A method and apparatus for controlling friction stir welding

Abstract

The present application relates to the technical field of welding, and discloses a control method and equipment for friction stir welding, which comprises the following steps: through the establishment of a multi-physical field model, real-time data acquisition, internal state estimation through data assimilation, future evolution prediction, model predictive control to determine and apply instructions, and online learning to continuously adjust model or control strategy parameters; the system comprises a model establishment module, a data acquisition module, a state estimation module, a state prediction module, a control decision and execution module, and an online learning module. The present application accurately estimates the internal state online through model and data assimilation, improves the sensing accuracy, realizes prospective adjustment based on state prediction and model predictive control, intelligently balances multi-dimensional performance by introducing a multi-objective cost function, enhances the adaptability and robustness of the system by integrating an online learning mechanism, and builds a complete intelligent closed-loop control system to improve the automation and intelligent level.

Description

Technical Field

[0001] This invention relates to the field of welding technology, and in particular to a control method and equipment for friction stir welding. Background Technology

[0002] Friction stir welding (FSW), as an important solid-state joining technology, is widely used in key manufacturing fields such as aerospace and automotive due to its advantages in avoiding melting-related defects and maintaining material properties. This technology achieves high-quality solid-state bonding by using a special rotating tool to generate heat through friction with the workpiece under axial force, and by violently stirring and plasticizing the material.

[0003] However, friction stir welding is essentially a complex multiphysics coupled dynamic system, and its weld quality is extremely sensitive to the precise control of process parameters. Currently, friction stir welding control methods largely rely on offline optimized fixed parameter sets or simple feedback control for single variables. These traditional methods are ill-suited to effectively address process uncertainties caused by raw material variations, tool wear, and environmental fluctuations in actual production. Due to the lack of real-time, in-depth perception of key states such as complex thermal interactions, material flow, and microstructure evolution within the welding zone, existing control strategies have limitations in proactively preventing defects and finely controlling weld quality, making it difficult to fully guarantee the stability of the welding process and the consistency of the final product.

[0004] Therefore, this invention proposes a control method and equipment for friction stir welding to overcome the shortcomings of the prior art. Summary of the Invention

[0005] The purpose of this invention is to provide a control method and equipment for friction stir welding, which solves the problems of insufficient perception of the internal state of the process, difficulty in adapting to changes in working conditions, and difficulty in controlling the welding quality in existing friction stir welding control methods.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for controlling friction stir welding, the method comprising the following steps:

[0007] S1. Based on the mathematical characterization of the physical phenomena and material properties of the friction stir welding process, an initial multiphysics model is established.

[0008] S2. During the friction stir welding process, multi-source sensor data of the welding process are collected in real time;

[0009] S3. Using the multiphysics model and the real-time collected multi-source sensor data, the internal physical state of the friction stir welding process is estimated online through a data assimilation method.

[0010] S4. Based on the internal physical state and using the multiphysics model, predict the state evolution of the friction stir welding process within a future time window;

[0011] S5. Based on the predicted future state evolution and the preset multi-objective cost function, a model predictive control strategy is used to determine the control command, and the determined control command is applied to the actuator of the friction stir welding equipment.

[0012] S6. Continuously monitor the performance feedback of the friction stir welding process and the multi-source sensor data, and adjust the parameters of the multiphysics model or the parameters of the model predictive control strategy through an online learning mechanism.

[0013] Preferably, establishing the initial multiphysics model in S1 includes:

[0014] The initial multiphysics model is defined to include a thermodynamic model module and a material constitutive model module. The thermodynamic model module is used to mathematically characterize the heat generation and transfer phenomena in the friction stir welding process, and the material constitutive model module is used to mathematically characterize the mechanical response and plastic flow characteristics of the friction stir welding material under welding conditions.

[0015] The material property parameters in the thermodynamic model module and the material constitutive model module are given initial values.

[0016] Preferably, the mathematical representation of the thermodynamic model module includes the energy conservation equation:

[0017] ;

[0018] in, For material density, For the specific heat capacity of the material, For temperature field variables, For time variables, For the thermal conductivity of the material, This refers to a heat source term per unit volume; and the heat source term per unit volume... The heat generated by friction and the heat generated by the plastic deformation of the material are determined by mathematical characterization.

[0019] Preferably, the multi-source sensor data acquired in real time during the welding process in step S2 includes:

[0020] The raw measurement signals acquired in real time by the sensor are preprocessed, including timestamp alignment, signal filtering, and outlier removal, in order to obtain sensor data.

[0021] The sensor data includes: axial force data. ; Lateral force data Stirring head torque data Temperature data at predetermined monitoring points for tools or weldments Actual welding speed data Actual rotation speed data of the stirring head Real-time spatial position data of the stirring head .

[0022] Preferably, the online estimation of the internal physical state of the friction stir welding process using a data assimilation method in step S3 includes:

[0023] At each discrete time step Using the multiphysics model and the previous discrete time step Posterior estimate of internal physical state The current discrete time step is calculated using the applied control input. Prior estimates of internal physical state ,in, The vector representing the internal physical field of the friction stir welding process;

[0024] The measurement vector formed by the multi-source sensor data The multiphysics model is based on the prior estimates of the internal physical states. Predicted corresponding observable Comparisons are made to generate new information;

[0025] A data assimilation algorithm is employed, namely Kalman filtering, extended Kalman filtering, unscented Kalman filtering, ensemble Kalman filtering, or particle filtering, to utilize the new information to estimate the prior value of the internal physical state. Perform correction to obtain the current discrete time step. Posterior estimate of internal physical state ,in To be in discrete time steps The measurement vector, An observation operator that maps a state vector to a measurement space.

[0026] Preferably, the prediction of the state evolution of the friction stir welding process within a future time window in S4 includes:

[0027] Using the internal physical state, along with its associated uncertainty information, as a discrete time step Initial conditions;

[0028] Using the aforementioned multiphysics model, for the preset prediction time domain Calculate the predicted future state evolution sequence in response to the candidate control input sequence. ,in, The length of the predicted time domain. The sequence of predicted future states;

[0029] Uncertainty information associated with the internal physical state and the multiphysics model is propagated within the prediction time domain to determine the evolution sequence of the predicted future state. Related uncertainties.

[0030] Preferably, the method of determining control commands using a model predictive control strategy in step S5 includes:

[0031] At each discrete time step Based on the predicted future state evolution and the associated uncertainties, a robust optimization problem is solved to determine the state in the preset control time domain. Optimal control input increment sequence within The sequence aims to minimize the preset multi-objective cost function. And satisfy the operational constraints, wherein, The length of the preset control time domain;

[0032] The optimal control input increment sequence Extract the first set of control input increments Based on the increment of the first set of control inputs, the control command for the current discrete time step is calculated. ,in This is an initial component of the optimal control input increment sequence.

[0033] Preferably, the preset multi-objective cost function A weighted combination including the following:

[0034] First cost item This is used to quantify the transient performance of the welding process over a short timescale.

[0035] Second cost item This is used to quantify the cumulative effects or potential defect formation trends of the welding process over a medium timescale.

[0036] Third cost item It is used to quantify the final quality or performance indicators of welds over a long time scale.

[0037] Preferably, adjusting the parameters of the multiphysics model or the parameters of the model predictive control strategy through an online learning mechanism in step S6 includes:

[0038] A performance evaluation signal is generated by comparing the real-time acquired multi-source sensor data with the corresponding output predicted by the multi-physics model based on the current parameter set, or by comparing the actual operating performance indicators of the friction stir welding process with the preset performance target value.

[0039] An online parameter estimation algorithm is used to adaptively adjust a set of adjustable model parameters in the multiphysics model based on the performance evaluation signal. Or a set of adjustable strategy parameters in the model predictive control strategy. ,in, This represents a pre-selected set of adjustable parameters in the multiphysics model. This represents a pre-selected set of adjustable parameters in the model predictive control strategy.

[0040] The present invention also provides a control device for friction stir welding, the device comprising the following modules:

[0041] The model building module establishes an initial multiphysics model based on the mathematical characterization of the physical phenomena and material properties of the friction stir welding process.

[0042] The data acquisition module collects multi-source sensor data in real time during the friction stir welding process.

[0043] The state estimation module uses the multiphysics model and combines it with the real-time acquired multi-source sensor data to estimate the internal physical state of the friction stir welding process online through a data assimilation method.

[0044] The state prediction module, based on the internal physical state and using the multiphysics model, predicts the state evolution of the friction stir welding process within a future time window.

[0045] The control decision and execution module determines control commands based on the predicted future state evolution and the preset multi-objective cost function using a model predictive control strategy, and applies the determined control commands to the actuator of the friction stir welding equipment.

[0046] The online learning module continuously monitors the performance feedback of the friction stir welding process and the multi-source sensor data, and adjusts the parameters of the multiphysics model or the parameters of the model predictive control strategy through the online learning mechanism.

[0047] In summary, the present invention has at least one of the following beneficial technical effects:

[0048] 1. This invention achieves accurate online estimation of key internal physical states that are difficult to measure directly during friction stir welding by constructing a sophisticated multiphysics model and combining it with real-time sensor data for data assimilation. This surpasses traditional control methods based on externally measurable parameters, significantly improving the understanding of the welding process mechanism and the precision of state perception, laying a solid foundation for achieving higher levels of process control.

[0049] 2. This invention employs a technical approach that combines model-based state prediction with model predictive control strategies, enabling the predictive adjustment of control inputs to address future state changes. By optimizing the control sequence for a future period in each control cycle, rather than merely responding to current errors, it can more proactively suppress disturbances, reduce process fluctuations, and effectively avoid the formation of potential welding defects, thereby significantly improving the control accuracy of the welding process and the quality stability of the final weld.

[0050] 3. This invention introduces a multi-objective cost function into the model predictive control framework. This cost function comprehensively considers multiple dimensions, including transient performance, mid-term cumulative effects, and long-term final quality of the welding process. It enables intelligent trade-offs and optimization among different performance indicators, balancing production efficiency or energy consumption while ensuring welding quality. This achieves more comprehensive and refined performance control of the friction stir welding process, meeting the comprehensive needs of complex welding tasks for multiple objectives.

[0051] 4. This invention integrates an online learning and adaptive adjustment mechanism, which can continuously monitor the actual performance feedback and sensor data of the welding process, and dynamically adjust the key parameters of the multiphysics model or the parameters of the model predictive control strategy accordingly. This enables it to have strong adaptability and robustness to uncertainties such as changes in working conditions, material batch differences, and tool wear, ensuring that the control system can maintain high efficiency and stability in long-term operation and continuously optimize its performance.

[0052] 5. This invention constructs a complete and highly intelligent closed-loop control system for friction stir welding through the collaborative work of model building, multi-source data acquisition and preprocessing, internal state estimation, future state prediction, model predictive control decision-making, and online learning modules. This significantly improves the automation and intelligence of the welding process, reduces reliance on operator experience, and enhances the reliability and consistency of complex welding tasks, providing strong technical support for achieving high-quality and high-efficiency friction stir welding. Attached Figure Description

[0053] Figure 1 This is a flowchart of the method of the present invention;

[0054] Figure 2 This is a diagram of the device architecture of the present invention. Detailed Implementation

[0055] The following is in conjunction with the appendix Figure 1 -Appendix Figure 2 The present invention will be further described in detail below.

[0056] This invention provides a control method for friction stir welding, the method comprising the following steps:

[0057] S1. Based on the mathematical characterization of the physical phenomena and material properties of the friction stir welding process, an initial multiphysics model is established.

[0058] In this embodiment, step S1 involves constructing an initial multiphysics model, which aims to mathematically characterize the key physical phenomena and material properties involved in friction stir welding. To achieve effective simulation of the friction stir welding process, the initial multiphysics model includes a thermodynamic model module and a material constitutive model module.

[0059] The thermodynamic model module is designed to describe and calculate the heat generation mechanism and transient transfer behavior in three-dimensional space during friction stir welding. The heat in friction stir welding mainly originates from two sources: first, the intense friction between the stirring tool (especially its shoulder and stirring pin) and the workpiece material due to high-speed relative rotation and forward movement; and second, the heat generated from plastic work as the workpiece material undergoes intense plastic deformation under the strong force of the stirring head. Heat transfer within the weldment and tool is primarily through conduction. Simultaneously, the effects of thermal convection and radiation with the environment must also be considered to ensure the accuracy of the model predictions.

[0060] To quantitatively describe the aforementioned thermal phenomena, the mathematical characterization of the thermodynamic model module preferably includes establishing and solving the energy conservation equation. The energy conservation equation takes the form shown below:

[0061] ;

[0062] in, For material density, For the specific heat capacity of the material, For temperature field variables, For time variables, For the thermal conductivity of the material, This refers to a heat source term per unit volume; and the heat source term per unit volume... The heat generated by friction and the heat generated by the plastic deformation of the material are determined by mathematical characterization.

[0063] The heat source term per unit volume The determination of heat generation is achieved by mathematically representing the sum of heat contributions from friction and plastic deformation of the material. For the friction-generated heat component, a model can be established based on the frictional state of the contact interface (such as adhesion or slippage), the coefficient of friction, the contact pressure, and the relative slippage velocity between the stirring tool and the workpiece.

[0064] For example, under slip friction conditions, the frictional heat flow rate per unit area It can be represented as ,in, The coefficient of friction, In order to contact the normal pressure, This represents the relative slip velocity.

[0065] For the heat generated during plastic deformation, its magnitude is related to the material's flow stress and plastic strain rate, and is called the Taylor-Quinney coefficient, which is the plastic work conversion factor. Related. Heat generation rate per unit volume from plastic deformation. It can be represented as:

[0066] ;

[0067] in, The equivalent flow stress of the material, It represents the equivalent plastic strain rate.

[0068] The material constitutive model module is designed to describe the complex mechanical response behavior and intense plastic flow characteristics of the materials being welded under the extreme physical conditions unique to friction stir welding—namely, high temperature, large strain, and high strain rate. During friction stir welding, the material undergoes a series of complex physical and metallurgical processes, including softening, stirring, transport, and forging. Its mechanical behavior directly determines the heat generation, the mixing effect of the material, and the final weld formation quality.

[0069] To characterize this complex material behavior, the material constitutive model module needs to select a mathematical model that reflects the stress-strain relationship of the material over a wide temperature and strain rate range.

[0070] For example, the Johnson-Cook model, Arrhenius-type flow stress models (such as the Sellars-Tegart model or the Zener-Hollomon parametric correlation model), or other advanced constitutive equations that comprehensively consider strain hardening, strain rate strengthening, and thermal softening effects can be employed. These models typically include a series of material constants that need to be calibrated based on experimental data for specific materials. The outputs of this module, such as flow stress, are not only used for mechanical analysis but also serve as important inputs for calculating the heat generated by plastic deformation.

[0071] After constructing the mathematical framework of the thermodynamic model module and the material constitutive model module, reasonable initial values ​​are assigned to the various material property parameters contained in these two modules. These parameters constitute the "skeleton" of the model, and their accuracy directly affects the reliability of the model's predictions. These material property parameters include, but are not limited to: density. Specific heat capacity Thermal conductivity (and its functional relationship with temperature), coefficient of thermal expansion, elastic modulus, Poisson's ratio, yield strength of the material, work hardening index, strain rate sensitivity coefficient, high-temperature creep parameter, and specific material constants in the aforementioned constitutive model, etc. These are the "initial values" provided here, which will be dynamically adjusted and optimized based on actual sensor data feedback during subsequent online learning, thereby enabling the multiphysics model to better approximate the actual welding process.

[0072] In some implementations, the initial multiphysics model can be further integrated with sub-models for describing microstructure evolution (e.g., prediction of dynamic recrystallization grain size, dissolution and precipitation of precipitates), or sub-models for more finely simulating complex interface behavior between the stirring head and the workpiece material (e.g., varying friction coefficients, contact state determination). Furthermore, a dynamic model library can be pre-established, containing multiple versions of physical models with different fidelity levels (e.g., simplified empirical or analytical models, computationally efficient two-dimensional finite element / finite difference models, and computationally accurate but resource-intensive three-dimensional high-precision finite element / finite volume models) and different physical coupling depths. In step S1, based on a preliminary trade-off between computational speed and prediction accuracy for the current application scenario, a suitable version is selected or constructed from this model library as the initial multiphysics model.

[0073] The initial multiphysics model is typically established using commercial finite element analysis software (such as ANSYS, ABAQUS, COMSOL, etc.) or a self-developed numerical calculation program. By discretizing the governing equations in space and time (e.g., using the finite element method to discretize space and the finite difference method to discretize time), the complex partial differential equations are transformed into a set of algebraic equations for solution, thereby simulating the spatiotemporal distribution of key physical quantities such as temperature field, stress-strain field, and material flow field during the welding process.

[0074] S2. During the friction stir welding process, multi-source sensor data of the welding process are collected in real time;

[0075] In this embodiment, step S2 involves real-time and synchronous acquisition of multi-source heterogeneous sensor data characterizing the welding process state during the dynamic process of friction stir welding. This step is a crucial bridge connecting the physical welding world and the digital twin model; the quality and real-time performance of the acquired data directly affect the accuracy of subsequent state estimation, the reliability of prediction, and the effectiveness of control decisions.

[0076] In the complex and dynamic environment of friction stir welding, by integrating non-contact measuring devices (such as infrared thermal imagers, machine vision systems, etc.) into key parts of the welding equipment (such as spindle systems, worktables, clamping systems, etc.) or deploying them near the welding area, raw measurement signals reflecting key mechanical, thermal, and kinematic characteristics of the welding process can be captured in real time. These raw signals are typically analog voltage or current signals or digitized data streams, directly derived from the sensor's response to physical quantities.

[0077] Raw measurement signals directly acquired from sensors often contain various interferences and uncertainties, making them difficult to use directly for precise analysis and control. Therefore, preprocessing these raw measurement signals aims to clean the data, eliminate redundancy, and improve data quality, thereby obtaining reliable sensor data. The preprocessing operation preferably includes the following key steps:

[0078] First, timestamp alignment is performed. Since multi-source sensor data typically originates from independent sensors or data acquisition systems with different sampling frequencies and internal clocks, their raw data streams are biased in terms of time reference. Timestamp alignment aims to assign a unified, high-precision time reference to all acquired data points. This is crucial for subsequent multi-sensor data fusion; for example, in the state estimation step S3, it is necessary to correlate the measurements of different physical quantities at the same time. Implementation methods include hardware-level synchronization using Network Time Protocol (NTP) or Precise Time Protocol (PTP), or post-calibration based on synchronization signals or common events using software algorithms.

[0079] Secondly, signal filtering is performed. Raw sensor signals often contain noise components introduced by electromagnetic interference, mechanical vibration, power fluctuations, or other environmental factors. If this noise is not processed, it will affect the accuracy of subsequent analysis and may lead to incorrect control decisions. The purpose of signal filtering is to suppress or remove this noise while preserving as much of the true signal characteristics as possible. Different types of filters are selected based on the characteristics of the noise and the type of signal. For example, low-pass filters are used to remove high-frequency noise, median filters are used to handle impulse interference, or more complex adaptive filtering techniques are employed in specific situations.

[0080] Next, outlier removal is performed. During data acquisition, due to factors such as momentary sensor malfunctions, sudden strong interference, data transmission errors, or atypical mutations in the measured object, some isolated and unreasonable values ​​deviating from the normal data variation range may occur; these are outliers or anomalies. If outliers are not removed, they will severely interfere with parameter identification and state estimation in the model. Outlier removal can be achieved by setting reasonable upper and lower thresholds for physical quantities, using statistical methods (such as the 3-sigma criterion and box plots), or using clustering methods based on the neighborhood information of data points to identify, remove, or correct these abnormal data points.

[0081] Through the above preprocessing operations, purer, more reliable, and time-synchronized sensor data is extracted from the original, flawed measurement signals, providing a solid data foundation for subsequent state updates and predictive control of the digital twin model.

[0082] The sensor data obtained after preprocessing comprehensively and multidimensionally depicts the complex state of the friction stir welding process, and its content preferably includes one or more of the following key data types:

[0083] Axial force data Axial force represents the force exerted by the stirring head on the workpiece surface along its rotation axis. Axial force is a crucial process parameter in friction stir welding, directly affecting the depth of penetration of the stirring head into the workpiece, the contact state between the tool and the workpiece, the frictional heat generated, and the plastic flow behavior of the material. Real-time monitoring of axial force data is essential. This helps to assess the stability of the welding process and can be used as a feedback signal for closed-loop control of axial force or indentation depth.

[0084] Lateral force data Transverse force data: Characterizes the force perpendicular to the welding direction and parallel to the workpiece surface, or its component in a specific direction (e.g., perpendicular to the weld centerline). Changes in the flow rate can reflect problems such as asymmetry in material flow, tool eccentricity, or improper workpiece clamping during the welding process. This is of indicative significance for monitoring welding stability and preventing certain types of welding defects (such as uneven lateral extrusion and weld offset).

[0085] Stirring head torque data This represents the magnitude of the torque required to drive the stirring head to rotate. Torque directly reflects the shear resistance of the workpiece material to the rotational motion of the stirring head, and is therefore related to the softening degree (i.e., temperature) of the material in the welding area, the resistance to plastic deformation, and the frictional state between the stirring head and the material. Real-time monitoring of stirring head torque data is crucial. It can indirectly infer the thermal state of the welding zone, and can also monitor tool wear or process abnormalities.

[0086] Temperature data at predetermined monitoring points for tools or weldments This data characterizes the real-time temperature values ​​of the welding tool (such as a specific point on the surface of a shoulder or stirring pin) or the workpiece (such as the weld surface or a critical location in the heat-affected zone) at one or more pre-defined monitoring points. Temperature is the most critical factor affecting the softening, plastic flow, microstructure evolution, and final joint performance of friction stir welding materials. (Acquiring temperature data is crucial.) This can be achieved using non-contact infrared thermometers, infrared thermal imagers, or by embedding thermocouples in tools or workpieces. Accurate temperature information is crucial for verifying thermal models, controlling peak temperatures, and ensuring welding quality.

[0087] Actual welding speed data This characterizes the actual translational speed of the stirring head relative to the workpiece, commonly referred to as welding speed. Welding speed is a key process parameter for controlling the heat input per unit length of weld and welding production efficiency. Actual welding speed data is obtained in real time through encoders or other displacement sensors. It can ensure that the welding process is executed precisely according to the preset process parameters, and can detect the deviation between the actual speed and the set value caused by drive system failure or external disturbance.

[0088] Actual rotation speed data of the stirring head This characterizes the actual rotational speed of the stirring head. The stirring head rotational speed directly affects the rate of frictional heat generation and the degree of material mixing. The actual rotational speed of the stirring head is monitored in real time using a motor encoder or speed sensor. This is used to ensure the stability of the rotational speed.

[0089] Real-time spatial position data of the stirring head This represents the three-dimensional real-time spatial coordinates of the stirring head (e.g., its tool coordinate system origin or stirring pin tip) in the working coordinate system of the welding equipment. It involves acquiring the real-time spatial position data of the stirring head. It is crucial for precisely controlling the weld trajectory, achieving complex path welding, and accurately defining the relative geometric relationship between the tool and the workpiece in a multiphysics model.

[0090] S3. Using the multiphysics model and the real-time collected multi-source sensor data, the internal physical state of the friction stir welding process is estimated online through a data assimilation method.

[0091] In this embodiment, step S3 dynamically integrates the theoretical multiphysics model constructed in step S1 with the multi-source sensor data collected in real time from the actual welding process in step S2. The purpose is to estimate online and accurately the key physical states inside the friction stir welding process that are difficult to obtain through direct measurement. The internal physical states include the three-dimensional transient temperature field, stress-strain field, complex plastic flow velocity field of the weld nugget region, and even the distribution of microstructure parameters (such as grain size and phase composition).

[0092] Specifically, the internal physical state of the friction stir welding process is estimated online using data assimilation methods, typically following an iterative prediction-correction framework, which mainly includes the following interrelated operational steps:

[0093] First, at each discrete time step Prior estimation of the internal physical state is performed. This is done using the multiphysics model established in step S1, which may have been adjusted online in step S6, and combined with the previous discrete-time step. The obtained posterior estimate of the internal physical state This serves as the initial condition for the current prediction. Simultaneously, at this time step... The actual control inputs applied to the welding equipment during this process (e.g., set stirring head speed, welding speed, axial force, etc.) are also used as boundary conditions or driving force inputs for the model. The current discrete time step is calculated by numerically solving the control equations of the multiphysics model. The prior estimate of the internal physical state of is denoted as . Here, the aforementioned It is a high-dimensional state vector that contains a series of variables that discretize and characterize the key physical fields inside the friction stir welding process. For example, it includes the temperature values ​​of all discrete units or nodes in the welding area, the components of the stress tensor, the components of the strain tensor, the components of the material flow velocity vector, and microstructure parameters such as average grain size and volume fraction of specific phases, which may be included in some advanced models. Indicates in Based on the time and all available information prior to that time Estimation of the state at time step. Therefore, This represents the current moment before integration. The model's prediction of the current state is based on the actual measured data.

[0094] Innovation or residuals are calculated. The purpose is to quantify the inconsistency between model predictions and actual observations. The multi-source sensor data, acquired in real-time and preprocessed in step S2, constitutes the current discrete-time step. The actual measurement vector, denoted as . The prior estimates of the internal physical state obtained in the previous step are used to... Through an observation operator Mapped to the sensor's measurement space.

[0095] The observation operator The function is to extract the internal state vector from a high-dimensional vector. Extract or calculate the actual sensor The model prediction value corresponding to the measured physical quantity. For example, if If it contains the temperature measurement value of a certain weld surface point, then From the internal three-dimensional temperature field The model predicts the temperature by interpolating or extracting the value at a specific point.

[0096] By comparing actual measurement vectors Corresponding observables predicted by the model based on prior state estimation This can form a new information vector. , is usually defined as: ;

[0097] The new information vector It intuitively reflects the degree and direction of the model's prediction deviating from the actual process, and is the key driving information for subsequent correction steps.

[0098] Finally, a suitable data assimilation algorithm is used to utilize the new information calculated in the previous step. Prior estimates of the internal physical state Perform corrections or updates to obtain the current discrete time step. The posterior estimate of the internal physical state of is denoted as . This a priori estimate is the best estimate of the system's internal state after incorporating the actual measurement information at the current moment.

[0099] The preferred data assimilation algorithms used in this invention include, but are not limited to, the following:

[0100] Kalman Filter (KF): As a classic optimal state estimation algorithm for linear Gaussian systems, it effectively fuses model predictions and noisy measurements in a recursive manner.

[0101] Extended-Kalman Filter (EKF): It achieves local linearization by performing a first-order Taylor series expansion on the nonlinear model equations and observation equations, thereby extending the Kalman filtering concept to nonlinear systems.

[0102] Unscented-Kalman Filter (UKF): Uses unscented-transform to approximate the probability distribution of state variables through the propagation of nonlinear functions. It usually has higher nonlinear approximation accuracy and better numerical stability than EKF.

[0103] Ensemble-Kalman Filter (EnKF): It uses a set of state samples (i.e., a set) to represent the probability distribution of the state, and propagates and updates these samples through the Monte Carlo method. It is particularly suitable for high-dimensional nonlinear systems and is easy to implement.

[0104] Particle-Filter (PF): A sequential importance sampling method based entirely on Monte Carlo simulation. It uses a set of weighted random samples (i.e., particles) to approximate the posterior probability density function. It can handle arbitrary forms of nonlinearity and non-Gaussian noise, but the computational cost is usually large.

[0105] These data assimilation algorithms are based on the uncertainty of the model prediction (derived from the process noise covariance matrix). The uncertainty of the measurement data (characterized by the measurement noise covariance matrix) and the measurement noise covariance matrix Characterization) is used to dynamically determine the level of confidence in model predictions and actual measurements (e.g., by calculating the Kalman gain matrix) and accordingly to make weighted corrections to prior state estimates.

[0106] S4. Based on the internal physical state and using the multiphysics model, predict the state evolution of the friction stir welding process within a future time window;

[0107] In this embodiment, based on the best estimate of the internal physical state of the current friction stir welding process obtained in step S3, and by making full use of the multiphysics model constructed in step S1, the dynamic behavior and state evolution of the welding process within a preset time window in the future are deduced and predicted.

[0108] Specifically, the prediction of the state evolution of the friction stir welding process within a future time window in S4 rigorously includes the following key operational steps:

[0109] The primary step is to set the initial conditions for the predicted evolution. This involves using the data assimilation method obtained in step S3 to determine the current discrete-time step. Posterior estimate of internal physical state This serves as the starting point for initiating future state prediction calculations. The posterior estimate of this internal physical state... Closely related uncertainties must also be included as part of the initial conditions for prediction. These uncertainties can be specifically manifested, for example, as the posterior state covariance matrix obtained when using Kalman filtering algorithms (such as EKF, UKF), or as the set of samples and their statistical properties that characterize the state probability distribution obtained when using ensemble Kalman filtering (EnKF) or particle filtering (PF).

[0110] Secondly, forward dynamic prediction is performed using the multiphysics model. Utilizing the multiphysics model constructed in step S1 and possibly optimized online by step S6, a prediction is performed for a pre-defined prediction time domain with practical physical meaning. Numerical simulations were performed. This prediction time domain... The system's "look-forward" timeframe is defined, and its selection requires a trade-off between the need for predictive accuracy and computational efficiency. Within this prediction time domain, the system will calculate and respond to at least one or more candidate future control input sequences, denoted as the predicted future state evolution sequence of the friction stir welding process. .in, This represents the current discrete time step; The length of the preset prediction time domain can be set, for example, to several seconds to tens of seconds in the future, covering several control cycles; and Then, within the prediction time domain, it is a sequence consisting of a series of predicted state vectors arranged in chronological order, i.e.: Here Representative based on the deadline All information, for the future The predicted values ​​of the physical state within a given time step. These candidate control input sequences represent a series of possible combinations of control actions attempted to optimize future performance within the model predictive control framework, such as a sequence of target stirring head rotation speed, target welding speed, or target axial force for several future steps. This prediction process typically involves iteratively solving a multiphysics model in the time dimension.

[0111] Uncertainty information associated with the internal physical state and the multiphysics model itself is carefully propagated within the prediction time domain. While performing forward dynamic prediction, the uncertainty of the initial state (originating from...) The covariance (or sample distribution) of the model, along with the uncertainty of the model parameters and the inherent process noise, will propagate forward in the prediction time domain as the nonlinear dynamics of the multiphysics model evolve, and may change (e.g., amplified or attenuated). Therefore, this step aims not only to predict the expected trajectory of future states, but also to simultaneously estimate and quantify the uncertainties of these predicted states.

[0112] The method for achieving uncertainty propagation is closely related to the data assimilation algorithm used in step S3. For example, if ensemble Kalman filtering or particle filtering is used, a set of predicted state samples can be obtained at each future time point in the prediction time domain by performing forward integration of the initial set of state samples through a deterministic multiphysics model, and the statistical distribution of the samples (such as mean, variance, confidence interval, etc.) directly characterizes the uncertainty of the predicted state. If extended Kalman filtering or unscented Kalman filtering is used, the covariance matrix of the initial state can be recursively derived through the corresponding covariance propagation equation to obtain the covariance matrix of the predicted state at each future time point.

[0113] Through this uncertainty propagation process, the final determination is made of the entire predicted future state evolution sequence. Each predicted state vector The associated uncertainty measure is expressed as the probability distribution, expected value, and variance, or the confidence interval of the prediction, for key predicted physical quantities (such as the peak temperature of the weld nugget at a future time, the maximum width of the heat-affected zone, or the stress level of a region).

[0114] S5. Based on the predicted future state evolution and the preset multi-objective cost function, a model predictive control strategy is used to determine the control command, and the determined control command is applied to the actuator of the friction stir welding equipment.

[0115] In this embodiment, step S5, based on the accurate prediction of the future state evolution of the friction stir welding process in step S4, and combined with the multi-objective cost function of the welding process quality and efficiency requirements, employs an advanced Model-Predictive-Control (MPC) strategy to calculate the optimal control command sequence online. Subsequently, the first control command in this sequence is applied to each actuator of the friction stir welding equipment, thereby achieving closed-loop, optimized, and adaptive adjustment of the welding process.

[0116] Specifically, in S5, the process of determining control commands using a model predictive control strategy occurs at each discrete control time step. The following key operations will be executed iteratively:

[0117] First, a robust optimization problem is constructed and solved. Based on the information provided in step S4 in the prediction time domain... Predicting future state evolution sequences within (in The search is conducted within a preset control time domain, along with its associated uncertainty information. The optimal control input increment sequence within is denoted as Here, This represents how many future control actions the controller optimizes at the current moment; its length is typically less than or equal to the prediction time domain. The aim is to indirectly optimize the future system state trajectory by optimizing the sequence of future control actions.

[0118] The optimal control input increment sequence Specifically represented as ,in, Represents the current moment Decisions and plans for the future The incremental vector of control input applied at each time step.

[0119] The determination of this sequence aims to achieve two objectives: first, to minimize a comprehensive, pre-defined multi-objective cost function. First, the cost function quantifies the expected performance of the welding process in the prediction time domain; second, it strictly satisfies a series of predefined process and equipment operation constraints throughout the optimization process to ensure the safety and feasibility of the welding process.

[0120] Due to the predicted future state evolution The associated uncertainties are explicitly taken into account, and the optimization problem here is a robust optimization problem. There are several specific technical approaches to achieve this robustness. For example, a penalty term for the uncertainty of the predicted state or output can be strategically added to the cost function, or the original deterministic constraints can be transformed into probabilistic constraints or constraints based on worst-case analysis.

[0121] The preset multi-objective cost function This allows for a comprehensive and hierarchical measurement of the dynamic performance of the friction stir welding process and the quality of the final product. In this embodiment, the cost function... Preferably, the design includes a weighted combination of the following items to reflect a comprehensive consideration of different time scales and different performance dimensions:

[0122] The first cost term is denoted as... Its main function is to quantify and penalize transient performance deviations in the welding process over a short timescale. This cost term typically focuses on the tracking error of the welding state variables to their desired reference trajectory, as well as the magnitude and drastic change of the control input required to achieve this tracking. For example, it can be specificd as a quadratic cost function of the following form:

[0123] ;

[0124] in, Representing the future The expected system output or state setpoint vector at each discrete time step; The representation is based on the current moment. The information and the applied candidate control sequence, predicted by the multiphysics model in the future... The system output vector for each discrete time step; That is, in the future A discrete-time step plan applies an incremental vector of control inputs; and These are user-defined nonnegative definite weight matrices used to adjust the relative importance between the output tracking error and the control input increment magnitude. By minimizing this... This feature enables the welding process to respond quickly and smoothly to control commands, effectively suppresses unnecessary overshoot or oscillation, and avoids overly drastic control actions.

[0125] The second cost term is denoted as This approach focuses on quantifying and controlling the cumulative effects or potential defect formation trends of the welding process over intermediate timescales. For example, it can include quantifying key characteristics of the thermal cycling in the weld zone and comparing them to desired ranges; these characteristics are closely related to the final microstructure evolution and residual stress distribution of the material. This can be achieved through... The optimization aims to proactively prevent or suppress the emergence and development of these potential defects from the perspective of process control.

[0126] The third cost term is denoted as Its goal is to quantify and optimize the final quality or key performance indicators of the weld over a longer timescale. This cost term is directly or indirectly related to the final macroscopic mechanical properties, microscopic metallographic structure, and weld geometry. Since these final, macroscopic quality indicators are typically difficult to measure directly in real-time online during the welding process, It is usually characterized by indirect physical quantities or comprehensive indices that are highly correlated with the final quality and can be predicted by multiphysics models in the prediction time domain.

[0127] The cost terms at the three different time scales mentioned above , ,and Through a set of pre-defined, non-negative weighting factors , , We perform weighted combination to form the overall multi-objective cost function:

[0128] ;

[0129] The specific values ​​of the weighting factors reflect the trade-offs and relative importance considerations among different control objectives, and can be flexibly adjusted according to the specific welding task requirements and the emphasis on quality standards.

[0130] The operational constraints are necessary to ensure the safe and stable operation of the welding process and to protect expensive equipment. These constraints include: upper and lower limits on the amplitude of control input variables; constraints on the rate of change of control inputs to avoid excessive dynamic shocks or premature fatigue damage to the actuators; constraints on the predicted internal state of the system; and constraints on the predicted system output.

[0131] The solution covering the entire control time domain was obtained through numerical optimization algorithms. Optimal control input increment sequence Subsequently, this invention strictly adheres to the "receding-horizon" principle unique to model predictive control. That is, the entire sequence of future control actions is calculated, and the first set of control input increments is extracted from this optimal sequence, denoted as... . It is the optimal control input increment sequence The initial components.

[0132] Subsequently, based on the first set of control input increments The calculation yields the result for the current discrete time step. Actual control commands Typically, the controller directly outputs the control setpoint; therefore, the calculation method for this instruction is as follows:

[0133] ;

[0134] in, This refers to the control command actually applied to the actuator in the previous time step. Incremental control helps achieve a smooth control transition and includes integral action to eliminate steady-state errors. The resulting control command... It is then sent and applied to the corresponding actuator of the friction stir welding equipment to complete the closed-loop operation of the current control cycle.

[0135] In the next discrete time step When the time comes, the entire control process will be repeated as a complete loop. By using a rolling optimization approach where the control input is re-optimized based on the latest process feedback information and predictions of future behavior in each control cycle, model predictive control can effectively address the challenges of model mismatch, unmodeled dynamics, random disturbances, and time-varying characteristics in friction stir welding.

[0136] S6. Continuously monitor the performance feedback of the friction stir welding process and the multi-source sensor data, and adjust the parameters of the multiphysics model or the parameters of the model predictive control strategy through an online learning mechanism;

[0137] In this embodiment, step S6 establishes an online learning mechanism to continuously monitor the actual operational performance feedback of the friction stir welding process and the multi-source sensor data collected in real time by step S2. Based on this information, it dynamically adjusts the key parameters of the multiphysics model constructed in step S1, or the relevant parameters of the model predictive control strategy adopted in step S5. The adaptive adjustment mechanism ensures that the control system can maintain its high performance and robustness even when faced with situations such as batch differences in raw materials, tool wear, fluctuations in environmental conditions, or imperfect initial model calibration, and gradually improves its accuracy in recognizing the real physical process and its control efficiency.

[0138] In step S6, adjusting the parameters of the multiphysics model or the parameters of the model prediction and control strategy through an online learning mechanism specifically includes:

[0139] Generate a performance evaluation signal that quantifies the current system performance or model accuracy. This signal is the fundamental basis for driving subsequent parameter adjustments. Preferably, its generation method includes one or a combination of the following approaches:

[0140] Firstly, it involves directly comparing the multi-source sensor data acquired in real time during step S2. The corresponding output quantity predicted by the multiphysics model under the same conditions The difference between these two, namely prediction error or innovation, constitutes a direct measure of the model's current accuracy. For example, this performance evaluation... It can be represented as:

[0141] ;

[0142] in, In the current discrete time step The sensor measurement vectors that have been actually acquired and preprocessed; For multiphysics models to utilize their current parameter set These are the predicted values ​​of these observables based on information from the previous time step. They are persistent and statistically significant. This indicates that the model parameters may have deviated from reality and need to be adjusted.

[0143] Secondly, the difference between the macroscopic performance indicators exhibited by the friction stir welding process in actual operation and the pre-set or expected performance target values ​​is compared. The actual performance indicators can be extracted directly from sensor data or obtained indirectly. The deviation formed by comparing the actual performance indicators with the pre-set performance target values ​​can also constitute an important performance evaluation signal. .

[0144] Using one or more online parameter estimation algorithms or learning algorithms, based on the performance evaluation signal generated in the aforementioned steps (e.g. or It adaptively adjusts a pre-selected set of adjustable parameters. The adjustable parameters are mainly divided into two categories:

[0145] One type is a set of adjustable model parameters in the multiphysics model, denoted as... This parameter set This represents the physical parameters or empirical coefficients in the initial multiphysics model established in step S1 that have a significant impact on the model's accuracy, may contain uncertainties or drift over time, and are suitable for identification and optimization through online learning. For example, This can include the thermophysical properties of the material (such as thermal conductivity as a function of temperature). Or specific heat capacity (coefficients in the model), friction model parameters of the tool-workpiece interface (such as friction coefficient) The parameters include adhesive shear factor, key parameters in the material constitutive relation (such as material constants in the Johnson-Cook model or Arrhenius model), and parameters describing heat exchange boundary conditions (such as convective heat transfer coefficient). The purpose of online adjustment of these parameters is to enable the multiphysics model to more closely "track" the behavior of actual physical processes, thereby improving its accuracy in state estimation in step S3 and state prediction in step S4.

[0146] Another type is a set of adjustable policy parameters in the model predictive control (MPC) strategy, denoted as... This parameter set These represent the key adjustable parameters within the MPC controller designed in step S5 that influence its decision-making behavior and performance. For example, This can be included in a multi-objective cost function. The various weighting factors (such as tracking accuracy, control energy consumption, process stability, final quality, etc.) used to balance different control objectives are used to balance different control objectives. or matrix (elements in the data), prediction time domain of MPC Or control time domain These include boundary values ​​of constraints, or parameters of uncertainty models or risk aversion levels used in robust MPC design. The purpose of online adjustment of strategy parameters is to enable the MPC controller to self-optimize its control logic based on actual operating results and continuously evolving process characteristics, in order to better achieve predetermined process objectives or adapt to changing external demands.

[0147] Online parameter estimation algorithms for adaptive parameter adjustment include those that depend on the nature of the parameter to be adjusted, the characteristics of the performance evaluation signal, and the available computational resources.

[0148] For example: If the goal is to adjust model parameters To reduce model prediction error The method employs recursive least squares (RLS) and gradient descent-based optimization algorithms (e.g., by minimizing a certain norm of the prediction error). To update the parameters: ;in, (e.g., learning rate), or more complex sequential Monte Carlo methods based on Bayesian inference (e.g., particle filtering for parameter tracking).

[0149] If the goal is to adjust the MPC strategy parameters To improve macroscopic process performance indicators We can draw on the ideas of reinforcement learning (RL), treating the MPC controller as an agent, the friction stir welding process and its environment as the environment, and the performance evaluation signals as... (or its transformed form) serves as a reward signal, and policy parameters that maximize cumulative reward are learned online using RL algorithms (such as Q-learning, policy gradient method, etc.). Furthermore, for certain scalar performance metrics, a model-free extremum-seeking-control method is used to iteratively optimize the policy parameters.

[0150] Please see the appendix Figure 2 A control device for friction stir welding, the device comprising the following modules:

[0151] The model building module establishes an initial multiphysics model based on the mathematical characterization of the physical phenomena and material properties of the friction stir welding process.

[0152] The goal is to establish the initial digital twin foundation for the friction stir welding process. By mathematically abstracting and parameterizing key physical phenomena such as thermo-mechanical coupling, material plastic flow, and microstructure evolution during welding, a multi-physics coupling model is established. This model serves as the theoretical basis for subsequent state perception, prediction, and control decisions.

[0153] The data acquisition module collects multi-source sensor data in real time during the friction stir welding process.

[0154] This module is responsible for capturing multi-source heterogeneous sensor signals, such as force, torque, temperature, speed, and position, from the welding equipment and welding area in real time and synchronously. The acquired raw data undergoes preprocessing such as timestamp alignment, filtering, and outlier removal to transform it into a high-quality, analyzable sensor data stream, providing the system with direct observation of the actual welding process.

[0155] The state estimation module uses the multiphysics model and combines it with the real-time acquired multi-source sensor data to estimate the internal physical state of the friction stir welding process online through a data assimilation method.

[0156] This module integrates multiphysics models with real-time sensor data and employs data assimilation techniques (such as Kalman filtering and particle filtering) to infer key physical states that are difficult to measure directly inside the friction stir welding process, such as the three-dimensional temperature field and stress-strain field of the weld nugget.

[0157] The state prediction module, based on the internal physical state and using the multiphysics model, predicts the state evolution of the friction stir welding process within a future time window.

[0158] Based on the current best internal state estimate and driving a multiphysics model, this module proactively extrapolates the state evolution trajectory of the welding process within a preset future time window. It not only predicts the desired state but also quantifies the uncertainty of the prediction, providing future dynamic information for model predictive control.

[0159] The control decision and execution module determines control commands based on the predicted future state evolution and the preset multi-objective cost function using a model predictive control strategy, and applies the determined control commands to the actuator of the friction stir welding equipment.

[0160] This module is the core decision-making unit of the control system. It employs a model predictive control (MPC) strategy, based on future state predictions and a multi-objective cost function (comprehensively considering welding quality, efficiency, and energy consumption), to solve the constrained optimization problem online and determine the optimal control input sequence. Only the first control command of the sequence is applied to the equipment actuator, achieving rolling optimization closed-loop control.

[0161] The online learning module continuously monitors the performance feedback of the friction stir welding process and the multi-source sensor data, and adjusts the parameters of the multiphysics model or the parameters of the model predictive control strategy through the online learning mechanism.

[0162] This module endows the system with adaptive and self-optimizing capabilities. By continuously monitoring the deviation between process performance feedback and sensor data and model predictions, it dynamically adjusts the key parameters of the multiphysics model or the parameters of the MPC control strategy using online parameter estimation algorithms or machine learning techniques. This enables the system to cope with changes in operating conditions and model mismatch, continuously improving control performance.

[0163] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for controlling friction stir welding, characterized in that, The method includes the following steps: S1. Based on the mathematical characterization of the physical phenomena and material properties of the friction stir welding process, an initial multiphysics model is established. S2. During the friction stir welding process, multi-source sensor data of the welding process are collected in real time; S3. Using the aforementioned multiphysics model and combining it with the real-time acquired multi-source sensor data, the internal physical state of the friction stir welding process is estimated online through a data assimilation method; wherein, at each discrete time step... Using the multiphysics model and the previous discrete time step Posterior estimate of internal physical state The current discrete time step is calculated using the applied control input. Prior estimates of internal physical state ,in, The vector representing the internal physical field of the friction stir welding process; The measurement vector formed by the multi-source sensor data The multiphysics model is based on the prior estimates of the internal physical states. Predicted corresponding observable Comparison leads to new information; A data assimilation algorithm is employed, namely Kalman filtering, extended Kalman filtering, unscented Kalman filtering, ensemble Kalman filtering, or particle filtering, to utilize the new information to estimate the prior value of the internal physical state. Perform correction to obtain the current discrete time step. Posterior estimate of internal physical state ,in To be in discrete time steps The measurement vector, An observation operator that maps a state vector to a measurement space; S4. Based on the internal physical state and using the multiphysics model, predict the state evolution of the friction stir welding process within a future time window; wherein, the internal physical state, along with its associated uncertainty information, is used as the basis for predicting the state evolution of the friction stir welding process within a discrete time step. Initial conditions; Using the aforementioned multiphysics model, for the preset prediction time domain Calculate the predicted future state evolution sequence in response to the candidate control input sequence. ,in, The length of the predicted time domain. The sequence of predicted future states; Uncertainty information associated with the internal physical state and the multiphysics model is propagated within the prediction time domain to determine the evolution sequence of the predicted future state. Related uncertainties; S5. Based on the predicted future state evolution and the preset multi-objective cost function, a model predictive control strategy is used to determine the control command, and the determined control command is applied to the actuator of the friction stir welding equipment. S6. Continuously monitor the performance feedback of the friction stir welding process and the multi-source sensor data, and adjust the parameters of the multiphysics model or the parameters of the model predictive control strategy through an online learning mechanism.

2. The control method for friction stir welding according to claim 1, characterized in that, The initial multiphysics model established in S1 includes: The initial multiphysics model is defined to include a thermodynamic model module and a material constitutive model module. The thermodynamic model module is used to mathematically characterize the heat generation and transfer phenomena in the friction stir welding process, and the material constitutive model module is used to mathematically characterize the mechanical response and plastic flow characteristics of the friction stir welding material under welding conditions. The material property parameters in the thermodynamic model module and the material constitutive model module are given initial values.

3. The control method for friction stir welding according to claim 2, characterized in that, The mathematical representation of the thermodynamic model module includes the energy conservation equation: ； in, For material density, For the specific heat capacity of the material, For temperature field variables, For time variables, For the thermal conductivity of the material, This refers to a heat source term per unit volume; and the heat source term per unit volume... The heat generated by friction and the heat generated by the plastic deformation of the material are determined by mathematical characterization.

4. The control method for friction stir welding according to claim 1, characterized in that, The multi-source sensor data acquired in real time during the welding process in S2 includes: The raw measurement signals acquired in real time by the sensor are preprocessed, including timestamp alignment, signal filtering, and outlier removal, in order to obtain sensor data. The sensor data includes: axial force data. ; Lateral force data Stirring head torque data Temperature data at predetermined monitoring points for tools or weldments Actual welding speed data Actual rotation speed data of the stirring head Real-time spatial position data of the stirring head .

5. The control method for friction stir welding according to claim 1, characterized in that, The model predictive control strategy used in S5 to determine control commands includes: At each discrete time step Based on the predicted future state evolution and the associated uncertainties, a robust optimization problem is solved to determine the state in the preset control time domain. Optimal control input increment sequence within The sequence aims to minimize the preset multi-objective cost function. And satisfy the operational constraints, wherein, The length of the preset control time domain; The optimal control input increment sequence Extract the first set of control input increments Based on the increment of the first set of control inputs, the control command for the current discrete time step is calculated. ,in This is an initial component of the optimal control input increment sequence.

6. The control method for friction stir welding according to claim 5, characterized in that, The preset multi-objective cost function A weighted combination including the following: First cost item This is used to quantify the transient performance of the welding process over a short timescale. Second cost item This is used to quantify the cumulative effects or potential defect formation trends of the welding process over a medium timescale. Third cost item It is used to quantify the final quality or performance indicators of welds over a long time scale.

7. The control method for friction stir welding according to claim 1, characterized in that, The parameters of the multiphysics model or the parameters of the model prediction control strategy adjusted through the online learning mechanism in step S6 include: A performance evaluation signal is generated by comparing the real-time acquired multi-source sensor data with the corresponding output predicted by the multi-physics model based on the current parameter set, or by comparing the actual operating performance indicators of the friction stir welding process with the preset performance target value. An online parameter estimation algorithm is used to adaptively adjust a set of adjustable model parameters in the multiphysics model based on the performance evaluation signal. Or a set of adjustable strategy parameters in the model predictive control strategy. ,in, This represents a pre-selected set of adjustable parameters in the multiphysics model. This represents a pre-selected set of adjustable parameters in the model predictive control strategy.

8. A control device for friction stir welding, applied to the method described in any one of claims 1-7, characterized in that, The device includes the following modules: The model building module establishes an initial multiphysics model based on the mathematical characterization of the physical phenomena and material properties of the friction stir welding process. The data acquisition module collects multi-source sensor data in real time during the friction stir welding process. The state estimation module uses the multiphysics model and combines it with the real-time acquired multi-source sensor data to estimate the internal physical state of the friction stir welding process online through a data assimilation method. The state prediction module, based on the internal physical state and using the multiphysics model, predicts the state evolution of the friction stir welding process within a future time window. The control decision and execution module determines control commands based on the predicted future state evolution and the preset multi-objective cost function using a model predictive control strategy, and applies the determined control commands to the actuator of the friction stir welding equipment. The online learning module continuously monitors the performance feedback of the friction stir welding process and the multi-source sensor data, and adjusts the parameters of the multiphysics model or the parameters of the model predictive control strategy through the online learning mechanism.

Images