculate the node heat conduction attenuation value based on the heat effect superposition weight matrix, and construct a set of feature vectors for the superposition of successive heat effects; The superimposed feature vector group of sequential thermal effects is input into the neural network prediction model, and the local thermal strain evolution sequence is output. Nonlinear regression analysis was performed on the local thermal strain evolution sequence to obtain the predicted value of asynchronous deformation trend.

[0028] In one implementation, the convolution operation in this embodiment focuses on generating a precise thermal impact superposition weight matrix to quantify the cumulative effect of thermal inputs at different times. A one-dimensional convolution kernel is used to perform sliding convolution processing on the thermal deformation lag time series. The kernel size is set according to the thermal deformation propagation period; for example, for a scenario where the thermal impact lasts for about 5 seconds, a kernel with a length of 50 (corresponding to a sampling frequency of 10Hz) is selected. During the convolution process, by adjusting the kernel parameters, the weight distribution exhibits a decaying pattern over time, meaning that recent thermal inputs have a higher weight on the current node, while the weight of distant thermal inputs gradually decreases. Through this convolution operation, the one-dimensional thermal deformation lag time series is transformed into a two-dimensional thermal impact superposition weight matrix. The rows of the matrix correspond to different nodes, and the columns correspond to different times. Each element represents the weight of the thermal input at a certain time on the corresponding node, clearly presenting the superposition pattern of thermal impacts across multiple time scales.

[0029] It is worth noting that the convolution operation aims to extract the local superposition pattern of thermal effects over time; the length of the convolution kernel should be set according to the physical time constant of thermal effect propagation, rather than the length of the entire sequence. For example, if the analysis shows that the main interaction of thermal effects between adjacent solder joints occurs within a time window of about 1 second, then the convolution kernel length can be set to 10 (covering 1 second of data) corresponding to a sampling frequency of 10 Hz. By using multiple convolution kernels of different lengths, both short-term and long-term thermal effect dependencies can be captured simultaneously.

[0030] It should be noted that the calculation of node thermal conduction attenuation values ​​and the construction of feature vector groups focus on integrating thermal impact information in both spatial and temporal dimensions. Based on the thermal impact superposition weight matrix and combined with the 3D mesh model of the skeleton structure, the thermal conduction attenuation value for each node is calculated. This value is determined by the material thickness, thermal conductivity, and topology of the node's region. For example, nodes in thick-walled regions have higher thermal conduction attenuation values, indicating greater heat loss during propagation and a weaker impact on distant nodes; nodes in thin-walled regions have lower thermal conduction attenuation values, indicating a longer heat propagation distance and a stronger impact. Based on the thermal impact superposition weight matrix and the node thermal conduction attenuation values, a set of sequential thermal impact superposition feature vectors is constructed. Each feature vector corresponds to a node and includes the thermal impact superposition weight, thermal conduction attenuation value, and corresponding thermal deformation hysteresis data for that node across multiple consecutive time steps.

[0031] In another implementation, the training of the neural network prediction model and the output of the local thermal strain evolution sequence preferably utilize a long short-term memory network to capture the long-range dependencies of the thermal process. The model structure includes an input layer, hidden layers, and an output layer. The dimension of the input layer matches the dimension of the superimposed feature vector group of sequential thermal effects. Three hidden layers are set, each containing 256 neurons. The dimension of the output layer corresponds to the length of the local thermal strain evolution sequence. Training data comes from a database of 1000 historical welding faults and simulation-generated sample data (4000 samples from the historical welding fault database and 6000 samples from simulation), covering thermal deformation data of different skeleton structures, welding processes, and fault types. Data augmentation techniques are used to expand the sample size to ensure the model's generalization ability. During training, an adaptive moment estimation optimization algorithm is used to minimize the mean square error between the predicted and actual values, while a dropout mechanism is introduced to prevent overfitting. The constructed feature vector group of sequential thermal effects is input into the trained model. The model selectively memorizes key thermal effect information and forgets irrelevant noise through a gating mechanism, and outputs the local thermal strain evolution sequence of each node. This sequence contains the thermal strain values ​​of the node at different times, clearly showing the change law of strain over time. For example, the thermal strain of a thin-walled node rises rapidly when the heat source is close and slowly decreases after the heat source leaves, with the peak appearing about 1.2 seconds after the heat source leaves.

[0032] It should be noted that the core of the nonlinear regression analysis lies in extracting the predicted value of asynchronous deformation trends. Gradient calculations are performed on the local thermal strain evolution sequence to obtain the strain rate curve. The speed of deformation development is determined by analyzing the slope of the curve. Based on the function predicting the deformation of nodes over a future period, and combined with the geometric accuracy requirements of the skeleton structure, the predicted value of asynchronous deformation trends is extracted. This predicted value not only includes the final predicted deformation of the node but also covers key parameters such as deformation initiation time and development speed, accurately reflecting the asynchronous nature of deformation in different regions. For example, the predicted deformation of the node in the weld initiation section is 0.3 mm, with an early deformation initiation time but slow development; the predicted deformation of the node in the arc termination region is 0.6 mm, with a later deformation initiation time but rapid development, both exhibiting significant asynchronous deformation characteristics.

[0033] In step S14, if the predicted value of the asynchronous deformation trend exceeds a preset trend threshold, an adjustment signal is obtained from the neural network prediction model. Based on the adjustment signal, geometric influence data is fused to obtain a real-time compensation vector, including: Compare the difference between the predicted value of asynchronous deformation trend and the preset trend threshold, and extract the adjustment signal that reflects the deviation gradient; Based on the adjustment signal and the geometric influence data, a geometric correction adjustment tensor is obtained. The geometric influence data includes the local stiffness distribution information and geometric characteristic parameters of the component. A dynamic correction factor is obtained by considering the historical attenuation function of the thermal superposition effect introduced by the geometric correction adjustment tensor. The historical attenuation function is constructed based on the attenuation law of historical welding heat input data. By combining the dynamic correction factor with the geometric correction adjustment tensor, a real-time compensation vector for offsetting the superposition effect of thermal effects is obtained.

[0034] It should be noted that obtaining the adjustment signal from the neural network prediction model specifically means that while the neural network prediction model outputs the asynchronous deformation trend prediction value, it also simultaneously outputs a confidence parameter reflecting the reliability or potential error range of the prediction value. The generation of the adjustment signal is a comprehensive judgment process. First, the asynchronous deformation trend prediction value output by the model is compared with a preset trend threshold. Second, the comparison result is weighted in conjunction with the confidence parameter output by the model. An adjustment signal containing compensation direction and intensity indication is generated based on the weighted deviation. This signal is not directly output by the model, but is calculated based on the model's output result by subsequent processing methods.

[0035] In one implementation, this embodiment constructs a real-time offset compensation vector generation scheme with multi-dimensional collaborative correction. In the deviation gradient and adjustment signal extraction stage, the focus is on quantifying deformation deviations and generating targeted adjustment instructions. The preset trend threshold is determined based on the geometric accuracy requirements of the skeleton structure and welding process standards. Specifically, based on the final product's geometric accuracy tolerances of the skeleton structure, such as the maximum allowable deformation of key dimensions, the upper limit of the allowable instantaneous deformation trend increment at each key node along the welding path is calculated through reverse derivation using thermo-elastic-plastic simulation. Secondly, combining historical welding process databases, the average mapping coefficient between the predicted deformation trend value and the measured final deformation amount under similar process parameters (current, voltage, speed) is statistically analyzed. Finally, the preset trend threshold is set as the maximum deformation amount divided by the average mapping coefficient, multiplied by a safety factor, typically 0.6~0.8, used to trigger compensation in the early stages of deformation trend and reserve control margin. This threshold can be differentiated according to the sensitivity of different welding areas, such as arc initiation, arc termination, and thickness transition zones. The difference between the asynchronous deformation trend prediction value and the preset trend threshold is calculated; the absolute value of the difference is the deformation deviation, and the sign of the difference reflects the deformation direction (positive for outward expansion deformation, negative for inward contraction deformation). Based on the deformation deviation, a deviation gradient is constructed, and a linear mapping algorithm is used to convert the deviation gradient into a standardized adjustment signal, with the signal strength positively correlated with the deviation gradient.

[0036] It's important to note that the core of constructing the geometric correction adjustment tensor lies in integrating geometric influence data to achieve spatially targeted compensation. Geometric influence data is obtained through a combination of finite element simulation and actual measurement: using a 3D mesh model of the skeleton structure, the local stiffness distribution information of each region is calculated; for example, the local stiffness near stiffeners is 3-5 times that of thin-walled regions. Geometric feature parameters of the component surface, including curvature, slope, and locations of abrupt changes in wall thickness, are collected using a laser scanner. After normalizing the local stiffness distribution information and geometric feature parameters, a multi-dimensional geometric influence factor matrix is ​​constructed, with each matrix element corresponding to the geometric influence weight of a mesh node. The geometric influence factor matrix is ​​then weighted and corrected based on the adjustment signal strength to generate the geometric correction adjustment tensor. The dimension of this tensor is consistent with the number of nodes in the 3D mesh model of the skeleton, and each element represents the amount of compensation correction required for the corresponding node due to geometric differences.

[0037] In another implementation, this embodiment considers the temporal decay law of the superposition effect of thermal effects during the dynamic correction factor generation stage. The historical decay function of the superposition effect of thermal effects is constructed based on fitting a large amount of historical welding data. By analyzing the degree of influence of heat input at different times on subsequent deformation, the decay coefficient and decay period are determined. For example, after the heat source passes through a certain node, the thermal effect decays by 50% within 1 second, 80% within 3 seconds, and below 20% after 5 seconds. Based on this, an exponential historical decay function is constructed. The historical data of heat input during the welding process is extracted in real time, including the heat input intensity, location, and duration at each time. This data is substituted into the historical decay function to calculate the residual thermal effect at the current time. The residual thermal effect is multiplied by a preset weighting coefficient to obtain the dynamic correction factor, which is controlled within the range of 0.2-0.9. For example, if the residual thermal effect at a node is 0.4 two seconds after the heat source passes, and the preset weighting coefficient is 0.8, then the dynamic correction factor is 0.32; if the residual effect is 0.2 five seconds after the heat source passes, the dynamic correction factor is 0.16, ensuring that the dynamic correction factor can reflect the temporal changes of the superposition of thermal effects in real time.

[0038] It should be noted that the dynamic correction factor range (0.2-0.9) is set based on engineering experience and physical limitations; the lower limit of 0.2 ensures that even when the residual effect of thermal action is very weak, a basic correction amount is still retained to cope with uncertainties; the upper limit of 0.9 avoids overcompensation and prevents the introduction of new instability or oscillation due to the compensation action itself; this range was determined after regression analysis of a large number of historical welding cases, and adjusting the compensation intensity within this range can effectively cover more than 95% of working conditions; for special materials or structures, fine-tuning can be made within this benchmark range.

[0039] Specifically, this embodiment generates the final compensation command through multi-dimensional parameter collaborative computation in the real-time offset compensation vector fusion stage. Matrix multiplication is used to fuse the geometric correction adjustment tensor and the dynamic correction factor; the compensation vector component of each node is equal to the product of the geometric correction adjustment tensor element and the dynamic correction factor of that node. Subsequently, delay information from the thermal deformation lag time series is introduced to perform time-series calibration on the fusion result. For example, for thick-walled regions with long thermal deformation lag times, the time weight of the compensation vector component is appropriately increased to ensure that the compensation action is synchronized with the actual deformation response. The compensation vector components of each node are arranged in spatial coordinate order to form a complete real-time offset compensation vector. This vector not only includes the compensation displacement of each node but also the compensation direction and compensation execution time. For example, the compensation vector of a node at the beginning of a weld is (-0.4 mm, 0.2 seconds), indicating that an inward contraction of 0.4 mm will be performed after 0.2 seconds; the compensation vector of a node in the arc termination region is (-0.6 mm, 0.1 seconds), corresponding to a faster compensation response and greater compensation intensity, ensuring accurate offsetting of the accelerated deformation trend caused by the superposition of thermal effects.

[0040] In step S15, the welding path sequence data is updated according to the real-time offset compensation vector, and the heat distribution mapping is recalculated using the finite element simulation method to obtain the optimized thermal deformation hysteresis time series, including: Based on the real-time offset compensation vector, a corrected welding path sequence containing compensation information is generated using a topology reconstruction algorithm; Based on the modified welding path sequence, the heat input is calculated in a virtual simulation environment, and a heat distribution mapping matrix reflecting the heat conduction delay characteristics is established. The heat distribution mapping matrix is ​​coupled to the thermo-elastic-plastic constitutive model to form a dynamic deformation field; By comparing the dynamic deformation field with the theoretical design model, the cumulative deviation gradient is calculated, and the optimized thermal deformation hysteresis time series is determined based on the cumulative deviation gradient.

[0041] In one implementation, this embodiment employs a graph-based topology reconstruction algorithm during the generation of topology reconstruction and corrected welding path sequences. First, it analyzes key information such as displacement compensation amount, compensation direction, and execution time in the real-time compensation vector. Combined with a three-dimensional mesh model of the skeleton structure, it identifies high-risk areas with high deformation prediction values ​​(such as weld termination sections, thin-walled areas, and stiffener connections). The algorithm divides the original continuous welding path into several sub-path segments and assigns differentiated compensation intensities based on the compensation needs of different areas: for areas with large deformation trends, short path segments with reverse offsets are generated, and reverse constraints are introduced in advance to offset thermal deformation; for areas with gentler deformation, path continuity is maintained and the path direction is fine-tuned. Simultaneously, the welding sequence of the sub-paths is adjusted according to the compensation execution time to ensure that the compensation action is synchronized with the rhythm of thermal deformation evolution, ultimately forming a corrected welding path sequence containing compensation information.

[0042] It should be noted that this embodiment focuses on accurately calculating the heat input and heat conduction delay characteristics during virtual simulation and the construction of the heat distribution mapping matrix. The corrected welding path sequence is imported into the ABAQUS virtual simulation environment, and a heat conduction simulation model is established by combining the thermal properties of the skeleton material (thermal conductivity, specific heat capacity, etc., as they change with temperature). Based on the length and offset of the sub-paths and the welding process parameters (current, voltage, welding speed), the heat input corresponding to each sub-path is calculated. The heat input intensity is appropriately reduced in deformation-risk areas, while normal heat input is maintained in areas with high rigidity. During the simulation, temperature response data of each grid node is collected at a frequency of 100Hz, recording the temperature changes at different times after the heat source passes, and a heat distribution mapping matrix is ​​established with node coordinates-time-temperature values ​​as the dimensions. This matrix clearly presents the heat conduction delay law in different regions. For example, the heat source reaches its temperature peak 1.2 seconds after passing a thin-walled node, while the temperature peak of the adjacent thick-walled node is delayed to 2.1 seconds. The matrix quantifies this delay gradient distribution numerically, providing an accurate basis for subsequent deformation field calculations.

[0043] In another implementation, the thermo-elastic-plastic constitutive model is coupled with the dynamic deformation field generation process, taking into account the thermo-mechanical coupling effect. The heat distribution mapping matrix is ​​used as the thermal load input and coupled into the thermo-elastic-plastic constitutive model, which has pre-imported mechanical parameters such as the stress-strain curves and temperature-dependent yield strength of the skeleton material. During the calculation, large deformation and geometric nonlinearity options are enabled to simulate the plastic deformation and geometric changes of the structure during welding, while also considering the influence of the asynchronous thermal expansion and contraction on the deformation. Through multiphysics coupling calculations, a dynamic deformation field is generated. This deformation field evolves in real time and outputs displacement data for each moment and node, clearly presenting the complete deformation process of the structure from the start of welding to cooling and stabilization.

[0044] It should be noted that when determining the thermal deformation lag time series after calculating and optimizing the cumulative deviation gradient, the deviation must be quantified and the timing parameters dynamically adjusted. The dynamic deformation field is compared node-by-node with the theoretical design model, and the difference between the actual displacement and the design displacement of each node at different times is calculated. The distribution pattern of the difference is statistically analyzed along the weld length to obtain the cumulative deviation gradient. For example, the cumulative deviation gradient of a certain casing weld is 0.38 mm / m at the beginning, rising to 0.62 mm / m in the middle, and reaching 0.91 mm / m in the arc-end region due to heat accumulation. Priorities are assigned based on the magnitude of the cumulative deviation gradient: areas with a deviation gradient exceeding 0.50 mm / m are high priority, requiring an extended thermal deformation lag judgment time of 2.4 seconds to ensure the compensation action covers the entire deformation process; areas with a deviation gradient between 0.30 and 0.50 mm / m are medium priority, maintaining a lag judgment time of 1.5 seconds; and areas with a deviation gradient below 0.30 mm / m are low priority, requiring a shortened lag judgment time of 0.9 seconds to improve efficiency. The lag time adjustment results of different regions are integrated in the order of timestamps to form an optimized thermal deformation lag time series. This series can be dynamically adapted according to the actual deviation distribution, making subsequent compensation actions more accurate and effectively reducing cumulative deviations.

[0045] It should be noted that the priority threshold for the cumulative deviation gradient is set proportionally based on the total length of the skeleton structure and the overall geometric accuracy requirements, such as the total deformation tolerance. For example, the high-priority threshold is set as the total deformation tolerance divided by the total length multiplied by β1, and the medium-priority threshold is set as the total deformation tolerance divided by the total length multiplied by β2, where β1 and β2 are proportional coefficients, typically β1 is 0.7~0.8 and β2 is 0.4~0.5. The 0.50mm / m and 0.30mm / m mentioned in the example are example values ​​calculated based on a typical casing component with a total length of about 2 meters and a total deformation tolerance of about 1.2 millimeters. This setting method ensures that the threshold is associated with the accuracy control target of the specific workpiece.

[0046] In step S16, the optimized thermal deformation hysteresis time series is input into the neural network prediction model for iterative verification. If the superposition parameters of the sequential thermal effects converge, the final predicted deformation distribution map is obtained, including: The optimized thermal deformation hysteresis time series is processed using the neural network prediction model to analyze the superposition parameters of the sequential thermal effects. Substitute the superimposed parameters of the sequential thermal effects into the preset transfer equation to calculate the predicted values ​​of the grid node displacements. A residual vector is generated based on the predicted displacement values ​​of the grid nodes, and it is determined whether the residual vector satisfies the preset convergence condition. If the preset convergence condition is met, the model parameters are locked, and the final predicted deformation distribution map is output. The node offset is quantified based on the final predicted deformation distribution map, and the overall torsional deformation control level is reflected by the node offset.

[0047] In one implementation, this embodiment groups the optimized thermal deformation lag time series by node and inputs it into a trained Long Short-Term Memory (LSTM) prediction model. This model selectively extracts thermal impact information at different time steps through a gating mechanism, accurately identifying the interaction between successive weld heat inputs. For example, the model can analyze the influence weight of thermal deformation at the weld initiation stage on the middle stage, the superposition effect of the middle stage on the arc termination stage, and quantify the transmission intensity and attenuation law of thermal impact in different regions. The analyzed superposition parameters of successive thermal impacts cover key indicators such as the thermal impact weights between nodes, time delay coefficients, and superposition attenuation rates.

[0048] It should be noted that a precise mapping from thermal superposition parameters to displacement is achieved through a pre-defined transfer equation. This pre-defined transfer equation is constructed based on thermo-elastic-plastic mechanics theory, integrating mechanical parameters such as the elastic modulus, Poisson's ratio, and coefficient of thermal expansion of the skeleton material, as well as temperature data from the optimized heat distribution mapping matrix. The analytically obtained sequential thermal influence superposition parameters are substituted into the equation, and the predicted displacement values ​​are calculated one by one for each mesh node, including linear displacement and angular displacement in the X, Y, and Z directions. During the calculation, a nonlinear correction term is enabled to consider the influence of plastic deformation after material yielding, ensuring that the predicted displacement values ​​are consistent with the actual welding deformation.

[0049] In another implementation, this embodiment uses a laser tracker to collect real-time actual displacement data of some key grid nodes during the welding process, or calls measured displacement data of similar structures from a historical welding database, comparing the actual values ​​with the predicted displacement values ​​of the grid nodes point by point. The difference between the predicted and actual values ​​for each node is calculated and arranged in order of node number to form a residual vector, with the dimension of the residual vector matching the number of grid nodes. A preset convergence condition is set based on the geometric accuracy requirements of the skeleton structure. If the residual vector meets the convergence condition, it is determined that the superposition parameters of the preceding and following thermal effects have converged, and the weights, biases, and other parameters of the current neural network model are locked to ensure the stability of subsequent predictions. If the convergence condition is not met, the residual vector is fed back to the model input, the weight allocation of the optimized thermal deformation lag time series is adjusted, and parameter analysis and displacement prediction are performed again until the convergence condition is met.

[0050] The preset convergence condition is the Euclidean norm of the residual vector, i.e., the square root of the sum of squares of all node residuals is less than or equal to the target accuracy. The target accuracy is usually set to 10% to 20% of the tolerance of the key dimensions of the skeleton structure. For example, if the tolerance is ±0.5mm, the target accuracy can be 0.05mm to 0.1mm. The maximum absolute value of all components in the residual vector is less than or equal to 2 to 3 times the target accuracy, which is used to control the local maximum deviation. When the above convergence conditions are met in three consecutive iterations, the parameter is determined to be converged.

[0051] It should be noted that the final predicted deformation distribution map output and the node offset quantification process enable the visualization of deformation patterns and the assessment of control levels. Based on the converged model parameters, the final predicted displacement values ​​of all mesh nodes are calculated, imported into visualization software, and a final predicted deformation distribution map in the form of a color cloud map is generated. The distribution map uses different colors to distinguish the magnitude of deformation, clearly presenting the spatial distribution pattern of deformation, such as the deformation concentrated on the inner side of the weld initiation section and the deformation shifting outward in the arc termination area. Based on the distribution map, the node offset of each region is quantified, and the mean, maximum, and standard deviation of node offsets over the entire weld length are statistically analyzed. For example, if the mean node offset is controlled within 0.3 mm and the maximum offset does not exceed 0.6 mm, it indicates that the overall torsional deformation is effectively constrained. At the same time, the deformation uniformity coefficient is calculated to assess the consistency of deformation in different regions, providing a quantitative basis for subsequent process optimization and ensuring that the geometric accuracy of the skeleton structure welding meets the design requirements.

[0052] In step S17, based on the final predicted deformation distribution map, a welding process control command sequence is generated, and the control command sequence is transmitted to the execution system to obtain the geometric accuracy indicators of the skeleton structure welding, including: A geometric compensation model is constructed based on the final predicted deformation distribution map. The geometric compensation model includes full node correction information calculated using the inverse mapping algorithm. The geometric compensation model is registered with the preset welding torch trajectory to generate a delay compensation process parameter matrix. The delay compensation process parameter matrix is ​​obtained by adjusting the energy input rate according to the thermal deformation hysteresis time series. The delay cancellation process parameter matrix is ​​encoded to generate a welding process control command sequence; The control command sequence is transmitted to the execution system, which executes the commands and feeds back the actual welding geometry data. Based on the actual welding geometry data, the geometric accuracy index of the skeleton structure welding is determined to offset the delay characteristics and superposition effects.

[0053] In one implementation, this embodiment targets complex skeleton structures such as aero-engine combustion chamber casings and large engineering machinery frames, ensuring precise offsetting of the delayed characteristics and superposition effects of thermal deformation during the welding process to ultimately achieve the preset geometric accuracy indicators. In the geometric compensation model construction stage, a reverse mapping algorithm is used to obtain full node correction information. Based on the final predicted deformation distribution map, key data such as the predicted deformation offset, deformation direction, and deformation rate of all mesh nodes are extracted. Using the reverse mapping algorithm, the predicted deformation trend of each node is reversed to calculate the geometric correction amount that can offset the deformation, including displacement compensation values ​​and rotation compensation values ​​in the X, Y, and Z directions. For example, if a node in the weld initiation section is predicted to expand outward by 0.4 mm, the reverse mapping algorithm calculates a displacement compensation value of 0.4 mm for inward contraction; if a node in the arc termination region is predicted to have a rotation deformation of 0.03 radians, a corresponding rotation compensation value of -0.03 radians is generated. The correction information of all nodes is integrated to construct a geometric compensation model. This model fully covers all nodes of the skeleton structure, ensuring the comprehensiveness and specificity of the compensation action.

[0054] It is important to note that the key to successful registration between the geometric compensation model and the preset welding torch trajectory is achieving precise alignment between the compensation information and the welding path. Preset welding torch trajectory data, including the spatial coordinates, speed, and attitude angles of the welding torch, is imported. A coordinate transformation algorithm maps the node correction information from the geometric compensation model to the welding torch trajectory coordinate system, ensuring consistency in their spatial references. Based on the thermal deformation lag time series, the execution timing and parameters of the welding torch trajectory are dynamically adjusted: for thick-walled areas with long thermal deformation lag times, the dwell time of the welding torch in that area is appropriately delayed, and the energy input rate is simultaneously reduced; for thin-walled areas with short thermal deformation lag times, the welding torch speed is increased, and the energy input rate is moderately increased to match the deformation response rhythm. For example, if the thermal deformation lag time at the beginning of welding is 2.4 seconds, the energy input rate of the welding torch in that area is reduced by 10% from the standard value; in the middle section, the lag time is shortened to 1.5 seconds, and the energy input rate is increased to 105% of the standard value. Through the above registration and adjustment, a delay compensation process parameter matrix is ​​generated. The matrix contains key parameters such as energy input rate, movement speed, displacement compensation, and rotation angle compensation for each welding torch trajectory segment.

[0055] In another implementation, this embodiment converts the process parameter matrix into executable instructions through encoding conversion. Using an industry-standard code encoding protocol, each parameter in the delay cancellation process parameter matrix is ​​encoded, converting the energy input rate into welding current and voltage control instructions, the motion speed into welding torch feed speed instructions, and the displacement and rotation angle compensation into welding torch position correction instructions. After encoding, all instructions are arranged in welding sequence to form a complete welding process control instruction sequence, ensuring that the instruction execution order matches the welding path order. Simultaneously, checksums and fault-tolerant instructions are added to the instruction sequence. When the execution system detects abnormal parameters, it can trigger a pause or emergency adjustment mechanism to improve the stability of the welding process.

[0056] It should be noted that the precise transmission of commands and the acquisition of actual data are achieved through industrial Ethernet. The control command sequence is transmitted to the welding execution system at a transmission rate of 100Mbps to ensure real-time command performance. The execution system controls the welding torch movement, energy output, and path correction according to the command sequence. Simultaneously, it acquires real-time welding geometric data, including the actual position of the welding torch, weld formation dimensions, and node displacement, using sensors such as laser trackers and vision cameras. The sampling frequency is 50Hz to ensure comprehensive capture of geometric changes during the welding process. The data acquired by the sensors is filtered and noise-reduced before being fed back to the control center, forming a closed-loop feedback data chain.

[0057] Specifically, a quantitative assessment of welding quality is achieved through comparative analysis. The actual welding geometric data is compared node-by-node with the theoretical design model data, calculating key indicators such as the deviation between actual and design displacements, weld dimensional tolerances, and overall form and position tolerances. For example, the mean, maximum, and standard deviation of node offsets along the entire weld length are statistically analyzed. If the mean is controlled within 0.3 mm, the maximum offset does not exceed 0.6 mm, and the standard deviation is less than 0.1 mm, the geometric accuracy is deemed satisfactory. The straightness and roundness tolerances of the weld are calculated; if they all meet the tolerance levels specified in the design drawings, it is confirmed that the thermal deformation delay characteristics and superposition effects have been effectively offset. Finally, all assessment data are integrated to determine the geometric accuracy indicators for the skeleton structure welding, forming a complete quality assessment report to provide data support for subsequent process optimization.

[0058] In summary, this invention discloses a real-time compensation method for skeleton welding deformation based on dynamic feedback. The method includes acquiring three-dimensional geometric model data and material thickness distribution data of the skeleton structure; constructing a mesh model containing thickness gradient information through finite element simulation; obtaining an initial heat distribution mapping through unsteady-state heat conduction calculation, heat flux vector analysis, and thermal resistance calculation; and completing delay characteristic analysis by combining phase lag parameters. Next, welding path sequence data is acquired, and a time-stamped weld point coordinate sequence is generated through spatial interpolation. Transient temperature field evolution data is obtained through nonlinear transient thermal structure coupling calculation, and the heat input propagation path is determined by tracing the heat flux density vector trajectory. Deformation response delay data is analyzed to construct a thermal deformation lag time series. Subsequently, the superposition parameters of sequential thermal effects are extracted from this time series, and a feature vector is constructed through convolution operations and numerical calculations of nodal heat conduction attenuation. The system takes a group of parameters and inputs them into a neural network prediction model to obtain the predicted value of asynchronous deformation trend. If the predicted value exceeds the preset threshold, the system extracts the adjustment signal and fuses the geometric influence data. It then introduces the thermal effect superposition historical attenuation function to generate a real-time offset compensation vector. Based on this vector, the welding path sequence data is updated, and the thermal distribution mapping is recalculated to obtain the optimized thermal deformation lag time series. The optimized time series is input into the neural network model for iterative verification. After the parameters of the superposition of the sequential thermal effects converge, the final predicted deformation distribution map is output. Finally, a geometric compensation model is constructed based on this distribution map. It is registered with the preset welding torch trajectory to generate a delay offset process parameter matrix. After encoding, it forms a welding process control command sequence and is transmitted to the execution system. The geometric accuracy index is determined by the feedback of the actual welding geometric data, thereby achieving accurate offsetting of the thermal deformation delay characteristics and the superposition effect of the sequential thermal effects.

[0059] Reference Figure 2 The second embodiment of the present invention provides a real-time compensation system for skeleton welding deformation based on dynamic feedback, comprising: The initial thermal distribution mapping module is used to acquire the three-dimensional geometric model data and material thickness distribution data of the skeleton structure, and process the three-dimensional geometric model data and the material thickness distribution data through the finite element simulation method to obtain the initial thermal distribution mapping. The thermal deformation hysteresis time series module is used to acquire welding path sequence data, and interpolate the welding path sequence data according to the initial heat distribution mapping to obtain the thermal deformation hysteresis time series. The asynchronous deformation prediction module is used to extract the superposition parameters of sequential thermal effects from the thermal deformation lag time series, input the superposition parameters of sequential thermal effects into a preset neural network prediction model, and obtain the asynchronous deformation trend prediction value. The real-time offset compensation vector module is used to acquire geometric shape influence data. If the asynchronous deformation trend prediction value exceeds the preset trend threshold, an adjustment signal is obtained from the neural network prediction model. The geometric shape influence data is fused according to the adjustment signal to obtain the real-time offset compensation vector. The optimized lag time series module is used to update the welding path sequence data according to the real-time offset compensation vector, and recalculate the heat distribution mapping using the finite element simulation method to obtain the optimized thermal deformation lag time series. The final deformation prediction module is used to input the optimized thermal deformation lag time series into the neural network prediction model for iterative verification. If the superposition parameters of the sequential thermal effects converge, the final predicted deformation distribution map is obtained. The control command generation and execution module is used to generate a welding process control command sequence based on the final predicted deformation distribution map, transmit the control command sequence to the execution system, and obtain the geometric accuracy index of the skeleton structure welding.

[0060] It should be noted that the real-time compensation system for skeleton welding deformation based on dynamic feedback provided in this embodiment of the invention is used to execute all the process steps of the real-time compensation method for skeleton welding deformation based on dynamic feedback in the above embodiment. The working principle and beneficial effect of the two are one-to-one, so they will not be described again.

[0061] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0062] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for real-time compensation of skeleton welding deformation based on dynamic feedback, characterized in that, include: The three-dimensional geometric model data and material thickness distribution data of the skeleton structure are obtained, and the three-dimensional geometric model data and the material thickness distribution data are processed by the finite element simulation method to obtain the initial thermal distribution mapping. Obtain welding path sequence data, and interpolate the welding path sequence data according to the initial heat distribution mapping to obtain the thermal deformation hysteresis time series. The parameters of the superposition of sequential thermal effects are extracted from the thermal deformation hysteresis time series, and the parameters of the superposition of sequential thermal effects are input into a preset neural network prediction model to obtain the predicted value of asynchronous deformation trend. Obtain geometric shape influence data; if the predicted value of asynchronous deformation trend exceeds a preset trend threshold, obtain an adjustment signal from the neural network prediction model; and fuse the geometric shape influence data according to the adjustment signal to obtain a real-time offset compensation vector. Based on the real-time offset compensation vector, update the welding path sequence data, and recalculate the heat distribution mapping using the finite element simulation method to obtain the optimized thermal deformation hysteresis time sequence. The optimized thermal deformation lag time series is input into the neural network prediction model for iterative verification. If the superposition parameters of the sequential thermal effects converge, the final predicted deformation distribution map is obtained. Based on the final predicted deformation distribution map, a welding process control command sequence is generated, and the control command sequence is transmitted to the execution system to obtain the geometric accuracy index of the skeleton structure welding.

2. The method for real-time compensation of skeleton welding deformation based on dynamic feedback according to claim 1, characterized in that, The process of acquiring the three-dimensional geometric model data and material thickness distribution data of the skeleton structure, and processing the three-dimensional geometric model data and the material thickness distribution data through the finite element simulation method to obtain the initial thermal distribution mapping includes: Obtain the three-dimensional geometric model data and material thickness distribution data of the skeleton structure, and construct a three-dimensional mesh model of the skeleton containing thickness gradient information based on the three-dimensional geometric model data and the material thickness distribution data. Unsteady-state heat conduction calculations were performed using the aforementioned skeleton 3D mesh model to obtain transient response temperature field data. Perform heat flux vector analysis on the transient response temperature field data to obtain the initial heat distribution mapping; The phase lag parameter is calculated based on the initial heat distribution mapping, and the delay time corresponding to the phase lag parameter is determined to complete the delay characteristic analysis of the skeleton structure.

3. The method for real-time compensation of skeleton welding deformation based on dynamic feedback according to claim 1, characterized in that, The process of acquiring welding path sequence data and interpolating the welding path sequence data according to the initial heat distribution mapping to obtain a thermal deformation hysteresis time series includes: Obtain welding path sequence data, process the welding path sequence data using spatial interpolation, and generate a sequence of weld point coordinates with timestamps; Nonlinear transient thermal structure coupling calculations are performed based on the timestamped weld point coordinate sequence to obtain transient temperature field evolution data; By tracing the heat flux density vector flow trajectory through the transient temperature field evolution data, a set of heat input propagation process paths can be obtained; Deformation response delay data is calculated along the set of heat input propagation paths. The deformation response delay data is then analyzed, and the heat accumulation time and heat release time are calculated based on the deformation response delay data to obtain the thermal deformation hysteresis time series.

4. The method for real-time compensation of skeleton welding deformation based on dynamic feedback according to claim 1, characterized in that, The step of extracting the superposition parameters of sequential thermal effects from the thermal deformation hysteresis time series and inputting the superposition parameters of sequential thermal effects into a neural network prediction model to obtain the predicted value of asynchronous deformation trend includes: The thermal deformation hysteresis time series is convolved to obtain the thermal effect superposition weight matrix; Calculate the node heat conduction attenuation value based on the heat effect superposition weight matrix, and construct a set of feature vectors for the superposition of successive heat effects; The superimposed feature vector group of sequential thermal effects is input into the neural network prediction model, and the local thermal strain evolution sequence is output. Nonlinear regression analysis was performed on the local thermal strain evolution sequence to obtain the predicted value of asynchronous deformation trend.

5. The method for real-time compensation of skeleton welding deformation based on dynamic feedback according to claim 1, characterized in that, If the predicted value of the asynchronous deformation trend exceeds a preset trend threshold, an adjustment signal is obtained from the neural network prediction model. Based on the adjustment signal, geometric influence data is fused to obtain a real-time compensation vector, including: Compare the difference between the predicted value of asynchronous deformation trend and the preset trend threshold, and extract the adjustment signal that reflects the deviation gradient; Based on the adjustment signal and the geometric influence data, a geometric correction adjustment tensor is obtained. The geometric influence data includes the local stiffness distribution information and geometric characteristic parameters of the component. A dynamic correction factor is obtained by considering the historical attenuation function of the thermal superposition effect introduced by the geometric correction adjustment tensor. The historical attenuation function is constructed based on the attenuation law of historical welding heat input data. By combining the dynamic correction factor with the geometric correction adjustment tensor, a real-time compensation vector for offsetting the superposition effect of thermal effects is obtained.

6. The method for real-time compensation of skeleton welding deformation based on dynamic feedback according to claim 1, characterized in that, The step of updating the welding path sequence data based on the real-time offset compensation vector and recalculating the heat distribution mapping using the finite element simulation method to obtain the optimized thermal deformation hysteresis time series includes: Based on the real-time offset compensation vector, a corrected welding path sequence containing compensation information is generated using a topology reconstruction algorithm; Based on the modified welding path sequence, the heat input is calculated in a virtual simulation environment, and a heat distribution mapping matrix reflecting the heat conduction delay characteristics is established. The heat distribution mapping matrix is ​​coupled to the thermo-elastic-plastic constitutive model to form a dynamic deformation field; By comparing the dynamic deformation field with the theoretical design model, the cumulative deviation gradient is calculated, and the optimized thermal deformation hysteresis time series is determined based on the cumulative deviation gradient.

7. The method for real-time compensation of skeleton welding deformation based on dynamic feedback according to claim 1, characterized in that, The optimized thermal deformation lag time series is input into the neural network prediction model for iterative verification. If the superposition parameters of the sequential thermal effects converge, the final predicted deformation distribution map is obtained, including: The optimized thermal deformation hysteresis time series is processed using the neural network prediction model to analyze the superposition parameters of the sequential thermal effects. Substitute the superimposed parameters of the sequential thermal effects into the preset transfer equation to calculate the predicted values ​​of the grid node displacements. A residual vector is generated based on the predicted displacement values ​​of the grid nodes, and it is determined whether the residual vector satisfies the preset convergence condition. If the preset convergence condition is met, the model parameters are locked, and the final predicted deformation distribution map is output. The node offset is quantified based on the final predicted deformation distribution map, and the overall torsional deformation control level is reflected by the node offset.

8. The method for real-time compensation of skeleton welding deformation based on dynamic feedback according to claim 1, characterized in that, The step of generating a welding process control command sequence based on the final predicted deformation distribution map, transmitting the control command sequence to the execution system, and obtaining the geometric accuracy indicators of the skeleton structure welding includes: A geometric compensation model is constructed based on the final predicted deformation distribution map. The geometric compensation model includes full node correction information calculated using the inverse mapping algorithm. The geometric compensation model is registered with the preset welding torch trajectory to generate a delay compensation process parameter matrix. The delay compensation process parameter matrix is ​​obtained by adjusting the energy input rate according to the thermal deformation hysteresis time series. The delay cancellation process parameter matrix is ​​encoded to generate a welding process control command sequence; The control command sequence is transmitted to the execution system, which executes the commands and feeds back the actual welding geometry data. The geometric accuracy index of the skeleton structure welding is determined based on the actual welding geometry data to offset the delay characteristics and superposition effects.

9. A real-time compensation system for skeleton welding deformation based on dynamic feedback, characterized in that, include: The initial thermal distribution mapping module is used to acquire the three-dimensional geometric model data and material thickness distribution data of the skeleton structure, and process the three-dimensional geometric model data and the material thickness distribution data through the finite element simulation method to obtain the initial thermal distribution mapping. The thermal deformation hysteresis time series module is used to acquire welding path sequence data, and interpolate the welding path sequence data according to the initial heat distribution mapping to obtain the thermal deformation hysteresis time series. The asynchronous deformation prediction module is used to extract the superposition parameters of sequential thermal effects from the thermal deformation lag time series, input the superposition parameters of sequential thermal effects into a preset neural network prediction model, and obtain the asynchronous deformation trend prediction value. The real-time offset compensation vector module is used to acquire geometric shape influence data. If the asynchronous deformation trend prediction value exceeds the preset trend threshold, an adjustment signal is obtained from the neural network prediction model. The geometric shape influence data is fused according to the adjustment signal to obtain the real-time offset compensation vector. The optimized lag time series module is used to update the welding path sequence data according to the real-time offset compensation vector, and recalculate the heat distribution mapping using the finite element simulation method to obtain the optimized thermal deformation lag time series. The final deformation prediction module is used to input the optimized thermal deformation lag time series into the neural network prediction model for iterative verification. If the superposition parameters of the sequential thermal effects converge, the final predicted deformation distribution map is obtained. The control command generation and execution module is used to generate a welding process control command sequence based on the final predicted deformation distribution map, transmit the control command sequence to the execution system, and obtain the geometric accuracy index of the skeleton structure welding.

Images