A twin model adaptive synchronization method based on real-time error feedback

Abstract

A twin model adaptive synchronization method based on real-time error feedback is provided, which realizes high-precision and self-evolution synchronization control between the physical entity and the twin model by introducing a dynamic gain matrix and an incremental learning mechanism; high-precision and low-delay synchronization control of the physical entity and its twin model in a dynamic environment is realized by introducing an adaptive control algorithm; by designing a control framework that fuses model predictive control (MPC) and an adaptive gain adjustment mechanism, the system can correct synchronization errors in real time during the movement process. In addition, by introducing an error feedback learning mechanism, the controller can accumulate experience and gradually optimize the control performance in multiple task executions, continuously improve the synchronization accuracy, ensure real-time performance, automatically correct the response of the twin model according to the state change of the physical entity, and automatically adjust the control parameters in real time according to the environmental change and error feedback, thereby having the characteristics of adaptability, robustness and continuous learning ability.

Description

Technical Field

[0001] This invention relates to the field of adaptive synchronization methods, and more particularly to an adaptive synchronization method based on a twin model with real-time error feedback. Background Technology

[0002] With the rapid development of digital twin technology, the real-time mapping and interaction between real physical entities and their digital models in virtual space has become an important research direction in fields such as intelligent manufacturing, robot control, and virtual simulation. The core of digital twins lies in establishing a high-precision virtual copy, achieving bidirectional information interaction and dynamic synchronization between the physical and digital worlds through sensing, transmission, modeling, and control.

[0003] In intelligent manufacturing and robotics systems, physical entities (such as industrial robotic arms and mobile robots) collect their operational status data, including joint angles, speed, acceleration, posture, load, and force information, through sensors. This data is then transmitted to a virtual model to drive real-time updates of the digital twin model. The digital twin model can perform motion prediction, path planning, collision detection, and control strategy optimization in the virtual environment, thereby providing auxiliary decision-making and feedforward control for the physical entity.

[0004] However, existing technologies still have the following shortcomings in achieving high-precision synchronization and adaptive control between physical entities and twin models: Limited synchronization accuracy: Most current synchronization control methods rely on static modeling and preset parameter control. A set of fixed control parameters is obtained through experimental calibration during system deployment to describe the kinematic and dynamic characteristics of the physical entity. However, in actual operation, the motion state of the mechanical system is affected by various complex factors, such as changes in joint friction coefficients, link inertia deviations, load changes, and temperature fluctuations. These nonlinear and time-varying characteristics cause the preset model to fail to accurately reflect the true dynamic behavior of the physical entity, resulting in accumulated errors between the twin model and the actual motion state, manifesting as synchronization trajectory deviation, attitude inaccuracies, or motion instability. Lack of adaptive adjustment mechanisms: Current control algorithms typically set the control gain and model parameters once before deployment, lacking a mechanism to automatically adjust parameters based on environmental changes and state deviations during operation. Control performance significantly degrades when external disturbances or system states change. Insufficient feedback and learning mechanisms: Most existing synchronization systems are feedforward control architectures, making it difficult to correct synchronization errors in real time or optimize control strategies using historical data, resulting in a lack of long-term self-learning and evolution capabilities. Summary of the Invention

[0005] The purpose of this invention is to provide a twin model adaptive synchronization method based on real-time error feedback, which enables the twin model to automatically correct its response according to changes in the physical entity state while ensuring real-time performance.

[0006] The technical solution adopted by this invention to solve the above-mentioned technical problems is: an adaptive synchronization method for a twin model based on real-time error feedback, comprising the following steps: S1. Establish a twin model of the physical entity and initialize the configuration of the physical entity and the twin model; S2. Introduce a multi-sensor fusion sampling architecture at the acquisition end to collect data through multiple sensors and obtain physical entity status information in real time. Each sensor automatically records its local sampling time when acquiring data. Generate data record unit Each The sensor's real-time observations, unique identifier, and corresponding timestamp are represented as follows: ; in, Indicates that the i-th sensor is in The raw measurement data collected at that time; Sensor type and number used to identify the source of data; S3. Perform timestamp synchronization and clock correction. Use the Network Time Protocol (NTP) and timestamp synchronization algorithm for time synchronization. Each sensor calculates the clock deviation based on the NTP four-way handshake. ; in: This indicates the time difference between the device's local clock and the target sensor's clock. When to send a synchronization request to the local device. For the time when the sensor receives the request; Send response time to the sensor. For the local device to receive the response time; Based on time difference Correct all sensor timestamps to a unified reference time: ; in: For a unified timestamp after synchronization, This is the original sampling timestamp; S4. Multi-source data fusion using Kalman filtering. 4.1 Establish a discrete linear model of the system state: ; ; Let k be a state variable, representing the true state of the system at discrete time k; For control input, it represents the control command applied to the system at time k; These are multi-sensor observations, representing the data measured by one or more sensors at time k. For process noise, The noise level represents the measurement error of the sensor. 4.2 Predicting the state estimate at the current moment: ; Predict the covariance at the current moment: ; in, The process noise covariance matrix; Calculate the Kalman gain: ; in, The covariance matrix of the observation noise; Updated state estimate: ; Update the covariance matrix: ; The final output is the updated state variable. As input for subsequent steps; S5. Define the synchronization error between the physical entity and the twin entity as: ; in, The state of a physical entity. The state of the twin model; Establish a dynamic error model: ; in, The current synchronization error, To control the input, Represents environmental change; S6, Definition The parameter vector consists of all learnable parameters of the controller at time t. The parameter gradient is calculated after each task execution. ; Loss function of the corresponding process for: ; The controller adaptively adjusts its parameters and performs incremental updates based on the error gradient information. ; Learning rate According to the decay mechanism, it gradually decreases over time; ; Where β is the attenuation coefficient, 0 < β < 1; Historical error sequences A moving average is applied, and the average error over a time window of length N is taken as the historical trend error of the system, expressed as: ; Using historical trend errors to update parameters in slow-layer models achieves long-term stability optimization: ; in, Slow learning rate; S7. Input the corrected MPC model parameters, corrected error weights, and corrected future prediction error terms into the MPC prediction control module to construct the MPC linear discrete model: ; Using the state estimate based on the Kalman filter output as the initial condition = ; ; in, It is about predicting the state of the next i steps from the current time k. It is the control quantity for the future i-step; Construct the cost function, which includes a penalty term for state error, a penalty term for changes in control input, and a compensation term for external disturbances: ; in, For the physical entity state, Q is the error weight, and R is the control change weight; S8. Construct and solve the MPC quadratic programming optimization problem. 8.1 Construct constraints, including control quantity constraints, state constraints, and control change constraints; Control constraints: ; State constraints: ; Controlling change constraints: ; 8.2 The quadratic programming optimization problem is expressed as: ; Among them, optimization variables H and f are derived from the cost function expansion, G and b are derived from the constraints, and the solver is a custom QP solver for ADMM; 8.3 Solve the quadratic programming optimization problem in each sampling period to obtain the predicted state for the next few steps and the optimized control sequence for the next few steps: ; Execute the first control action of the optimized control sequence. ,Will As the control input signal for the system at time t: ; The quadratic programming optimization problem is solved again in the next sampling period to achieve rolling optimization of MPC predictive control and output the result. As the feedforward control quantity obtained from MPC optimization; S9. Construct a fast gain adaptive update mechanism based on error-driven mechanism, with the following update rules: ; in, Here is the gain matrix. For a fast learning rate; A low-pass filter is used as a first-order exponential smoothing mechanism for the gain matrix. Smoothing is performed to obtain the smoothed gain matrix. : ; in, This is the gain smoothing coefficient, with a value ranging from 0.7 to 0.95; S10. Based on the smoothed gain matrix The fast error feedback term is obtained and fused with the MPC predictive control result to generate the final control signal. : ; Drive the physical entity to perform synchronous control based on the final control signal, and output the control quantity to the actuator to realize the real-time following of the twin trajectory by the physical entity or the high-precision synchronization of the twin with the movement of the physical entity. S11. Monitor the system's operating status, collect feedback data, monitor synchronization error, control input changes, energy consumption, disturbance amount, and tracking stability in real time, and send the monitoring data to the strategy optimization module; S12. The strategy optimization module performs long-term adaptive and multi-condition optimization to construct the system performance evaluation function: ; in, For external disturbances or energy consumption indicators, w e w u w p These are the weighting coefficients; computing systems in the past The average overall performance over a period of time is used as a long-term performance indicator. : ; in, This represents the performance cost at time i. For the set of policy parameters Calculate the gradient and the gradient of the policy parameters. : ; Based on gradient trends, with policy-level learning rates For strategy parameters Slow adaptive updates are performed to achieve continuous optimization of the control strategy in the long term. The adaptive parameter update strategy is as follows: ; Based on performance evaluation values Based on the changing trend, the weight matrix of the MPC prediction model is dynamically adjusted, and the error weight Q and control change weight R are updated: ; ; in, and Update the step size for the strategy.

[0007] Preferably, in S6, the attenuation coefficient β ranges from 0.95 to 0.99.

[0008] According to the above technical solution, the beneficial effects of the present invention are: This invention achieves high-precision, self-evolving synchronization control between a physical entity and its twin model by introducing a dynamic gain matrix and an incremental learning mechanism. It also achieves high-precision, low-latency synchronization control of the physical entity and its twin model in dynamic environments by introducing an adaptive control algorithm. Furthermore, by designing a control framework that integrates model predictive control (MPC) and adaptive gain adjustment mechanisms, the system can correct synchronization errors in real time during motion. In addition, an error feedback learning mechanism is introduced, enabling the controller to accumulate experience and gradually optimize control performance through multiple task executions, achieving continuous improvement in synchronization accuracy. Compared to the fixed-parameter control of existing technologies, this invention, while ensuring real-time performance, allows the twin model to automatically correct its response based on changes in the physical entity's state and automatically adjust control parameters in real time according to environmental changes and error feedback, exhibiting adaptability, robustness, and continuous learning capabilities.

[0009] Unlike traditional MPC, which only adjusts the control input during the prediction phase, this invention achieves dynamic co-optimization of the prediction model parameters, gain matrix, and feedback error, forming a dynamic adaptive closed-loop system driven by error. The control parameters and prediction model can be self-learned and updated based on real-time errors. During the control process, a dynamic gain matrix K(t) is introduced to achieve two-layer linkage optimization of the controller parameters and the prediction model, thereby realizing high-precision self-evolving synchronous control of the physical entity and the twin model. An error model is established using historical operating data, and the adaptive controller parameters are updated through incremental learning, enabling the system to continuously optimize control performance in multiple runs, achieving improved synchronization accuracy through "self-evolution." Through the above process and method, the system structure of this invention is clear, with close connections between its components, facilitating practical deployment. The adaptive control module, through its two-layer design and feedback learning mechanism, achieves high-precision synchronous control in complex dynamic environments, significantly outperforming traditional fixed-gain control strategies. Attached Figure Description

[0010] Figure 1 This is a model framework diagram of the present invention; Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0011] like Figure 1 , 2 As shown in the embodiment, an adaptive synchronization method for a Siamese model based on real-time error feedback is provided, including the following steps: S1. Establish a twin model of the physical entity and initialize the configuration of the physical entity and the twin model.

[0012] S2. Introduce a multi-sensor fusion sampling architecture at the acquisition end to collect data through multiple sensors and obtain physical entity status information in real time.

[0013] Each sensor automatically records its local sampling time when acquiring data. Generate data record unit Each The sensor's real-time observations, unique identifier, and corresponding timestamp are represented as follows: ; in, Indicates that the i-th sensor is in The raw measurement data collected at that time; Sensor type and number used to identify the source of data.

[0014] S3. Perform timestamp synchronization and clock correction. Use the Network Time Protocol (NTP) and timestamp synchronization algorithm for time synchronization. Each sensor calculates the clock deviation based on the NTP four-way handshake. ; in: This indicates the time difference between the device's local clock and the target sensor's clock. When to send a synchronization request to the local device. For the time when the sensor receives the request; Send response time to the sensor. The response time received by the local device.

[0015] Based on time difference Correct all sensor timestamps to a unified reference time: ; in: For a unified timestamp after synchronization, This is the original sampling timestamp.

[0016] S4. Multi-source data fusion using Kalman filtering. 4.1 Establish a discrete linear model of the system state: ; ; Let k be a state variable, representing the true state of the system at discrete time k; For control input, it represents the control command applied to the system at time k; These are multi-sensor observations, representing the data measured by one or more sensors at time k. For process noise, The observation noise represents the measurement error of the sensor.

[0017] 4.2 Predicting the state estimate at the current moment: .

[0018] Predict the covariance at the current moment: ; in, Let be the process noise covariance matrix.

[0019] Calculate the Kalman gain: ; in, Let be the covariance matrix of the observation noise.

[0020] Updated state estimate: .

[0021] Update the covariance matrix: .

[0022] The final output is the updated state variable. As input for subsequent steps.

[0023] S5. Define the synchronization error between the physical entity and the twin entity as: ; in, The state of a physical entity. This represents the state of the twin model.

[0024] Establish a dynamic error model: ; in, The current synchronization error, To control the input, It represents environmental change.

[0025] S6, Definition The parameter vector consists of all learnable parameters of the controller at time t. The parameter gradient is calculated after each task execution. .

[0026] Loss function of the corresponding process for: .

[0027] The controller adaptively adjusts its parameters and performs incremental updates based on the error gradient information. .

[0028] Learning rate According to the decay mechanism, it gradually decreases over time; ; Where β is the attenuation coefficient, 0 < β < 1, and its typical value range is 0.95~0.99.

[0029] Historical error sequences A moving average is applied, and the average error over a time window of length N is taken as the historical trend error of the system, expressed as: .

[0030] Using historical trend errors to update parameters in slow-layer models achieves long-term stability optimization: ; in, This is the slow learning rate.

[0031] S7. Input the corrected MPC model parameters, corrected error weights, and corrected future prediction error terms into the MPC prediction control module to construct the MPC linear discrete model: ; Using the state estimate based on the Kalman filter output as the initial condition = .

[0032] MPC is used to predict the next N steps: ; in, It is about predicting the state of the next i steps from the current time k. It is the control quantity for the future i-step.

[0033] Construct the cost function, which includes a penalty term for state error, a penalty term for changes in control input, and a compensation term for external disturbances: ; in, Let Q be the physical entity state, Q be the error weight, and R be the control change weight.

[0034] S8. Construct and solve the MPC quadratic programming optimization problem. 8.1 Construct constraints, including control quantity constraints, state constraints, and control change constraints.

[0035] Control constraints: .

[0036] State constraints: .

[0037] Controlling change constraints: .

[0038] 8.2 The quadratic programming optimization problem is expressed as: ; Among them, optimization variables H and f are derived from the cost function expansion, G and b are derived from the constraints, and the solver is a custom QP solver for ADMM.

[0039] 8.3 Solve the quadratic programming optimization problem in each sampling period to obtain the predicted state for the next few steps and the optimized control sequence for the next few steps: .

[0040] Execute the first control action of the optimized control sequence. ,Will As the control input signal for the system at time t: .

[0041] The quadratic programming optimization problem is solved again in the next sampling period to achieve rolling optimization of MPC predictive control and output the result. As the feedforward control quantity obtained from MPC optimization.

[0042] S9. Construct a fast gain adaptive update mechanism based on error-driven mechanism, with the following update rules: ; in, Here is the gain matrix. For a fast learning rate.

[0043] A low-pass filter is used as a first-order exponential smoothing mechanism for the gain matrix. Smoothing is performed to obtain the smoothed gain matrix. : ; in, This is the gain smoothing coefficient, with a value ranging from 0.7 to 0.95.

[0044] S10. Based on the smoothed gain matrix The fast error feedback term is obtained and fused with the MPC predictive control result to generate the final control signal. : .

[0045] The physical entity is driven to perform synchronous control based on the final control signal, and the control quantity is output to the actuator to realize the real-time following of the twin trajectory by the physical entity or the high-precision synchronization of the twin with the movement of the physical entity.

[0046] S11. Monitor the system's operating status, collect feedback data, monitor synchronization error, control input changes, energy consumption, disturbance amount, and tracking stability in real time, and transmit the monitoring data to the strategy optimization module.

[0047] S12. The strategy optimization module performs long-term adaptive and multi-condition optimization to construct the system performance evaluation function: ; in, For external disturbances or energy consumption indicators, w e w u w p These are the weighting coefficients.

[0048] computing systems in the past The average overall performance over a period of time is used as a long-term performance indicator. : ; in, Let be the performance cost at time i.

[0049] For the set of policy parameters Calculate the gradient and the gradient of the policy parameters. : .

[0050] Based on gradient trends, with policy-level learning rates For strategy parameters Slow adaptive updates are performed to achieve continuous optimization of the control strategy in the long term. The adaptive parameter update strategy is as follows: .

[0051] Based on performance evaluation values Based on the changing trend, the weight matrix of the MPC prediction model is dynamically adjusted, and the error weight Q and control change weight R are updated: ; ; in, and Update the step size for the strategy.

[0052] This embodiment forms a closed-loop self-evolving structure for the control system and completes overall synchronous control, forming a top-down optimization link: strategy layer → model layer → control layer → fast layer feedback, and a bottom-up learning link: fast layer feedback → incremental learning → strategy layer; ultimately achieving an overall control process of real-time synchronization + adaptive adjustment + long-term self-evolving optimization.

Claims

1. An adaptive synchronization method based on a twin model using real-time error feedback, characterized in that, Includes the following steps: S1. Establish a twin model of the physical entity and initialize the configuration of the physical entity and the twin model; S2. Introduce a multi-sensor fusion sampling architecture at the acquisition end to collect data through multiple sensors and obtain physical entity status information in real time. Each sensor automatically records its local sampling time when acquiring data. Generate data record unit Each The sensor's real-time observations, unique identifier, and corresponding timestamp are represented as follows: ； in, Indicates that the i-th sensor is in The raw measurement data collected at that time; Sensor type and number used to identify the source of data; S3. Perform timestamp synchronization and clock correction. Use the Network Time Protocol (NTP) and timestamp synchronization algorithm for time synchronization. Each sensor calculates the clock deviation based on the NTP four-way handshake. ； in: This indicates the time difference between the device's local clock and the target sensor's clock. When to send a synchronization request to the local device. For the time when the sensor receives the request; Send response time to the sensor. For the local device to receive the response time; Based on time difference Correct all sensor timestamps to a unified reference time: ； in: For a unified timestamp after synchronization, This is the original sampling timestamp; S4. Multi-source data fusion using Kalman filtering. 4.1 Establish a discrete linear model of the system state: ； ； Let k be a state variable, representing the true state of the system at discrete time k; For control input, it represents the control command applied to the system at time k; These are multi-sensor observations, representing the data measured by one or more sensors at time k. For process noise, The noise level represents the measurement error of the sensor. 4.2 Predicting the state estimate at the current moment: ； Predict the covariance at the current moment: ； in, The process noise covariance matrix; Calculate the Kalman gain: ； in, The covariance matrix of the observation noise; Updated state estimate: ； Update the covariance matrix: ； The final output is the updated state variable. As input for subsequent steps; S5. Define the synchronization error between the physical entity and the twin entity as: ； in, The state of a physical entity. The state of the twin model; Establish a dynamic error model: ； in, The current synchronization error, To control the input, Represents environmental change; S6, Definition The parameter vector consists of all learnable parameters of the controller at time t. The parameter gradient is calculated after each task execution. ； Loss function of the corresponding process for: ； The controller adaptively adjusts its parameters and performs incremental updates based on the error gradient information. ； Learning rate According to the decay mechanism, it gradually decreases over time; ； Where β is the attenuation coefficient, 0 < β < 1; Historical error sequences A moving average is applied, and the average error over a time window of length N is taken as the historical trend error of the system, expressed as: ； Using historical trend errors to update parameters in slow-layer models achieves long-term stability optimization: ； in, Slow learning rate; S7. Input the corrected MPC model parameters, corrected error weights, and corrected future prediction error terms into the MPC prediction control module to construct the MPC linear discrete model: ； Using the state estimate based on the Kalman filter output as the initial condition = ; ； in, It is about predicting the state of the next i steps from the current time k. It is the control quantity for the future i-step; Construct the cost function, which includes a penalty term for state error, a penalty term for changes in control input, and a compensation term for external disturbances: ； in, For the physical entity state, Q is the error weight, and R is the control change weight; S8. Construct and solve the MPC quadratic programming optimization problem. 8.1 Construct constraints, including control quantity constraints, state constraints, and control change constraints; Control constraints: ； State constraints: ； Controlling change constraints: ； 8.2 The quadratic programming optimization problem is expressed as: ； in, Optimize variables H and f are derived from the cost function expansion, G and b are derived from the constraints, and the solver is a custom QP solver for ADMM; 8.3 Solve the quadratic programming optimization problem in each sampling period to obtain the predicted state for the next few steps and the optimized control sequence for the next few steps: ； Execute the first control action of the optimized control sequence. ,Will As the control input signal for the system at time t: ; The quadratic programming optimization problem is solved again in the next sampling period to achieve rolling optimization of MPC predictive control and output the result. As the feedforward control quantity obtained from MPC optimization; S9. Construct a fast gain adaptive update mechanism based on error-driven mechanism, with the following update rules: ； in, Here is the gain matrix. For a fast learning rate; A low-pass filter is used as a first-order exponential smoothing mechanism for the gain matrix. Smoothing is performed to obtain the smoothed gain matrix. : ； in, This is the gain smoothing coefficient, with a value ranging from 0.7 to 0.95; S10. Based on the smoothed gain matrix The fast error feedback term is obtained and fused with the MPC predictive control result to generate the final control signal. : ； Drive the physical entity to perform synchronous control based on the final control signal, and output the control quantity to the actuator to realize the real-time following of the twin trajectory by the physical entity or the high-precision synchronization of the twin with the movement of the physical entity. S11. Monitor the system's operating status, collect feedback data, monitor synchronization error, control input changes, energy consumption, disturbance amount, and tracking stability in real time, and send the monitoring data to the strategy optimization module; S12. The strategy optimization module performs long-term adaptive and multi-condition optimization to construct the system performance evaluation function: ； in, For external disturbances or energy consumption indicators, w e w u w p These are the weighting coefficients; computing systems in the past The average overall performance over a period of time is used as a long-term performance indicator. : ； in, This represents the performance cost at time i. For the set of policy parameters Calculate the gradient and the gradient of the policy parameters. : ； Based on gradient trends, with policy-level learning rates For strategy parameters Slow adaptive updates are performed to achieve continuous optimization of the control strategy in the long term. The adaptive parameter update strategy is as follows: ； Based on performance evaluation values Based on the changing trend, the weight matrix of the MPC prediction model is dynamically adjusted, and the error weight Q and control change weight R are updated: ； ； in, and Update the step size for the strategy.

2. The adaptive synchronization method for a twin model based on real-time error feedback according to claim 1, characterized in that: In S6, the attenuation coefficient β ranges from 0.95 to 0.99.

Images