A method and system for welding a crystallizer copper tube
By detecting residual stress before welding the copper tubes in the crystallizer, decomposing displacement signals in real time, and performing differentiated control, the problems of deformation and stress release during the copper tube welding process were solved, achieving high-precision welding quality and improved reliability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHENYANG HETAI METALLURGICAL EQUIP CO LTD
- Filing Date
- 2026-04-01
- Publication Date
- 2026-06-09
AI Technical Summary
During the welding process of copper tubes in crystallizers, it is difficult to establish a molten pool when welding copper and steel, which easily leads to incomplete fusion defects. Brittle compounds are generated at the interface, and the large coefficient of thermal expansion of the slender structure makes it difficult to control deformation and stress release. Traditional welding cannot make real-time adjustments, resulting in deformation accumulation and cracking.
By detecting residual stress before welding, a stress release prediction model is established. The displacement signal is decomposed into thermal deformation and stress release components in real time. Differential control is carried out by combining the support force feedforward curve. The welding parameters are adjusted in real time by using Kalman filtering and wavelet transform to fuse the signal. After welding, the model parameters are corrected by machine learning.
It achieves precise control over the welding process of the crystallizer copper tubes, avoiding misjudgment and stress release deformation, improving welding quality and service reliability, and possessing continuous optimization capabilities.
Smart Images
Figure CN121957253B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of welding engineering technology, and specifically to a welding method and system for copper tubes in a crystallizer. Background Technology
[0002] The crystallizer copper tube is a core component of continuous casting equipment, and the geometric accuracy of its internal cavity directly affects the quality of the cast billet. The crystallizer copper tube is usually composed of a high thermal conductivity chromium-zirconium copper inner tube and a structural steel back plate, and it must withstand repeated thermal shocks from high-temperature molten steel during service.
[0003] However, welding copper to steel presents a series of unique challenges. First, copper has extremely high thermal conductivity, causing welding heat to dissipate rapidly along the copper tube base material, making it difficult to establish a molten pool and easily leading to incomplete fusion defects. Second, copper and steel have poor miscibility at high temperatures, and a brittle intermetallic compound layer easily forms at the interface, resulting in decreased weld toughness. More importantly, crystallizer copper tubes are mostly slender structures, and copper has a large coefficient of linear expansion, resulting in significant thermal expansion and contraction during welding. Coupled with the presence of initial residual stress, this easily leads to bending or twisting deformation of the copper tube.
[0004] In existing technologies, rigid clamps are typically used for forced restraint to suppress deformation. However, forced restraint introduces significant residual welding stress, which can lead to secondary deformation or even cracking during subsequent service. Furthermore, traditional constant-parameter welding cannot dynamically adjust based on real-time deformation during the welding process, making it difficult to control the cumulative deformation effect of long welds. Therefore, how to identify the causes of deformation in real time during welding and adopt differentiated control strategies to simultaneously suppress thermal deformation and stress-release deformation is a critical problem that urgently needs to be solved in the field of crystallizer copper tube welding.
[0005] Therefore, the present invention provides a welding method and system for copper tubes of crystallizers. Summary of the Invention
[0006] The purpose of this invention is to provide a welding method and system for copper tubes in a crystallizer to solve the aforementioned background problems.
[0007] The objective of this invention can be achieved through the following technical solution: a method for welding copper tubes for crystallizers, comprising the following steps:
[0008] Before welding, residual stress was detected in the copper tube of the crystallizer, an initial stress distribution map was established, and based on the initial stress distribution map, a stress release prediction model was constructed through finite element simulation, and a feedforward curve of the support force along the weld was generated.
[0009] During the welding process, displacement, local stress and temperature data are collected simultaneously, and the displacement is decomposed into thermal deformation component and stress release deformation component. Combined with the stress release prediction model, the proportion of thermal deformation and stress release deformation in the current deformation is estimated in real time.
[0010] The proportion of stress release deformation is detected in real time. The current welding mode is judged by comparing and analyzing the proportion of stress release deformation. Combined with the feedforward curve of support force along the weld, differentiated thermo-mechanical coordinated control is carried out according to the welding mode.
[0011] After welding, the stress distribution of the copper tube in the crystallizer is detected a second time. The actual stress release deformation is compared and analyzed with the predicted value of the stress release prediction model. The parameters of the stress release prediction model are corrected using machine learning algorithms.
[0012] Furthermore, the stress relief prediction model is constructed as follows:
[0013] A grid of measuring points was arranged along the length and circumference of the copper tube. The axial and circumferential residual stresses were measured point by point using a non-destructive stress testing device. The average value was taken for each measuring point. The spatial coordinates of the measuring points were recorded during the testing process.
[0014] The stress data from discrete measuring points are extended to the entire surface of the copper tube through an interpolation algorithm to generate a continuous three-dimensional stress distribution and obtain an initial stress distribution map.
[0015] A thermo-mechanical coupling model of the copper tube in the crystallizer was established based on finite element simulation. The geometric model of the copper tube was created and meshed. Material properties were assigned and temperature-related mechanical properties were set. The initial stress distribution spectrum was imported as a predefined field.
[0016] Define a welding heat source model, set heat source parameters, apply boundary conditions, perform transient thermo-mechanical coupling analysis, simulate the welding process, calculate the temperature field, stress field and displacement field at any time, and extract the deformation caused by stress release at each time.
[0017] A proxy model is constructed for real-time control. Multiple simulations with different initial stress distributions and different process parameters are run to generate a sample library. The initial stress peak value, location, and heat input are used as inputs, and the stress release deformation curve is used as output to train a neural network model to obtain a stress release prediction model.
[0018] Furthermore, the support force feedforward curve is constructed as follows:
[0019] The copper tube is simplified into a beam model. The temperature field distribution generated by the movement of the heat source is calculated based on the welding process parameters. The thermal bending moment is calculated from the temperature field, and the theoretical thermal deformation curve is obtained by solving the bending differential equation.
[0020] The stiffness of the copper tube at different locations was calibrated by static loading tests. The support force feedforward curve is equal to the stiffness curve multiplied by the theoretical thermal deformation curve.
[0021] Furthermore, the process of decomposing the displacement into thermal deformation components and stress-relieving deformation components is as follows:
[0022] Welding is performed using reference power and reference speed. Displacement sensors collect displacement at high frequency to generate a real-time deformation field, while stress sensors simultaneously collect local stress.
[0023] By selecting the wavelet basis and the number of decomposition levels, wavelet decomposition is performed to obtain the approximation coefficients and detail coefficients of each level. The smoothing component is reconstructed to obtain the thermal deformation component, and the pulse component is reconstructed to obtain the stress release deformation component.
[0024] Furthermore, the calculation method for the ratio of thermal deformation to stress relief deformation is as follows:
[0025] Construct a state-space model, integrate displacement and stress data, and estimate thermal deformation and stress release deformation in real time;
[0026] Construct a state vector that includes thermal deformation components, stress-relief deformation components, thermal deformation rate, and stress-relief deformation rate;
[0027] A state transition matrix is constructed based on the heat conduction equation and stress release prediction model;
[0028] Construct an observation vector that includes macroscopic displacement measured by displacement sensors and local stress values measured by stress sensors;
[0029] Construct an observation matrix where displacement observations are the sum of thermal deformation and stress-release deformation, and stress observations correspond to stress-release deformation.
[0030] Kalman filtering is performed recursively. The current state and error covariance are predicted based on the state equation. The Kalman gain is calculated, and the state is updated in combination with the measured values. At each step, the estimated values of thermal deformation and stress release deformation at the current moment are output.
[0031] And calculate the proportion of thermal deformation and the proportion of stress release deformation: proportion of thermal deformation = thermal deformation at the current moment / (thermal deformation at the current moment + estimated value of stress release deformation), proportion of stress release deformation = 1 - proportion of thermal deformation.
[0032] Furthermore, the method for determining the current welding mode is as follows:
[0033] The stress reduction rate per unit time is calculated in real time. If both the stress release deformation ratio and the stress reduction rate per unit time meet the requirements, it is determined to be a stress release mode. Otherwise, if either the stress release deformation ratio or the stress reduction rate per unit time does not meet the requirements, it is a normal welding mode.
[0034] Furthermore, the process of differentiated thermo-mechanical coordinated control based on the welding mode is as follows:
[0035] If the current welding mode is determined to be normal:
[0036] The feedforward support force is obtained by querying the feedforward curve based on the current welding position. The thermal deformation component is then high-pass filtered to obtain the high-frequency component of thermal deformation. The support force feedback is equal to the support force proportional coefficient multiplied by the high-frequency component of thermal deformation.
[0037] The thermal deformation component is low-pass filtered to obtain the low-frequency component of thermal deformation. If the absolute value of the low-frequency component of thermal deformation meets the requirements, the adaptive adjustment of the heat source parameters is initiated.
[0038] The power adjustment is the reference power minus the power adjustment coefficient multiplied by the absolute value of the low-frequency component of thermal deformation; the speed adjustment is the reference speed plus the speed adjustment coefficient multiplied by the absolute value of the low-frequency component of thermal deformation.
[0039] Using the current state as input, predict the deformation development at N future time points, construct a cost function, solve the optimization problem in each control cycle, obtain the optimal control sequence at future time points, and apply the first control variable of the optimization solution to the system.
[0040] Furthermore, the process of differentiated thermo-mechanical coordinated control based on the welding mode also includes:
[0041] If it is determined that the current mode is stress relief mode, keep the current power and speed unchanged, reduce the support force feedback gain, and the support force command is the feedforward support force plus the reduced support force feedback gain multiplied by the thermal deformation component at the current moment.
[0042] The stress reduction rate is calculated in real time. If the stress reduction rate meets the requirements, the local heating device inside the mandrel is activated to provide auxiliary heating to the stress release area.
[0043] Furthermore, the method for correcting the parameters of the stress relief prediction model is as follows:
[0044] The residual stress on the surface of the copper tube after welding was measured using the same equipment and measuring point layout. The estimated stress release deformation along the weld was extracted from the welding process record. The actual stress release deformation was compared with the stress release prediction value output by the stress release prediction model point by point to calculate the error.
[0045] Error data from multiple batches of welding are collected to form a corrected dataset. A machine learning algorithm is used to establish an error prediction model. The input is process parameters and initial stress characteristics, and the output is the prediction error. The error prediction model is superimposed with the original prediction model to form a corrected stress release prediction model.
[0046] A welding system for copper tubes in a crystallizer includes the following modules:
[0047] Stress detection and model building module: Before welding, residual stress is detected on the copper tube of the crystallizer, an initial stress distribution map is established, and based on the initial stress distribution map, a stress release prediction model is built through finite element simulation, and a feedforward curve of the support force along the weld is generated.
[0048] Welding process sensing and signal decoupling module: Simultaneously collects displacement, local stress and temperature data during the welding process, and decomposes the displacement into thermal deformation component and stress release deformation component. Combined with the stress release prediction model, it estimates the proportion of thermal deformation and stress release deformation in the current deformation in real time.
[0049] Welding mode judgment and differentiated control module: Real-time detection of stress release deformation ratio, judgment of current welding mode by comparing and analyzing stress release deformation ratio, and differentiated thermo-mechanical coordinated control based on welding mode by combining the support force feedforward curve along the weld.
[0050] Post-weld stress detection and model correction module: After welding, the stress distribution of the crystallizer copper tube is detected again, and the actual stress release deformation is compared and analyzed with the predicted value of the stress release prediction model. The parameters of the stress release prediction model are corrected using machine learning algorithms.
[0051] The beneficial effects of this invention are as follows:
[0052] By constructing residual stress detection and stress release prediction models, the originally uncertain and uncontrollable initial stress factors are transformed into quantifiable predictive inputs, providing a physical benchmark for subsequent real-time control. At the same time, the feedforward curve enables the pre-preparation of active correction.
[0053] By fusing wavelet transform and Kalman filtering, the original displacement signal is decomposed into thermal deformation and stress release deformation, and the precise proportion of the two deformations is output in real time. This overcomes the problems of insufficient information from a single sensor and noise interference, enabling the system to accurately determine the cause of deformation.
[0054] The welding mode is determined in real time by the stress release deformation ratio. In normal mode, thermal-mechanical synergy is used to actively suppress thermal deformation, while in stress release mode, flexible following is used to allow stress to be released naturally. The two modes perform their respective functions, which not only ensures welding accuracy, but also avoids the introduction of new stress due to misinterpretation, thus achieving optimal control of form and property synergy.
[0055] By using post-weld inspection and machine learning model correction, the system continuously accumulates data and optimizes the prediction model during actual operation, enabling it to continuously evolve and steadily improve welding quality over the long term. Attached Figure Description
[0056] The invention will now be further described with reference to the accompanying drawings.
[0057] Figure 1 This is a flowchart of a method for welding copper tubes for a crystallizer according to Embodiment 1 of the present invention;
[0058] Figure 2 This is the logic diagram for determining the current welding mode in Embodiment 1 of the present invention;
[0059] Figure 3 This is a functional block diagram of a welding system for a crystallizer copper tube in Embodiment 2 of the present invention. Detailed Implementation
[0060] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.
[0061] Example 1: Please refer to Figure 1 As shown in the embodiment of the present invention, a method for welding a crystallizer copper tube specifically includes the following steps:
[0062] Step 1: Before welding, residual stress is detected on the copper tube of the crystallizer, an initial stress distribution map is established, and based on the initial stress distribution map, a stress release prediction model is constructed through finite element simulation, and a feedforward curve of the support force along the weld is generated.
[0063] In step one, the process of detecting residual stress in the copper tube of the crystallizer includes:
[0064] Arrange a grid of measuring points along the length and circumference of the copper tube (e.g., axial spacing of 50-100mm, circumferential spacing of 45° or 60°). Use a non-destructive stress testing device to measure the axial and circumferential residual stress point by point. Repeat the measurement 2-3 times for each measuring point and take the average value. Record the spatial coordinates of the measuring points during the testing process.
[0065] In step one, the process of establishing the initial stress distribution map includes:
[0066] The stress data from discrete measurement points are extended to the entire surface of the copper tube using interpolation algorithms (such as Kriging interpolation and radial basis function interpolation) to generate a continuous three-dimensional stress distribution field.
[0067] For internal stresses that cannot be directly measured, the finite element inverse fitting method is used, that is, the internal stress distribution is inferred from the surface measurement values through the elasticity equation, or the stress in the thickness direction is assumed to be linearly distributed.
[0068] In step one, the process of constructing the stress relief prediction model includes:
[0069] A thermo-mechanical coupling model of the copper tube in the crystallizer was established based on finite element simulation, specifically as follows:
[0070] Create a geometric model of the copper tube and mesh it, refine the mesh near the weld, assign material properties, and set temperature-dependent mechanical properties (such as the decrease in yield strength at high temperatures).
[0071] Import the initial stress distribution map as a predefined field, define the welding heat source model (such as a double ellipsoidal heat source or a Gaussian heat source), and set the heat source parameters, including power, efficiency, and shape parameters.
[0072] Apply boundary conditions (such as clamp constraints and convective heat dissipation) to perform transient thermo-mechanical coupling analysis, simulate the welding process, and calculate the temperature field, stress field, and displacement field at any time.
[0073] Extract the deformation caused by stress release at each time point;
[0074] Since complete finite element calculation is time-consuming, a proxy model needs to be built for real-time control. Intrinsic orthogonal decomposition (POD) or artificial neural network methods can be used: run multiple sets of simulations with different initial stress distributions and different process parameters to generate a sample library. Use key parameters, such as initial stress peak, location, and heat input, as inputs and stress release deformation curves as outputs to train a neural network model and obtain a stress release prediction model.
[0075] In step one, the process of constructing the support force feedforward curve along the weld seam includes:
[0076] Based on the theoretical mechanics model, the feedforward support force required to counteract the expected thermal deformation at each location along the weld is calculated, specifically:
[0077] The copper tube is simplified into a beam model, and the temperature field distribution generated by the movement of the heat source is calculated based on the welding process parameters.
[0078] The thermal bending moment is calculated from the temperature field, and the theoretical thermal deformation curve is obtained by solving the bending differential equation.
[0079] The stiffness of the copper tube at different locations was calibrated by static loading test: standard force was applied at different locations of the copper tube, displacement changes were measured, and stiffness curves were fitted.
[0080] Support force feedforward curve = stiffness curve × theoretical thermal deformation curve;
[0081] It should be noted that the role of the pre-weld residual stress detection and stress release prediction model is to provide the original basis for distinguishing between welding hot deformation and stress release deformation, and to transform the initial stress field into a predictable law of stress release behavior (when, where, and how much) during the welding process. This enables the system to predict sudden deformation in advance, calculate the support force required to counteract the expected hot deformation in advance, and enable the system to actively apply the reverse action at the beginning of welding, thereby reducing the burden on subsequent feedback control.
[0082] Step 2: During the welding process, displacement, local stress and temperature data are collected simultaneously, and the displacement is decomposed into thermal deformation component and stress release deformation component. Combined with the stress release prediction model, the proportion of thermal deformation and stress release deformation in the current deformation is estimated in real time.
[0083] In step two, the process of decomposing the displacement signal into thermal deformation and stress relief deformation includes:
[0084] Welding is performed at a reference power and reference speed. The heat source moves along a predetermined path. The displacement sensor collects the displacement at high frequency to generate a real-time deformation field. The stress sensor collects the local stress synchronously. The thermal imager collects the temperature field to assist in the adjustment of the heat source.
[0085] Choose an appropriate wavelet basis and the number of decomposition layers, which are determined based on the sampling frequency and the distorted frequency band. For example, when sampling at 100Hz, decompose into 5 layers, and divide the frequency band into 1-1.56Hz, 1.56Hz-3.125Hz, etc.
[0086] Wavelet decomposition is performed to obtain the approximation coefficients (low frequency) and detail coefficients (high frequency) for each level.
[0087] Reconstructing the smooth component: Select the coefficients of the low-frequency sub-band (such as 0-1Hz) containing the dominant frequency of the heat source and perform inverse wavelet transform to obtain the thermal deformation component;
[0088] Reconstructing the pulse component: Select the coefficients of the high-frequency subband (e.g., >5Hz) containing the burst signal, perform inverse wavelet transform, and obtain the stress release deformation component after thresholding (noise filtering).
[0089] In step two, the estimation process for the proportion of thermal deformation and stress-relief deformation in the current deformation includes:
[0090] Construct a state-space model, integrate displacement and stress data, and estimate thermal deformation and stress relief deformation in real time, specifically as follows:
[0091] The first point to clarify is that the construction of the state vector and state transition matrix is as follows:
[0092] State vector x k These are a set of variables describing the internal state of the system at time k: ,in, For thermal deformation component, For stress-relieving deformation components, Let be the rate of change of thermal deformation, i.e., the derivative of thermal deformation with respect to time. The rate of change of stress-release deformation is used to capture the severity of stress-release events;
[0093] It should be noted that Kalman filtering requires the state to fully describe the dynamic behavior of the system. Thermal deformation and stress release deformation are two components with different physical mechanisms, which evolve independently. Introducing a rate term allows the state transition model to be a first-order or second-order system, thereby improving prediction accuracy.
[0094] The state transition matrix F is constructed based on the heat conduction equation and stress release prediction model:
[0095] The state transition matrix describes how the system state evolves over time without new measurement information, i.e., the predictive relationship from time k-1 to time k: F is a 4×4 matrix, the specific form of which is determined by the physical laws of thermal deformation and stress release deformation;
[0096] The thermal deformation part can be simplified to a first-order inertial model of thermal input; the stress release part is based on the stress release prediction model, and the stress release deformation at the current moment is related to the value at the previous moment and the welding position.
[0097] Secondly, it should be noted that the construction of the observation vector and observation matrix is as follows:
[0098] Observation vector z k This is the data actually measured by the sensor at time k: ,in, The macroscopic displacement measured by the displacement sensor is a superposition of thermal deformation and stress-relief deformation. The local stress value measured by the stress sensor;
[0099] The observation matrix H maps the system state to observed values, that is, it establishes the relationship between state variables and sensor readings. The observation matrix H includes:
[0100] Displacement observation = thermal deformation + stress relief deformation + noise;
[0101] Stress observation = f (stress release deformation) + noise, that is, there is a corresponding relationship between stress release and stress value, which can be calibrated through offline testing;
[0102] Thirdly, it should be noted that the Kalman filtering recursion specifically includes:
[0103] Understandably, Kalman filtering is a recursive state estimation algorithm. Its core idea is to first predict the system state using a physical model (prediction step), and then correct this prediction using sensor measurements (update step). These two steps are executed alternately, and each step makes the estimation more accurate.
[0104] The purpose of the prediction step is to infer the current state of the system using physical laws in the absence of new measurement data. The formula is:
[0105] This means: based on the optimal estimate from the previous time step. The state transition matrix F is used to predict the prior state estimate at the current time step. ;
[0106] Where P is the covariance matrix of the state estimation error, representing the uncertainty of the estimation. Let $\mathbf{k-1}$ be the estimation error covariance at the previous time step $k-1$. It represents the uncertainty of the predicted state. F is the state transition matrix, and Q is the process noise covariance matrix, which represents the inaccuracy of the model itself. The larger Q is, the less we trust the model's predictions and the more we will rely on the measured values.
[0107] For example, assuming the estimated thermal deformation at the previous moment was 0.1 mm and the rate of change was 0.01 mm / s, the prediction step calculates based on the transition state matrix F that the thermal deformation at this moment should be 0.1 + 0.01 × =0.101mm (if =0.1s), and since the model may not be completely accurate (for example, the heat source may have slight fluctuations), we add uncertainty Q to this prediction, indicating that although the prediction is reasonable, there may be errors;
[0108] The purpose of the update step is to use the current sensor measurements to correct the results of the prediction step, thereby obtaining a more reliable state estimate. The formula is:
[0109] , where K k Kalman gain is a 4×2 matrix. Its physical meaning is to weigh the prediction and measurement. The larger the gain, the stronger the effect of the measurement on state correction. H is the observation matrix, and R is the measurement noise covariance matrix, representing the sensor's measurement error. The larger R is, the less reliable the sensor data is, and the gain will decrease accordingly.
[0110] ,in, The information is the difference between the actual measured value and the estimated value calculated based on the predicted value. This is the posterior estimate for the current time step;
[0111] For example, in the scenario of welding copper tubes in a crystallizer, the prediction step tells us that the thermal deformation at this moment should be 0.101 mm. The displacement sensor actually measures a total deformation of 0.12 mm. The innovation = 0.12 mm - 0.101 mm = 0.019 mm, indicating that the prediction is too small and the actual deformation is larger. At the same time, the stress sensor may also provide information (such as stress drop indicating possible stress release). Kalman gain K k Based on the uncertainty of prediction and the uncertainty of measurement, a decision is made on how much information to use for correction. The final output is: the thermal deformation may be 0.108 mm, the stress relief deformation may be 0.012 mm, and the sum of the two is equal to 0.12 mm, while the uncertainty is reduced.
[0112] The predicted state is corrected by multiplying the Kalman gain by the innovation to obtain the posterior estimate. This correction takes into account both displacement measurement and stress measurement, thus fusing information from the two sensors.
[0113] Predict the current state and error covariance based on the state equation, calculate the Kalman gain, and update the state using the measured values. Each step outputs the estimated values of thermal deformation and stress release deformation at the current moment, and calculates the proportion of thermal deformation and the proportion of stress release deformation: thermal deformation proportion = thermal deformation at the current moment / (thermal deformation at the current moment + estimated value of stress release deformation), stress release deformation proportion = 1 - thermal deformation proportion;
[0114] It should be noted that the role of real-time sensing and signal decoupling in the welding process is to avoid misjudgment caused by insufficient information from a single sensor, decompose the original displacement signal into smooth components (thermal deformation) and pulse components (stress release), initially separate the deformation of two different physical mechanisms, provide feature separation for subsequent accurate estimation, fuse displacement and stress measurement data, combine with physical models, and output the optimal estimated values of thermal deformation and stress release deformation and their proportions in real time, overcoming noise interference and model uncertainty.
[0115] Step 3: Real-time detection of the stress release deformation ratio; judgment of the current welding mode by comparing and analyzing the stress release deformation ratio; and differentiated thermo-mechanical coordinated control based on the welding mode by combining the feedforward curve of the support force along the weld.
[0116] Please see Figure 2 As shown, in step three, the process of determining the current welding mode includes:
[0117] Set a threshold for the proportion of stress-relieving deformation and a threshold for the rate of stress drop per unit time. Calculate the rate of stress drop per unit time. If the proportion of stress-relieving deformation is greater than the threshold for the proportion of stress-relieving deformation, and the local stress sensor detects that the rate of stress drop per unit time is greater than the threshold for the rate of stress drop per unit time, then it is determined to be a stress-relieving mode. Otherwise, if the proportion of stress-relieving deformation is less than or equal to the threshold for the proportion of stress-relieving deformation, or the rate of stress drop per unit time is less than or equal to the threshold for the rate of stress drop per unit time, then it is a normal welding mode.
[0118] In step three, the process of differentiated thermo-mechanical coordinated control based on the welding mode includes:
[0119] If the current welding mode is determined to be normal:
[0120] The first point to clarify is that the support force feedback is based on frequency band segmentation, specifically as follows:
[0121] The feedforward support force can be obtained by querying the feedforward curve based on the current welding position.
[0122] High-pass filtering is applied to the thermal deformation component to obtain the high-frequency component of thermal deformation. Support force feedback = support force proportionality coefficient × high-frequency component of thermal deformation.
[0123] Among them, the support force proportionality coefficient is obtained through static calibration test. Different support forces are applied, the corresponding displacement changes are measured, the force-displacement transfer function is established, and the slope of its linear segment is taken.
[0124] Secondly, it should be noted that the heat source parameters are adaptively adjusted, specifically as follows:
[0125] The thermal deformation component is low-pass filtered to obtain the low-frequency component of thermal deformation. A dead zone (e.g., 0.05mm) is set. If the absolute value of the low-frequency component of thermal deformation is greater than the dead zone and continues for more than a preset time, then adaptive adjustment of the heat source parameters is initiated.
[0126] The adjustable power is the reference power minus the power adjustment coefficient multiplied by the absolute value of the low-frequency component of thermal deformation.
[0127] The adjustment speed is the base speed + speed adjustment coefficient × absolute value of the low-frequency component of thermal deformation;
[0128] The process for determining the power regulation coefficient and the speed regulation coefficient is as follows:
[0129] Through process testing and calibration, under the same conditions, power and speed were changed respectively, the changes in deformation were recorded, and the sensitivity coefficient was fitted, which is the power adjustment amount corresponding to unit deformation.
[0130] Thirdly, it should be noted that model predictive control coordination specifically involves:
[0131] Using the current state (deformation components, temperature, and position) as input, predict the deformation development at the next N time moments, construct a cost function, weigh deformation deviation, support force change, and heat source adjustment amplitude, solve the optimization problem in each control cycle, and obtain the optimal control sequence for the future time moments, including support force fine-tuning and heat source fine-tuning, but only execute the values at the current time moment;
[0132] Based on offline simulation, a simplified model is established to predict the deformation at N future moments according to the current state and control inputs (fine-tuning of support force, power, and speed).
[0133] Each control cycle solves the optimization problem, minimizing the sum of future deformation deviation and control change. Quadratic programming is used to solve the problem, and the control quantity at the first moment of the optimal solution is superimposed on the current value.
[0134] Within each control cycle:
[0135] Obtain the current thermal deformation component and stress relief deformation component from the Kalman filter;
[0136] Using a predictive model, based on the current state and candidate control sequences, the deformation amount in the next N steps is calculated, and the cost function is minimized. This function includes the sum of squares of future deformation deviations and the sum of squares of changes in control quantities. Constraints include the upper limit of support force and the power range.
[0137] Take the first control variable of the optimized solution and apply it to the system;
[0138] If the current mode is determined to be stress relief mode:
[0139] The first point to clarify is that the heat source is frozen, meaning that the current power and speed are kept constant and no longer respond to thermal deformation adjustment, so as to avoid interfering with the stress release process.
[0140] Secondly, it should be noted that the support force follows flexibly, specifically as follows:
[0141] Reduce the support force feedback gain, for example, reduce the support force feedback gain to 1 / 10 of the normal value, and the support force command is the feedforward support force + the reduced support force feedback gain × the thermal deformation component at the current moment. Only maintain the workpiece posture and do not actively resist stress release deformation.
[0142] Thirdly, it should be noted that the auxiliary heating intervention specifically includes:
[0143] The stress drop rate is calculated in real time and compared with the preset rate. If the stress drop rate is greater than or equal to the preset rate, the local heating device inside the mandrel is activated to provide auxiliary heating to the stress release area (e.g., heating to 300°C) to promote uniform stress release and prevent local strain concentration.
[0144] It should be noted that the purpose of differentiated thermo-mechanical coordinated control based on the welding mode is as follows: by judging the mode in real time by the stress release deformation ratio, the system can adopt completely different control strategies for different deformation causes, avoiding misoperation. In normal mode, thermo-mechanical coordination achieves active suppression; in stress release mode, flexible following allows natural release while preventing instability. The combination of the two ensures welding accuracy while avoiding the introduction of new stress.
[0145] Step 4: After welding, the stress distribution of the copper tube in the crystallizer is detected a second time. The actual stress release deformation is compared and analyzed with the predicted value of the stress release prediction model. The parameters of the stress release prediction model are corrected using machine learning algorithms.
[0146] In step four, the process of correcting the parameters of the stress release prediction model includes:
[0147] Using the same equipment and measuring point layout, the residual stress on the surface of the copper tube after welding was measured.
[0148] The estimated stress release deformation distribution along the weld can be extracted from the welding process record. At the same time, the actual released stress value can be estimated based on the stress difference before and after welding.
[0149] The actual stress release deformation is compared point by point with the stress release deformation prediction value output by the stress release prediction model. The error is calculated and the error data is used as a new sample to correct the prediction model.
[0150] Error data from multiple batches of welding were collected to form a correction dataset;
[0151] An error prediction model is established using machine learning algorithms (such as gradient boosting trees, support vector regression, or neural networks). The inputs are process parameters and initial stress characteristics, and the output is the prediction error.
[0152] The error prediction model is superimposed on the original prediction model to form the corrected stress release prediction model.
[0153] It should be noted that the role of post-weld stress detection and model correction is to train the error prediction model using multiple batches of welding data, and then add error compensation to the original prediction model to form a more accurate correction model.
[0154] The technical solution and advantages of this application embodiment are as follows: Before welding, residual stress is detected on the copper tube of the crystallizer, an initial stress distribution map is established, and based on the initial stress distribution map, a stress release prediction model is constructed through finite element simulation, and a support force feedforward curve along the weld is generated; during the welding process, displacement, local stress, and temperature data are collected simultaneously, and the displacement is decomposed into thermal deformation components and stress release deformation components. Combined with the stress release prediction model, the proportion of thermal deformation and stress release deformation in the current deformation is estimated in real time; the proportion of stress release deformation is detected in real time, and the current welding mode is judged by comparing and analyzing the proportion of stress release deformation. Combined with the support force feedforward curve along the weld, differentiated thermo-mechanical coordinated control is performed according to the welding mode. This invention establishes an initial stress distribution map and constructs a stress release prediction model and generates a support force feedforward curve by detecting residual stress before welding and building a model. During welding, displacement, local stress, and temperature data are collected simultaneously. The displacement signal is decomposed into thermal deformation and stress release deformation, and the ratio of the two is estimated in real time. The ratio of stress release deformation is detected in real time, and the welding mode is determined based on the ratio, and differentiated thermo-mechanical coordinated control is implemented. After welding, the stress distribution is detected again, and machine learning is used to correct the prediction model. This invention effectively solves the problems of difficult deformation control and prominent contradiction between stress and precision in the welding of copper tubes in crystallizers through a closed-loop strategy of pre-weld prediction, in-weld perception and decoupling, differentiated coordinated control, and post-weld self-learning correction, significantly improving welding quality and service reliability.
[0155] Example 2: Please refer to Figure 3 As shown in the embodiment of the present invention, a welding system for a crystallizer copper tube includes the following modules:
[0156] Stress detection and model building module: Before welding, residual stress is detected on the copper tube of the crystallizer, an initial stress distribution map is established, and based on the initial stress distribution map, a stress release prediction model is built through finite element simulation, and a feedforward curve of the support force along the weld is generated.
[0157] Welding process sensing and signal decoupling module: Simultaneously collects displacement, local stress and temperature data during the welding process, and decomposes the displacement into thermal deformation component and stress release deformation component. Combined with the stress release prediction model, it estimates the proportion of thermal deformation and stress release deformation in the current deformation in real time.
[0158] Welding mode judgment and differentiated control module: Real-time detection of stress release deformation ratio, judgment of current welding mode by comparing and analyzing stress release deformation ratio, and differentiated thermo-mechanical coordinated control based on welding mode by combining the support force feedforward curve along the weld.
[0159] Post-weld stress detection and model correction module: After welding, the stress distribution of the crystallizer copper tube is detected again, and the actual stress release deformation is compared and analyzed with the predicted value of the stress release prediction model. The parameters of the stress release prediction model are corrected using machine learning algorithms.
[0160] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.
Claims
1. A method for welding copper tubes for a crystallizer, characterized in that: Includes the following steps: Before welding, residual stress was detected in the copper tube of the crystallizer, an initial stress distribution map was established, and based on the initial stress distribution map, a stress release prediction model was constructed through finite element simulation, and a feedforward curve of the support force along the weld was generated. During the welding process, displacement, local stress and temperature data are collected simultaneously, and the displacement is decomposed into thermal deformation component and stress release deformation component. Combined with the stress release prediction model, the proportion of thermal deformation and stress release deformation in the current deformation is estimated in real time. The proportion of stress release deformation is detected in real time. The current welding mode is judged by comparing and analyzing the proportion of stress release deformation. Combined with the feedforward curve of support force along the weld, differentiated thermo-mechanical coordinated control is carried out according to the welding mode. After welding, the stress distribution of the copper tube in the crystallizer is detected a second time. The actual stress release deformation is compared and analyzed with the predicted value of the stress release prediction model. The parameters of the stress release prediction model are corrected using machine learning algorithms.
2. The welding method for a crystallizer copper tube according to claim 1, characterized in that: The stress relief prediction model is constructed as follows: A grid of measuring points was arranged along the length and circumference of the copper tube. The axial and circumferential residual stresses were measured point by point using a non-destructive stress testing device. The average value was taken for each measuring point. The spatial coordinates of the measuring points were recorded during the testing process. The stress data from discrete measuring points are extended to the entire surface of the copper tube through an interpolation algorithm to generate a continuous three-dimensional stress distribution and obtain an initial stress distribution map. A thermo-mechanical coupling model of the copper tube in the crystallizer was established based on finite element simulation. The geometric model of the copper tube was created and meshed. Material properties were assigned and temperature-related mechanical properties were set. The initial stress distribution spectrum was imported as a predefined field. Define a welding heat source model, set heat source parameters, apply boundary conditions, perform transient thermo-mechanical coupling analysis, simulate the welding process, calculate the temperature field, stress field and displacement field at any time, and extract the deformation caused by stress release at each time. A proxy model is constructed for real-time control. Multiple simulations with different initial stress distributions and different process parameters are run to generate a sample library. The initial stress peak value, location, and heat input are used as inputs, and the stress release deformation curve is used as output to train a neural network model to obtain a stress release prediction model.
3. The welding method for a crystallizer copper tube according to claim 2, characterized in that: The support force feedforward curve is constructed as follows: The copper tube is simplified into a beam model. The temperature field distribution generated by the movement of the heat source is calculated based on the welding process parameters. The thermal bending moment is calculated from the temperature field, and the theoretical thermal deformation curve is obtained by solving the bending differential equation. The stiffness of the copper tube at different locations was calibrated by static loading tests. The support force feedforward curve is equal to the stiffness curve multiplied by the theoretical thermal deformation curve.
4. The welding method for a crystallizer copper tube according to claim 1, characterized in that: The process of decomposing displacement into thermal deformation components and stress relief deformation components is as follows: Welding is performed using reference power and reference speed. Displacement sensors collect displacement at high frequency to generate a real-time deformation field, while stress sensors simultaneously collect local stress. By selecting the wavelet basis and the number of decomposition levels, wavelet decomposition is performed to obtain the approximation coefficients and detail coefficients of each level. The smoothing component is reconstructed to obtain the thermal deformation component, and the pulse component is reconstructed to obtain the stress release deformation component.
5. The welding method for a crystallizer copper tube according to claim 4, characterized in that: The calculation method for the ratio of thermal deformation to stress relief deformation is as follows: Construct a state-space model, integrate displacement and stress data, and estimate thermal deformation and stress release deformation in real time; Construct a state vector that includes thermal deformation components, stress-relief deformation components, thermal deformation rate, and stress-relief deformation rate; A state transition matrix is constructed based on the heat conduction equation and stress release prediction model; Construct an observation vector that includes macroscopic displacement measured by displacement sensors and local stress values measured by stress sensors; Construct an observation matrix where displacement observations are the sum of thermal deformation and stress-release deformation, and stress observations correspond to stress-release deformation. Kalman filtering is performed recursively. The current state and error covariance are predicted based on the state equation. The Kalman gain is calculated, and the state is updated in combination with the measured values. At each step, the estimated values of thermal deformation and stress release deformation at the current moment are output. And calculate the proportion of thermal deformation and the proportion of stress release deformation: proportion of thermal deformation = thermal deformation at the current moment / (thermal deformation at the current moment + estimated value of stress release deformation), proportion of stress release deformation = 1 - proportion of thermal deformation.
6. The welding method for a crystallizer copper tube according to claim 1, characterized in that: The method for determining the current welding mode is as follows: The stress reduction rate per unit time is calculated in real time. If both the stress release deformation ratio and the stress reduction rate per unit time meet the requirements, it is determined to be a stress release mode. Otherwise, if either the stress release deformation ratio or the stress reduction rate per unit time does not meet the requirements, it is a normal welding mode.
7. The welding method for a crystallizer copper tube according to claim 6, characterized in that: The process of differentiated thermo-mechanical coordinated control based on welding mode is as follows: If the current welding mode is determined to be normal: The feedforward support force is obtained by querying the feedforward curve based on the current welding position. The thermal deformation component is then high-pass filtered to obtain the high-frequency component of thermal deformation. The support force feedback is equal to the support force proportional coefficient multiplied by the high-frequency component of thermal deformation. The thermal deformation component is low-pass filtered to obtain the low-frequency component of thermal deformation. If the absolute value of the low-frequency component of thermal deformation meets the requirements, the adaptive adjustment of the heat source parameters is initiated. The adjustment power is the reference power minus the power adjustment coefficient multiplied by the absolute value of the low-frequency component of thermal deformation, and the adjustment speed is the reference speed plus the speed adjustment coefficient multiplied by the absolute value of the low-frequency component of thermal deformation. Using the current state as input, predict the deformation development at N future time points, construct a cost function, solve the optimization problem in each control cycle, obtain the optimal control sequence at future time points, and apply the first control variable of the optimization solution to the system.
8. The welding method for a crystallizer copper tube according to claim 7, characterized in that: The process of differentiated thermo-mechanical coordinated control based on welding mode also includes: If it is determined that the current mode is stress relief mode, keep the current power and speed unchanged, reduce the support force feedback gain, and the support force command is the feedforward support force plus the reduced support force feedback gain multiplied by the thermal deformation component at the current moment. The stress reduction rate is calculated in real time. If the stress reduction rate meets the requirements, the local heating device inside the mandrel is activated to provide auxiliary heating to the stress release area.
9. The welding method for a crystallizer copper tube according to claim 1, characterized in that: The method for correcting the parameters of the stress release prediction model is as follows: The residual stress on the surface of the copper tube after welding was measured using the same equipment and measuring point layout. The estimated stress release deformation along the weld was extracted from the welding process record. The actual stress release deformation was compared with the stress release deformation prediction value output by the stress release prediction model point by point to calculate the error. Error data from multiple batches of welding are collected to form a corrected dataset. A machine learning algorithm is used to establish an error prediction model. The input is process parameters and initial stress characteristics, and the output is the prediction error. The error prediction model is superimposed with the original prediction model to form a corrected stress release prediction model.
10. A welding system for copper tubes in a crystallizer, characterized in that: Includes the following modules: Stress detection and model building module: Before welding, residual stress is detected on the copper tube of the crystallizer, an initial stress distribution map is established, and based on the initial stress distribution map, a stress release prediction model is built through finite element simulation, and a feedforward curve of the support force along the weld is generated. Welding process sensing and signal decoupling module: Simultaneously collects displacement, local stress and temperature data during the welding process, and decomposes the displacement into thermal deformation component and stress release deformation component. Combined with the stress release prediction model, it estimates the proportion of thermal deformation and stress release deformation in the current deformation in real time. Welding mode judgment and differentiated control module: Real-time detection of stress release deformation ratio, judgment of current welding mode by comparing and analyzing stress release deformation ratio, and differentiated thermo-mechanical coordinated control according to welding mode by combining the support force feedforward curve along the weld. Post-weld stress detection and model correction module: After welding, the stress distribution of the crystallizer copper tube is detected again, the actual stress release deformation is compared and analyzed with the predicted value of the stress release prediction model, and the parameters of the stress release prediction model are corrected using machine learning algorithms.