Self-evolution digital twin modeling method for unmanned vehicle power module

Abstract

The application discloses a self-evolution digital twin modeling method for unmanned vehicle power modules, step one: signal acquisition and preprocessing; step two: construction of an offline linear mechanism model; step three: construction of a residual sequence and analysis of autocorrelation; step four: establishment of an offline residual model; step five: during online operation, an online event-triggered evolution mechanism is executed; step six: a mechanism-residual digital twin model is obtained after evolution update through fusion, and a twin model prediction output is output. The application has the following advantages: 1) a complete mechanism model and a nonlinear residual compensation model are designed, and a construction and evolution method is provided; 2) the twin model is self-evolved on demand during multi-working-condition operation, and modeling accuracy and calculation efficiency are improved; 3) while ensuring the self-adaptive ability of the model, the stability and engineering implementability of online operation are improved, and the application is suitable for real-time state estimation and intelligent control decision of unmanned vehicle power modules.

Description

Technical Field

[0001] This invention is a self-evolving digital twin modeling method for the power module of unmanned vehicles, belonging to the field of unmanned system modeling and control technology. Background Technology

[0002] With the rapid development of autonomous vehicles, drones, and intelligent robot systems, higher demands are being placed on the modeling accuracy and real-time control performance of power modules. As the core execution unit of an autonomous vehicle, the dynamic characteristics of the power module directly affect the vehicle's motion response, trajectory tracking accuracy, and system stability. However, due to the complex and variable actual operating environment, power modules are often affected by factors such as frictional nonlinearity, drive dead zone, load variations, motor saturation, and power supply fluctuations, making it difficult for traditional modeling methods to balance accuracy and stability.

[0003] In existing technologies, dynamic modeling methods mainly fall into two categories: mechanistic modeling methods and data-driven modeling methods. Mechanistic modeling methods typically construct system models based on dynamic equations or equivalent linear models, offering advantages such as clear physical meaning and strong interpretability. However, during actual operation, model parameters are prone to drift with changes in operating conditions, making it difficult to characterize complex nonlinear characteristics and leading to the accumulation of prediction errors. To improve model adaptability, some methods introduce online parameter identification mechanisms; however, frequent updates can easily cause model oscillations, and stability is insufficient under multi-condition operation.

[0004] On the other hand, purely data-driven methods, such as neural network-based dynamic modeling, can fit complex nonlinear relationships, but they often lack physical constraints and are prone to overfitting or unstable outputs. Furthermore, online continuous learning suffers from catastrophic forgetting, where training under new operating conditions leads to the loss of historical knowledge, thus affecting long-term operational stability. In addition, existing online update methods often employ continuous training strategies without establishing reasonable triggering mechanisms, frequently updating the model even when the dynamic state is stable, increasing computational burden and reducing system robustness.

[0005] In recent years, digital twin technology has been applied to the field of unmanned system modeling. However, most digital twin models are still in the offline modeling or static calibration stage and lack a complete online self-evolution mechanism. At the same time, under the condition of continuous operation under multiple working conditions, how to maintain the stability and memory capacity of the model while ensuring modeling accuracy remains an important challenge for the current technology.

[0006] To address the aforementioned issues, this invention proposes a self-evolving digital twin modeling method for the powertrain module of autonomous vehicles. This method is based on a structured linear mechanism model and achieves self-evolving updates of dynamic parameters through online Bayesian recursion. On this basis, a residual prediction model based on a gated recurrent unit (GRU) is constructed to characterize unmodeled nonlinear dynamic behavior. To improve the initialization quality of the residual model, this invention introduces an improved pigeon flock optimization method to globally search and optimize the GRU network parameters and training parameters, enabling the model to exhibit better convergence performance and stability in the online phase.

[0007] Furthermore, this invention designs an event-triggered mechanism based on residual statistical characteristics, initiating the online update process only when the model prediction error increases significantly, thus avoiding parameter oscillations caused by frequent updates under stable operating conditions. During the online evolution process, a continuous learning constraint based on elastic weights is introduced, and the regularization coefficient is adaptively adjusted through a gating factor to achieve a dynamic balance of learning intensity across different operating conditions, thereby enhancing the model's adaptive capability while suppressing catastrophic forgetting.

[0008] Furthermore, this invention achieves the coordinated output of linear mechanistic models and nonlinear residual models through a gating fusion mechanism, and combines stability assurance strategies such as warm-start and rollback mechanisms to construct a complete engineering-grade self-evolving digital twin framework. This method can achieve high-precision, stable, and continuous online modeling and control support for the power module of unmanned vehicles in complex and ever-changing operating environments. Summary of the Invention

[0009] Purpose of the Invention: To address the problems in existing autonomous vehicle dynamics modeling methods, such as the difficulty in characterizing complex nonlinear characteristics by mechanistic models, the lack of structural constraints and insufficient stability of purely data-driven models, and frequent oscillations and catastrophic forgetting during online updates, this invention aims to provide a self-evolving digital twin modeling method for autonomous vehicle dynamics modules, such as... Figure 1 As shown, by constructing a hybrid twin framework that combines a linear mechanism model and a data-driven residual model, and introducing event-triggered and continuous learning constraint mechanisms, the dynamic model can achieve stable, high-precision, and adaptive evolution updates under multiple operating conditions, thereby improving the reliability and real-time performance of dynamic state estimation and control decisions for unmanned vehicles.

[0010] This invention addresses the problem of twin modeling of the power module in unmanned vehicles, such as... Figure 2 As shown, a self-evolving digital twin modeling method for the power module of an autonomous vehicle is designed. This method mainly includes an offline construction and online evolution update method for a data-driven mechanism-residual hybrid twin model, and an event-triggered online adaptive evolution framework. The specific steps of this method are as follows: Step 1: Collect and preprocess data S11. Collect the control input and output response signals of the unmanned vehicle's power module. Under the preset sampling period, Continuously collect control input signals from the autonomous vehicle's power module With the corresponding power output response signal The control input signal includes drive motor control command, voltage or equivalent drive quantity; the output response signal includes motor angular velocity, wheel speed or equivalent dynamic state variable.

[0011] S12, Data Preprocessing To ensure the accuracy of the modeling data, the acquired signals undergo time alignment, outlier removal, and necessary filtering; simultaneously, based on the control input signal from the previous moment... Constructing control input variation It is used to characterize the dynamic changes in the input.

[0012] (1) Will , as well as A discrete data sequence arranged in chronological order is formed for subsequent linear mechanism model building and residual modeling analysis. The sampling period is set according to the control frequency of the autonomous vehicle's power module to ensure that the discrete modeling satisfies the Nyquist sampling condition.

[0013] This step enables the standardized construction of power module operation data, laying the foundation for establishing a structured twin model.

[0014] Step 2: Construct an offline linear mechanism model The recursive Bayesian method is used to construct a data-driven linear mechanism model. The specific process is as follows: Based on the discrete time series data obtained in Step 1, a first-order discrete linear mechanism model of the autonomous vehicle's power module is established to characterize the dominant dynamic behavior of the power module. This linear model uses historical output and historical input as regression variables, and the model parameters are obtained through parameter estimation methods. The model is defined as follows: (2) in, Indicates in The output response of the linear model at time step 1. and For the first The system outputs response signals and control input signals at specific times. This represents the combined effect of inertia and damping of the system. This represents the input gain coefficient for rotational speed.

[0015] During the operation of an autonomous vehicle, considering that the dynamic characteristics may drift slowly or abruptly with changes in operating conditions, this invention employs a Bayesian recursive estimation method to update the parameters of the linear model online. By continuously correcting the posterior distribution of the parameters through a recursive approach, the adaptive evolution of the equivalent dynamic gain and time constant is achieved.

[0016] Constructing a regression vector : The parameter vector formed by the linear model parameters in equation (2) is obtained by using the Bayesian recursive estimation method. mean covariance matrix Perform recursive updates. Calculate the Kalman gain. Parameter mean With covariance The updated formula is: (3) (4) (5) in, This is the recursive Kalman gain matrix, used to adjust the parameter update magnitude; This represents the average value of the parameters at the previous time step.

[0017] according to Parameter mean at time The predicted output of the linear twin model can be obtained. for: (6) Step 3: Construct residual sequences and analyze correlations S31. Calculate the linear model prediction residual sequence based on the output predicted by the linear mechanism model. Based on the linear twin model prediction output obtained in step two and the actual system output The linear prediction residual of the mechanistic model can be expressed as: (7) S32. Analyze the autocorrelation of the residual series. Statistical analysis is performed on the residual sequence to calculate its autocorrelation function, mean, and variance characteristics. When the residual sequence exhibits significant time-series correlation or non-stationary characteristics, it is determined that the power module has unmodeled nonlinear characteristics, including but not limited to friction dead zones, saturation characteristics, input asymmetry, or dynamic hysteresis effects.

[0018] Based on this, a feature vector is constructed for subsequent nonlinear residual modeling. The feature vector includes at least the current and historical residual values, the current and historical control inputs, and the change in control inputs.

[0019] This step enables a structured decomposition of the sources of linear error, transforming the dynamic modeling problem from a complete black-box fitting to a hierarchical modeling framework that combines mechanistic components with residual compensation.

[0020] Step 4: Establish an offline residual model S41. Optimize the network parameters of the gated cyclic unit by improving the pigeon flock optimization method. The Gated Recurrent Unit (GRU) achieves adaptive modeling of temporal dependencies through a gating mechanism, making it suitable for residual modeling with temporal correlations. The core computations of the GRU network are performed by update gates, reset gates, and candidate hidden states. To improve the quality of model initialization, an improved pigeon flocking optimization method is used to optimize the GRU network parameters.

[0021] To address the strategy switching process during the exploration and development phases, a dynamically changing balancing factor is introduced. Used to uniformly regulate the intensity of population renewal: (8) in, The maximum number of iterations, This represents the current iteration number; and It is control Balance factor The hyperparameters for the range are typically set to 0.7 and 0.3, respectively. The balancing factor is set to a larger value in the early stages of iteration and gradually decreases in the later stages, thus achieving a smooth transition from strong exploration to strong exploitation.

[0022] To handle boundary violations during the exploration process, a rollback-style boundary mapping strategy is adopted: (9) in, Indicates the first During the first iteration update process, the... The individual in the first The parameter value of the dimension; Indicates the first Individuals that go out of bounds during the iterative update process Parameter values ​​after rollback mapping; The boundary elastic constant is ; the lower bound of the optimized parameter is . The upper boundary is .

[0023] The improved pigeon flock optimization method described above is used to perform offline global optimization of GRU network parameters. Parameters such as GRU time window length, hidden layer dimension, and learning rate are encoded as individual optimization particles, and a fitness function based on mean squared error is constructed. (10) in, This represents the position of an individual in the flock, i.e., the encoded network parameters to be optimized; The length of the training dataset; and These represent the actual residuals of the linear model and the predicted residuals of the GRU network, respectively.

[0024] In the exploration and development phase with a balance factor, the first Individual pigeons in a flock speed The update formula is expressed as: (11) in, and They represent the first Only one pigeon in the 1st Velocity and position in the next iteration The position representing the pigeon with the highest current fitness level. It is a natural constant. For map factor constants, It is a random number between 0 and 1; No. Individual pigeons in a flock Location The update formula is expressed as: (12) By calculating the fitness of each individual and ranking the fitness, the half of the pigeons with the highest fitness are retained, and their positions are placed in the center. For reference, the position update formula for the second stage is as follows: (13) After completing all iterations, the best individual is selected as the optimization result, and the GRU network is trained based on the optimized network structure parameters to establish a residual prediction model.

[0025] S42. Perform offline training of the GRU residual model. Using the network structure parameters and training parameters obtained in step S41, an offline GRU network is trained on the residual sequence obtained in step three. The GRU network takes a fixed-length time window as input, and its input features include current and historical residual values, current and historical control inputs, and changes in control input. A time-series sample sequence is formed using a sliding window method, and the predicted residual value for the next time step is output.

[0026] GRU network training can employ backpropagation, with parameters updated using gradient descent. (14) in, for Time-bound GRU network weight parameters; The loss cost calculated using the cross-entropy loss function; Represents the loss function Weights The partial derivatives of , i.e., the corresponding gradient; The learning rate hyperparameter is obtained in step S41 by improving the pigeon flock optimization method.

[0027] S43. Adjusting the output of the residual model based on the gating factor. Further design of gating factors This is used to adaptively adjust the residual compensation strength. The gate factor is adaptively determined based on the current residual statistical characteristics. (15) in, The linear model residuals obtained through equation (7); The smoothing parameters are used to avoid instability in the gating factor value when the error approaches zero and to control the sensitivity of the gating response.

[0028] The output of the linear twin model and the residual prediction output are combined through a gating fusion mechanism to obtain the fused digital twin model of the autonomous vehicle's power module. Predicted output at time step for: (16) in, Predict the output for a linear model; This is the residual prediction output of the GRU network.

[0029] Using gating factors When the linear model already has high accuracy, the participation of the residual network can be suppressed; when the error of the linear model increases, the residual compensation capability is gradually released to enhance the model's ability to express complex dynamics.

[0030] Through the aforementioned gating fusion mechanism, adaptive collaboration between the linear mechanism model and the data-driven residual model is achieved.

[0031] Step 5: During online runtime, execute the online event-triggered evolution mechanism. S51. Conduct online model consistency assessment Specifically, this includes: collecting real data and the prediction window data sequence of the twin model online, calculating the mean square error of the residual sequence, and determining whether the trigger threshold has been reached. Based on the residual sequence constructed in step three, its value at a fixed length is calculated. The statistical indicators within the evolution time window, using root mean square error. As a consistency assessment indicator: (17) The consistency between digital twin simulation data and measured data is periodically evaluated online by calculating the RMSE (Real-Time Sequence Error). If the RMSE value of two data sequences exceeds a set threshold, the reliability of the digital twin is considered to have significantly decreased, triggering a self-evolutionary update method. When the statistical index falls below the threshold, the current model parameters remain unchanged, thus avoiding unnecessary parameter oscillations under stable operating conditions. Through this event-triggered mechanism, an update strategy for the digital twin model to evolve on demand during operation is implemented, improving model stability and computational efficiency while ensuring modeling accuracy.

[0032] S52. When the triggering condition is met, enter the warm start phase. To ensure model stability during online evolution, the method further introduces a warm-start mechanism: after triggering online update, a certain number of window sample data are first accumulated before parameter update, in order to avoid erroneous updates caused by a single abnormal data.

[0033] S53. Perform online updates to model parameters; S531. Update the linear mechanism model online based on the recursive Bayesian method in step two; When the triggering condition in step five is met, the parameters of the linear mechanism model are updated using the recursive Bayesian method from step two.

[0034] S532. Online parameter updates for the GRU residual model are performed using continuous learning constraints based on elastic weights. The GRU residual prediction model is updated online, and the online update adopts a window-based training mechanism, that is, a set of sequence samples is constructed within a fixed-length time window: (18) in, For window length, It is a time-series feature vector that includes historical residuals, control inputs, and input increments. This represents the true residual of the linear model's prediction results.

[0035] The task loss function is constructed based on the prediction error of samples within the window. : (19) in, This represents the prediction residuals of the GRU network.

[0036] To suppress catastrophic forgetting in the model during continuous evolution across multiple operating conditions, elastic weights based on gating factors are introduced to stabilize the GA-EWC constraint terms, and then combined with the basic construction of the task loss function. The loss function for the GRU network parameter optimization problem can be obtained. for: (20) in, The current GRU parameters; Historical parameter values; GRU network parameters The corresponding values ​​in the Fisher information matrix are simplified to a diagonal information matrix; The total number of parameters required for a twin model to dynamically evolve is... The gating factor obtained from step S43 Adaptive adjustment of the canonical intensity coefficient; (twenty one) in, This represents the initial canonical intensity coefficient.

[0037] In practice, to reduce computational complexity, a diagonal approximation is typically used to estimate the importance of each parameter. : (twenty two) in, For the sample size, For the first The task loss corresponding to each sample.

[0038] because As time changes, the GRU network parameter optimization problem becomes an online optimization problem under time-varying parameter constraints, and its optimal solution can be expressed as: (twenty three) S54. After the parameters are updated, perform a model evolution evaluation and determine whether to roll back the parameters. The updated model is used to re-predict the output, and the updated mean squared error of the prediction is calculated. If the updated window error is greater than the original error, the model parameters from the previous time step are restored. The specific process is as follows: After completing a window-based online evolution update, the effectiveness of the update results needs to be evaluated. Specifically, the updated model parameters are used to re-predict the samples within the current sliding window, the corresponding window prediction error is calculated, and compared with the error level before the update.

[0039] Let the window prediction error before the update be... The updated window prediction error is The error is still calculated according to the window mean square error defined in equation (17). If the updated condition satisfies... If the updated parameters meet the requirements, it indicates that the parameter update effectively improves the model's fitting ability to the current working conditions. Therefore, the updated model parameters are retained, and the model continues to the subsequent running stage. Conversely, if the updated parameters do not meet the requirements, it indicates that the online update has failed to improve performance and may even lead to a decrease in model prediction performance. To ensure the stability of the model during online evolution, this invention introduces a parameter rollback mechanism: when the updated model parameters do not meet the requirements... When the update is complete, the model parameters are restored to their state before the update, that is, restored to the parameters saved at the previous moment, and the current update process is terminated.

[0040] By employing the aforementioned update evaluation and rollback strategies, we can effectively avoid undesirable parameter updates caused by local noise data, abnormal samples, or gradient oscillations, thereby improving the stability and reliability of the model during continuous evolution under multiple operating conditions.

[0041] Step Six: Fuse the evolved and updated mechanism-residual digital twin model to obtain the predicted response output value of the dynamic module twin model. Combining the linear mechanism model and the residual model yields the model expression for hybrid mechanism-residual modeling: (twenty four) in, This represents the evolution result of the linear mechanism model in step two. This is the evolution result of the GRU network residual model in step four.

[0042] During real-time operation, the linear mechanistic twin model first generates basic prediction results based on the current system input and historical states. The prediction results reflect the main dynamic structural characteristics of the system; simultaneously, the residual prediction network generates residual compensation terms based on historical error sequences and relevant input features. It is used to characterize nonlinear effects and the impact of environmental disturbances that are difficult to express by mechanistic models.

[0043] The prediction output of the linear mechanism twin model is combined with the output of the residual prediction network through the gating fusion mechanism in step S43 to obtain the final hybrid twin model prediction result: (25) in, This is the predicted output of the GRU network residual model.

[0044] Through the aforementioned fusion mechanism, the mechanistic model and the data-driven model can work collaboratively, enabling the hybrid twin model to possess both the stability and interpretability of the mechanistic model and the expressive power of the data model for complex nonlinear characteristics. The resulting hybrid prediction results can be further used for tasks such as dynamic state estimation, operational performance evaluation, and control decision optimization for autonomous vehicles, providing high-precision prediction support for the stable operation of the system under various operating conditions.

[0045] Compared with existing technologies, the self-evolving digital twin modeling method based on mechanism-residual co-modeling proposed in this invention has the following beneficial effects: (1) This invention integrates the mechanism twin model and the residual prediction model to model the system, so that while maintaining the physical interpretability and stability of the mechanism model, the model introduces a data-driven model to dynamically compensate for complex nonlinearities, thereby effectively improving the prediction accuracy of the dynamic system model.

[0046] (2) This invention constructs an event-triggered update mechanism based on sliding window error statistics, which only initiates model evolution update when the prediction error exceeds a set threshold, thereby reducing unnecessary model training processes, reducing computational overhead, and improving the efficiency and stability of online operation of digital twin models.

[0047] (3) This invention introduces an elastic weight-stabilizing constraint mechanism based on Fisher information matrix and combines a gating factor to adaptively adjust the regularization strength, so that the model can effectively maintain the modeling ability of historical working conditions during the continuous learning process of multiple working conditions, while improving the adaptability to new working conditions, thereby alleviating the problem of catastrophic forgetting.

[0048] (4) The present invention introduces a prediction error evaluation and parameter rollback mechanism in the online model update process. When the window prediction error of the updated model is not improved, the parameters before the update are automatically restored, thereby avoiding unstable updates caused by abnormal data or noise interference and improving the safety and reliability of the model evolution process.

[0049] (5) This invention uses a gating fusion mechanism to adaptively weight and combine the prediction output of the mechanism model with the prediction output of the residual, so that the model can dynamically adjust the residual compensation intensity according to the system operating state, thereby maintaining good prediction performance under different operating conditions.

[0050] (6) The self-evolving hybrid digital twin model constructed in this invention can provide high-precision prediction information for the dynamic state estimation, operation performance evaluation and control decision of unmanned vehicles, thereby improving the system's operation stability and control performance in complex working conditions.

[0051] In summary, this invention achieves high-precision prediction and long-term stable operation of digital twin models under multiple working conditions by combining mechanistic models and data-driven models with online self-evolution mechanisms, and has high engineering application value. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the self-evolution framework of the twin model of an autonomous vehicle.

[0053] Figure 2 The flowchart shows the self-evolutionary modeling method for mechanism-residual digital twin models.

[0054] Figure 3 This is a graph showing the autocorrelation results of the residual data.

[0055] Figure 4 This is a schematic diagram comparing the prediction and error results of the mechanism-residual model.

[0056] Figure 5 This diagram illustrates the relationship between update frequency and prediction accuracy under different evolution trigger thresholds. Detailed Implementation

[0057] The effectiveness of the self-evolving twin modeling method for unmanned vehicle power modules proposed in this invention is verified through a specific example below. This embodiment uses the Mecanum wheel unmanned vehicle drive motor power module as an example for illustration, but the invention is not limited to this structural form. The code running environment of this example is configured with an Intel i9-13980 processor, 2.20GHz clock speed, 16GB memory, and the dependencies are as follows: Operating System: Windows 11; Development Environment: PyCharm; Programming Language: Python; The specific steps of this method are as follows: Step 1: Collect the control input signals and output response signals of the unmanned vehicle's power module, and perform data preprocessing. In this embodiment, the unmanned vehicle's power module is a DC brushless motor drive unit, and the control input is the motor drive voltage command. The output response is the motor angular velocity. The data sampling period is 10 ms, and the data acquisition and recording duration is no less than 200 s. The input signal uses a combination of a step signal and random excitation, with the addition of a wideband excitation design. The acquired raw data is first processed for time synchronization to ensure that the input and output signals are aligned at the same sampling time; then, a first-order low-pass filter is used to remove high-frequency measurement noise, and obvious outliers are thresholded and removed.

[0058] Under a preset sampling period, the control input signals of the unmanned vehicle's power module are collected. With the corresponding power output response signal The control input signal includes drive motor control command, voltage or equivalent drive quantity; the output response signal includes motor angular velocity, wheel speed or equivalent dynamic state variable.

[0059] To ensure the accuracy of the modeling data, the acquired signals were time-aligned, outlier removed, and filtered as necessary; simultaneously, control input changes were constructed. It is used to characterize the dynamic changes in the input.

[0060] The discrete data sequence arranged in chronological order is used for subsequent linear mechanism model establishment and residual modeling analysis; wherein, the sampling period is set according to the control frequency of the unmanned vehicle power module to ensure that the discrete modeling meets the Nyquist sampling condition.

[0061] Step 2: Construct a data-driven linear mechanism model using the recursive Bayesian method. Based on the discrete time series data obtained in step one, a first-order discrete linear mechanism model of the unmanned vehicle's power module is established, as shown in equation (2). In the initial stage, the least squares method is used for parameter estimation to obtain the initial parameters of the model. During the operation of the unmanned vehicle, to address parameter drift caused by load changes or motor temperature rise, a recursive Bayesian update method is used to correct the parameters online. This update process is executed within each sampling period, enabling the linear twin model to continuously track the slow changes in power characteristics.

[0062] This step enables the standardized construction of power module operational data. Prediction results obtained through model identification experiments on historical datasets show an RMSE of 0.64 rad / s, a MAE of 0.51 rad / s, and a maximum absolute error of 3.6 rad / s. These results demonstrate that the linear model can stably track the dynamically driven processes with a low overall error level, serving as an interpretable mechanistic component of the hybrid twin framework.

[0063] Step 3: Construct residual sequences and analyze autocorrelation. Based on the linear twin model prediction output obtained in step two And calculate the residuals Statistical analysis was performed on the residual series to calculate its autocorrelation function, mean, and variance characteristics. The autocorrelation coefficient results are shown below. Figure 3 As shown in the figure. Experimental results show that under multiple load conditions, the residual autocorrelation function decays slowly within 1 second, indicating a significant time correlation. Therefore, it is suitable to use a time-series neural network for modeling and to use a time window of length 30 for sequence organization.

[0064] This step enables a structured decomposition of the sources of linear error, transforming the dynamic modeling problem from a holistic black-box fitting to a hierarchical modeling framework that combines mechanistic components with residual compensation.

[0065] Step 4: Establish a residual prediction model based on Gated Recurrent Units (GRUs), and optimize the network parameters by improving the pigeon flock optimization method to complete the offline training of the GRU residual model. In this embodiment, to compensate for the nonlinear dynamic behavior that the linear twin model fails to characterize, a residual prediction model based on gated recurrent units is constructed, and the network parameters are initialized and trained offline by an improved pigeon flock optimization method.

[0066] The GRU residual network uses a single-layer GRU architecture with a 3-dimensional input feature, a time window length of 30, 32 GRU hidden units, and a tanh activation function. The output layer is a fully connected layer with a 1-dimensional output. The network is trained using the Adam optimizer with a batch size of 128 and a learning rate of 1e-3.

[0067] To improve the convergence speed and stability of the GRU model in the online phase, this embodiment employs an improved pigeon flock optimization method for global parameter search. The optimization targets parameters such as the GRU time window length, hidden layer dimension, and learning rate; and the residual mean square is used as the fitness function. The parameters of the pigeon flock optimization method are set as follows: population size of 30, maximum number of iterations of 80, and inertia weight of 0.7.

[0068] Once the maximum number of iterations or the fitness function convergence threshold is reached, the parameters corresponding to the optimal individual are selected as the network structure and initialization parameters of the GRU residual model.

[0069] In this embodiment, a small amount of gradient fine-tuning is performed after model initialization to further improve the model fitting accuracy. This offline initialization process ensures that the GRU residual model has a good predictive capability and stability foundation before entering the online update stage, thereby reducing the risk of oscillations during the online phase.

[0070] Step 5: During the operation of the autonomous vehicle, construct an event triggering mechanism based on the mean square error of the residual sequence. Based on the residual sequence constructed in step three, its statistical index within a fixed-length W evolution time window is calculated, using the root mean square error as the statistical measure. As a consistency assessment indicator.

[0071] An evolutionary sliding time window of 100 was selected, and the RMSE trigger threshold was set to 0.7. The consistency between the digital twin simulation data and the measured data was periodically evaluated online by calculating the RMSE. If the RMSE value of two sequence data points exceeds the set threshold, the reliability of the digital twin is considered to have significantly decreased, triggering the self-evolutionary update method. When the statistical index is below the threshold, the current model parameters remain unchanged. Under the current parameter settings, the event-triggered evolutionary strategy reduces the number of updates on the dataset by approximately 90%, while the predicted RMSE value is only slightly lower than that of periodic triggering by about 5%. In particular, it can be seen that event triggering significantly reduces the online update frequency, greatly saving computational costs without sacrificing accuracy, thus meeting the requirements for online deployment.

[0072] In the early stages of evolution, a warm-start mechanism is implemented, with a warm-start duration of 200 sampling periods. After triggering an online update, a certain amount of window sample data is first accumulated before parameter updates are performed to avoid erroneous updates caused by a single abnormal data point.

[0073] The aforementioned stability guarantee mechanism enhances the robustness and engineering feasibility of the twin model during its online evolution.

[0074] When the triggering condition is met, the parameters of the mechanistic model and the GRU residual model are updated online, and a continuous learning constraint based on elastic weights is introduced. When the triggering condition in step five is met, the recursive Bayesian method described in step two is used to update the parameters of the linear mechanism model. A window-based online parameter update is then performed on the GRU residual prediction model. First, a sequence sample set is constructed. The window length is set to 30, and the update rounds use mini-batch gradient updates with 5 batches. The online learning rate is set to 1e-4. Unlike traditional time-triggered online training, this embodiment uses an event-triggered optimization method to avoid meaningless parameter drift during the model's stable phase.

[0075] To address the catastrophic forgetting problem during continuous operation under multiple operating conditions, this embodiment introduces a regularization constraint based on elastic weights. The basic regularization coefficient is set to 0.5, and adaptive regularization is achieved by calculating a gating adjustment factor. Finally, the GRU network is updated under EWC constraints according to the composite loss function in equation (20).

[0076] Determine whether to perform parameter rollback After completing a window-based online evolution update, the effectiveness of the update results needs to be evaluated. Specifically, the updated model parameters are used to re-predict the samples within the current sliding window, the corresponding window prediction error is calculated, and compared with the error level before the update.

[0077] Step Six: Fuse the evolved and updated mechanism-residual digital twin model to obtain the predicted response output value of the dynamic module twin model. The output of the linear twin model and the residual prediction output are fused through a gating fusion mechanism to obtain the prediction results of the self-evolving hybrid twin model. Figure 4 The prediction results and error comparison results of the linear mechanism model and the mechanism residual mixture model are presented. Figure 5 The curves showing update frequency versus prediction accuracy under different event trigger thresholds are presented.

Claims

1. A self-evolving digital twin modeling method for the power module of an unmanned vehicle, characterized in that: The implementation steps of this method are as follows: Step 1: Collect the control input signals and system output response signals of the unmanned vehicle's power module, and preprocess the collected data to obtain an offline dataset for initial model construction; Step 2: Based on the offline dataset, construct the initial linear mechanism model of the autonomous vehicle's power module. Specifically, the initial linear mechanism model is constructed using the recursive Bayesian method. Step 3: Calculate the system's predicted output using a linear mechanism model, and construct the system residual sequence based on the difference between the predicted output and the actual output; Step 4: Based on the residual sequence, establish an initial residual prediction model, including: S41. Select the network structure parameters of the gated cyclic unit by improving the pigeon flock optimization method; S42. Perform offline training of the gated recurrent unit network residual model; S43. Based on the gating factor, the output of the residual model is controlled and fused to obtain the initial mechanism-residual digital twin model; Step 5: During the system operation phase, execute the online event-triggered evolution mechanism, including: S51. Conduct online model consistency assessment; S52. When the triggering condition is met, enter the warm start phase; S53. Perform online updates to model parameters; S54. After the parameters are updated, evaluate the model evolution effect and determine whether to roll back the parameters. Step 6: Fuse to obtain the updated mechanism-residual digital twin self-evolutionary model, and output the predicted response value of the digital twin model.

2. The self-evolving digital twin modeling method for unmanned vehicle power modules according to claim 1, characterized in that: The specific steps for data acquisition and processing from the power module in step one include: Under the preset sampling period, Continuously collect control input signals from the autonomous vehicle's power module With the corresponding power output response signal The power output response signal includes, but is not limited to, motor angular velocity, wheel speed, or equivalent power state variables; The input and output signals are synchronized in time and denoised, and the control input signal from the previous moment is used as the basis for the synchronization. Construct the input incremental signal: ； Will , as well as The discrete data sequence is arranged in chronological order and used for subsequent linear mechanism model establishment and residual modeling analysis. The sampling period is set according to the control frequency of the unmanned vehicle's power module to ensure that the discrete modeling meets the Nyquist sampling condition.

3. The self-evolving digital twin modeling method for unmanned vehicle power modules according to claim 1, characterized in that: The specific steps for constructing the initial linear mechanism model in step two include: Based on the discrete-time data sequence constructed in step one, a first-order discrete linear mechanism model of the unmanned vehicle's power module is established, expressed in the following form: ； in, Indicates in The predicted output of the time-mapping linear model. and For the first The system outputs response signals and control input signals at specific times. , The parameters to be estimated are denoted as the combined effect of inertia and damping of the system and the input gain coefficient with respect to rotational speed, respectively. Constructing a regression vector : ； The Kalman gain is calculated using the Bayesian recursive estimation method for the linear model parameter vector. mean covariance matrix renew: ； ； ； in, This is the recursive Kalman gain matrix, used to adjust the parameter update magnitude; This represents the average value of the parameters at the previous time step. according to Parameter mean at time The linear twin model prediction output is obtained. for: 。 4. The self-evolving digital twin modeling method for unmanned vehicle power modules according to claim 3, characterized in that: Step 3, constructing and analyzing the residual sequence, includes the following steps: Based on the linear twin model prediction output obtained in step two and the actual system output The linear prediction residual of the linear mechanism model is expressed as: ； Statistical analysis is performed on the residual sequence to calculate the autocorrelation function, mean, and variance variation characteristics. When the residual sequence exhibits significant time-series correlation or non-stationary characteristics, it is determined that the power module has unmodeled nonlinear characteristics. Based on this, feature vectors are constructed for subsequent nonlinear residual modeling. The feature vectors include at least the current and historical residual values, the current and historical control inputs, and the change in control input.

5. The self-evolving digital twin modeling method for the power module of an unmanned vehicle according to claim 1, characterized in that: Step S41, the step of selecting the structural parameters of the gated recurrent unit network by improving the pigeon flock optimization method, specifically includes: To address the strategy switching process during the exploration and development phases of pigeon flock optimization methods, a dynamically changing balance factor is introduced. Used to uniformly regulate the intensity of population renewal: ； in, The maximum number of iterations, This represents the current iteration number; and It is a control balance factor The hyperparameters for the value range are set to 0.7 and 0.3 respectively; To handle boundary violations during the exploration process, a rollback-style boundary mapping strategy is adopted: ； in, Indicates the first During the first iteration update process, the... The individual in the first The parameter value of the dimension; Indicates the first Individuals that go out of bounds during the iterative update process Parameter values ​​after rollback mapping; The boundary elastic constant is ; the lower bound of the optimized parameter is . The upper boundary is ; To improve the quality of model initialization, an improved pigeon flock optimization method was used to perform offline global optimization of the GRU network parameters; the optimization process included: The GRU time window length, hidden layer dimension, and learning rate are used. The parameters are encoded as individual optimized particles, and a fitness function based on mean square error is constructed: ； in, This represents the position of an individual in the flock, i.e., the encoded network parameters to be optimized; The length of the training dataset; and These are the true residuals of the linear model and the predicted residuals of the GRU network in the training dataset, respectively. Entering the exploration and development phase with a balance factor; in this process, the first... Individual pigeons in a flock speed The update formula is expressed as: ； in, and They represent the first Only one pigeon in the 1st Velocity and position in the next iteration The position representing the pigeon with the highest current fitness level. It is a natural constant. For map factor constants, It is a random number between 0 and 1; No. Individual pigeons in a flock Location The update formula is expressed as: ； By calculating the fitness of each individual and ranking the fitness, the half of the pigeons with the highest fitness are retained, and their positions are placed in the center. For reference, the formula for updating the individual positions of pigeons in the second stage is obtained: ； After completing all iterations, the best individual is selected as the optimization result, and the GRU network is trained based on the optimized network structure parameters to establish a residual prediction model.

6. The self-evolving digital twin modeling method for the power module of an unmanned vehicle according to claim 1, characterized in that: In step S42, the offline training of the GRU residual model specifically includes: The core operations of the GRU network are performed by the update gate, the reset gate, and the candidate hidden states. GRU network training uses backpropagation and gradient descent to update parameters. ； in, for Time-bound GRU network weight parameters; The loss cost calculated using the cross-entropy loss function; Represents the loss function Weights The partial derivatives of , i.e., the corresponding gradient; Improve the learning rate obtained from pigeon flock optimization in step S41; A time-series residual prediction network based on gated recurrent units is used to perform nonlinear modeling on the obtained residual sequence. The GRU network takes a fixed-length time window as input, and the input features include the current and historical residual values, the current and historical control inputs, and the change in control input. A time-series sample sequence is formed by a sliding window method, and the predicted residual value for the next time step is output.

7. The self-evolving digital twin modeling method for the power module of an unmanned vehicle according to claim 1, characterized in that: In step S43, the specific design and usage methods of the gating factor include: Design gating factors It is used to adaptively adjust the residual compensation strength; the gating factor is adaptively determined based on the current residual statistical characteristics: ； in, For linear model residuals; To smooth out parameters, to avoid instability of the gating factor value when the error is close to zero, and to control the sensitivity of the gating response; The output of the linear twin model and the residual prediction output are combined through a gating fusion mechanism to obtain the fused digital twin model of the autonomous vehicle's power module. Predicted output at time step for: ； in, Predict the output for a linear model; This is the residual prediction output of the GRU network; This is the gating factor.

8. The self-evolving digital twin modeling method for the power module of an unmanned vehicle according to claim 1, characterized in that: Step S51, specifically the online model consistency assessment, includes: exist At time 1, based on the residual sequence obtained from the prediction error, the calculation is performed on a fixed-length sequence. The statistical indicators within the evolution time window, using root mean square error. As a consistency assessment indicator: ； The consistency between digital twin simulation data and measured data is evaluated periodically online by calculating RMSE. If the RMSE value of two sequence data is greater than the set threshold, the credibility of the digital twin is considered to have decreased significantly, triggering the execution of the self-evolutionary update method. When the statistical index is lower than the threshold, the current model parameters are kept unchanged to avoid unnecessary parameter oscillations under stable conditions.

9. The self-evolving digital twin modeling method for the power module of an unmanned vehicle according to claim 1, characterized in that: In step S53, the online update of model parameters specifically includes: using the Bayesian recursive estimation method to update the parameters of the linear mechanism model, and using the continuous learning constraint based on elastic weights to update the parameters of the GRU residual model online. Specifically, during the online update of the residual model, a continuous learning constraint based on stable elastic weights is introduced to reduce the damage to the original model parameters caused by learning new tasks; this includes: exist At any given time, for parameter updates of the GRU residual prediction model, online updates employ a window-based training mechanism, which involves constructing a sequence sample set within a fixed-length time window. : ； in, The time window length, It is a time-series feature vector that includes historical residuals, control inputs, and input increments. The true residuals of the linear model prediction results; The task loss function is constructed based on the prediction error of samples within the window. : ； in, The prediction residuals of the GRU network; To suppress catastrophic forgetting in the model during continuous evolution across multiple operating conditions, elastic weights based on gating factors are introduced to stabilize the GA-EWC constraint terms, and then combined with the basic construction of the task loss function. The loss function for the GRU network parameter optimization problem is obtained. for: ； in, These are the parameters in the current GRU network; Historical parameter values; GRU network parameters The corresponding values ​​in the Fisher information matrix are simplified to a diagonal information matrix; The total number of parameters required for the dynamic self-evolution of the GRU network model is [not specified]. The gating factor obtained from step S43 Adaptive adjustment of canonical intensity coefficient: ； in, The initial canonical intensity coefficient; because As time changes, the GRU network parameter optimization problem becomes an online optimization problem under time-varying parameter constraints, and the optimal solution is expressed as: 。 10. The self-evolving digital twin modeling method for the power module of an unmanned vehicle according to claim 1, characterized in that: In step S54, after the parameters are updated, the specific steps for evaluating the model evolution effect are as follows: the updated model parameters are used to re-predict the samples in the current sliding window, the corresponding window prediction error is calculated, and it is compared with the error level before the update: when the updated prediction error is less than or equal to the error before the update, the updated model parameters are retained and the subsequent running stage continues; when the updated prediction error is greater than the error before the update, the model parameters before the update are restored.

Images