Self-balancing two-wheeled vehicle simulation design method based on digital twinborn model

Abstract

The invention relates to the technical field of vehicle simulation design and automatic control, and discloses a self-balancing two-wheeled vehicle simulation design method based on a digital twin model, which comprises the following steps: S1, constructing and calibrating the digital twin model, and calculating to obtain a model credibility inverse index for quantifying the uncertainty of the digital twin model; s2, adaptively adjusting and optimizing algorithm parameters according to the credibility inverse index so as to search a robust design space meeting constraints; s3, implementing bidirectional evolution of the model and the optimization target, namely updating the optimization target according to the difference between the model and actually measured data, and expanding the model function according to the newly added target; and S4, outputting a robust design space and recommending a scheme, and performing final verification by using the evolved digital twinborn model. According to the self-balancing two-wheeled vehicle simulation design method, a digital twinning optimization closed loop capable of sensing the credibility of the self-balancing two-wheeled vehicle and self-evolving is constructed, so that the efficiency and the adaptability of the self-balancing two-wheeled vehicle simulation design and the robustness and the reliability of a final scheme are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of vehicle simulation design and automatic control technology, in particular to a simulation design method for a self-balancing two-wheeled vehicle based on a digital twin model. BACKGROUND

[0002] Self-balancing two-wheeled vehicles, especially electric vehicle models with a front-rear dual-wheel layout, have the advantages of compact structure and flexible design, and integrate a self-balancing system based on sensor fusion and intelligent control algorithms to achieve automatic balance control of the vehicle in static, low-speed, medium-speed, and complex road conditions, greatly improving the stability and safety of driving, and showing broad prospects in personal transportation and special application fields. However, this type of vehicle is a typical multivariable, strongly coupled, and nonlinear underactuated system, and its dynamic balance completely depends on active control, which poses high requirements for the structural parameters and control system design of the vehicle. In order to achieve optimal comprehensive performance under the premise of ensuring safety and stability, a simulation design method based on a model has become a key link in the research and development process, and the application of digital twin technology provides strong support for achieving high-fidelity simulation and optimization.

[0003] In the prior art, a typical simulation design method based on a digital twin model generally includes the following steps: first, a high-fidelity multi-physics domain simulation model is established based on the physical principles of the vehicle; then, the model is calibrated once for all parameters based on experimental data collected from a physical prototype to improve the model accuracy as much as possible; finally, the calibrated model is used in combination with intelligent optimization algorithms (such as genetic algorithms, particle swarm optimization algorithms, etc.) to optimize the design parameters in order to find a set of design solutions that can make a specific performance indicator optimal.

[0004] Although the prior art improves design efficiency to some extent by constructing a high-precision model and combining optimization algorithms, there are still some deficiencies:

[0005] The existing design process generally regards the digital twin model as a static, absolutely reliable true value substitute, which leads to a lack of awareness of model uncertainty in the optimization process, affecting efficiency and reliability. The fundamental reason is that any model is only an approximation of physical reality, and initial calibration can only guarantee accuracy under specific working conditions. The Model-Reality Gap is objectively existing. However, traditional optimization algorithms do not know the confidence level of the current model, and they treat all simulation results equally, which may result in a large amount of computing resources being wasted on over-fine optimization of an inaccurate model, or even converging to a false optimal solution that does not conform to physical reality.

[0006] In addition, the existing design process is usually a linear, one-way execution process, lacking the ability to dynamically adapt and self-improve. This is because the model and optimization objectives are fixed before optimization begins. When physical phenomena that the initial model fails to capture (such as specific vibrations) are discovered during the later stages of design, or new design requirements arise (such as the need to increase comfort considerations), the entire process must be interrupted, requiring engineers to manually analyze, modify the model, or reconfigure the optimization problem, and then restart. This rigid workflow cannot automatically respond to unknown problems and changes in requirements, and is inefficient and lacks adaptability.

[0007] Finally, most existing technologies focus on finding a single design point that optimizes performance indicators, and this optimization is often fragile. The reason for this is that this point optimization method ignores the uncertainties that are ubiquitous in the real world, such as manufacturing tolerances, material property variations, and environmental disturbances. SUMMARY

[0008] In view of the shortcomings of the prior art, the present application provides a self-balancing scooter simulation design method based on a digital twin model, which solves the problem of the simulation design method based on a static digital twin model in the prior art, which is blind in the optimization process and cannot dynamically adapt to unknown working conditions and new requirements, resulting in insufficient reliability and robustness of the design scheme.

[0009] To achieve the above purpose, the present application is implemented by the following technical scheme: a self-balancing scooter simulation design method based on a digital twin model, comprising the following steps:

[0010] S1, constructing and calibrating a digital twin model with credibility quantification, calibrating the digital twin model using the running data of a physical prototype, and calculating a model credibility inverse index quantifying the uncertainty of the digital twin model;

[0011] S2, performing adaptive optimization based on model credibility and robust design space exploration, dynamically adjusting the running parameters of the optimization algorithm according to the model credibility inverse index, and searching for a robust design space that satisfies the constraints of the pre-set key performance indicators under pre-set test scenarios;

[0012] S3, implementing bidirectional dynamic evolution of the twin model and optimization objectives, dynamically updating the optimization objectives according to the differences between the digital twin model and the physical prototype running data, and dynamically expanding the functions of the digital twin model according to the new optimization objectives;

[0013] S4, outputting and verifying the final design scheme, outputting the robust design space and recommending a design scheme, and using the dynamically expanded digital twin model to perform final verification on the recommended design scheme.

[0014] Preferably, the step of constructing and calibrating the digital twin model with credibility quantification comprises:

[0015] The Extended Kalman Filter algorithm is used to estimate the state and parameters of the digital twin model online according to the running data of the physical prototype, and output the posterior covariance matrix;

[0016] The trace of the posterior covariance matrix is taken as the inverse indicator of model credibility.

[0017] Preferably, the digital twin model is a multi-physical-domain unified model, including a mechanical system dynamics model, a driving system model, and a sensing system model.

[0018] Preferably, the step of dynamically adjusting the operating parameters of the optimization algorithm comprises:

[0019] The particle swarm optimization algorithm is adopted, and the inertia weight and / or learning factor of the particle swarm optimization algorithm are adaptively adjusted according to the value of the model credibility inverse indicator.

[0020] Preferably, the robust design space is a set of design parameters, and any set of design parameters in the parameter set can make the key performance indicators of the self-balancing two-wheeled vehicle meet the preset threshold in all test scenarios.

[0021] Preferably, the step of dynamically updating the optimization objective according to the difference between the digital twin model and the running data of the physical prototype comprises:

[0022] The innovation sequence in the Extended Kalman Filter algorithm is monitored;

[0023] When there is a feature representing unmodeled abnormal dynamics in the innovation sequence, a penalty term is automatically generated and added to the optimization objective.

[0024] Preferably, the feature representing unmodeled abnormal dynamics is a sustained energy peak in a specific frequency band after spectral analysis of the innovation sequence.

[0025] Preferably, the step of dynamically expanding the function of the digital twin model according to the new optimization objective comprises:

[0026] When the key performance indicators relied on by the new optimization objective require physical quantities that cannot be provided by the current digital twin model, corresponding sub-models are called from a preset sub-model library and integrated into the digital twin model.

[0027] Preferably, the step of outputting and verifying the final design scheme comprises:

[0028] perform dimensionality reduction analysis and visual presentation on the robust design space;

[0029] perform virtual acceptance test on the recommended design scheme by using the digital twin model finally formed after the bidirectional dynamic evolution steps of the twin model and the optimization target.

[0030] Preferably, the method further comprises:

[0031] defining the test scene through a human-computer interaction interface, the test scene including boundary test working conditions, fault injection scenes and / or user experience scenes.

[0032] The application provides a simulation design method for self-balancing two-wheeled vehicles based on a digital twin model.

[0033] 1. The application improves the efficiency and reliability of simulation design by constructing a digital twin model with credibility quantification and dynamically adjusting the operation parameters of the optimization algorithm based on the credibility index.

[0034] 2. The application establishes a bidirectional dynamic evolution mechanism of the twin model and the optimization target, greatly enhancing the adaptability and automation level of the design system.

[0035] 3. The application effectively improves the robustness and reliability of the final design scheme by exploring the robust design space as the optimization target. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 The method flowchart of the application is shown;

[0037] Figure 2 The adaptive optimization flowchart of the application is shown;

[0038] Figure 3 This is a schematic diagram of the bidirectional dynamic evolution mechanism of the twin model and the optimization target in this invention;

[0039] Figure 4 This is a schematic diagram of the robust design space visualization and final verification process of the present invention. Detailed Implementation

[0040] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Please refer to the appendix. Figure 1 , Figure 1 This is a schematic diagram of a method flow according to an embodiment of the present invention. The present invention provides a simulation design method for a self-balancing two-wheeled vehicle (specifically referring to an electric vehicle with a front and rear dual-wheel layout) based on a digital twin model. This method aims to establish a design system with deep coupling and dynamic evolution between the physical entity and the virtual model, so as to systematically improve design efficiency and the robustness of vehicle performance.

[0042] The system upon which the method of this invention relies may include a physical system, a data interaction module 10, and a digital twin and optimization system 20. The physical system is a self-balancing two-wheeled prototype vehicle used for data acquisition, equipped with IMU sensors, wheel speed encoders, motor drivers, steering drivers, and brake drivers, etc., to collect operational data from the physical prototype vehicle. This operational data includes vehicle tilt angle and angular velocity data collected by the inertial measurement unit (IMU), wheel speed and position data collected by the wheel speed encoders, and motor current data collected by the motor drivers, etc. In this embodiment, the basic physical parameters of the prototype vehicle are set as follows: total vehicle weight (excluding battery) 55 kg, initial estimated center of gravity height 0.5 m, front and rear wheelbase 1.3 m, and wheel diameter 0.4 m. The data interaction module 10 uses a Controller Area Network (CAN) bus to perform real-time data transmission between the physical prototype vehicle and the digital twin and optimization system 20 with a data update cycle of 100 Hz. The digital twin and optimization system 20 is the core execution unit of the method of this invention and is deployed on a computer running software such as MATLAB / Simulink.

[0043] like Figure 1 As shown, a specific embodiment of the present invention may include the following steps:

[0044] Step S1: Construct and calibrate the digital twin model with quantified credibility. This step aims to establish a high-fidelity virtual mapping body for the physical prototype vehicle and to establish a mechanism to quantify the accuracy and credibility of the virtual mapping body. First, based on the basic physical parameters of the prototype vehicle, a multi-physical domain unified simulation model including mechanical, electrical and sensing systems is established in the Simulink environment. Then, through the data interaction module 10, the actual running data of the physical prototype vehicle under a specific excitation signal (such as a small amplitude thrust disturbance) is collected, and the extended Kalman filter algorithm is used to dynamically calibrate the key parameters (such as the actual center of mass height) in the model.

[0045] During the calibration process, the algorithm not only outputs the estimated value of the parameter, but also outputs a posteriori estimation covariance matrix. The trace of this matrix is used by the present invention to derive an inverse indicator of model credibility. For example, before the calibration starts, due to the uncertainty of the actual center of mass height, its initial variance corresponding to the covariance matrix can be set to a larger value, such as 0.05 square meters. After 10 seconds of effective running data are collected and processed, the EKF algorithm converges, the estimated value of the center of mass height stabilizes, and its corresponding variance decreases to a smaller value, such as 0.0001 square meters.

[0046] At this time, the trace of the entire covariance matrix, i.e. the inverse indicator of model credibility, also decreases from a larger initial value (such as 0.1) to a smaller stable value (such as 0.01). The smaller the value of this indicator, the smaller the overall variance of the model parameter estimation, i.e. the higher the consistency of the digital twin model with the physical reality, and the higher the credibility of the model.

[0047] Step S2: Perform adaptive optimization and robust design space exploration based on model credibility. The digital twin and optimization system 20 performs optimization of design parameters based on the digital twin model with quantified credibility established in step S1. The optimization strategy of this step can be dynamically adjusted according to the credibility of the model.

[0048] Specifically, this step uses the particle swarm optimization algorithm to optimize the proportional (Kp), integral (Ki) and derivative (Kd) gains of the controller. The inertia weight and learning factor of this algorithm are dynamically related to the inverse indicator of model credibility obtained in step S1. For example, a threshold can be set, when the credibility inverse indicator is greater than 0.1 (indicating that the model is not accurate), the inertia weight is set to 0.9 to enhance the global exploration ability; when the indicator is less than 0.01 (indicating high model credibility), the inertia weight is set to 0.4 to enhance the local search ability.

[0049] In addition, the optimization goal of this step is to explore and output a robust design parameter space. To this end, a set of test scenarios are predefined, for example: scenario one, the vehicle passes over a 5 cm high step at a speed of 3 m / s; scenario two, a lateral force of 20 N is applied at the center of mass for 0.2 seconds when the vehicle is in static equilibrium. The constraint condition for optimization is that the maximum inclination angle of the vehicle body should not exceed 15 degrees under all the above scenarios. The goal of the optimization algorithm is to find a space of Kp, Ki, Kd parameter combinations that satisfy this constraint condition.

[0050] Step S3: Implementing bidirectional dynamic evolution of twin models and optimization goals. This step aims to establish a self-improving design system. The evolution process includes two directions:

[0051] First, during the model calibration process in step S1, by performing spectral analysis on the innovation sequence of the EKF algorithm, if a persistent energy peak is found in the 20-25 Hz frequency band, the system determines that there is an unmodeled vehicle frame resonance. At this time, the system will automatically add a penalty term to the optimization goal in step S2, which is used to suppress the energy in the controller output in this frequency band, thereby guiding the optimization process to avoid control parameters that will excite this resonance.

[0052] Second, when the designer introduces a new goal to improve ride comfort in step S2, the system analyzes and finds that the existing rigid body model cannot evaluate this goal. At this time, the system will automatically call a parameterized tire spring-damper model from the model library and integrate it into the digital twin model, making the comfort (such as vehicle body vertical acceleration) evaluatable and optimizable.

[0053] Step S4: Output and verify the final design scheme. After the optimization process converges, the digital twin and optimization system 20 outputs the final robust design parameter space (such as in the form of a feasible region map on the Kp-Kd plane), and according to the secondary criterion of lowest energy consumption, recommends a set of optimal design parameter combinations, such as Kp=50, Ki=5, Kd=15. Finally, the digital twin model containing the tire model obtained through the final evolution in step S3 is used to perform a comprehensive virtual acceptance test on the recommended scheme, confirming that the maximum inclination angle is less than 15 degrees under all the test scenarios defined in step S2, thereby completing the design verification.

[0054] In summary, the present application forms a complete technical closed loop from model construction, credibility quantification, adaptive optimization to bidirectional evolution through the above steps, thereby achieving efficient and highly robust simulation design of self-balancing two-wheeled vehicles.

[0055] The first step of the method of the present application is S1: constructing and calibrating a credibility-quantified digital twin model. The purpose of this step is to establish a digital twin for the physical prototype vehicle that can accurately reproduce its multi-physical domain dynamic characteristics in a virtual environment, and to establish a mechanism that can correct the twin online and in real time, while quantifying its degree of conformity with the physical reality.

[0056] Specifically, this step can be decomposed into the following sub-steps:

[0057] S110, perform multi-physical domain unified modeling. This step aims to describe the dynamic behavior of different physical domains such as mechanics, electricity, and sensors involved in the self-balancing two-wheeled vehicle in a unified mathematical framework to form an integrated simulation model with high fidelity.

[0058] S111, establish a mechanical system dynamics model. For a self-balancing two-wheeled vehicle with a front and rear wheel layout, the core motion is the balance in the longitudinal plane. This embodiment uses the Lagrange equation to model it systematically. First, define the generalized coordinate vector that describes the pose of the system, which can include the longitudinal inclination of the vehicle body relative to the vertical direction, the longitudinal displacement of the vehicle on the ground, the rotation angle of the front and rear wheels relative to the vehicle body, etc. The dynamic characteristics of the system are described by the following nonlinear matrix equation:

[0059] ;

[0060] where, , , are the generalized coordinate vector, generalized velocity vector and generalized acceleration vector of the system, respectively; is the symmetric positive definite mass inertia matrix of the system, whose elements are functions of the generalized coordinates and are uniquely determined by the design parameters of the vehicle, such as the mass, center of mass position, moment of inertia tensor of each component, etc.; is the vector containing the Coriolis force and centrifugal force, which describes the coupling effect between the components due to rotational motion; is the gravity term vector, which represents the force or torque generated by gravity on each generalized coordinate; is the torque vector applied to the system, which is mainly derived from the output of the drive system.

[0061] S112, establish a drive and sensor system model. To make the model complete, the driving system that provides power and the sensing system that provides state feedback need to be mathematically described. For a typical DC motor drive system, its electrical equation and electromechanical conversion relationship can be represented as:

[0062] ;

[0063] where, , , , are the armature voltage, armature current, armature resistance and inductance of the motor, respectively; , are the back EMF coefficient and torque coefficient of the motor, respectively; is the angular velocity of the motor rotor, which is directly related to the wheel speed; is the electromagnetic torque output by the motor, which, after passing through the transmission system, constitutes the main driving term of the force moment vector in the above dynamics equation.

[0064] For the sensing system, take the inertial measurement unit (IMU) as an example. Its output is not the true value, but contains various errors. Its measurement model can be expressed as:

[0065] ;

[0066] where, , are the angular velocity measured by the gyroscope and the inclination angle calculated by the accelerometer, respectively; , are the true angular velocity and true inclination angle of the vehicle body inclination angle, respectively; , are the zero offset errors of the gyroscope and accelerometer, which will change slowly over time; , are the random measurement noises that follow a certain statistical distribution (such as Gaussian distribution).

[0067] S113, integrate the coupling of each subsystem model. The mechanical dynamics model of S111 and the driving and sensing model of S112 are coupled to form a complete system model composed of a set of ordinary differential equations and algebraic equations. This integrated model can comprehensively describe the whole process from the input of control instructions to the output of vehicle dynamic response.

[0068] S120, implement data-driven model dynamic calibration and credibility evaluation. This step is the key to ensure that the digital twin model maintains high fidelity throughout the vehicle's life cycle. It uses real data collected from the physical sample vehicle to continuously correct model parameters and quantify the uncertainty of the correction results.

[0069] S121, construct an augmented state space model. In order to estimate the dynamic state of the system and correct the static parameters of the model at the same time, an augmented state vector This vector not only contains traditional state variables such as tilt angle and angular velocity, but also model parameters that need to be identified online, such as the actual center of gravity position of the vehicle, the moment of inertia of key components, and the friction coefficient between the tires and the ground—parameters that are uncertain during initial modeling or may change over time. The state-space model of the system can be written in the following nonlinear discrete-time form:

[0070] ;

[0071] In the formula, for The augmented state vector at time step; For the previous time step The augmented state vector; For the previous time step The control input vector applied to the system; This is the process noise vector; for The control input vector constantly applied to the system, such as the motor drive voltage; for The vector of actual measurement data constantly collected from sensors (such as IMUs) on the physical prototype vehicle; It is a nonlinear state transition function, obtained by discretizing the complete system model established in S110; It is a nonlinear measurement function that describes the mathematical relationship between the augmented state vector and the sensor measurement value; , These are process noise and measurement noise, respectively, used to describe the uncertainty of the model and the interference of the measurement.

[0072] S122, performs extended Kalman filtering (EKF) and quantifies model confidence. This incorporates real measurement data from the physical prototype. As input, the Extended Kalman Filter (EKF) algorithm is used to recursively estimate the state of the aforementioned state-space model. The EKF algorithm iterates through both the prediction and update phases by linearizing the nonlinear function, resulting in an estimated value of the augmented state vector. It can best approximate the true value.

[0073] During this process, not only are the model parameters corrected online, but the EKF algorithm also calculates and outputs the posterior estimated covariance matrix. The diagonal elements of this matrix represent the variance of the estimated value of each component in the augmented state vector.

[0074] This invention utilizes the covariance matrix To quantify the overall credibility of current digital twin models, specifically, we define an inverse credibility index:

[0075] ;

[0076] This is the inverse index of model credibility. This scalar value is used to quantify the overall uncertainty of the digital twin model; the smaller the value, the higher the model credibility. The trace operator is a matrix operator, defined as the sum of the elements on the main diagonal of a square matrix. for The posterior estimated covariance matrix at time t.

[0077] In summary, this indicator By calculating the trace of the covariance matrix, the variances of all states and parameter estimates are synthesized, thus providing a clear and computable quantitative measure of the overall reliability of the model. This quantitative metric will serve as a key input for adaptive optimization in subsequent steps.

[0078] Please refer to the appendix. Figure 2 , Figure 2 This is a schematic diagram of an adaptive optimization process according to an embodiment of the present invention. The second step of the method of the present invention is S2: performing adaptive optimization and robust design space exploration based on model credibility. In this step, based on the quantified digital twin model of credibility constructed in step S1, an intelligent optimization algorithm is used to search for the optimal design parameters.

[0079] The core of this step is that its optimization process is not static, but can dynamically adjust its search strategy according to the degree of conformity between the digital twin model and physical reality. Furthermore, its optimization goal is to obtain a range of design parameters that can perform robustly under various harsh operating conditions, namely, a robust design space.

[0080] Specifically, this step can be broken down into the following sub-steps:

[0081] S210, Interactively define boundary test, fault injection scenarios, and / or user experience scenarios. To ensure the robustness and overall performance of the design, optimization needs to be performed in a virtual test environment that covers potential operational risks and typical usage environments. This embodiment allows designers to parameterize one or more sets of test scenarios based on design specifications, regulatory requirements, engineering experience, or user needs through a human-computer interaction interface. These scenarios may include boundary test conditions, fault injection scenarios, and / or user experience scenarios. This set of scenarios constitutes a collection. :

[0082] ;

[0083] This is the set of test scenarios, which is the total number of virtual test cases upon which subsequent robustness optimization and evaluation are based; (in = 1,2,..., is the i-th independent boundary test or fault injection scenario in the set , each of which corresponds to a specific set of parameterized input conditions to simulate the vehicle's running state under a certain extreme working condition or fault mode; is the total number of scenarios in the set , which is a positive integer.

[0084] The human-computer interaction interface can provide graphical tools to enable designers to intuitively set various test conditions. For example:

[0085] For boundary test working conditions, the vehicle can be set to pass a 5 cm high step at a speed of 3 m / s;

[0086] For fault injection scenarios, a scenario can be set in which a sensor signal is lost, with parameters including the start time and duration of signal loss;

[0087] For user experience scenarios, the vehicle can be set to travel at a specific speed on a B-class virtual road surface that meets the ISO 8608 standard to evaluate the ride comfort of the scheme.

[0088] S220, construct the objective function of the robust design space. The present application expands the optimization objective to find a multi-dimensional design parameter space, i.e. the robust design space . This space is composed of key design parameters of the vehicle (such as controller gain parameters, etc.). Any point in this space, i.e. any set of design parameters, must satisfy a robustness constraint condition.

[0089] This constraint condition requires that the scheme using this set of design parameters, when subjected to all test scenarios defined in S210 , the values of all pre-set key performance indicators (KPIs) must not exceed their respective pre-set performance or safety thresholds. These key performance indicators are quantitative descriptions of vehicle performance, for example, the stability indicator can be quantified as the maximum tilt angle of the vehicle body after being subjected to an impact, and the ride comfort indicator can be quantified as the root mean square (RMS) value of the vehicle body vertical acceleration under a specific user experience scenario.

[0090] Therefore, the optimization objective can be formally described as maximizing the volume or boundary of the robust design space, where the definition is:

[0091] ;

[0092] In the formula, is the robust design space, which is a set of design parameter vectors that satisfy all constraint conditions; is a design parameter vector, representing a specific design scheme or design point; is a set of pre-defined boundary test cases and fault injection scenarios; is a set of specific test scenarios in ; is a performance index function, whose calculated value is the quantified result of the th key performance indicator (KPI) when the scheme with design parameter vector experiences the test scenario ; is the safety or performance threshold preset for the th key performance indicator; is a universal quantifier in logic, representing for all or any one; is the quantified result of the key performance indicator; is a set theory symbol, representing belonging to.

[0093] S230, performing model confidence based adaptive particle swarm optimization. To efficiently search and delineate the boundary of the robust design space in the huge multi-dimensional parameter space, this embodiment adopts an adaptive particle swarm optimization (PSO) algorithm. This algorithm simulates the social behavior of particle swarm to find the optimal solution.

[0094] The key of this adaptive PSO algorithm lies in that its internal parameters are not fixed but dynamically associated with the model confidence inverse indicator calculated in step S122. In each iteration of the PSO algorithm, the speed and position update of each particle (representing a design scheme) follows the following rules:

[0095] ;

[0096] wherein is the speed vector of the particle in the th iteration (i.e. the next iteration); is the index number of the particle in the particle swarm; is the current iteration number of the optimization algorithm; is the inertia weight, which is a function of the model confidence inverse indicator , used to adjust the influence of the previous speed on the current speed; is the model confidence inverse indicator calculated in step S1; is the position of the particle in the velocity vector at the current iteration (i.e., the next iteration); is a cognitive learning factor whose value is a function of to adjust the tendency of the particle to fly towards its own historical best position; is a uniformly distributed random number in the interval [0, 1]; is the historical best position vector of the particle experienced by the particle from the start to the current iteration; is the historical best position vector of the particle at the first iteration; is the current position vector of the particle at the i-th iteration, which is a set of candidate design parameters; is a social learning factor whose value is a function of to adjust the tendency of the particle to fly towards the historical best position of the entire swarm; is a uniformly distributed random number in the interval [0, 1] and independent of ; is the global historical best position vector found by the entire swarm from the start to the current iteration.

[0097] The design principle of this function relationship is as follows: when the value of the current digital twin model is high, and the simulation results can well reflect the physical reality. At this time, the parameter combination output by the function , , will enhance the local search ability of the algorithm

[0098] For example, a smaller inertia weight and a larger learning factor will prompt the particle swarm to conduct fine and deep mining in the discovered excellent area, so as to accelerate the convergence. On the contrary, when the value of the model is large, indicating that there is a large uncertainty in the model, and the simulation results may deviate from the reality.

[0099] At this time, the parameter combination output by the function will enhance the global exploration ability of the algorithm (for example, a larger inertia weight), so that the particle can explore in the design space with a larger step, avoiding the algorithm from prematurely falling into a false optimal region on the basis of an inaccurate model. This adaptive mechanism enables the optimization process to intelligently balance exploration and utilization, thereby improving the efficiency and reliability of optimization under high uncertainty.

[0100] Please refer to the attached Figure 3 , Figure 3This is a schematic diagram of the bidirectional dynamic evolution mechanism of the twin model and the optimization objective according to an embodiment of the present invention. The third step of the method of the present invention is S3: implementing the bidirectional dynamic evolution of the twin model and the optimization objective. This step elevates the entire design process from a linear, unidirectional execution process to a dynamic closed-loop system that can self-improve and continuously evolve.

[0101] The core of this step is that the digital twin model itself and the optimization objective function used to guide the design are no longer static and fixed, but can evolve bidirectionally and collaboratively during the design process based on new information and requirements.

[0102] Specifically, this step can be broken down into the following two parallel sub-steps:

[0103] S310 implements the optimization objective evolution driven by model evolution. This sub-step aims to enable the design system to automatically discover and address unknown physical phenomena not considered in the initial model, ensuring the comprehensiveness of the optimization direction. This process builds upon the Extended Kalman Filter (EKF) calibration mechanism in step S122. During the update phase of the EKF algorithm, a key intermediate quantity, the Innovation Sequence, is calculated. .

[0104] New sequence The defining formula is:

[0105] ;

[0106] In the formula, yes Real sensor measurements collected from the physical prototype vehicle at all times; It is to utilize State estimates at time t Predictions made from time sensor measurements; Index of the discrete time step; It is a nonlinear measurement function; for The prior state estimation vector at time t.

[0107] In an ideal scenario where the model is perfectly accurate, the innovation sequence It should be a zero-mean white noise sequence. However, when the physical prototype exhibits dynamic behaviors that the digital twin model fails to describe (e.g., high-frequency structural resonance of the chassis at specific speeds), this unmodeled dynamics are systematically reflected in the innovation sequence. In this way, it no longer exhibits white noise characteristics.

[0108] This embodiment includes a news sequence monitoring module, which continuously monitors... online signal processing and analysis. One specific implementation is to take the Fourier Transform (FFT) of the innovation sequence within a time window, and obtain its Power Spectral Density (PSD). If there is a persistent energy peak in a certain frequency band of the PSD that is far beyond the normal noise level, the system determines that an un-modeled abnormal dynamic has been detected.

[0109] Once the abnormality is detected, the system will automatically trigger an optimization objective augmentation mechanism. This mechanism will generate a new penalty term and dynamically add it to the integrated performance index function in step S220. The penalty term aims to suppress the control action that can excite this abnormal dynamic. For example, the penalty term can be defined as the energy integral of the controller output signal in the abnormal frequency band:

[0110] ;

[0111] where, is the dynamic penalty term; this term is a scalar value that quantifies the adverse effect of the un-modeled abnormal dynamic detected in step S310, and incorporates it into the optimization objective function; is the angular frequency; this is the independent variable in signal spectral analysis; is the lower limit angular frequency of the abnormal frequency band; this is the starting frequency of the frequency band where the persistent energy peak is detected when performing spectral analysis on the innovation sequence; is the upper limit angular frequency of the abnormal frequency band; this is the ending frequency of the frequency band where the persistent energy peak is detected when performing spectral analysis on the innovation sequence; is the Power Spectral Density (PSD) of the controller output signal ; the PSD function describes how the power of the signal is distributed over frequency; is the controller output signal; this signal is a function of time, and is the physical command (e.g. voltage or current signal) that the control system computes and outputs to the actuator (e.g. motor driver); is the differential of the angular frequency; this is the infinitesimal of the integration operation.

[0112] By adding to the optimization objective, the adaptive optimization algorithm in step S2 will naturally tend to find design parameters (especially controller parameters) that can avoid exciting the resonance in this frequency band during the subsequent search process, thus avoiding potential physical problems in the design phase.

[0113] S320, model evolution driven by the evolution of the optimization objective is implemented. This sub-step aims to endow the design system with the ability to respond to new design requirements, even if these requirements are beyond the description of the initial model. The process begins with the designer introducing a new performance optimization objective through the human-machine interface of step S210.

[0114] For example, the designer may add an optimization requirement about ride comfort at the later stage of design, and quantify it as a new key performance indicator , which may be directly related to the vertical acceleration of the vehicle body. At this time, the system first triggers a model dependency checking module. The model dependency checking module parses the necessary physical quantities (in this example, the vertical acceleration of the vehicle body) and checks whether the digital twin model currently constructed in step S110 has the ability to output these physical quantities.

[0115] If the initial model is a rigid body model, it may not be able to directly calculate the vertical vibration caused by road roughness. At this time, the result of the dependency check is missing. The system then triggers a model dynamic expansion module. The dynamic expansion module connects a pre-constructed, parameterized sub-model library. The library stores a variety of pluggable physical effect models, such as spring-damper tire models, simple suspension system models, motor torque ripple models, etc.

[0116] The model dynamic expansion module automatically selects one or more appropriate sub-models (for example, adding spring-damper contact models to the front and rear wheels) in the library according to the missing physical quantities. Then, the model dynamic expansion module automatically handles the interface matching and coupling problem between the new and old models, seamlessly integrating the new sub-models into the existing multi-body dynamics equation set. In this way, the capabilities of the digital twin model are dynamically expanded, making the new performance indicator become computable and evaluable. At this point, the adaptive optimization algorithm of step S2 can include this new design dimension in its optimization process to find a design solution that best balances comfort and other performance.

[0117] Through the collaborative work of the two bidirectional evolution paths of S310 and S320, the invention builds a design environment that can learn and grow, enhancing its ability to deal with complexity and uncertainty.

[0118] Please refer to the attached Figure 4 , Figure 4 is a robust design space visualization and final verification process diagram according to an embodiment of the invention. The fourth step of the method of the invention is S4: output and verify the final design solution.

[0119] ​This step is the convergence and delivery stage of the whole design process, aiming to present the complex, multi-dimensional optimization results in previous steps to the designers in a clear and usable form, and to conduct a comprehensive and high-confidence virtual acceptance test on the finally selected design scheme.

[0120] Specifically, this step can be decomposed into the following sub-steps:

[0121] S410, output the robust design space and recommended scheme. After the adaptive optimization process in step S2 reaches the preset termination condition (for example, reaches the maximum number of iterations, the design space boundary converges stably or the performance improvement is lower than the threshold), the system outputs the final design result.

[0122] Since the optimization goal of the present application is to explore a robust design space , not a single design point, the primary part of the output result is the description and visualization of the multi-dimensional space . Directly presenting a high-dimensional space is difficult for designers to understand. Therefore, the present embodiment uses dimension reduction analysis and visualization techniques to present the space. A specific implementation is to select two or three core design parameters that are most concerned by the designers as coordinate axes (for example, controller proportional gain and differential gain ), and then fix other parameters at the center value of their optimal region.

[0123] The system will calculate and plot the feasible region or safety zone that satisfies all robustness constraints in the plane or solid space formed by the two or three parameters. The region map can intuitively tell the designers how large the adjustable range of these core parameters is, i.e., how large the design margin is, under the premise of ensuring the safety and stability of the vehicle under all boundary tests.

[0124] At the same time of outputting the robust design space, the system also gives one or more sets of recommended optimal design schemes . All parameter points in the robust design space can meet the basic safety and stability requirements, but they may differ in other secondary performance indicators (such as energy consumption, response speed, manufacturing cost, etc.). The system can further search in the space according to the secondary optimization criteria preset by the designers (for example, pursuing the lowest energy consumption under the premise of meeting robustness), and output the optimal parameter combination under the criteria.

[0125] S420, perform closed-loop verification of the final scheme. After the designers select the final design scheme Afterwards, to ensure the quality of delivery, a final virtual validation is needed. This validation process is called Virtual Sign-off, whose purpose is to confirm the performance of the design solution to the maximum extent before putting into physical prototype manufacturing and field testing.

[0126] The unique thing about this Virtual Sign-off is that it uses the most complete and highest fidelity digital twin model that has been evolved in step S3. This model has not only been continuously calibrated with physical data, but also possibly dynamically integrated with extended sub-models to cope with abnormal dynamics or new design requirements. Therefore, the validation test based on this model has the highest confidence in its results.

[0127] The procedure of Virtual Sign-off is to configure the final selected design parameters into the final version of the digital twin model, and then let the virtual prototype run through all the boundary test and fault injection scenario sets defined in step S210 . The system will automatically record and generate a detailed test report, which clearly lists the specific values of the vehicle's key performance indicators (KPIs) under each severe test scenario, and compares them with the safety thresholds set during design .

[0128] When all indicators under all scenarios meet the requirements, the design solution is considered to have passed the Virtual Sign-off test and can enter the next stage of physical implementation. This step constitutes the final closed-loop verification of the entire design process, ensuring the reliability and robustness of the design output.

[0129] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and variations can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A simulation design method for a self-balancing two-wheeled vehicle based on a digital twin model, characterized in that, Includes the following steps: S1. Construct and calibrate a digital twin model for quantifying credibility, calibrate the digital twin model using the operating data of the physical prototype vehicle, and calculate the model credibility inverse index that quantifies the uncertainty of the digital twin model. S2. Perform adaptive optimization and robust design space exploration based on model credibility. According to the inverse index of model credibility, dynamically adjust the running parameters of the optimization algorithm, and search for a robust design space that meets the constraints of preset key performance indicators under the preset test scenario. S3. Implement bidirectional dynamic evolution of the digital twin model and optimization objectives. Based on the differences between the digital twin model and the physical prototype vehicle's operating data, dynamically update the optimization objectives and dynamically expand the functionality of the digital twin model based on the newly added optimization objectives. S4. Output and verify the final design scheme, output the robust design space and recommend the design scheme, and use the dynamically expanded digital twin model to finally verify the recommended design scheme.

2. The simulation design method for a self-balancing two-wheeled vehicle based on a digital twin model according to claim 1, characterized in that, The steps for constructing and calibrating a digital twin model with reliable metric quantification include: Using the extended Kalman filter algorithm, the state and parameters of the digital twin model are estimated online based on the operating data of the physical prototype vehicle, and the posterior estimated covariance matrix is ​​output. The physical prototype's operating data includes body tilt angle and angular velocity data collected using an inertial measurement unit, wheel speed and position data collected using a wheel speed encoder, and motor current data collected using a motor driver. The trace of the posterior estimated covariance matrix is ​​used as the inverse index of the model's credibility.

3. The simulation design method for a self-balancing two-wheeled vehicle based on a digital twin model according to claim 1, characterized in that, The digital twin model is a unified multi-physics domain model, including: a mechanical system dynamics model and a drive and sensing system model.

4. The simulation design method for a self-balancing two-wheeled vehicle based on a digital twin model according to claim 1, characterized in that, The steps for dynamically adjusting the running parameters of the optimization algorithm include: The particle swarm optimization algorithm is employed, and the inertia weights and / or learning factors of the particle swarm optimization algorithm are adaptively adjusted based on the value of the inverse confidence index of the model.

5. The simulation design method for a self-balancing two-wheeled vehicle based on a digital twin model according to claim 1, characterized in that, The robust design space is a set of design parameters. Any set of design parameters in the parameter set can ensure that the key performance indicators of the self-balancing two-wheeled vehicle meet the preset threshold values ​​in all the test scenarios.

6. The simulation design method for a self-balancing two-wheeled vehicle based on a digital twin model according to claim 1, characterized in that, The step of dynamically updating and optimizing the target based on the differences between the digital twin model and the physical prototype vehicle's operating data includes: Monitoring the innovation sequence in the extended Kalman filter algorithm; When the innovation sequence contains features that represent unmodeled anomalous dynamics, a penalty term is automatically generated and added to the optimization objective.

7. The simulation design method for a self-balancing two-wheeled vehicle based on a digital twin model according to claim 6, characterized in that, The characteristic of the unmodeled anomalous dynamics is the persistent energy peak that appears in a specific frequency band after spectral analysis of the innovation sequence.

8. The simulation design method for a self-balancing two-wheeled vehicle based on a digital twin model according to claim 1, characterized in that, The step of dynamically expanding the functionality of the digital twin model based on the newly added optimization objectives includes: When the key performance indicators upon which the newly added optimization objective depends require physical quantities that the current digital twin model cannot provide, the corresponding sub-model is called from the preset sub-model library and integrated into the digital twin model.

9. The simulation design method for a self-balancing two-wheeled vehicle based on a digital twin model according to claim 1, characterized in that, The steps for outputting and verifying the final design scheme include: The robust design space is subjected to dimensionality reduction analysis and visualization. Using the digital twin model that is finally formed after the bidirectional dynamic evolution steps of the twin model and the optimization target, a virtual acceptance test is performed on the recommended design scheme.

10. The simulation design method for a self-balancing two-wheeled vehicle based on a digital twin model according to claim 1, characterized in that, The method further includes: The test scenarios are defined through a human-computer interaction interface, including boundary test conditions, fault injection scenarios, and / or user experience scenarios.

Images