A control method and system for thermal-assisted flow drill riveting based on digital twin

Abstract

The present invention discloses a control method and system for thermal-assisted flow drill riveting based on digital twin. The method includes: collecting temperature, pressure, displacement and surface topography data; constructing a comprehensive temperature-pressure-topography index through weighted fusion; performing real-time processing through an FPGA chip and a lightweight AI model; generating a Pareto front solution set by combining the Monte Carlo-NSGA-II hybrid algorithm; constructing a thermo-mechanical-phase change coupling unified constitutive model for multi-physical field inverse optimization; realizing dynamic update of process parameters through a digital twin case library; and dynamically modulating PID parameters by using deep reinforcement learning. The system includes a multi-source data acquisition terminal, an intelligent decision-making module and a dynamic control module. The present invention solves the problems of parameter lag and poor adaptability of traditional methods and can efficiently and accurately complete processing tasks.

Description

Technical Field

[0001] The present invention relates to the field of precision machining control technology, and specifically to a thermal-assisted flow drill riveting control method and system based on digital twin, which is applicable to the connection machining of metal parts. Background Art

[0002] Traditional thermal-assisted flow drill riveting faces many challenges in the actual production and machining process. First of all, sensor data is often independent of each other, lacking multi-modal fusion judgment, resulting in the inability to accurately and real-time detect abnormal situations in the production and machining process. Secondly, empirical parameters are adopted in the control algorithm, but when facing complex machining material structures (such as the laminated structure of carbon fiber material and titanium alloy), its adaptability significantly decreases. Moreover, although the cloud centralized operation mode provides powerful data processing capabilities, the resulting lag of control instructions cannot be ignored, and the lag phenomenon caused during high-speed machining reduces the machining stability. Finally, the process highly depends on the trial-and-error method, which can only be achieved at a relatively high cost and a long adjustment cycle, and it is difficult to meet the requirements of modern manufacturing for efficient and flexible production. Summary of the Invention

[0003] In order to overcome the deficiencies of the prior art, the purpose of the present invention is to provide a thermal-assisted flow drill riveting control method and system based on digital twin. The present invention can efficiently and accurately complete machining tasks, and at the same time has good maintainability and scalability, providing a new technical solution for the field of intelligent manufacturing.

[0004] The present invention is realized through the following technical solutions: A thermal-assisted flow drill riveting control method based on digital twin, including the following steps:

[0005] Step 1: Collect the machining data of the thermal-assisted flow drill riveting equipment, including temperature, pressure, displacement, and surface topography image data in the machining area;

[0006] Step 2: Preprocess the temperature, pressure, displacement, and surface topography image data, perform standardization processing on the preprocessed temperature, pressure, and surface topography image data, and then perform weighted fusion on the standardized temperature, pressure, and surface topography image data to obtain temperature-pressure-topography weighted fusion data;

[0007] Step 3: Real-time process the standardized data and temperature-pressure-topography weighted fusion data through the FPGA chip and the lightweight AI model, dynamically optimize the process parameters, and synchronously generate temperature prediction and defect detection results;

[0008] Step 4: Based on the process parameters, temperature prediction, and defect detection results obtained in Step 3, further optimize through the Monte Carlo-NSGA-II hybrid algorithm to generate candidate process parameters, screen the Pareto front solution set, and transmit it to the digital twin case library in real time;

[0009] Step 5: Construct a unified constitutive model of thermal-mechanical-phase transformation coupling, and use the multi-physics field coupling numerical analysis method to perform reverse optimization on the Pareto front solution set. Re-optimize the Pareto front solution set through the digital twin case library and then enter the lightweight AI model. Input the real-time monitored processing data back into the simulation model for closed-loop verification;

[0010] Step 6: The verified process parameters are classified into the digital twin case library, and the digital twin case library pushes a new Pareto front solution set to the lightweight AI model;

[0011] Step 7: Dynamically modulate the PID parameters through the deep reinforcement learning model. Combine the verified process parameters and control the thermal-assisted flow drill riveting equipment through the PID control circuit.

[0012] Further preferably, the temperature-pressure-topography weighted fusion data , is the normalized temperature data, F * is the normalized pressure data, I * is the binary surface topography image data.

[0013] Further preferably, in Step 7, when dynamically modulating the PID parameters through the deep reinforcement learning model, the PID parameters are corrected in real time in combination with the tensile strength-energy consumption balance reward function, and hardware-level protection is triggered when exceeding the limit. The control result of the thermal-assisted flow drill riveting equipment is fed back to the digital twin case library.

[0014] Further preferably, the state space of the deep reinforcement learning model (DRL) includes real-time processing data and digital twin inputs, and the action space is divided into continuous actions and discrete actions.

[0015] Further preferably, the tensile strength-energy consumption balance reward function is: R = 10×S 达标 + 5×E 降低 − 20×T 超限 − 15×V 超标 , where R is the reward value, S 达标 represents that the connection strength meets the standard, E 降低 represents the reduction of energy consumption, T 超限 represents the temperature overrun flag; V 超标 represents the vibration amplitude overrun flag.

[0016] Further preferably, the process of Step 3 is as follows:

[0017] Step b1: Build a lightweight AI model architecture;

[0018] Step b2: Split the lightweight AI model layer into parallel pipeline tasks, match the FPGA chip, and use the FPGA chip cache to pre-store data;

[0019] The loss function is obtained by summing the mean square error and the structural similarity according to weights;

[0020] Step b4: Dynamically optimize the process parameters;

[0021] Step b5: Generate temperature prediction and defect detection results.

[0022] Further preferably, according to the normalized temperature data and the normalized pressure data predict the target temperature T opt ; Use the lightweight AI model to classify the defect types according to the binary surface topography image data I * output the defect coordinates, types, and severity levels, and provide real-time feedback to the parameter optimizer.

[0023] The present invention also provides a digital-twin-based thermal-assisted flow drill riveting control system, including:

[0024] A multi-source data acquisition terminal for capturing temperature, pressure, displacement, and surface topography image data of the processing area in real time;

[0025] An intelligent decision-making module, which consists of a lightweight AI model, a parameter optimizer, a digital-twin case library, and a digital-twin model library. The digital-twin model library is used to store the thermal-mechanical-phase change coupling unified constitutive model and process data through the Monte Carlo-NSGA-II hybrid algorithm;

[0026] A dynamic control module, which consists of a safety monitoring module, a PID control module, and the servo motor and heating module of the thermal-assisted flow drill riveting equipment; The PID control module consists of a PID control circuit and a deep reinforcement learning model, and dynamically modulates the PID parameters through the deep reinforcement learning model.

[0027] Further preferably, the multi-source data acquisition terminal includes an infrared thermal imager, a multi-axis force sensor, a laser displacement meter, and an industrial camera.

[0028] Compared with the existing control methods, the present invention has the following technical effects:

[0029] Collect temperature, pressure, displacement, and surface topography data through devices such as infrared thermal imagers and multi-axis force sensors. After preprocessing such as Gaussian filtering and FFT transformation, weighted fusion is adopted to achieve a comprehensive perception of the processing process and improve the precision control level.

[0030] Based on FPGA chips and lightweight AI models, combined with neural architecture search and hardware co-optimization, temperature prediction and defect detection with a latency of ≤10 ms are achieved.

[0031] The Monte Carlo - NSGA-II algorithm is used to generate the Pareto front solution set (tensile strength ≥800 MPa, energy consumption ≤500 J / time), and combined with the thermo-mechanical-phase transformation coupling unified constitutive model for multi-physical field simulation to reduce the prediction error rate of the morphology in the heat affected zone.

[0032] The digital twin case library stores optimized process parameters and simulation data, shortening the process debugging cycle of new products and supporting rapid process reuse and expansion.

[0033] The deep reinforcement learning model dynamically modulates the PID parameters, combined with the tensile strength - energy consumption balance reward function, to achieve high-precision closed-loop control during the processing. Description of the Drawings

[0034] Figure 1 It is a schematic diagram of the heat-assisted flow drill riveting control system based on digital twin;

[0035] Figure 2 It is a flowchart of the heat-assisted flow drill riveting control method based on digital twin;

[0036] Figure 3 It is a circuit design diagram of the PID control module. Detailed Implementation Modes

[0037] To facilitate the understanding and implementation of the present invention by those of ordinary skill in the art, the present invention will be further described in detail and in depth below with reference to the accompanying drawings.

[0038] As Figure 1 shown, the heat-assisted flow drill riveting control system based on digital twin includes:

[0039] The multi-source data acquisition terminal includes an infrared thermal imager, a multi-axis force sensor, a laser displacement meter and an industrial camera, which are respectively responsible for capturing temperature, pressure, displacement and surface topography image data in the processing area in real time. The configured infrared thermal imager is used for temperature field monitoring, monitoring temperature at a sampling rate of 100 Hz, with a spatial resolution of 0.1 mm² / pixel, a temperature range of 400 - 700 °C, and an accuracy of ±1 °C. The multi-axis force sensor is integrated on the main shaft of the drill riveting mechanism for monitoring pressure and vibration, with a range of ±5 kN and a bandwidth of 10 kHz, and transmits data through the Ether CAT protocol. The laser displacement meter is used to measure the deformation and displacement of the workpiece surface, using triangulation technology with an accuracy of ±0.02 mm. The industrial camera has a frame rate of 120 fps and a resolution of 2048×1536 pixels, and is used to capture surface topography image data.

[0040] The intelligent decision-making module consists of a lightweight AI model, a parameter optimizer, a digital twin case library, and a digital twin model library; the digital twin model library is used to store numerical models such as the thermal-mechanical-phase change coupling unified constitutive model and the Monte Carlo-NSGA-II hybrid algorithm.

[0041] The dynamic control module consists of a safety monitoring module, a PID control module, a servo motor of the thermal-assisted flow drill riveting equipment, and a heating module. The PID control module dynamically modulates the PID parameters in combination with a deep reinforcement learning model (DRL); real-time detection of temperature and pressure overlimit events triggers an exception handling mechanism.

[0042] As shown in the Figure 2 accompanying figure, the digital twin-based thermal-assisted flow drill riveting control method proposed by the present invention adopts a modular hierarchical architecture, and divides the implementation process into a "perception-decision-execution" three-ring closed-loop system, specifically divided into: a multi-modal perception layer (steps 1-step 3), an intelligent decision-making layer (steps 4-step 6), and a dynamic control layer (step 7).

[0043] Step 1: Data collection is carried out through a multi-source data collection terminal, and the processing data of the thermal-assisted flow drill riveting equipment is collected, including temperature, pressure, displacement, and surface topography image data in the processing area;

[0044] Step 2: Preprocess the temperature, pressure, displacement, and surface topography image data, perform standardization processing on the preprocessed temperature, pressure, and surface topography image data, and then perform weighted fusion on the standardized temperature, pressure, and surface topography image data to obtain temperature-pressure-topography weighted fusion data;

[0045] Preprocessing refers to preprocessing each group of processing data through noise suppression, feature extraction, smoothing trajectories, and edge detection to obtain high-quality basic data. Preprocessing can improve the credibility of single-sensor data, and through multi-source data fusion, it realizes the comprehensive perception and precise control of the processing process, promoting the improvement of process efficiency (the debugging cycle is shortened by 40%) and the optimization of processing quality (the defect rate ≤ 1%). The specific preprocessing methods are as follows:

[0046] Temperature data: Gaussian filtering (window size: 3×3) is used to remove noise and retain effective thermal field features;

[0047] Pressure data: Fast Fourier transform (FFT) is used to extract the characteristic spectrum in the frequency band of 0.1-5 kHz to identify the vibration characteristic frequency band;

[0048] Displacement data: Kalman filtering is used to smooth the trajectory and reduce the measurement jitter error;

[0049] Surface topography image data: Perform Canny edge detection through the OpenCV library and label the defect areas.

[0050] The normalization process includes:

[0051] Normalize the temperature data: ∈[0, 1];

[0052] Normalize the pressure data: ∈[−1, 1];

[0053] Binarize the surface topography image data: I * ∈{0, 1} (defect areas are labeled as 1, others are labeled as 0).

[0054] In the formula, is the normalized temperature data, F * is the normalized pressure data, I * is the binarized surface topography image data, T max 、T min are the set maximum and minimum temperatures respectively, F max is the set maximum pressure, T is the preprocessed temperature data, and F is the preprocessed pressure data.

[0055] Weighted fusion means fusing the data after normalization processing to obtain a comprehensive index that can comprehensively characterize the dynamic characteristics of the processing process, namely the temperature-pressure-topography weighted fusion data, which is beneficial for subsequent analysis and closed-loop control. The temperature-pressure-topography weighted fusion data , is used to characterize the dynamic characteristics of the processing process.

[0056] Step 3: Perform real-time processing on the normalized data and the temperature-pressure-topography weighted fusion data FdV through the FPGA chip and the lightweight AI model, dynamically optimize the process parameters, and synchronously generate temperature prediction and defect detection results, so as to achieve intelligent drill riveting control with low latency (≤10 ms), high precision (error ±1 °C), and energy efficiency improvement (energy consumption reduced by 40%) in the subsequent steps. The specific processing steps are as follows:

[0057] Step b1: Build the lightweight AI model architecture: Based on the MobileViT model and the convolutional neural network (CNN), through optimization techniques such as neural architecture search (NAS), knowledge distillation, and channel pruning, improve the feature extraction ability of the lightweight AI model for the temperature field and defect areas.

[0058] Step b2: Hardware co-optimization: Split the lightweight AI model layers into parallel pipelined tasks, match the FPGA chip, and use the FPGA chip cache to pre-store data (the normalized data and the temperature-pressure-topography weighted fusion data FdV) to reduce external memory access latency.

[0059] Step b3: Improve the loss function: Combine the mean squared error (MSE) with the structural similarity (SSIM) to enhance the ability to retain local details in temperature prediction and balance multi-objective conflicts. The loss function is improved to:

[0060] ;

[0061] The weight allocation (0.7:0.3) is verified through ablation experiments. The 0.7 MSE weight can balance numerical accuracy and structural fidelity, reducing MSE by 20% while increasing SSIM by 15%.

[0062] Step b4: Dynamically optimize process parameters: Update the temperature-pressure-topography weighted fusion data FdV every 10 ms to respond to changes in the processing state (such as temperature fluctuations and material deformation).

[0063] Step b5: Generate temperature prediction and defect detection results: Based on the normalized temperature data and the normalized pressure data predict the target temperature T opt ; Use CNN to classify defect types (cracks, pores, deformation) based on the binary surface topography image data I * and output the defect coordinates, types, and severity levels, which are fed back to the parameter optimizer in real time.

[0064] Step 4: Based on the process parameters, temperature prediction, and defect detection results obtained in Step 3, further optimize through the Monte Carlo-NSGA-II hybrid algorithm to generate candidate process parameters (temperature, pressure, displacement, and rotational speed), screen the Pareto front solution set (tensile strength ≥ 800 MPa, energy consumption ≤ 500 J / time, a set of solutions that cannot be further optimized by adjusting parameters without sacrificing other objectives for flow drill riveting performance), and transmit it to the digital twin case library in real time through the EtherCAT protocol. In this way, digital twin-driven parameter optimization is achieved.

[0065] Step 5: Construct a unified constitutive model for thermal-mechanical-phase transformation coupling, and use the multi-physics field coupling numerical analysis method to perform reverse optimization on the Pareto front solution set. Re-optimize the Pareto front solution set through the digital twin case library and then enter it into the lightweight AI model to find better processing control process parameters; at the same time, input the real-time monitored processing data back into the simulation model for closed-loop verification, realizing the closed-loop verification of digital twin virtual-reality, which can reduce the prediction error rate of the heat-affected zone (HAZ) topography by 30%.

[0066] Through multi-physics coupling modeling and simulation, based on the phase transformation characteristics of titanium alloy in thermal-assisted machining, a coupled thermo-mechanical-phase transformation unified constitutive model is established, covering aspects such as high-temperature plasticity of materials, martensitic phase transformation kinetics, and residual stress evolution, so as to reveal the drilling and riveting mechanism and facilitate the optimization of the drilling and riveting process. The coupled thermo-mechanical-phase transformation unified constitutive model combines the Johnson-Cook plastic model ( ) and the Avrami phase transformation equation ( ), and the specific equations are as follows:

[0067] ;

[0068] where σ is the flow stress, A is the yield stress, B is the hardening modulus, , are the corrections of the phase transformation volume fraction X to the initial yield stress (A0) and the initial hardening modulus (B0), , are the phase transformation softening coefficients of the yield stress and the hardening modulus respectively, ε is the equivalent plastic strain, n is the hardening index, C is the strain rate sensitivity coefficient, is the normalized strain rate, is the normalized temperature, m is the temperature softening index, α is the phase transformation softening factor, X is the phase transformation volume fraction, k is the phase transformation rate, equivalent plastic strain rate, t is the time, ρ is the density, is the specific heat capacity at constant pressure, is the partial derivative of temperature with respect to time, divergence, is the thermal conductivity, is the temperature gradient, η is the proportion of plastic work converted into heat, ΔH is the latent heat of phase transformation, is the phase transformation rate, is the change rate of residual stress, E is the elastic modulus, is the coefficient of thermal expansion, is the temperature change, is the phase transformation expansion coefficient, is the change amount of the phase transformation volume fraction.

[0069] Step 6: The verified process parameters are classified into the digital twin case library, and the digital twin case library pushes the new Pareto front solution set to the lightweight AI model, thereby realizing dynamic update, which has been proven to shorten the process debugging cycle of new products by 40%.

[0070] The digital twin case library contains various cases and relevant information on process parameter optimization, such as simulation analysis, Pareto front solution set, process parameter optimization data, and optimization control strategies, etc.

[0071] Step 7: Dynamically modulate the PID parameters through a deep reinforcement learning model (DRL). Combine the verified process parameters and control the thermal-assisted flow drill riveting equipment through the PID control circuit (including driving the servo motor and heating module, etc.) to achieve high-precision closed-loop control. Combine the tensile strength - energy consumption balance reward function to correct the PID parameters in real time and trigger hardware-level protection when exceeding the limit to ensure high processing precision and system stability. Then, feedback the control results of the thermal-assisted flow drill riveting equipment to the digital twin case library.

[0072] The state space of the described deep reinforcement learning model (DRL) includes real-time processing data (temperature, pressure, displacement, and surface topography image data) and digital twin inputs (prediction error of the heat-affected zone (HAZ) topography, martensite ratio, residual stress, etc.). Its action space is divided into continuous actions (ΔK p (proportional gain adjustment), ΔK i (integral time constant adjustment), ΔK d (derivative gain adjustment)) and discrete actions (path correction, exception handling).

[0073] The tensile strength - energy consumption balance reward function is: R = 10×S 达标 + 5×E 降低 - 20×T 超限 - 15×V 超标 , where R is the reward value, S 达标 represents that the connection strength meets the standard, E 降低 represents the reduction of energy consumption, T 超限 represents the temperature overrun flag; V 超标 represents the vibration amplitude overrun flag.

[0074] The described PID control circuit, as shown in the appendix Figure 3 , includes a proportional control part, an integral control part, a derivative control part, and an adder circuit.

[0075] The proportional control part includes an operational amplifier and a resistor (R p ). The input signal (V in ) and the reference signal (V ref ) are input to the inverting input terminal of the operational amplifier through a resistor network. The proportional gain (K p ) is determined by the ratio of the feedback resistor (R f ) to the input resistor (R in ). The proportional gain determines the system response speed. Too large may cause oscillation, and too small may result in slow response. The output signal (V p ) is the product of the input deviation (V in - V ref ) and K p .

[0076] The integral control part includes an operational amplifier, a capacitor, and a resistor (R i ). The input signal (V in ) is input to the inverting input terminal of the operational amplifier through the resistor (R i ). A capacitor is connected in the feedback loop to accumulate the integral value of the input deviation. The integral gain (K i ) is determined by the resistor (R i ) and the capacitor. The integral gain can eliminate the steady-state error. If it is too large, it may cause overshoot; if it is too small, the error elimination is slow. The output signal (V i ) is the product of the integral of the input deviation (V in - V ref ) and K i .

[0077] The differential control part includes an operational amplifier, a capacitor, and a resistor (R d ). The input signal (V in ) is input to the inverting input terminal of the operational amplifier through the capacitor. A resistor (R d ) is connected in the feedback loop to detect the change rate of the input deviation. The differential gain (K d ) is determined by the resistor (R d ) and the capacitor. The differential gain can suppress system oscillation. If it is too large, it may introduce noise; if it is too small, the suppression effect is insufficient. The output signal (V d ) is the product of the differential of the input deviation (V in - V ref ) and K d .

[0078] The adder circuit of the PID control module inputs the proportional (V p ), integral (V i ), and differential (V d ) control outputs to the adder circuit through a resistor network. The adder linearly superimposes the three parts of the outputs to generate the total control signal (V out ). The total control signal (V out ) is used to drive the controlled object (such as a servo motor, a heating module, etc.). The calculation formula of the total control signal (V out ) is as follows:

[0079] ;

[0080] The dynamic adjustment process of the PID control is as follows:

[0081] 1. Set the initial parameters

[0082] Digital twin pre-configuration: Provide the initial parameters (Kp = 2.0, K i = 0.1, K d = 0.5);

[0083] Safety margin: Set the range of PID parameter changes (K p ∈ [1.0, 3.0], K i ∈ [0.05, 0.15], K d ∈ [0.3, 0.7]).

[0084] 2. Real-time control loop steps (period ≤ 1 ms), the steps are as follows:

[0085] c1: Read the status data from the digital twin case library;

[0086] c2: The DRL policy network outputs ΔK p , ΔK i , ΔK d ;

[0087] c3: Update the PID parameters: K p + = K p + ΔK p , K i + = K i + ΔK i , K d + = K d + ΔK d ;

[0088] c4: Calculate the control quantity: ;

[0089] c5: Output the control signal to the actuator;

[0090] c6: Feed back the actual processing data to the digital twin case library;

[0091] c7: Update the weights of the DRL policy network every 10 minutes to optimize the long-term control strategy.

[0092] 3. Establish an exception handling mechanism

[0093] Overlimit threshold protection mechanism: When it is detected that the environmental parameters exceed the safety threshold (temperature ≥ 650 °C or pressure ≥ 1500 N), immediately trigger the "emergency power reduction" control strategy of deep reinforcement learning (DRL), and synchronously activate the hardware-level redundant protection circuit;

[0094] Model failure fallback mechanism: When the PID output deteriorates continuously (e.g., the reward value is lower than the preset threshold for 10 consecutive times), it automatically switches to the conservative PID parameters preset in the digital twin.

[0095] The above describes the present invention and its implementation manners. Such description is not restrictive. What is shown in the drawings is only one of the implementation manners of the present invention, and the actual structure is not limited thereto. All in all, if those of ordinary skill in the art are inspired by it and, without departing from the gist of the present invention, creatively design structural manners and embodiments similar to the technical solution, they shall fall within the protection scope of the present invention.

Claims

1. A control method for thermal-assisted flow drill riveting based on digital twin, characterized in that, It includes the following steps: Step 1: Collect the processing data of the thermal-assisted flow drill riveting equipment, including the temperature, pressure, displacement, and surface topography image data of the processing area; Step 2: Preprocess the temperature, pressure, displacement, and surface topography image data, standardize the preprocessed temperature, pressure, and surface topography image data, and then perform weighted fusion on the standardized temperature, pressure, and surface topography image data to obtain temperature-pressure-topography weighted fusion data; Step 3: Real-time process the standardized data and temperature-pressure-topography weighted fusion data through the FPGA chip and the lightweight AI model, dynamically optimize the process parameters, and synchronously generate temperature prediction and defect detection results; Step 4: Based on the process parameters, temperature prediction, and defect detection results obtained in Step 3, further optimize through the Monte Carlo-NSGA-II hybrid algorithm, generate candidate process parameters, screen the Pareto front solution set, and transmit it to the digital twin case library in real time; Step 5: Construct a thermal-mechanical-phase change coupling unified constitutive model, and use the multi-physical field coupling numerical analysis method to perform reverse optimization on the Pareto front solution set. Re-optimize the Pareto front solution set through the digital twin case library and the lightweight AI model again, and reverse input the real-time monitored processing data into the simulation model for closed-loop verification; Step 6: The verified process parameters are classified into the digital twin case library, and the digital twin case library pushes a new Pareto front solution set to the lightweight AI model; Step 7: Dynamically modulate the PID parameters through the deep reinforcement learning model, combine with the verified process parameters, and control the thermal-assisted flow drill riveting equipment through the PID control circuit.

2. The hot-assisted flow drill riveting control method according to claim 1, wherein The temperature-pressure-topography weighted fusion data , is the normalized temperature data, F * is the normalized pressure data, I * is the binarized surface topography image data.

3. The hot-assisted flow drill riveting control method according to claim 1, characterized in that In Step 7, when dynamically modulating the PID parameters through the deep reinforcement learning model, the PID parameters are corrected in real time in combination with the tensile strength-energy consumption balance reward function, and hardware-level protection is triggered when exceeding the limit. The control result of the thermal-assisted flow drill riveting equipment is fed back to the digital twin case library.

4. The hot-assisted flow drill riveting control method according to claim 1, wherein The state space of the deep reinforcement learning model includes real-time processing data and digital twin input, and the action space is divided into continuous actions and discrete actions.

5. The hot-assisted flow drill riveting control method according to claim 3, wherein The tensile strength - energy consumption balance reward function is: R = 10×S 达标 + 5×E 降低 − 20×T 超限 − 15×V 超标 , where R is the reward value, S 达标 indicates that the connection strength meets the standard, E 降低 indicates that the energy consumption is reduced, T 超限 indicates the temperature over - limit flag; V 超标 indicates the vibration amplitude over - limit flag.

6. The hot-assisted flow drill riveting control method according to claim 1, wherein The process of Step 3 is as follows: Step b1: Build the lightweight AI model architecture; Step b2: Split the lightweight AI model layer into parallel pipeline tasks, match the FPGA chip, and use the FPGA chip to cache and pre-store data; Step b3: The loss function is obtained by summing the mean square error and the structural similarity by weight; Step b4: Dynamically optimize the process parameters; Step b5: Generate temperature prediction and defect detection results.

7. The hot-assisted flow drill riveting control method according to claim 6, wherein In step b5, based on the normalized temperature data and the normalized pressure data predict the target temperature T opt ; use a lightweight AI model to classify the defect type according to the binary surface topography image data I * output the defect coordinates, type, and severity level, and provide real-time feedback to the parameter optimizer.

8. A thermal-assisted flow drill riveting control system for implementing the auxiliary flow drill riveting control method according to any one of claims 1-7, characterized in that, It includes: A multi-source data acquisition terminal for real-time capturing the temperature, pressure, displacement, and surface topography image data of the processing area; An intelligent decision-making module, which consists of a lightweight AI model, a parameter optimizer, a digital twin case library, and a digital twin model library. The digital twin model library is used to store the thermal-mechanical-phase change coupling unified constitutive model and the Monte Carlo-NSGA-II hybrid algorithm; The dynamic control module consists of a safety monitoring module, a PID control module, a servo motor of the thermal-assisted flow drill riveting equipment, and a heating module; the PID control module consists of a PID control circuit and a deep reinforcement learning model, and dynamically modulates the PID parameters through the deep reinforcement learning model.

9. The hot-assisted flow drill riveting control system according to claim 8, wherein The multi-source data acquisition terminal includes an infrared thermal imager, a multi-axis force sensor, a laser displacement meter, and an industrial camera.

Images