Online monitoring system for friction stir welding process and control method

Abstract

The invention relates to the technical field of intelligent control of the welding process, and discloses an online monitoring system and control method for the friction stir welding process, and the system comprises a multi-mode sensor array module which is used for collecting temperature, pressure and displacement data in the welding process; the edge computing node module preprocesses the collected data and extracts spatial-temporal characteristics; the space-time diagram convolutional network module is used for mining a space-time association relationship between the sensors; the hierarchical reinforcement learning controller generates a control instruction according to the associated features; the execution mechanism module receives the control instruction and outputs a welding signal; the digital twinborn verification module performs simulation feedback on the control effect; and the cloud model optimization module fuses the multi-source feedback information to continuously optimize the control strategy. The welding process self-adaptive control technology based on reinforcement learning is adopted, the technical effect of adjusting the welding parameters in real time to optimize the welding quality is achieved, and the problem that in a traditional method, the welding quality is unstable due to working condition changes is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent control of the welding process, and specifically to an on-line monitoring system and control method for the friction stir welding process. Background Art

[0002] In industrial scenarios such as high-end manufacturing, aerospace, rail transit, and new energy equipment, which have extremely high requirements for welding quality and process stability, traditional welding processes are difficult to meet the refined control requirements under complex working conditions. Especially in solid-state welding processes such as friction stir welding, the welding process is extremely sensitive to the temperature field, mechanical state, and control accuracy. Once the control parameters are not adjusted in real time, it is extremely easy to cause weld defects, uneven organization, or deterioration of joint performance.

[0003] The existing welding process monitoring and control technologies have achieved certain results, mainly focusing on real-time acquisition of sensors and signal processing. Some systems can achieve a certain degree of abnormal alarm of welding parameters through setting fixed threshold control logic. In some automated welding platforms, the input parameters can also be automatically corrected by using a preset process window, thereby improving the repeatability of the welding process. In addition, existing systems have also explored off-line optimization schemes based on historical data, which can make empirical improvements to conventional processes. These technologies have indeed improved the consistency and efficiency of welding under standard working conditions and stable load conditions, and have practicality.

[0004] However, the lack of adaptability of existing technologies under complex working conditions is still relatively obvious. First of all, traditional control systems rely on static parameter tables and are difficult to respond to disturbances in the dynamic welding process, such as material changes, temperature fluctuations, or external mechanical interference. They often show hysteresis and cannot achieve real-time adjustment. Secondly, existing monitoring means mainly focus on the surface data of sensors and lack the simulation and judgment of deep physical processes such as the thermal field and stress field, resulting in a lack of depth in feedback control and inaccurate response. Thirdly, most system optimization processes lack the ability of remote modeling, the strategy update is lagging and depends on manual intervention, the system cannot achieve self-evolution and iterative update, and it is difficult to support long-term operation under complex working conditions. Finally, most current solutions do not form an edge-cloud collaborative closed-loop structure, making it difficult to quickly apply the optimization results to actual control, the control strategy is disconnected from the on-site state, and the practicality is limited. For this reason, those skilled in the art have proposed an on-line monitoring system and control method for the friction stir welding process to solve the above problems. Summary of the Invention

[0005] Aiming at the deficiencies of the existing technology, the present invention provides an on-line monitoring system and control method for the friction stir welding process, which solves the problems in the existing technology such as the inability to adjust the control strategy in real time during the welding process, the lack of a deep feedback mechanism, and the lag of optimization and update.

[0006] To achieve the above objectives, the present invention is realized through the following technical solutions: An on-line monitoring system for friction stir welding process, comprising: A multi-modal sensor array module, configured to collect multiple physical parameters during the welding process, where the physical parameters include temperature, pressure, and displacement; An edge computing node module, configured to receive the output data of the multi-modal sensor array module and perform preprocessing to generate spatio-temporal features of the sensor data; A spatio-temporal graph convolutional network module, configured to construct a graph structure based on the spatio-temporal features generated by the edge computing node module, and extract spatio-temporal correlation features between multiple sensor nodes through graph convolution operations; A hierarchical reinforcement learning controller module, configured to generate welding parameter control instructions based on the spatio-temporal correlation features, where the control instructions include rotational speed adjustment amount, pressure adjustment amount, and displacement adjustment amount; An actuator module, configured to receive the control instructions and output corresponding welding control signals; A digital twin verification module, configured to perform physical simulation on the welding process based on the control signals and feedback the simulation results to the hierarchical reinforcement learning controller module; A cloud model optimization module, configured to receive feedback information from the edge computing node module and the digital twin verification module, generate an optimized control strategy, and update it to the hierarchical reinforcement learning controller module.

[0007] Preferably, the multi-modal sensor array module includes: A temperature sensing unit, configured to obtain real-time temperature data of different nodes in the welding area; A pressure sensing unit, configured to obtain the contact pressure between the stirring pin or the stirring head and the workpiece; A displacement sensing unit, configured to detect the displacement change of the welding equipment along the weld track.

[0008] Preferably, the edge computing node module includes: A data receiving unit, configured to receive multi-channel data inputs of the sensor array; A preprocessing unit, configured to perform filtering, normalization, and time series reconstruction; A feature generation unit, configured to construct a set of feature vectors including weight relationships between nodes.

[0009] Preferably, the spatio-temporal graph convolutional network module includes: A graph structure construction unit, configured to define edge weights in the graph according to the physical distance or functional coupling relationship between multi-modal sensor nodes; A graph convolution operation unit, configured to perform graph convolution calculations on the input features and output an updated feature map; A feature fusion unit for fusing multi-layer convolution results to output a spatio-temporal feature vector.

[0010] Preferably, the hierarchical reinforcement learning controller module includes: A bottom-layer controller unit for receiving the spatio-temporal feature vector and outputting a basic control action; A meta-controller unit for optimizing the policy network parameters and reward function structure of the bottom-layer controller unit according to the feedback result; A policy selection unit for selecting an output that satisfies the optimal condition from the set of control actions.

[0011] Preferably, the actuator module includes: A rotational speed adjustment unit for controlling the rotational speed of the welding head; A pressure loading unit for adjusting the normal force of the welding head on the workpiece surface; A displacement driving unit for adjusting the translational speed and trajectory displacement of the welding head.

[0012] Preferably, the digital twin verification module includes: A physical modeling unit for constructing a welding physical model that couples the thermal field and the stress field; A simulation execution unit for running a physical simulation based on the control instruction; A verification feedback unit for calculating the difference value between the simulation result and the actual welding data and outputting it to the controller module.

[0013] Preferably, the cloud model optimization module includes: A feedback receiving unit for receiving feedback information from the edge computing node and the digital twin verification module; A control policy update unit for updating the model parameters in the hierarchical reinforcement learning controller module; A policy synchronization unit for synchronizing the updated control policy to the edge controller.

[0014] Preferably, the calculation method of the graph convolution operation unit satisfies the following expression: ; Where: represents the node feature matrix of the th layer; represents the node feature matrix of the th layer; represents the number of convolution kernels; represents the th layer, the th convolution kernel corresponding trainable weight matrix; represents the The normalized graph Laplacian matrix corresponding to a convolutional kernel is used to describe the connection weight relationship between nodes in the graph; ReLU(·) represents the rectified linear unit activation function, which is used to introduce non-linear processing.

[0015] A control method for on-line monitoring of friction stir welding process includes the following steps: S1. Collect temperature, pressure and displacement data during the welding process through a multi-modal sensor array module; S2. Receive and preprocess the data through an edge computing node module to construct node features; S3. Input the node features into a spatio-temporal graph convolutional network module to construct a graph structure and extract spatio-temporal feature vectors; S4. Receive the spatio-temporal feature vectors through a hierarchical reinforcement learning controller module and output welding parameter control instructions; S5. Input the control instructions into an actuator module to realize the adjustment of the rotation speed, pressure and displacement of the welding equipment; S6. Input the control instructions into a digital twin verification module at the same time for welding process simulation; S7. Feed back the simulation results to the hierarchical reinforcement learning controller module for policy optimization; S8. Send the welding data and simulation feedback to a cloud model optimization module to update the control strategy; S9. Resynchronize the updated strategy to the hierarchical reinforcement learning controller module to form a closed-loop control.

[0016] In summary, the present application includes at least one of the following beneficial technical effects: 1. The present invention adopts an adaptive control technology for the welding process based on reinforcement learning, achieving the technical effect of adjusting welding parameters in real time to optimize welding quality. Compared with the control methods based on fixed parameter settings in the prior art, the present invention can dynamically adjust the control strategy according to real-time feedback, avoiding the problem of unstable welding quality caused by working condition changes in traditional methods.

[0017] 2. The present invention introduces a digital twin verification module. Through the comparison between real-time simulation and the actual welding process, it can monitor and verify the changes in the thermal field and stress field during the welding process in real time. Compared with traditional welding monitoring systems, the present invention provides higher accuracy and reliability. Especially in complex welding environments, it effectively solves the deficiency that traditional technologies cannot accurately capture process changes.

[0018] 3. The cloud model optimization module of the present invention collects feedback information from different devices and conducts remote training and optimization, making the update of the control strategy more intelligent and efficient. Compared with the optimization method that relies on manual intervention and adjustment in the prior art, the present invention realizes full-automatic and continuous optimization, greatly reducing the labor cost and improving the adaptability and self-optimization ability of the system at the same time.

[0019] 4. The present invention adopts a control strategy update method based on KL divergence and reward function, optimizing the convergence speed and stability of the policy network. Compared with the method that simply relies on empirical rules or manual parameter adjustment in the prior art, the present invention introduces meta-learning and regularization strategies, enabling the controller to quickly adapt and operate stably in a complex and changing welding environment, effectively improving the accuracy and efficiency of the welding process. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of the system architecture of the present invention; Figure 2 It is a schematic diagram of the architecture of the multi-modal sensor array module of the present invention; Figure 3 It is a schematic diagram of the architecture of the edge computing node module of the present invention; Figure 4 It is a schematic diagram of the architecture of the spatio-temporal graph convolutional network module of the present invention; Figure 5 It is a schematic diagram of the architecture of the hierarchical reinforcement learning controller module of the present invention; Figure 6 It is a schematic diagram of the architecture of the actuator module of the present invention; Figure 7 It is a schematic diagram of the architecture of the digital twin verification module of the present invention; Figure 8 It is a schematic diagram of the architecture of the cloud model optimization module of the present invention; Figure 9 It is a schematic diagram of the method flow of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0022] Please refer to the attached Figure 1 - attached Figure 8 , the embodiments of the present invention provide an on-line monitoring system for the friction stir welding process, including: Multimodal sensor array module, used to collect multiple physical parameters during welding, including temperature, pressure and displacement; Specifically, in order to achieve high-precision real-time acquisition of key physical quantities in the friction stir welding process and ensure the timeliness and data quality of the subsequent data processing module, it is necessary to deploy sensor units with multi-source perception capabilities in the welding operation area. As the first-stage functional component of the system of the present invention, the multimodal sensor array module is not only responsible for completing the acquisition of physical quantities, but also for providing the initial information structure of space and time for the data structure construction of the edge node.

[0023] In this embodiment, the multimodal sensor array module is used to synchronously collect multiple physical parameter information during the welding process, including but not limited to key process variables such as temperature, pressure and displacement.

[0024] Generally, various sensors are distributed in arrays along the welding path and its adjacent areas to ensure spatial coverage and redundant sampling, and reduce the impact of single-point failures on system robustness.

[0025] In a possible implementation, the temperature sensing unit may be arranged at the front and rear ends of the welding track using a thermocouple array or an infrared thermal imager assembly to record the temperature rise behavior caused by plastic deformation and friction heat during welding. In order to ensure data stability, in some embodiments, the thermocouple type may be a K-type thermocouple with a response time better than 0.1 seconds and a sampling frequency of not less than 1kHz.

[0026] Specifically, the pressure sensing unit is usually arranged at the contact interface between the stirring needle and the workpiece, using a micro piezoelectric film or strain gauge array to obtain the contact force changes under different welding conditions. For example, at high speeds, the pressure curve shows obvious periodic fluctuation characteristics, which will be subsequently converted into key node characteristics of the space-time graph structure.

[0027] The displacement sensing unit is used to capture the dynamic displacement behavior of the welding head along the weld track, and a laser displacement sensor or an inductive encoder can be used. In some embodiments, by configuring a linear slide system and combining it with a high-precision encoder, the system can achieve a displacement detection accuracy of ±0.01mm and a sampling frequency of 2kHz.

[0028] In this embodiment, in order to adapt to the requirements of graph structure construction, the sensor nodes are recorded as a set: ; in: Represents the total number of sensor nodes, each node Corresponding to a certain layout position of a type of physical quantity sensing unit.

[0029] As an option, to enhance the adaptability and diversity of the system, the sensor array can switch the sensing frequency or adaptively sample the area according to different working conditions. For example, when the system detects that the welding speed increases by more than the set threshold, the temperature sampling frequency automatically increases by 20% to match the dynamic thermal change frequency.

[0030] It should be noted that this module is not limited to the acquisition of raw physical signals, but also includes underlying support structures such as front-end interface design, anti-interference filtering circuit integration, and timing synchronization mechanism, which are used to ensure the time consistency and stability of cross-source data acquisition.

[0031] The edge computing node module is used to receive the output data of the multi-modal sensor array module and perform preprocessing to generate spatio-temporal features of the sensor data; Specifically, in the online monitoring system for friction stir welding process, the edge computing node module plays a key role in real-time data processing and feature extraction. It is responsible for receiving data from the multi-modal sensor array module, performing preliminary preprocessing on it, and generating spatio-temporal features. This module is responsible for local data processing and real-time feedback at the front end of the data stream, ensuring that the system can achieve fast response without relying on cloud processing, and improving the real-time performance and robustness of the system.

[0032] In this embodiment, the edge computing node module consists of a data receiving unit, a preprocessing unit, and a feature generating unit. The role of this module is to effectively process the data output by the sensor and provide valuable feature inputs for the subsequent spatio-temporal graph convolution network module.

[0033] Generally, the data receiving unit is responsible for accessing the multiple signals of the sensor array module and managing the data stream, ensuring that the sensing data of different physical quantities can be synchronously collected according to the set time window. To improve data transmission efficiency and processing speed, the data receiving unit can use a high-performance data acquisition card, which supports multi-channel input and high-frequency signal transmission. In some embodiments, the data transmission rate can reach 100 Mbps, meeting the high-frequency sampling requirements during the welding process.

[0034] Specifically, the preprocessing unit is mainly used to perform necessary filtering, denoising, normalization, and timing reconstruction on the received raw data. Through this process, the system can eliminate the influence of environmental interference on the data and ensure that the data has a certain accuracy and consistency when transmitted to the subsequent module. For example, for temperature sensing data, the preprocessing unit can use a mean filtering algorithm to remove high-frequency noise in the signal to reduce external interference. In addition, the normalization operation can ensure the dimensional consistency of data between different sensors, facilitating subsequent feature fusion.

[0035] In some embodiments, the normalization operation can be performed in the following ways: ; Wherein: is the original signal of the sensor; is the mean value of the signal; is the standard deviation; is the signal after normalization.

[0036] This process can ensure the comparability of different sensor data for subsequent processing modules under the same dimension.

[0037] As an option, the feature generation unit converts the preprocessed data into spatio-temporal feature vectors and constructs the relationship weights between nodes. In this process, the feature generation unit calculates the weight parameters between nodes according to the coupling relationship between sensor nodes (such as physical location or functional correlation) to construct a graph structure. Through this structure, the system can provide features based on the relative space and temporal dynamics of sensor nodes for the graph convolutional network.

[0038] In some embodiments, the relationship weight between nodes can be defined by calculating the Euclidean distance between nodes, for example: ; Wherein, and respectively represent the th and th sensor node positions or feature vectors, is a parameter for adjusting the distance sensitivity, usually determined through experiments.

[0039] In this weight calculation method, the similarity between nodes is quantified through distance metrics, so that the generated graph structure can truly reflect the interaction relationship between physical fields. This graph structure will play a crucial role in the subsequent spatio-temporal graph convolutional network.

[0040] Through the refined processing of the original data by the edge computing node module, the system can quickly generate features suitable for subsequent spatio-temporal graph convolutional processing without fully relying on cloud resources. These processed features not only ensure the data transmission efficiency and computational stability, but also provide the necessary guarantee for the real-time performance and self-adaptability of the entire system.

[0041] The spatio-temporal graph convolutional network module is used to construct a graph structure based on the spatio-temporal features generated by the edge computing node module and extract the spatio-temporal correlation features between multiple sensor nodes through graph convolutional operations; Specifically, in the in - process monitoring system for friction stir welding of the present invention, the spatio - temporal graph convolutional network module plays a key role. It is responsible for constructing a graph structure based on the spatio - temporal features generated by the edge computing node module, and extracting the spatio - temporal correlation features between sensor nodes through graph convolution operations. Through this process, the system can capture the spatio - temporal variation rules during the welding process, thus providing strong input support for the subsequent hierarchical reinforcement learning control module.

[0042] In this embodiment, the spatio - temporal graph convolutional network module mainly includes a graph structure construction unit, a graph convolution operation unit, and a feature fusion unit. The graph structure construction unit constructs a graph structure with spatio - temporal attributes according to the physical distance or functional coupling relationship between sensor nodes. The graph convolution operation unit calculates the spatio - temporal features based on graph convolution operations and updates the feature map. The feature fusion unit is used to fuse the outputs of multiple convolutional layers to generate the final spatio - temporal feature vector.

[0043] Generally, the graph structure construction unit determines the connection weights between nodes by analyzing information such as the physical adjacency degree and sensor response synchronization between sensor nodes. These weights are used to define the relationships between nodes in the graph, ensuring that the graph convolutional network can reflect the true coupling between sensor nodes. Through the graph structure generated by the weight calculation of the edge computing node module, the relationships between nodes can be modeled according to their physical or functional correlations.

[0044] Specifically, the graph convolution operation unit processes the input node feature matrix using classical graph convolution operations. The graph convolution operation can aggregate information from adjacent nodes and weight it through a convolution kernel. The general form of the graph convolution operation can be expressed as: ; Where: represents the node feature matrix of the th layer; represents the node feature matrix of the th layer; represents the number of convolution kernels; represents the trainable weight matrix corresponding to the th convolution kernel in the th convolution kernel; represents the normalized graph Laplacian matrix corresponding to the th convolution kernel in the graph structure, which is used to describe the connection weight relationship between nodes in the graph; ReLU(·) represents the rectified linear unit activation function, which is used to introduce non - linear processing.

[0045] In some embodiments, the graph convolution operation unit extracts spatio-temporal features at different scales through a multi-layer graph convolution structure. These features can capture the spatio-temporal correlations between different sensor nodes during the welding process, such as the spatial distribution and temporal evolution of physical quantities like temperature changes and pressure fluctuations.

[0046] As an option, the feature fusion unit fuses the output features of the multi-layer graph convolution. Specifically, the spatio-temporal features extracted by multiple graph convolution layers can be fused through weighted averaging, concatenation, or other methods to obtain a spatio-temporal feature vector containing more abundant information. This feature vector can better reflect the complex spatio-temporal dependence relationships during the welding process and provide a more accurate and comprehensive input for the subsequent hierarchical reinforcement learning controller module.

[0047] For example, in certain embodiments, the feature fusion can adopt the form of weighted summation, expressed as: ; where: is the final spatio-temporal feature vector; is the weight of each layer of features; is the total number of convolutional layers; is the output feature of the

[0048] The spatio-temporal graph convolution network module can not only process static sensor data but also consider the temporal characteristics of the data. Through graph convolution operations, the system can effectively capture the dynamic coupling relationships between different regions in space and time. For example, during the welding process, the responses between sensor nodes may have strong temporal dependencies, and this characteristic can be fully reflected through the design of the graph convolution network.

[0049] The hierarchical reinforcement learning controller module is used to generate welding parameter control instructions based on spatio-temporal correlation features, and the control instructions include rotational speed adjustment amount, pressure adjustment amount, and displacement adjustment amount; Specifically, in the online monitoring system for friction stir welding process of the present invention, the hierarchical reinforcement learning controller module plays an important role in realizing dynamic optimization and welding process control. Based on the spatio-temporal feature vector generated by the spatio-temporal graph convolution network module, this module generates key control instructions during the welding process, such as rotational speed adjustment amount, pressure adjustment amount, and displacement adjustment amount, thereby effectively adjusting the welding process parameters and realizing the optimization and fine control of the welding process.

[0050] In this embodiment, the hierarchical reinforcement learning controller module consists of a bottom-layer controller unit, a meta-controller unit, and a policy selection unit. The bottom-layer controller unit is responsible for generating basic control actions based on the input spatio-temporal feature vector. The meta-controller unit optimizes the policy of the bottom-layer controller unit according to the feedback results. The policy selection unit is responsible for selecting the output that meets the optimal conditions from the set of control actions and adjusting the welding parameters.

[0051] Generally, the bottom-layer controller unit adopts the TD3 (Twin Delayed Deep Deterministic Policy Gradient) algorithm in the reinforcement learning algorithm. This algorithm generates precise control instructions by optimizing the control policy. The control instructions include the rotational speed adjustment , pressure adjustment , and displacement adjustment . These adjustment amounts will directly act on the welding equipment, thus realizing the dynamic adjustment of the welding process.

[0052] Specifically, the state space of the bottom-layer controller includes spatio-temporal features , feature gradients , and time steps . These information provide the spatio-temporal change characteristics of different physical quantities during the welding process, ensuring that the bottom-layer controller can make reasonable control actions based on real-time feedback. The output of the control action ; where: , , and are the adjustment amounts of rotational speed, pressure, and displacement respectively. The goal of the bottom-layer controller is to optimize the quality of the welding process by adjusting these parameters.

[0053] As an option, the meta-controller unit further optimizes the control effect during the welding process by adjusting the policy network parameters and reward function structure of the bottom-layer controller. The meta-controller optimizes the reward function through the exploration and exploitation strategy. The specific update process can be expressed as: ; where: is the control policy optimized by the meta-learning process; is the basic policy; is the empirical data set; is the Kullback-Leibler divergence, which measures the difference between the new policy and the basic policy; is the reward function; is the regularization coefficient.

[0054] In this way, the meta - controller can continuously optimize the control strategy according to the feedback information, thereby improving the stability and quality of the welding process.

[0055] In some embodiments, the policy networks of the underlying controller and the meta - controller adopt deep neural network models to better adapt to the complex non - linear welding process. The output of the neural network model is processed by an activation function (such as ReLU) to introduce non - linear features, ensuring that the model can more accurately map the relationship between input features and control actions.

[0056] Specifically, the policy selection unit selects a set of control actions that meet the optimal conditions based on the outputs of the underlying controller and the meta - controller. In this way, the system can dynamically adjust the control parameters of the welding equipment according to the real - time welding feedback. The adjustment of the control parameters is based on the real - time input of spatio - temporal feature vectors to ensure that various parameters in the welding process can be optimized collaboratively, avoiding over - adjustment or unreasonable control strategies.

[0057] For example, in some embodiments, the policy selection unit selects control actions by maximizing a certain form of reward function: ; where: is the optimal control action; is the set of control actions; is the reward function based on the current state and the control action .

[0058] The actuator module is used to receive control instructions and output corresponding welding control signals; Specifically, in the online monitoring system for friction stir welding process of the present invention, the actuator module is responsible for adjusting various operating parameters of the welding equipment according to the control instructions output by the hierarchical reinforcement learning controller module, specifically including parameters such as rotational speed, pressure, and displacement. Through precise parameter adjustment, the actuator module can ensure the precise control of the welding process and achieve dynamic adjustment, thereby optimizing the welding quality and stability.

[0059] In this embodiment, the actuator module includes a rotational speed adjustment unit, a pressure loading unit, and a displacement driving unit. Each unit can adjust the corresponding parameters of the welding equipment in real - time according to the input of control instructions to meet the real - time requirements in the welding process.

[0060] Generally, the rotational speed adjustment unit is used to adjust the rotational speed of the welding head , ensure its adaptation to other physical quantities (such as pressure and displacement) during the welding process, thereby maintaining the stability of the welding process. Specifically, the rotation speed adjustment unit can drive the rotation of the welding head through a motor, and the adjustment amount of the rotation speed comes from the output of the hierarchical reinforcement learning controller module.

[0061] In a possible implementation, the rotation speed adjustment unit can adopt a closed-loop control strategy and dynamically adjust the rotation speed of the welding head according to the real-time collected temperature, pressure, and displacement data. Through real-time feedback, the system can automatically optimize the rotation speed according to the changes during the welding process, thereby ensuring the continuity and stability of the welding process.

[0062] The pressure loading unit is used to adjust the normal force of the welding head on the workpiece surface. This unit controls the force application of the welding head through a hydraulic or electric drive device to ensure that an appropriate pressure is maintained throughout the welding process to achieve a sufficient stirring effect. The adjustment amount of the pressure is calculated by the hierarchical reinforcement learning controller module based on spatio-temporal features to ensure that an appropriate pressure can be provided at different welding stages to achieve the best welding quality.

[0063] Specifically, the adjustment range and accuracy of the pressure loading unit are determined according to the welding process and material characteristics used. In some embodiments, the pressure adjustment can be performed in real time based on the contact state between the welding head and the workpiece to ensure that the contact pressure is always maintained within the ideal range, thereby avoiding welding defects.

[0064] The displacement drive unit is used to control the translation speed and trajectory of the welding head. This unit realizes the precise positioning and movement of the welding head through an electric or hydraulic drive device to ensure the accuracy of the welding path. The adjustment amount of the displacement is output by the controller and dynamically adjusted in combination with real-time feedback. For example, the welding equipment may adjust the translation speed and trajectory of the welding head along the weld according to the geometric shape of the workpiece and the weld requirements.

[0065] As an option, the displacement drive unit can also be combined with a position sensor or encoder for feedback control to ensure that the welding head always moves along the predetermined path throughout the welding process to avoid welding quality problems caused by path deviation. In some embodiments, the position sensor is used in combination with the feedback system to ensure the real-time precise control of the welding head position.

[0066] Specifically, the adjustment of the displacement drive unit also needs to consider factors such as the thermal expansion effect, pressure change, and material characteristics between the welding head and the workpiece during the welding process. Therefore, the adjustment of this unit not only requires a fast response but also needs to have a certain adaptability to cope with different working conditions during the welding process.

[0067] A digital twin verification module, which is used to perform physical simulation on the welding process based on control signals and feed back the simulation results to the hierarchical reinforcement learning controller module; Specifically, in the on-line monitoring system for friction stir welding process of the present invention, the digital twin verification module provides feedback support for the control and optimization of the welding process by real-time simulating and verifying the welding process. Through physical modeling and simulation technology, this module can real-time simulate physical phenomena such as the thermal field and stress field during the welding process, ensure that the welding process meets the quality requirements, and provide adjustment instructions to optimize the welding quality when abnormalities occur during the actual welding process.

[0068] In this embodiment, the digital twin verification module mainly includes a physical modeling unit, a simulation execution unit, and a verification feedback unit. The physical modeling unit is responsible for constructing a coupled model of the thermal field and stress field of the welding process according to real-time data; the simulation execution unit performs real-time simulation calculations based on this model; and the verification feedback unit calculates the verification error by comparing the simulation results with the actual welding data and transmits the feedback information to the control system.

[0069] Generally, the physical modeling unit first constructs a thermodynamic model of the welding process based on data such as temperature, pressure, and displacement provided by the multi-modal sensor array module. In some embodiments, the change in temperature distribution can be described by the heat conduction equation: ; Where: is the temperature; is the thermal diffusivity; is the heat source; is the material density; is the specific heat capacity; is the thermo-mechanical coupling coefficient; is the strain generated during the welding process.

[0070] Through this equation, the physical modeling unit can simulate the change in temperature field generated during the welding process.

[0071] Specifically, the physical modeling unit not only focuses on the temperature distribution, but also considers the stress field generated during the welding process. The modeling of the stress field can be described by the elasticity equation: ; Where: is the stress; is the elastic modulus of the material; is the strain; is the initial strain of the material.

[0072] In this way, the digital twin verification module can comprehensively reflect the coupling effect of heat and force during the welding process.

[0073] In some embodiments, the physical modeling unit may also combine changes in the welding process, such as welding speed, welding pressure, etc., to further optimize the simulation accuracy of the thermal field and stress field. By accurately simulating the physical phenomena during the welding process, the physical modeling unit provides reliable input data for subsequent simulation calculations.

[0074] The simulation execution unit is responsible for performing real-time simulation based on the thermo-mechanical coupling model generated by the physical modeling unit. This unit can quickly simulate the evolution process of factors such as temperature distribution and stress changes during the welding process. In some embodiments, the simulation execution unit may use the finite element method (FEM) for numerical simulation calculations of the welding process. Through this calculation, the simulation execution unit can evaluate the thermal stress distribution and its change trend during the welding process in real time.

[0075] In certain embodiments, the simulation execution unit also has an adaptive ability and can adjust simulation parameters according to different welding conditions. For example, during the welding process, if the welding speed changes, the simulation execution unit can adjust the heat source model and stress distribution model in real time to reflect the actual changes during the welding process.

[0076] The verification and feedback unit compares the simulation results of the simulation execution unit with the data from the sensor array and calculates the verification error. The calculation of the verification error can be expressed as: ; where: is the data actually measured by the sensor array; is the data obtained from the simulation; is the total number of data points; is the regularization coefficient.

[0077] Through this error calculation, the system can evaluate the difference between the simulation results and the actual welding process, thereby providing feedback information for optimizing the welding process.

[0078] As an option, when calculating the error, the verification and feedback unit may appropriately adjust the tolerance of the error according to different stages of welding (such as the preheating stage, welding stage, and cooling stage) to ensure that the system can provide timely and accurate feedback on the welding process for each stage.

[0079] The cloud model optimization module is used to receive the feedback information from the edge computing node module and the digital twin verification module, generate an optimization control strategy, and update it to the hierarchical reinforcement learning controller module.

[0080] Specifically, in the online monitoring system for friction stir welding process proposed by the present invention, the cloud model optimization module is responsible for continuously learning and optimizing the control strategy from a global perspective. This module receives feedback information from the edge computing node module and the digital twin verification module, and through the remote training and parameter iteration of the reinforcement learning model, realizes the high-frequency remote update of the policy network in the hierarchical reinforcement learning controller module, and synchronizes the optimized model to the edge side to form an effective closed loop.

[0081] In this embodiment, the cloud model optimization module mainly includes a feedback receiving unit, a control strategy updating unit, and a policy synchronization unit.

[0082] Generally, the feedback receiving unit continuously listens to the data streams transmitted from the edge computing node and the digital twin verification module. The received data content includes, but is not limited to: the spatio-temporal feature extraction results during the welding process, the error feedback between the control instruction and the actual execution, and the prediction deviation generated based on the simulation model. These data constitute the necessary training samples in the cloud optimization process to guide the update process of the policy network.

[0083] Specifically, the control strategy updating unit performs the optimization task of the control strategy based on the collected feedback data. In a possible implementation manner, the update of the control strategy adopts a policy regularization method based on KL divergence, and its goal is to balance the model update speed and the policy stability. The update objective function is defined as follows: ; Where: is the optimized policy function; is the base policy at the previous update; represents the Kullback-Leibler divergence between the two; is the state under the expected reward value; is the regularization factor; is the experience replay data set.

[0084] As an option, a trust region constraint can be introduced during the policy update process to prevent the policy parameters from changing drastically in the short term. For example, in some embodiments, the maximum change range of the policy parameters can be set to satisfy: ; Where: and respectively represent the policy network parameters before and after the update.

[0085] Based on the updated control strategy, the policy synchronization unit pushes the optimized model parameters to the hierarchical reinforcement learning controller module at the edge. Considering network latency and model stability during the synchronization process, an asynchronous incremental synchronization strategy is generally adopted to avoid edge control oscillations caused by inconsistent policies.

[0086] In some embodiments, to improve the edge-cloud collaboration efficiency, a parameter hash identification mechanism can be used to manage the model version for each round of synchronization. When the optimization result of the cloud model meets the preset gain threshold, the replacement operation of the edge node model will be triggered to ensure that the policy update has practical benefits.

[0087] In addition, to improve the system adaptability, the cloud model optimization module further includes a model verification sub-module for online simulation verification before model update. By substituting the candidate control strategy into the digital twin model to evaluate its performance under typical working conditions, verifying whether its stability and performance are superior to the original strategy, and then deciding whether to enter the policy synchronization stage.

[0088] A control method for online monitoring of the friction stir welding process described below can be corresponded and referred to with a system for online monitoring of the friction stir welding process described above.

[0089] Please refer to the attached Figure 9 , a control method for online monitoring of the friction stir welding process, comprising the following steps: S1. Collect temperature, pressure, and displacement data during the welding process through a multi-modal sensor array module; S2. Receive and preprocess the data through an edge computing node module to construct node features; S3. Input the node features into a spatio-temporal graph convolutional network module to construct a graph structure and extract spatio-temporal feature vectors; S4. Receive the spatio-temporal feature vectors through a hierarchical reinforcement learning controller module and output welding parameter control instructions; S5. Input the control instructions into an actuator module to adjust the rotation speed, pressure, and displacement of the welding equipment; S6. Input the control instructions into a digital twin verification module for welding process simulation simultaneously; S7. Feed back the simulation results to the hierarchical reinforcement learning controller module for policy optimization; S8. Send the welding data and simulation feedback to the cloud model optimization module to update the control strategy; S9. Resynchronize the updated policy to the hierarchical reinforcement learning controller module to form a closed-loop control.

[0090] The method of this embodiment can be used to implement the above system embodiment, and its principle and technical effects are similar, which will not be elaborated here.

[0091] The embodiments of this specific implementation manner are all preferred embodiments of this application, and do not limit the protection scope of this application accordingly. The same parts are represented by the same reference numerals. Therefore, all equivalent changes made according to the structure, shape, and principle of this application shall be covered within the protection scope of this application.

Claims

1. An on-line monitoring system for friction stir welding process, characterized in that Including: A multi-modal sensor array module for collecting multiple physical parameters during the welding process, where the physical parameters include temperature, pressure, and displacement; An edge computing node module for receiving the output data of the multi-modal sensor array module and performing preprocessing to generate spatio-temporal features of the sensor data; A spatio-temporal graph convolutional network module for constructing a graph structure based on the spatio-temporal features generated by the edge computing node module and extracting spatio-temporal correlation features between multiple sensor nodes through graph convolution operations; A hierarchical reinforcement learning controller module for generating welding parameter control instructions based on the spatio-temporal correlation features, where the control instructions include rotational speed adjustment amount, pressure adjustment amount, and displacement adjustment amount; An actuator module for receiving the control instructions and outputting corresponding welding control signals; A digital twin verification module for physically simulating the welding process based on the control signals and feeding back the simulation results to the hierarchical reinforcement learning controller module; A cloud model optimization module for receiving feedback information from the edge computing node module and the digital twin verification module, generating an optimized control strategy, and updating it to the hierarchical reinforcement learning controller module.

2. The on-line monitoring system for friction stir welding process according to claim 1, characterized in that, The multi-modal sensor array module includes: A temperature sensing unit for acquiring real-time temperature data of different nodes in the welding area; A pressure sensing unit for acquiring the contact pressure between the stirring pin or stirring head and the workpiece; A displacement sensing unit for detecting the displacement change of the welding equipment along the weld track.

3. The on-line monitoring system for friction stir welding process according to claim 1, characterized in that, The edge computing node module includes: A data receiving unit for receiving multi-channel data inputs of the sensor array; A preprocessing unit for performing filtering, normalization, and time series reconstruction; A feature generation unit for constructing a set of feature vectors including weight relationships between nodes.

4. The on-line monitoring system for friction stir welding process according to claim 1, characterized in that, The spatio-temporal graph convolutional network module includes: A graph structure construction unit for defining edge weights in the graph according to the physical distance or functional coupling relationship between multi-modal sensor nodes; A graph convolution operation unit for performing graph convolution calculations on the input features and outputting an updated feature map; A feature fusion unit for fusing multi-layer convolution results to output a spatio-temporal feature vector.

5. The on-line monitoring system for friction stir welding process according to claim 1, characterized in that, The hierarchical reinforcement learning controller module includes: A bottom-layer controller unit for receiving the spatio-temporal feature vector and outputting a basic control action; A meta-controller unit for optimizing the policy network parameters and reward function structure of the bottom-layer controller unit according to the feedback results; A policy selection unit for selecting the output that meets the optimal conditions from the set of control actions.

6. The on-line monitoring system for friction stir welding process according to claim 1, characterized in that The actuator module includes: A rotational speed adjustment unit for controlling the rotational speed of the welding head; A pressure loading unit for adjusting the normal force of the welding head on the workpiece surface; A displacement driving unit for adjusting the translational speed and trajectory displacement of the welding head.

7. The on-line monitoring system for friction stir welding process according to claim 1, characterized in that, The digital twin verification module includes: A physical modeling unit for constructing a welding physical model that couples the thermal field and stress field; A simulation execution unit for running physical simulations based on the control instructions; A verification feedback unit for calculating the difference value between the simulation results and the actual welding data and outputting it to the controller module.

8. The on-line monitoring system for friction stir welding process according to claim 1, characterized in that The cloud model optimization module includes: A feedback receiving unit, configured to receive feedback information from an edge computing node and a digital twin verification module; A control strategy updating unit, configured to update the model parameters in the hierarchical reinforcement learning controller module; A policy synchronization unit, configured to synchronize the updated control strategy to the edge controller.

9. The on-line monitoring system for friction stir welding process according to claim 4, characterized in that, The calculation method of the graph convolution operation unit satisfies the following expression: ； Wherein: represents the node feature matrix of the th layer; represents the node feature matrix of the th layer; represents the number of convolutional kernels; represents the trainable weight matrix corresponding to the th convolutional kernel in the th layer; represents the normalized graph Laplacian matrix corresponding to the th convolutional kernel in the graph structure, which is used to describe the connection weight relationship between nodes in the graph; ReLU(·) represents the rectified linear activation function, which is used to introduce non-linear processing.

10. A control method for on-line monitoring during friction stir welding, which is applied to an on-line monitoring system for friction stir welding process according to any one of claims 1-9, characterized in that, Including the following steps: S1. Collect temperature, pressure, and displacement data during the welding process through a multi-modal sensor array module; S2. Receive and preprocess the data through an edge computing node module to construct node features; S3. Input the node features into a spatio-temporal graph convolution network module to construct a graph structure and extract spatio-temporal feature vectors; S4. Receive the spatio-temporal feature vectors through a hierarchical reinforcement learning controller module and output welding parameter control instructions; S5. Input the control instructions into an actuator module to adjust the rotation speed, pressure, and displacement of the welding equipment; S6. Input the control instructions into a digital twin verification module simultaneously for welding process simulation; S7. Feed back the simulation results to the hierarchical reinforcement learning controller module for policy optimization; S8. Send the welding data and simulation feedback to a cloud model optimization module to update the control strategy; S9. Resynchronize the updated strategy to the hierarchical reinforcement learning controller module to form a closed-loop control.

Images