Steering system model construction method based on digital twinning

Abstract

The invention relates to the technical field of digital twinning and system health state prediction, in particular to a steering system model construction method based on digital twinning. Comprising the steps of 1, constructing a high-fidelity physical mechanism model representing the dynamic characteristics of the steering system; acquiring a working condition data flow of the steering system in real time; constructing a state observer; resolving a residual signal through a state observer; step 2, constructing an augmented state vector; performing recursive estimation on the augmented state vector by adopting a nonlinear filtering algorithm; 3, injecting the online identification parameter sequence into the high-fidelity physical mechanism model in real time to realize adaptive updating of the physical mechanism model; collecting historical data of the online identification parameter sequence; training a degradation prediction model by using historical data; and predicting and outputting a degradation trend prediction time domain of the steering system. According to the method, the fidelity and credibility of the model are remarkably improved, and a solid foundation is laid for subsequent analysis and prediction.

Description

Technical Field

[0001] This invention relates to the field of digital twin and system health status prediction technology, specifically a method for constructing a steering system model based on digital twin. Background Technology

[0002] As the complexity and reliability requirements of modern engineering systems continue to increase, health status management throughout their entire life cycle has become crucial. Currently, most widely used system simulation models are static physical models, whose model parameters are usually determined once during the design phase through calibration. Although these static models can reflect physical characteristics well in the early stages of system operation, they cannot capture the dynamic performance degradation of physical entities caused by wear, material aging, or changes in operating conditions during long-term service. Over time, the discrepancy between the static model and physical reality will gradually increase, resulting in a model-reality mismatch problem. This causes the monitoring and analysis based on the static model to gradually lose fidelity, making it difficult to accurately assess the current real health status of the system, let alone make forward-looking predictions of future performance degradation trends, thus affecting the accuracy and timeliness of predictive maintenance decisions. The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention discloses a method for constructing a steering system model based on digital twins. Specifically, the technical solution of this invention includes: Step 1: Construct a high-fidelity physical mechanism model characterizing the dynamic characteristics of the steering system; acquire the working condition data stream of the steering system in real time; construct a state observer, which is based on the working condition data stream and the high-fidelity physical mechanism model; calculate the residual signal through the state observer, where the residual signal characterizes the difference between the model prediction value and the sensor measurement value. Step 2: Construct an augmented state vector, which combines system state variables and physical parameters to be identified; use a nonlinear filtering algorithm to recursively estimate the augmented state vector, where the recursive estimation utilizes residual signals; identify and output the physical parameters to be identified, forming an online identified parameter sequence. Step 3: Inject the online identification parameter sequence into the high-fidelity physical mechanism model in real time to achieve adaptive updating of the physical mechanism model; collect historical data of the online identification parameter sequence; train the degradation prediction model using the historical data; compare the degradation prediction model with the preset physical failure threshold, predict and output the time domain of the degradation trend prediction of the steering system.

[0004] Preferably, the high-fidelity physical mechanism model is constructed using second-order nonlinear ordinary differential equations, including: Determine the equivalent moment of inertia of the system; Set the steering column angle as a state variable; Define the physical parameters to be identified, including the damping coefficient, friction coefficient, and transmission clearance. The physical mechanism model uses the motor assist torque and the road surface reaction torque as system inputs.

[0005] Preferably, step one also includes: The system collects operating condition data streams in real time via bus and sensors. The operating condition data streams include values ​​from the angle sensor, torque sensor, current sensor, and temperature sensor. Perform timestamp alignment on the operating condition data stream; Denoise the operating data stream; The operating condition data stream is resampled to form an observation vector with a uniform sampling rate.

[0006] Preferably, the residual signal is solved through the following steps: Obtain the observation vector and use it as the sensor's measured vector; The model is used to calculate the predicted values ​​by solving the observation function, which includes the estimated angle and the estimated sum of internal resistance torques. The residual signal is obtained by subtracting the model observation prediction value from the sensor's measured vector.

[0007] Preferably, the nonlinear filtering algorithm includes a prediction step and an update step; The prediction step includes: Based on the posterior estimate and input from the previous time step, time steps are advanced through a nonlinear state transition function; The nonlinear state transition function originates from the discretization of the high-fidelity physical mechanism model; By advancing through time steps, the prior estimate and the prior covariance matrix are calculated.

[0008] Preferably, the update step of the nonlinear filtering algorithm includes: The prior estimate is corrected using Kalman gain and residual signal to obtain the posterior estimate; The physical parameters to be identified are extracted from the posterior estimate to form an online identification parameter sequence.

[0009] Preferably, the adaptive update of the physical mechanism model is achieved through the following steps: The identified physical parameters are fed back in real time and the original parameter values ​​in the high-fidelity physical mechanism model are replaced. This allows the high-fidelity physical mechanism model to use the updated parameters to calculate the residual signal in the next time step, thus forming a parameter closed-loop injection.

[0010] Preferably, the degradation prediction model is trained through the following steps: A time series forecasting model is used; The time series prediction model was trained offline using prior degradation datasets obtained from similar equipment or accelerated aging tests. The degradation prediction model also receives assumptions about future operating conditions as input to adjust the prediction results.

[0011] Preferably, the prediction of degradation trend in the time domain is achieved through the following steps: Define a preset physical failure threshold as a parameter limit vector characterizing physical failure; The degradation prediction model is used to output the predicted trajectories of each parameter; Real-time checks are performed to determine if the predicted trajectory of any parameter exceeds the corresponding physical failure threshold. The moment when the degradation trend is first exceeded is defined as the time domain for predicting the degradation trend.

[0012] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention constructs a high-fidelity physical mechanism model and combines it with real-time acquired operating condition data streams, utilizing a state observer to calculate the residual signal. This driving method, based on the difference between actual data and model predictions, ensures that the digital twin model can accurately reflect the rapidly changing real operating state of the steering system, significantly improving the model's fidelity and reliability, and laying a solid foundation for subsequent analysis and prediction.

[0013] 2. This invention employs augmented state vectors and nonlinear filtering algorithms to perform online identification and recursive estimation of the system state and key physical parameters of the steering system. This frees the model from dependence on fixed parameters, enabling it to dynamically capture changes in physical characteristics caused by factors such as wear and aging, achieving adaptive tracking of physical entities, and improving the model's accuracy.

[0014] 3. This invention injects the parameter sequence identified online into the high-fidelity physical mechanism model in real time, replacing the original parameter values, thus forming a closed-loop parameter injection mechanism. This adaptive update method ensures that the model always uses the latest and most realistic physical parameters in the calculation of the next time step, realizing continuous self-correction and iterative optimization of the model and guaranteeing the long-term synchronization between the digital twin and the physical entity.

[0015] 4. This invention utilizes historical data from online parameter sequence identification to train a degradation prediction model, and then compares and analyzes it against a preset physical failure threshold. This method not only reflects the current state but also proactively predicts the future degradation trajectory of key parameters in the steering system, accurately outputting the degradation trend prediction time domain. This provides a reliable basis for condition-based maintenance and predictive maintenance, effectively avoiding sudden failures and ensuring system safety. Attached Figure Description

[0016] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0018] Example 1 Please see Figure 1 A method for constructing a steering system model based on digital twins, comprising: Step 1: Construct a high-fidelity physical mechanism model characterizing the dynamic characteristics of the steering system; acquire the working condition data stream of the steering system in real time; construct a state observer, which is based on the working condition data stream and the high-fidelity physical mechanism model; calculate the residual signal through the state observer, where the residual signal characterizes the difference between the model prediction value and the sensor measurement value. Step 2: Construct an augmented state vector, which combines system state variables and physical parameters to be identified; use a nonlinear filtering algorithm to recursively estimate the augmented state vector, where the recursive estimation utilizes residual signals; identify and output the physical parameters to be identified, forming an online identified parameter sequence. Step 3: Inject the online identification parameter sequence into the high-fidelity physical mechanism model in real time to achieve adaptive updating of the physical mechanism model; collect historical data of the online identification parameter sequence; train the degradation prediction model using the historical data; compare the degradation prediction model with the preset physical failure threshold, predict and output the degradation trend prediction time domain of the steering system; This embodiment provides a method for constructing a steering system model based on digital twins. This method is a closed-loop system that realizes adaptive evolution of the model and prediction of its health status. A high-fidelity physical mechanism model characterizing the dynamic characteristics of the steering system is constructed. The purpose of this model is to establish an interpretable and high-precision system dynamic simulation core based on prior physical knowledge. In this embodiment, the model is a second-order nonlinear ordinary differential equation based on the torque balance equation of Newton's second law. The purpose of acquiring the steering system's operating condition data stream in real time is to obtain the actual operating state of the physical entity as a realistic benchmark for model correction. In this embodiment, the raw data stream is acquired in real time through the CAN bus and sensors, and after timestamp alignment, noise reduction and resampling processing, an observation vector under a uniform sampling rate is formed. Construct a state observer, which is based on the operating condition data stream and a high-fidelity physical mechanism model; the state observer refers to an algorithm, such as the Kalman filter detailed in the following embodiments, which uses the model and incomplete measurement data to estimate the internal state of the system. The residual signal is calculated by the state observer. The residual signal represents the difference between the model prediction and the sensor measurement. The residual signal refers to the difference between the model prediction and the actual sensor measurement. Its function is to quantify the deviation between the model and reality, and to serve as the driving signal for subsequent parameter identification. Construct an augmented state vector, which combines system state variables and physical parameters to be identified. To achieve joint estimation of state and parameters, the augmented state vector is an extended vector that combines the original system state variables such as angle and angular velocity with physical parameters to be identified such as damping and friction. Its role is to transform the parameter identification problem into a standard state estimation problem. A nonlinear filtering algorithm is used to recursively estimate the augmented state vector, where the recursive estimation utilizes the residual signal. Nonlinear filtering algorithms, such as the Extended Kalman Filter (EKF) or the Unscented Kalman Filter (UKF), are state estimation algorithms suitable for nonlinear systems, and their recursive estimation uses the residual signal as a correction term. This algorithm continuously corrects the estimation of the augmented state vector through two steps: prediction and update. The system identifies and outputs the physical parameters to be identified, forming an online parameter identification sequence. By extracting the parameter part from the recursively estimated posterior augmented state vector, the system can output the current value of these parameters in real time. These identified parameters, such as the damping coefficient and friction coefficient that change with time, are accumulated to form an online parameter identification sequence. The online identified parameter sequence is injected into the high-fidelity physical mechanism model in real time to achieve adaptive updating of the physical mechanism model. Real-time injection means replacing the original or previous parameter values ​​in the physical mechanism model with the latest identified parameter values. This operation enables the adaptive updating of the physical mechanism model, forming a parameter closed loop, so that the model can reflect the wear, aging and other degradation processes of physical entities in real time. Collect historical data of online identification parameter sequences; historical data refers to, for example, This parameter time series fully records the degradation trajectory of key physical parameters; Historical data is used to train a degradation prediction model; a degradation prediction model refers to a time series prediction model such as ARIMA or Gaussian process regression (GPR), which predicts the future trend of parameter changes based on historical degradation trajectories. The degradation trend prediction time domain of the steering system is predicted and output by comparing the degradation prediction model with the preset physical failure threshold. The physical failure threshold refers to the parameter limit value defined by the design specifications or safety standards, such as the maximum allowable clearance. The degradation trend prediction time domain refers to the predicted remaining service life (RUL), that is, the time from the current moment until any key parameter first touches its failure threshold. This embodiment constructs a complete digital twin closed-loop system through the above steps; by fusing physical mechanism models with real-time data and online parameter identification, it solves the problem that traditional static models cannot reflect entity degradation, and achieves high-fidelity, adaptive tracking of physical entities by the model; through closed-loop parameter injection, it ensures the evolutionary fidelity of the twin model throughout its entire lifecycle; by predicting degradation trends based on identification results, it elevates the model's descriptive capabilities to predictive capabilities, realizing forward-looking management and predictive maintenance of the steering system's health status, and improving the system's reliability and safety.

[0019] Example 2 The high-fidelity physical mechanism model, constructed using second-order nonlinear ordinary differential equations, includes: Determine the equivalent moment of inertia of the system; Set the steering column angle as a state variable; Define the physical parameters to be identified, including the damping coefficient, friction coefficient, and transmission clearance. The physical mechanism model uses the motor assist torque and the road reaction torque as system inputs; Based on Example 1, this embodiment further defines the high-fidelity physical mechanism model. The model is explicitly constructed as a second-order nonlinear ordinary differential equation, the purpose of which is to accurately describe the dynamic response of the steering column under the combined action of internal and external torques from the perspective of physical mechanism. This equation, which is the high-fidelity physical mechanism model, is expressed as: ; In this equation, The equivalent rotational inertia of the system is used to characterize the system's inertia against changes in angular acceleration. It is obtained through calculation using a CAD model or calibration via offline bench tests. The steering column angle is set as a core state variable of the system; and These are its first and second derivatives, namely angular velocity and angular acceleration; The source can be measured in real time by an angle sensor. and Then it serves as an intermediate variable for state estimation and calculation; The physical parameters to be identified are the damping torque coefficient, friction torque-related parameters, and transmission clearance, respectively. These three parameters are core indicators characterizing the system's degradation state, such as wear and aging, and are the targets for online identification. Their initial values... It originates from the calibration of the new component bench test and is identified and updated in real time by a nonlinear filtering algorithm during operation; These are internal resistance torque functions, including viscous damping torque functions, friction torque functions (such as in the Stribeck model), and tooth flank clearance torque functions (such as in the dead zone model); their function is to describe the relationship between internal resistance torque and system state. and parameters to be identified Nonlinear relationship between them; The specific forms of these functions are derived from prior physical knowledge; the tooth flank clearance torque function can be expressed as: when hour, ,when hour, In other cases, it is 0, where For contact stiffness; Friction torque function The Stribeck model table can be used as follows: ,in The torque is the Coulomb friction coefficient. The static friction coefficient torque, For Stribeck speed, The coefficient of viscous friction torque. This represents the set of friction parameters. The sign function is used; this model is affected by temperature. Influence; Temperature serves as a key external variable affecting frictional characteristics, inputting it into the frictional torque function. Its source is the real-time measurement value from the temperature sensor; The current is the motor drive current, which serves as the input for calculating the motor's assist torque. Its source is the real-time measurement value from the current sensor. The assist torque for the motor serves as one of the system inputs, representing the active assistance provided by the EPS system. Its source is current. It is calculated through motor model functions; for example, a simplified linear motor model can be used. ,in This is the torque constant of the motor; The road surface reaction moment serves as another system input, characterizing external disturbances caused by road surface unevenness, etc. Its source can be estimated through other models such as vehicle dynamics models, or as a known input in specific bench tests. During online identification, if this moment cannot be estimated or measured in real time, it cannot be used as a known input. Instead, it should be regarded as an unmodeled disturbance of the system and incorporated into the process noise of subsequent nonlinear filtering algorithms; By employing second-order nonlinear ordinary differential equations based on explicit physical mechanisms, this invention ensures the physical interpretability of the digital twin model; unlike purely data-driven black-box models, each parameter to be identified in this embodiment... All of them have clear physical meaning; this makes subsequent identification results such as An increase in the value can be directly correlated with the specific degradation patterns of physical entities, such as gear backlash wear, thus providing a reliable and interpretable basis for subsequent degradation prediction and health management.

[0020] Example 3 Step one also includes: The system collects operating condition data streams in real time via bus and sensors. The operating condition data streams include values ​​from the angle sensor, torque sensor, current sensor, and temperature sensor. Perform timestamp alignment on the operating condition data stream; Denoise the operating data stream; The operating condition data stream is resampled to form an observation vector with a uniform sampling rate; Based on Example 1, this embodiment specifies the acquisition and processing of operating condition data streams. The purpose of this process is to convert the heterogeneous and noisy raw measurement values ​​from physical entities into clean, time-aligned observation data that can be used by state observers such as Kalman filters. Real-time data streams of operating conditions are acquired via a bus and sensors; in this embodiment, the bus refers to the CAN bus, and the sensors specifically include angle sensor values. Torque sensor value Used to measure the internal resistance torque and current sensor value of the steering column. and temperature sensor value The original data stream is denoted as ; Furthermore, a series of preprocessing steps are performed on the operating condition data stream, including timestamp alignment to resolve the time asynchrony issue between different sensor data and ensure that all data points are associated with a unified time reference; denoising, such as using a low-pass filter or moving average method, to remove high-frequency noise from sensor measurements and prevent noise interference with subsequent state estimation; and resampling to unify all data with different sampling rates to the uniform sampling rate required for the Kalman filter to operate, such as a 10ms cycle. ; The final output of the preprocessing is used to form the observation vector at a uniform sampling rate, denoted as... ; Observation vector It refers to A real-world benchmark used for comparison with model predictions; in this embodiment, this vector is selected as: ; Through the above preprocessing steps of the operating condition data stream, this embodiment ensures the high quality and consistency of the data input to the state observer; timestamp alignment, denoising, and resampling effectively eliminate the heterogeneity and noise interference of the original sensor data, providing key guarantees for the accurate calculation of the subsequent residual signal and the convergence and robustness of subsequent parameter identification.

[0021] Example 4 The residual signal is solved through the following steps: Obtain the observation vector and use it as the sensor's measured vector; The model is used to calculate the predicted values ​​by solving the observation function, which includes the estimated angle and the estimated sum of internal resistance torques. Subtract the model's observed predictions from the sensor's measured vectors to obtain the residual signal; This embodiment, based on Embodiment 1, specifies the computation process for the residual signal; residual signal It is the core input that drives the Kalman filter to perform state correction; Obtain the observation vector and use it as the sensor's measured vector. ;Right now ; The model is used to calculate the predicted values ​​through the observation function. ; Observation function This refers to a system's internal state and parameters Mapping to the observable measurement space The function in the space must have the same form as The structures are strictly matched; In this embodiment, ; for matching The two components, The function specifically includes the estimated angle, i.e. The first component directly corresponds to the angle estimate in the state vector. ; and the estimated sum of internal resistance torques, i.e. The second component; the technical motivation for this setup is the assumption of a torque sensor. The measurement is the sum of the internal resistance torques on the steering column, namely the sum of damping, friction, and clearance torques; therefore, this component is defined as: This definition is based on a technical assumption, namely, torque sensor. It is physically mounted in a specific location on the steering column, allowing for the isolated measurement of the sum of internal resistance torques caused by damping, friction, and clearance, unaffected by the motor's assist torque. or road surface reaction moment The direct impact; Therefore, the model observations predict the value for: ; The residual signal is obtained by subtracting the model's observed predictions from the sensor's measured vectors. The purpose of this calculation is to quantify the degree of deviation between the model's predictions and reality under the current parameters; this calculation is expressed as... In this formula, for The residual vector at time step 1 serves as the core input for the update step of the nonlinear filtering algorithm. for The sensor's measured vector at any given time; for The model observation predictions at time t are derived from the aforementioned observation function. Calculated; Through the above residual calculation steps, especially by using the observation function The second component, the estimated torque, is precisely defined as the sum of internal resistance torques that match the sensor's physical meaning, i.e., internal resistance torque. This embodiment improves the physical meaning and accuracy of the residual signal; this precise matching ensures the residual... It can highly sensitively reflect the parameters to be identified. The actual changes in the data improve the accuracy and convergence speed of subsequent parameter identification.

[0022] Example 5 Nonlinear filtering algorithms include a prediction step and an update step; The prediction step includes: Based on the posterior estimate and input from the previous time step, time steps are advanced through a nonlinear state transition function; The nonlinear state transition function originates from the discretization of the high-fidelity physical mechanism model; By advancing through time steps, the prior estimate and the prior covariance matrix are calculated and obtained. Based on Example 1, this embodiment specifically defines the internal flow of a nonlinear filtering algorithm such as the Extended Kalman Filter (EKF); the algorithm includes a prediction step and an update step. The purpose of the prediction step is to, based on the previous time step... The optimal state estimate and physical mechanism model are used to deduce the current moment. State prior estimation; Based on the posterior estimate and input from the previous time step, time steps are advanced through a nonlinear state transition function; the posterior estimate from the previous time step refers to... That is, augmented state vector The input at the previous moment refers to Include As assumed Process noise handling; nonlinear state transition function Its purpose is to describe the augmented state vector. How to Evolving to time; The nonlinear state transition function This stems from the discretization of high-fidelity physical mechanism models; for example, using numerical integration methods such as the Euler method, to... Using the step size, the second-order ordinary differential equation is discretized. For augmented vectors State section and parameter section The transition logic is different: the state part The transfer strictly follows the physical model; through calculation angular acceleration at time t Then update via Euler integral Angle and angular velocity at time; parameter section The transition is based on the assumption that parameters change slowly over time, i.e., the random walk model. Its predicted value is equal to the posterior value of the previous time step, for example... ; Progressing by time steps, i.e., execution Calculate the prior estimate for the current time. and the corresponding prior covariance matrix ; Through the design of the above prediction step, especially by using a state transition function based on a high-fidelity physical model discretization, This embodiment ensures that the prediction phase of state estimation strictly follows physical laws; at the same time, modeling the parameter part as a random walk allows the identification results to be corrected under the drive of residuals; this design makes prior estimation It can approximate the actual evolution trend of the system to the greatest extent, providing a high-quality prediction benchmark for accurate correction of subsequent update steps.

[0023] Example 6 The update steps of the nonlinear filtering algorithm include: The prior estimate is corrected using Kalman gain and residual signal to obtain the posterior estimate; The physical parameters to be identified are extracted from the posterior estimate to form an online identification parameter sequence; This embodiment, based on embodiment 5, further defines the update step of the nonlinear filtering algorithm; the purpose of the update step is to calculate the actual residual signal. To correct the prior estimate of the prediction Thus, the current time is obtained. optimal posterior estimate ; The core of the update step lies in using Kalman gain. and residual signal Correct the prior estimate to obtain the posterior estimate. This calculation is the standard update equation for Kalman filtering, and its purpose is to pull back and correct the prior estimate based on the magnitude of the residual. This calculation is expressed as ;in, for The posterior estimate at time 1 is the optimal estimate; for Prior estimates of time; for The residual signal at time step is specifically defined in the update step as follows: ; The Kalman gain is a dynamically calculated weight matrix used to balance the confidence of the prediction, determined by the prior covariance. The reliability of the measurement, and the measurement noise covariance, are reflected in the measurement noise covariance. This is reflected in the standard Kalman gain formula; its source is through this formula. The calculation yielded, where It is the observation function exist Jacobian matrix at the location; To ensure the stable convergence of the filtering algorithm, and The matrix needs to be properly calibrated; the noise covariance matrix needs to be measured. The value can be determined based on the angle sensor. and torque sensor The noise level can be determined by the specifications provided, or by collecting sensor data under static conditions and calculating its variance; the process noise covariance matrix. In, corresponding to the state of Some values ​​can be set to smaller values ​​to trust the physics model; Corresponding to parameters of The numerical value of this part reflects the expected physical degradation rate of the parameter, and can be optimized through offline simulation or during the system debugging phase to balance the response speed and noise resistance of parameter identification. To improve the robustness of the algorithm under strong nonlinearity and sensor outliers, this embodiment preferably uses unscented Kalman filtering instead of extended Kalman filtering, and can combine it with robust statistical methods such as M-estimation to analyze the residuals. and measurement noise covariance Make dynamic adjustments; The optimal augmented state vector posterior estimate is obtained through computation. Then, from this posterior estimate Extract the corresponding physical parameters to be identified. ; The extracted They are accumulated to form an online identification parameter sequence. The sequence It is the final output of online identification and serves as the input for subsequent adaptive updates and predictions; This embodiment utilizes residual signals. and dynamically calculated Kalman gain This enables real-time, optimal correction of the augmented state vector; this step not only estimates the system state. Furthermore, it successfully extracted physical parameters from noise and uncertainty. The true evolution value; this enables the present invention to quantify the degradation of physical entities, such as increased friction and widening gaps, online and in real time, which is the core data foundation for subsequent adaptive updates and degradation prediction.

[0024] Example 7 The adaptive update of the physical mechanism model is achieved through the following steps: The identified physical parameters are fed back in real time and the original parameter values ​​in the high-fidelity physical mechanism model are replaced. This enables the high-fidelity physical mechanism model to use the updated parameters to calculate the residual signal in the next time step, thus forming a parameter closed-loop injection. Based on Example 1, this embodiment further specifies the adaptive update process of the physical mechanism model. The purpose of this process is to break down the barrier between the model and the identification, and to build a closed-loop system that enables the model to evolve synchronously with the degradation of the entity. The physical parameters to be identified online Real-time feedback and replacement of the original parameter values ​​in the high-fidelity physical mechanism model or The parameter value at time; The technical significance of this replacement operation lies in: enabling the high-fidelity physical mechanism model to be implemented in the next time step. When it is used to calculate the residual signal, i.e., in the prediction step... And calculation in the update step At that time, the parameters that were just updated were used. ; This identification result Update model ,Model Used to calculate the residual at the next time step. residual Used for identification The process is designed to form a closed-loop parameter injection system; By implementing this closed-loop parameter injection, this embodiment ensures that the digital twin model, i.e., the high-fidelity physical mechanism model, is no longer a static, unchanging benchmark, but rather transforms into a dynamic, adaptive twin that evolves synchronously with the physical entity. This adaptive update capability improves the model's fidelity throughout its entire lifecycle, ensuring that the model remains valid even after the physical entity experiences wear or aging. Changes occur, and its predicted output changes. It can still closely track the sensor's measured values. This solves the problem of twin system failure caused by model-reality mismatch in static models.

[0025] Example 8 The degradation prediction model is trained through the following steps: A time series forecasting model is used; The time series prediction model was trained offline using prior degradation datasets obtained from similar equipment or accelerated aging tests. The degradation prediction model also receives assumptions about future operating conditions as input to adjust the prediction results; This embodiment, based on Embodiment 1, specifically defines the training process of the degradation prediction model; the purpose of this model is to establish a mathematical model capable of predicting future degradation trends based on the output historical degradation data. ; A time series forecasting model is employed; a time series forecasting model is a model specifically designed to process data points ordered by time, and its function is to uncover temporal dependencies and long-term trends in the data; in this embodiment, this model... It could be ARIMA (Autoregressive Integral Moving Average), Gaussian Process Regression (GPR), or Recurrent Neural Network (RNN), etc. In terms of training methods, this embodiment adopts an offline training and online inference strategy; the time series prediction model is trained offline using prior degradation datasets obtained from other similar devices or accelerated aging tests. The prior degradation dataset refers to a large amount of parameter degradation data collected prior to the deployment of this invention through accelerated aging tests in the laboratory or from similar steering systems already in service; to ensure the independence of variables, this dataset contains multiple sets of data under specific constant operating conditions, such as specific constant vibrations. and constant load The measured parameters characterizing the entire life cycle, such as... The logic of offline training lies in fitting the model using methods such as least squares or gradient descent. The internal parameters of the model, such as the weights of the neural network, make the model... Predicted output Degradation trajectory as actually collected The error between them is minimal; The reason for using offline training is that when a single device is first put into use, its historical data... This is insufficient to train a robust prediction model; prior datasets can be used to improve the model. It provides a strong initial generalization capability; subsequently, historical data collected online. It can be used to fine-tune the model; The degradation prediction model also accepts assumptions about future operating conditions. As an adjustable input to refine the prediction results; future operating condition assumptions This refers to a preset parameter that characterizes the intensity of future operation, for example, set as standard operating conditions, severe operating conditions, or light load operating conditions; Its function is as follows: because the degradation rate of parameters, such as wear rate, is strongly correlated with the intensity of working conditions, such as load and vibration; by introducing ,Model like It can output predicted trajectories under different working conditions, such as under harsh working conditions and gaps. It will exceed the limit after 500 hours; under standard operating conditions, it will exceed the limit after 1500 hours; in specific implementation, in order to make the model understand , Quantified into one or more values, such as the normalized expected average load. and vibration intensity These values ​​will serve as a degradation prediction model. Additional input features; model During offline training, the nonlinear effects of these working condition features on the degradation rate were learned based on a prior degradation dataset containing different working conditions. By combining offline training with prior data and inputting assumptions about future operating conditions, this embodiment solves two major challenges in degradation prediction: cold start and operating condition dependence. The use of prior data ensures the model has high-reliability predictive capabilities from the early stages of service, while the introduction of... This makes the prediction results no longer a single, static trajectory, but a dynamic, context-aware prediction, enhancing the flexibility and accuracy of predictive maintenance decisions.

[0026] Example 9 The time-domain prediction of degradation trends is achieved through the following steps: Define a preset physical failure threshold as a parameter limit vector characterizing physical failure; The degradation prediction model is used to output the predicted trajectories of each parameter; Real-time checks are performed to determine if the predicted trajectory of any parameter exceeds the corresponding physical failure threshold. The moment when the degradation trend is first exceeded is defined as the time domain for predicting the degradation trend. Based on Example 1, this embodiment specifically defines the prediction and calculation process for the degradation trend prediction time domain, i.e., the remaining useful life; the purpose of this process is to transform the parameter prediction trajectory into a clear failure time point. Define a preset physical failure threshold as a parameter limit vector characterizing physical failure; physical failure threshold It is a key, pre-defined engineering parameter whose source must be a physically interpretable design specification or safety standard; it is a parameter related to the parameter to be identified. The corresponding vector, for example ,in It may be set to a maximum permissible tooth flank clearance of 0.5 mm; Using degradation prediction models Output the predicted trajectory of each parameter; that is, the model. It will output a prediction vector This vector represents the time span from the current moment to any future moment. The predicted values ​​of the parameters; The system performs calculations to determine the final result. That is, the time domain for predicting degradation trends: real-time inspection to check whether any degradation trend exists. Predicted trajectory of parameters For example, gap prediction trajectory Exceeding its corresponding physical failure threshold ,For example The first time that occurs will exceed [a certain time]. The final degradation trend prediction time domain was determined. ; The purpose of this computational logic is to identify the weakest link in the system, that is, to determine the earliest parameter and the moment that leads to system failure; its mathematical expression is: ; In this expression, The remaining useful life is the time domain for predicting the degradation trend, which is the final prediction output of this embodiment; For future time; For example, parameter index ; For prediction models For the Parameters in the future The predicted value at any given time; For the first The physical failure threshold of each parameter is derived from a preset, physically meaningful safety specification. Through the above calculation steps, this embodiment predicts multi-dimensional parameter degradation. All of these changes converge to a single, clearly instructive scalar quantity: remaining useful life. This method checks whether any parameter exceeds its limit. This ensures the comprehensiveness and security of the predictions, adhering to the principle of the weakest link in the chain; the prediction results... With physically interpretable failure thresholds The combination of these features makes the final output predictive maintenance time domain highly valuable and reliable in engineering practice, realizing a functional closed loop from parameter identification to decision support.

[0027] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for constructing a steering system model based on digital twins, characterized in that, The specific steps include: Step 1: Construct a high-fidelity physical mechanism model characterizing the dynamic characteristics of the steering system; acquire the working condition data stream of the steering system in real time; construct a state observer, which is based on the working condition data stream and the high-fidelity physical mechanism model; calculate the residual signal through the state observer, where the residual signal characterizes the difference between the model prediction value and the sensor measurement value. Step 2: Construct an augmented state vector, which combines system state variables and physical parameters to be identified; use a nonlinear filtering algorithm to recursively estimate the augmented state vector, where the recursive estimation utilizes residual signals; identify and output the physical parameters to be identified, forming an online identified parameter sequence. Step 3: Inject the online identification parameter sequence into the high-fidelity physical mechanism model in real time to achieve adaptive updating of the physical mechanism model; collect historical data of the online identification parameter sequence; train the degradation prediction model using the historical data; compare the degradation prediction model with the preset physical failure threshold, predict and output the time domain of the degradation trend prediction of the steering system.

2. The method for constructing a steering system model based on digital twins according to claim 1, characterized in that, The high-fidelity physical mechanism model, constructed using second-order nonlinear ordinary differential equations, includes: Determine the equivalent moment of inertia of the system; Set the steering column angle as a state variable; Define the physical parameters to be identified, including the damping coefficient, friction coefficient, and transmission clearance. The physical mechanism model uses the motor assist torque and the road surface reaction torque as system inputs.

3. The method for constructing a steering system model based on digital twins according to claim 1, characterized in that, Step one also includes: The system collects operating condition data streams in real time via bus and sensors. The operating condition data streams include values ​​from the angle sensor, torque sensor, current sensor, and temperature sensor. Perform timestamp alignment on the operating condition data stream; Denoise the operating data stream; The operating condition data stream is resampled to form an observation vector with a uniform sampling rate.

4. The method for constructing a steering system model based on digital twins according to claim 1, characterized in that, The residual signal is solved through the following steps: Obtain the observation vector and use it as the sensor's measured vector; The model is used to calculate the predicted values ​​by solving the observation function, which includes the estimated angle and the estimated sum of internal resistance torques. The residual signal is obtained by subtracting the model observation prediction value from the sensor's measured vector.

5. The method for constructing a steering system model based on digital twins according to claim 1, characterized in that, Nonlinear filtering algorithms include a prediction step and an update step; The prediction step includes: Based on the posterior estimate and input from the previous time step, time steps are advanced through a nonlinear state transition function; The nonlinear state transition function originates from the discretization of the high-fidelity physical mechanism model; By advancing through time steps, the prior estimate and the prior covariance matrix are calculated.

6. The method for constructing a steering system model based on digital twins according to claim 5, characterized in that, The update steps of the nonlinear filtering algorithm include: The prior estimate is corrected using Kalman gain and residual signal to obtain the posterior estimate; The physical parameters to be identified are extracted from the posterior estimate to form an online identification parameter sequence.

7. The method for constructing a steering system model based on digital twins according to claim 1, characterized in that, The adaptive update of the physical mechanism model is achieved through the following steps: The identified physical parameters are fed back in real time and the original parameter values ​​in the high-fidelity physical mechanism model are replaced. This allows the high-fidelity physical mechanism model to use the updated parameters to calculate the residual signal in the next time step, thus forming a parameter closed-loop injection.

8. The method for constructing a steering system model based on digital twins according to claim 1, characterized in that, The degradation prediction model is trained through the following steps: A time series forecasting model is used; The time series prediction model was trained offline using prior degradation datasets obtained from similar equipment or accelerated aging tests. The degradation prediction model also receives assumptions about future operating conditions as input to adjust the prediction results.

9. The method for constructing a steering system model based on digital twins according to claim 1, characterized in that, The time-domain prediction of degradation trends is achieved through the following steps: Define a preset physical failure threshold as a parameter limit vector characterizing physical failure; The degradation prediction model is used to output the predicted trajectories of each parameter; Real-time checks are performed to determine if the predicted trajectory of any parameter exceeds the corresponding physical failure threshold. The moment when the degradation trend is first exceeded is defined as the time domain for predicting the degradation trend.

Images