Vehicle-mounted dynamic model identification device

Abstract

The invention relates to the technical field of vehicle dynamics, intelligent control, tire state monitoring and real-time identification, in particular to a vehicle-mounted dynamics model identification device, which comprises a state sensing module for dynamically adjusting data sampling frequency according to a vehicle longitudinal acceleration change rate and a yaw velocity change rate and outputting original sensor data. The adaptive preprocessing module adopts a variational auto-encoder to carry out depth signal reconstruction on the data, calculates a confidence coefficient weight and outputs a feature sequence. And the model evolution module processes the feature sequence through a hybrid model of a physical constraint branch and a data driving branch, and generates a tire kinetic parameter estimation value and a prediction error. And the collaborative optimization module adopts a distributed learning framework to optimize parameters, and aggregates and updates the parameters through a federated learning algorithm. And the decision fusion module calculates a parameter estimation confidence interval, and generates a vehicle stability control signal through evidence theory fusion. All the modules form a closed-loop system through a feedback mechanism, and real-time accurate identification of tire dynamic parameters is achieved.

Description

Technical Field

[0001] This invention relates to the fields of vehicle dynamics, intelligent control, tire condition monitoring and real-time identification, and in particular to an on-board dynamics model identification device. Background Technology

[0002] Tire dynamics model identification technology is a method to extract tire dynamic behavior models through experimental data or simulation. Its core lies in using the tire's input-output response under various operating conditions, such as the relationship between force and slip ratio, to inversely derive model parameters. Because the interaction between the tire and the road surface exhibits a high degree of nonlinearity, existing static models are insufficient to accurately describe the dynamic process. Therefore, a systematic identification process is needed, including data acquisition, model structure selection, and parameter estimation. Physically based modeling methods rely on tire mechanics theory to construct equations, while data-driven methods learn patterns from a large number of observations through machine learning algorithms. Both methods utilize optimization techniques to minimize the error between model predictions and measured data, thereby improving the model's generalization ability and practicality, and providing a more accurate foundation for vehicle dynamics control.

[0003] Existing tire dynamics model identification technology has the following technical pain points: it relies on offline test benches to calibrate parameters under fixed conditions, which cannot capture the dynamic changes of temperature, wear and extreme road conditions in real roads in real time, resulting in a lag in the estimation of tire mechanical parameters; when intelligent driving vehicles perform stability control, such as when the tire slip angle increases instantaneously during high-speed cornering, the offline model has difficulty accurately predicting tire force, causing the electronic stability program to make decision-making errors and increasing the risk of vehicle instability. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an on-board dynamics model identification device. This device solves the technical problem that existing tire dynamics model identification technologies rely on offline test benches and specific operating conditions, making it impossible to adapt to real-time dynamic changes in real roads (such as temperature, wear, and extreme road conditions). This results in inaccurate estimation of tire mechanical parameters and response delays in intelligent driving and vehicle stability control systems, leading to control strategy failure and reduced safety.

[0005] To solve the above-mentioned technical problems, the specific contents of the present invention are as follows: The present invention provides a vehicle dynamics model identification device, comprising: The state perception module is used to collect data on tire force, vehicle motion status, and tire pressure and temperature. It dynamically adjusts the data sampling frequency based on the rate of change of longitudinal acceleration and the rate of change of yaw rate, and outputs raw sensor data. An adaptive preprocessing module is used to receive the original sensor data, repair the original sensor data using deep signal reconstruction technology, calculate the confidence weight of each data point, and output a feature sequence composed of the repaired sensor data and the corresponding confidence weight. The model evolution module is used to receive the feature sequence and process it through a hybrid model including a physical constraint branch and a data-driven branch. The physical constraint branch calculates the theoretical tire force based on the parameterized magic formula, and the data-driven branch predicts the tire force residual based on the long short-term memory network. The outputs of the two branches are fused through a gated recurrent unit to generate tire dynamics parameter estimates and model prediction errors. The collaborative optimization module is used to receive the estimated values ​​of the tire dynamic parameters and the model prediction error, optimize the local model parameters corresponding to each tire using a distributed learning framework, and aggregate the local model parameter updates through a federated learning algorithm to output the optimized global tire model parameters. The decision fusion module is used to receive the global tire model parameters and vehicle bus environment information, calculate the confidence interval of the parameter estimates, and generate vehicle stability control signals through evidence theory fusion and output them to the vehicle stability control system. The decision fusion module also outputs vehicle stability state assessment results to the state perception module to dynamically adjust the data sampling frequency; the confidence weights output by the adaptive preprocessing module are used as input weights for the data-driven branch in the model evolution module; and the collaborative optimization module adjusts the exploration strategy in the distributed learning framework according to the model prediction error.

[0006] Furthermore, in the vehicle dynamics model identification device of the present invention, the state perception module is configured as follows: Calculate the rate of change of the vehicle's yaw rate at the current moment; The rate of change of yaw rate is compared with a preset first threshold. Based on the comparison results, when the rate of change of yaw angular velocity is less than the first threshold, sensor data is collected using a reference sampling frequency; When the rate of change of yaw rate is greater than or equal to the first threshold, sensor data is collected using a high-frequency sampling mode, wherein the frequency of the high-frequency sampling mode is higher than the reference sampling frequency.

[0007] Furthermore, in the vehicle dynamics model identification device of the present invention, the adaptive preprocessing module is configured as follows: The original sensor data is input into the encoder of the variational autoencoder, which maps the input data to a low-dimensional latent space and outputs a latent feature vector. The latent feature vector is input into the decoder of the variational autoencoder. The decoder integrates an attention mechanism to assign reconstruction weights to different features in the latent feature vector, and decodes the latent feature vector based on the reconstruction weights to output the reconstructed sensor data.

[0008] Furthermore, in the vehicle dynamics model identification device of the present invention, the adaptive preprocessing module is configured as follows: Calculate the residual between the reconstructed sensor data and the corresponding original sensor data; Obtain the historical error statistics of the sensor from which the original sensor data originates; Based on the residuals, the historical error statistics, and the consistency assessment results between the current data point and the preceding and following time series data, the confidence weights in the range of 0 to 1 are generated.

[0009] Furthermore, in the vehicle dynamics model identification device of the present invention, the model evolution module includes a gated loop unit, which is configured as follows: The gated loop unit receives the tire slip rate at the current moment; Using the slip ratio as input, a first weight for weighting the physical constraint branch output and a second weight for weighting the data-driven branch output are calculated; The output of the physical constraint branch is multiplied by the first weight to obtain the first weighted result; The output of the data-driven branch is multiplied by the second weight to obtain the second weighted result; The first weighted result is added to the second weighted result to output the estimated value of the tire dynamics parameter.

[0010] Furthermore, in the vehicle dynamics model identification device of the present invention, the collaborative optimization module employs a distributed learning framework to optimize the local model parameters corresponding to each tire, including: Create an independent local agent for each tire, and load the latest global tire model parameters as the initial model for each local agent; Each local agent uses a proximal policy optimization algorithm to update local model parameters within a trust region constraint defined by a physically reasonable range of parameter changes, with the goal of minimizing the difference between local observation data and model predictions.

[0011] Furthermore, in the vehicle dynamics model identification device of the present invention, the collaborative optimization module aggregating local model parameter updates through a federated learning algorithm includes: The central coordinator periodically collects updated local model parameters from each of the local agents; The central coordinator assigns aggregation weights to each collected local model parameter based on the amount of data corresponding to each local agent. The central coordinator uses the aggregated weights to calculate a weighted average of all local model parameters, generating a new round of global tire model parameters. The newly generated global tire model parameters are distributed to each of the local agents.

[0012] Furthermore, in the vehicle dynamics model identification device of the present invention, the confidence interval for calculating the parameter estimation in the decision fusion module includes: Receive the time series of tire model parameters from the collaborative optimization module; The time series of tire model parameters is smoothed using a Kalman filter to obtain smoothed parameter estimates. Calculate the innovative covariance of the Kalman filter during the smoothing process to quantify the uncertainty measure of the smoothed parameter estimates; The confidence interval for the parameter estimate is determined based on the uncertainty measure.

[0013] Furthermore, in the vehicle dynamics model identification device of the present invention, the decision fusion module calculates the confidence interval of the parameter estimates and generates a vehicle stability control signal through evidence theory fusion, which is then output to the vehicle stability control system. A recognition framework is constructed, the input of which includes the smoothed parameter estimates and vehicle bus environment information; The DS evidence theory is used to combine and calculate multiple pieces of evidence in the identification framework to obtain a comprehensive trust function; The comprehensive trust function is input into the fuzzy logic rule table, and the vehicle stability control signal is generated through table lookup mapping.

[0014] Furthermore, in the vehicle dynamics model identification device of the present invention, the initial values ​​of the model parameters of the model evolution module are loaded from a pre-established typical working condition parameter library through a transfer learning strategy, wherein the transfer learning strategy includes: Identify the similarity between the current vehicle operating condition and the typical operating conditions in the typical operating condition parameter library; Based on the similarity recognition results, the model parameters corresponding to the most similar typical working conditions are selected from the typical working condition parameter library as initial values; Starting from the initial value, the model parameters of the model evolution module are iteratively updated and optimized using the currently collected feature sequence as training data through the gradient descent algorithm.

[0015] Beneficial effects of this invention; This invention utilizes a state perception module to dynamically adjust the data sampling frequency based on the vehicle's longitudinal acceleration and yaw rate of change. Under transient conditions, a high-frequency sampling mode is employed to fully capture the dynamic response of tire forces, improving the real-time performance of data acquisition. An adaptive preprocessing module uses a variational autoencoder to perform deep signal reconstruction on the raw sensor data, generating feature sequences with confidence weights, enhancing the traceability of data quality. A model evolution module employs a hybrid model combining physical constraint branches and data-driven branches, combined with a gated loop unit, to adaptively fuse the output based on tire slip ratio, improving the accuracy of tire dynamics parameter estimation. (The last sentence appears to be incomplete and possibly refers to a collaborative mechanism.) The optimization module employs a distributed learning framework and federated learning algorithm. Through local agent updates and global parameter aggregation, it optimizes the model's generalization ability under varying operating conditions. The decision fusion module smooths the parameter sequence and calculates the confidence interval using a Kalman filter. It then combines DS evidence theory with environmental information to generate control signals. Simultaneously, it feeds back the vehicle stability state assessment results to the state perception module, forming a closed-loop adjustment mechanism. This achieves high-precision and robust real-time identification of tire dynamic parameters, effectively overcoming the shortcomings of offline model parameter estimation lag and response delay, and significantly improving the reliability and safety of the vehicle stability control system. Attached Figure Description

[0016] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the system architecture of the vehicle dynamics model identification device of the present invention. Detailed Implementation

[0018] To make the technical solution of the present invention clearer, the present invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. 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. The present invention provided by various embodiments will be described in detail below with reference to the accompanying drawings. To better understand the purpose of the present invention, the present invention will be described in further detail below.

[0019] Please see Figure 1 The present invention provides a vehicle dynamics model identification device, comprising: The state perception module is used to collect data on tire force, vehicle motion status, and tire pressure and temperature. It dynamically adjusts the data sampling frequency based on the rate of change of longitudinal acceleration and the rate of change of yaw rate, and outputs raw sensor data. An adaptive preprocessing module is used to receive the original sensor data, repair the original sensor data using deep signal reconstruction technology, calculate the confidence weight of each data point, and output a feature sequence composed of the repaired sensor data and the corresponding confidence weight. The model evolution module is used to receive the feature sequence and process it through a hybrid model including a physical constraint branch and a data-driven branch. The physical constraint branch calculates the theoretical tire force based on the parameterized magic formula, and the data-driven branch predicts the tire force residual based on the long short-term memory network. The outputs of the two branches are fused through a gated recurrent unit to generate tire dynamics parameter estimates and model prediction errors. The collaborative optimization module is used to receive the estimated values ​​of the tire dynamic parameters and the model prediction error, optimize the local model parameters corresponding to each tire using a distributed learning framework, and aggregate the local model parameter updates through a federated learning algorithm to output the optimized global tire model parameters. The decision fusion module is used to receive the global tire model parameters and vehicle bus environment information, calculate the confidence interval of the parameter estimates, and generate vehicle stability control signals through evidence theory fusion and output them to the vehicle stability control system. The decision fusion module also outputs vehicle stability state assessment results to the state perception module to dynamically adjust the data sampling frequency; the confidence weights output by the adaptive preprocessing module are used as input weights for the data-driven branch in the model evolution module; and the collaborative optimization module adjusts the exploration strategy in the distributed learning framework according to the model prediction error.

[0020] The state perception module integrates a tire six-component force sensor, an inertial measurement unit, and an embedded tire pressure and temperature sensor to form a physical sensing layer. This layer collects real-time data on the three-dimensional forces, torques, vehicle acceleration, angular velocity, and internal tire thermodynamic parameters generated by the interaction between the tire and the road surface. The module employs an event-triggered mechanism, dynamically adjusting the data sampling frequency based on real-time calculations of the vehicle's longitudinal acceleration and yaw rate. When the yaw rate is below a preset threshold, the system enters a low-frequency sampling mode to reduce power consumption. When transient conditions such as emergency braking or rapid steering are detected, it automatically switches to a high-frequency sampling mode to fully capture the dynamic response characteristics of tire forces. The collected raw sensor data, tagged with timestamps and sampling frequency identifiers, is output to the adaptive preprocessing module.

[0021] After receiving the raw sensor data, the adaptive preprocessing module uses a variational autoencoder for deep signal reconstruction. The encoder maps the high-dimensional sensor data to a low-dimensional latent space and extracts key feature vectors. The decoder uses an attention mechanism to perform context-aware reconstruction of the latent feature vectors, repairing missing or anomalous data points. The reconstructed sensor data is compared with the original data, residuals are calculated, and confidence weights for each data point are generated by combining historical sensor error statistics and time-series consistency assessments. The weights range from 0 to 1. The output is a feature sequence consisting of the repaired sensor data and the corresponding confidence weights, which is passed as high-quality input to the model evolution module.

[0022] After receiving the feature sequence, the model evolution module constructs a hybrid model based on a physical information neural network. The physical constraint branch embeds a parameterized magic formula to calculate theoretical tire forces, while the data-driven branch employs a long short-term memory network to learn tire dynamic hysteresis and nonlinear residual characteristics. The outputs of the two branches are adaptively fused through a gated recurrent unit. This unit takes the current tire slip ratio as input and dynamically adjusts the fusion weights of the physical and data-driven branches to generate estimated tire dynamic parameters and model prediction errors. Initial model parameters are loaded from a pre-established database of typical operating condition parameters using a transfer learning strategy and continuously evolved through an online learning mechanism.

[0023] After receiving the estimated tire dynamics parameters and model prediction errors, the collaborative optimization module introduces a multi-agent reinforcement learning framework based on federated learning. Each tire corresponds to a local agent, which uses a proximal policy optimization algorithm to update its local model parameters under the constraint of a trust region, defined by the physically reasonable range of parameter changes. A central coordinator periodically aggregates the parameter updates from all agents, applies a weighted average algorithm to generate globally consistent tire model parameters, and distributes the optimized parameters to each local agent. The optimization process dynamically adjusts the exploration strategy based on the model prediction error to improve the model's generalization ability.

[0024] After receiving global tire model parameters and vehicle bus environment information, the decision fusion module uses a Kalman filter to smooth the parameter time series, obtaining smoothed parameter estimates, and calculates the innovation covariance to quantify uncertainty. Based on the uncertainty metric, the confidence interval of the parameter estimates is determined. Subsequently, DS evidence theory is applied to fuse multi-source evidence, including the smoothed parameter estimates and environmental information, to generate a comprehensive trust function. This comprehensive trust function is mapped through a fuzzy logic rule table to generate a vehicle stability control signal, which is output to the vehicle stability control system. Simultaneously, the decision fusion module outputs the vehicle stability state assessment results to the state perception module for feedback adjustment of the data sampling frequency, forming a closed-loop control.

[0025] The specific process by which the state perception module dynamically adjusts the data sampling frequency is as follows: First, the instantaneous rate of change of the vehicle's yaw rate is calculated in real time, and this rate of change is compared with a preset first threshold. When the rate of change of the yaw rate is less than the first threshold, the system determines that the vehicle is in a steady-state driving condition, and at this time, a reference sampling frequency is used for data acquisition to reduce system power consumption; when the rate of change of the yaw rate is greater than or equal to the first threshold, the system recognizes that the vehicle has entered a transient condition, and immediately switches the sampling frequency to a high-frequency sampling mode. The frequency value of this sampling mode is significantly higher than the reference sampling frequency, thereby ensuring that the dynamic response characteristics of the tire force can be fully captured.

[0026] The signal reconstruction process of the adaptive preprocessing module is specifically manifested as follows: the original sensor data is input into the encoder network of a variational autoencoder, which maps the high-dimensional sensor data to a low-dimensional latent space and outputs a latent feature vector. This latent feature vector is then input into a decoder with an integrated attention mechanism. The decoder assigns differentiated reconstruction weights to different feature dimensions in the latent feature vector and decodes and reconstructs the latent feature vector based on each weight, finally outputting the repaired sensor data.

[0027] The confidence weight generation mechanism includes: calculating the numerical residual between the reconstructed sensor data and the original sensor data, and simultaneously obtaining historical error statistics of the sensor as prior information. Combining the numerical residual, historical error statistics, and the consistency evaluation results between the current data point and the preceding and following time series data, a confidence weight value in the range of 0 to 1 is generated through a weighted fusion algorithm. This weight value characterizes the reliability of the data point.

[0028] The fusion mechanism of the gated loop unit is as follows: The gated loop unit receives real-time collected tire slip ratio data as a gating signal, and calculates the first weighting coefficient of the physical constraint branch and the second weighting coefficient of the data-driven branch based on the slip ratio value. The theoretical tire force output by the physical constraint branch is multiplied by the first weighting coefficient to obtain the first weighted result, and the tire force residual predicted by the data-driven branch is multiplied by the second weighting coefficient to obtain the second weighted result. Finally, the two weighted results are added and fused to generate the estimated value of the tire dynamic parameters.

[0029] The distributed learning framework is implemented as follows: an independent local agent instance is created for each tire, and each agent loads the latest version of the global tire model parameters as the initial model. The local agent adopts a proximate policy optimization algorithm, with the optimization objective of minimizing the difference between local observation data and model predictions, and performs local model parameter update operations within a trust region constraint defined by the physically reasonable range of parameters.

[0030] The federated learning aggregation process includes: a central coordinator periodically collecting updated model parameters from each local agent, and assigning differentiated aggregation weights to the collected parameters based on the amount of data corresponding to each agent. The central coordinator uses each weight to perform a weighted average calculation on all local model parameters, generating a new round of global tire model parameters, and distributing the updated parameters to each local agent.

[0031] The confidence interval calculation method is as follows: Time series data of tire model parameters are received from the collaborative optimization module. A Kalman filter is used to smooth the time series data to obtain optimized parameter estimates. The innovation covariance matrix generated by the Kalman filter during the smoothing process is calculated. This matrix is ​​used to quantify the uncertainty measure of the smoothed parameter estimates. Finally, the confidence interval range of the parameter estimates is determined based on the uncertainty measure.

[0032] The evidence theory fusion process specifically includes: constructing an identification framework that includes smoothed parameter estimates and vehicle bus environment information; using DS evidence theory to combine and calculate multiple evidence sources within the framework to obtain a comprehensive trust function; inputting this trust function into a preset fuzzy logic rule table; and generating the final vehicle stability control signal through a lookup and mapping operation.

[0033] The implementation steps of the transfer learning strategy are as follows: First, identify the similarity index between the current vehicle operating condition and the operating conditions stored in the typical operating condition parameter library. Based on the similarity identification result, select the model parameters corresponding to the typical operating condition with the highest matching degree from the parameter library as initial values. Starting from these initial values, use the real-time collected feature sequences as the training dataset, and iteratively update and optimize the parameters of the model evolution module using the gradient descent algorithm.

[0034] This invention effectively solves the technical problem that existing tire dynamics model identification technologies rely on offline test benches and cannot adapt to real-time dynamic changes on real roads by constructing a closed-loop system that includes state perception, adaptive preprocessing, model evolution, collaborative optimization, and decision fusion modules.

[0035] The state perception module employs an event-triggered mechanism, dynamically adjusting the data sampling frequency based on the vehicle's longitudinal acceleration and yaw rate of change. When the vehicle is in a steady-state condition, a baseline sampling frequency is used to reduce power consumption. When transient conditions such as emergency braking or rapid steering are detected, it automatically switches to a high-frequency sampling mode, thus fully capturing the dynamic response characteristics of tire forces. This adaptive sampling strategy ensures both real-time data acquisition and condition matching.

[0036] The adaptive preprocessing module performs deep signal reconstruction on the raw sensor data using a variational autoencoder. The encoder maps high-dimensional data to a low-dimensional latent space to extract feature vectors, and the decoder reconstructs these latent feature vectors based on an attention mechanism, repairing missing or outlier data points. The residual between the reconstructed data and the original data is calculated, and confidence weights are generated by combining historical sensor error statistics and temporal consistency assessments. The resulting feature sequence with weighted labels provides a data foundation with traceable quality for downstream modules.

[0037] After receiving the feature sequence, the model evolution module constructs a hybrid model within a physical information neural network framework. The physical constraint branch calculates the theoretical tire force based on the parameterized magic formula, while the data-driven branch predicts the tire force residual through a long short-term memory network. The gated recurrent unit dynamically adjusts the fusion weights of the two branches using the tire slip ratio as input, thereby generating parameter estimates and model prediction errors. Initial model parameters are loaded from a pre-built typical working condition parameter library through a transfer learning strategy and continuously evolved through an online learning mechanism, significantly improving the model's adaptability to dynamic working conditions.

[0038] The collaborative optimization module employs a multi-agent reinforcement learning framework based on federated learning. Each tire corresponds to a local agent, which updates its local parameters using a proximate policy optimization algorithm under physical rationality constraints. A central coordinator periodically aggregates the parameter updates from all agents and generates globally consistent tire model parameters through weighted averaging. This distributed learning mechanism protects the privacy of individual tire data while leveraging collective experience to improve the model's generalization ability.

[0039] The decision fusion module uses a Kalman filter to smooth the parameter time series, calculates the innovation covariance to quantify uncertainty, and determines the confidence interval of the parameter estimates. Subsequently, DS evidence theory is applied to fuse the smoothed parameter estimates with vehicle bus environment information to generate a comprehensive trust function. This function is mapped to a vehicle stability control signal through a fuzzy logic rule table, directly driving the electronic stability program or the drive-by-wire chassis system.

[0040] The system feeds back the vehicle stability state assessment results output by the decision fusion module to the state perception module, dynamically adjusting the sampling frequency. The confidence weights of the adaptive preprocessing module serve as input weights for the data-driven branch of the model evolution module. The collaborative optimization module adjusts the exploration strategy based on the model prediction error. This multi-level feedback mechanism creates deep coupling between modules, resulting in overall adaptability that no single module possesses. This enables high-precision, robust, and real-time identification of tire dynamic characteristics, fundamentally solving the control strategy failure problem caused by the lag in offline model parameter estimation.

[0041] This invention relates to an onboard dynamics model identification device. This device achieves real-time identification of tire dynamic parameters through the collaborative work of multiple modules, thereby solving the problem that offline models cannot adapt to dynamic changes in real roads. The following description uses an application scenario such as stability control of intelligent driving vehicles as an example.

[0042] The state perception module integrates a tire six-component force sensor, an inertial measurement unit, and an embedded tire pressure and temperature sensor to collect real-time data on the three-dimensional forces, torques, vehicle acceleration, angular velocities, and internal tire thermodynamic parameters generated by the interaction between the tire and the road surface. This module dynamically adjusts the data sampling frequency based on the real-time calculated values ​​of the vehicle's longitudinal acceleration and yaw rate. When the yaw rate is below a preset threshold, the system uses a baseline sampling frequency to reduce power consumption. When transient conditions such as emergency braking or rapid steering are detected, it automatically switches to a high-frequency sampling mode to fully capture the dynamic response characteristics of tire forces. The collected raw sensor data, with timestamps and sampling frequency identifiers, is output to the adaptive preprocessing module.

[0043] After receiving the raw sensor data, the adaptive preprocessing module uses a variational autoencoder (VAE) for deep signal reconstruction. The VAE maps the high-dimensional sensor data to a low-dimensional latent space, outputting a latent feature vector. The decoder integrates an attention mechanism, assigning reconstruction weights to different features within the latent feature vector, and decodes the vector based on these weights, outputting reconstructed sensor data to repair missing or outliers. Subsequently, the module calculates the residual between the reconstructed data and the original data, and, combining historical sensor error statistics and consistency assessments between the current data point and preceding and following time-series data, generates confidence weights within the range of 0 to 1. Finally, the module outputs a feature sequence composed of the repaired sensor data and the corresponding confidence weights, which is then passed as high-quality input to the model evolution module.

[0044] After receiving the feature sequence, the model evolution module constructs a hybrid model based on a physical information neural network. The physical constraint branch calculates the theoretical tire force based on the parameterized magic formula, while the data-driven branch uses a long short-term memory network to predict the tire force residual. The gated recurrent unit receives the tire slip ratio at the current moment as input, calculates the first weight for weighting the output of the physical constraint branch and the second weight for weighting the output of the data-driven branch, and generates tire dynamic parameter estimates and model prediction errors through weighted fusion. The initial values ​​of the model parameters are loaded from a pre-established typical operating condition parameter library through a transfer learning strategy. The transfer learning strategy includes identifying the similarity between the current vehicle operating condition and typical operating conditions, selecting the model parameters with the highest matching degree as initial values, and using the feature sequence to iteratively update and optimize using a gradient descent algorithm, thereby improving the model's adaptability under dynamic operating conditions.

[0045] After receiving the estimated tire dynamics parameters and model prediction errors, the collaborative optimization module employs a distributed learning framework to optimize the local model parameters for each tire. An independent local agent is created for each tire. Each agent loads the latest global tire model parameters as its initial model and uses a proximate policy optimization algorithm to minimize the difference between local observations and model predictions, updating the local model parameters within a trust region constrained by a physically reasonable range of parameter changes. A central coordinator periodically collects the updated parameters from each local agent, assigns aggregation weights to each parameter based on the amount of data, and generates a new round of global tire model parameters through weighted averaging, distributing them to all agents. This federated learning mechanism leverages collective experience to improve model generalization ability while protecting data privacy.

[0046] After receiving global tire model parameters and vehicle bus environment information, the decision fusion module uses a Kalman filter to smooth the time series of tire model parameters, obtaining smoothed parameter estimates. It then calculates the innovation covariance to quantify uncertainty and determine the confidence interval of the parameter estimates. Subsequently, an identification framework is constructed. The framework inputs include the smoothed parameter estimates and vehicle bus environment information. DS evidence theory is used to combine multiple pieces of evidence to obtain a comprehensive trust function. This function is then input into a fuzzy logic rule table, where a lookup mapping is used to generate a vehicle stability control signal, which is output to the vehicle stability control system. Simultaneously, the decision fusion module outputs the vehicle stability state assessment results to the state perception module for dynamically adjusting the data sampling frequency, forming a closed-loop control.

[0047] The device of this invention achieves real-time identification through deep coupling between modules: the confidence weight of the adaptive preprocessing module serves as the input weight of the data-driven branch in the model evolution module; the collaborative optimization module adjusts the exploration strategy according to the model prediction error; and the feedback of the decision fusion module adjusts the data acquisition. This closed-loop system can adapt to changes in temperature, wear, and extreme road conditions, improve the accuracy of parameter estimation and response speed, and ultimately enhance the safety and reliability of vehicle stability control.

[0048] Embodiment 1 of this invention: In high-speed cornering scenarios of intelligent driving vehicles, the tire slip angle increases instantaneously, placing stringent requirements on the real-time identification of the tire dynamics model. The state perception module collects tire force, vehicle motion state, and tire pressure and temperature data in real time through tire six-component force sensors and an inertial measurement unit, and dynamically adjusts the sampling frequency based on the yaw rate change rate; when the change rate exceeds a preset threshold, it switches to a high-frequency sampling mode to capture transient responses. The adaptive preprocessing module uses a variational autoencoder to deeply reconstruct the original data. The encoder extracts low-dimensional latent features, the decoder repairs outliers based on an attention mechanism, and generates feature sequences with confidence weights. The model evolution module calculates theoretical tire force through physical constraint branches, predicts residuals through data-driven branches, and a gated loop unit dynamically fuses the outputs of the two branches based on slip ratio to generate parameter estimates and prediction errors. The collaborative optimization module creates a local agent for each tire, updates parameters using a proximal policy optimization algorithm, and aggregates the global model through federated learning. The decision fusion module uses a Kalman filter to smooth the parameter sequence, calculates the confidence interval, and then fuses environmental information through DS evidence theory to generate control signals that directly drive the electronic stability program. Meanwhile, the stability assessment results are fed back to the state perception module to form a closed-loop adjustment, effectively addressing tire force prediction deviations during cornering and improving vehicle handling stability.

[0049] Embodiment 2 of this invention: In complex road conditions such as low-temperature and wet conditions, the friction coefficient between the tire and the road surface changes dynamically, requiring the model to quickly adapt to the effects of temperature and wear. The state perception module monitors thermodynamic parameters through tire pressure and temperature sensors and dynamically adjusts the sampling frequency in conjunction with motion data, increasing the sampling density when a sudden change in the road surface adhesion coefficient is detected. After reconstructing the data, the adaptive preprocessing module generates a high-confidence feature sequence by combining historical error statistics, highlighting the signal variation caused by temperature. The model evolution module's transfer learning strategy loads initial values ​​from a typical working condition parameter library and optimizes them online using a gradient descent algorithm, enabling the physics branch and the data-driven branch to collaboratively capture nonlinear characteristics under low temperatures. The collaborative optimization module integrates multi-tire data using a distributed framework, ensuring the physical rationality of parameters through trust region constraints, and quickly converges to the global optimum. The decision fusion module incorporates environmental information such as road surface humidity, reduces uncertainty through evidence theory, and outputs robust control signals. The system identifies and compensates for temperature drift in real time, reducing response latency and enhancing safety under extreme road conditions.

Claims

1. A vehicle-mounted dynamics model identification device, characterized in that, include: The state perception module is used to collect data on tire force, vehicle motion status, and tire pressure and temperature. It dynamically adjusts the data sampling frequency based on the rate of change of longitudinal acceleration and the rate of change of yaw rate, and outputs raw sensor data. An adaptive preprocessing module is used to receive the original sensor data, repair the original sensor data using deep signal reconstruction technology, calculate the confidence weight of each data point, and output a feature sequence composed of the repaired sensor data and the corresponding confidence weight. The model evolution module is used to receive the feature sequence and process it through a hybrid model including a physical constraint branch and a data-driven branch. The physical constraint branch calculates the theoretical tire force based on the parameterized magic formula, and the data-driven branch predicts the tire force residual based on the long short-term memory network. The outputs of the two branches are fused through a gated recurrent unit to generate tire dynamics parameter estimates and model prediction errors. The collaborative optimization module is used to receive the estimated values ​​of the tire dynamic parameters and the model prediction error, optimize the local model parameters corresponding to each tire using a distributed learning framework, and aggregate the local model parameter updates through a federated learning algorithm to output the optimized global tire model parameters. The decision fusion module is used to receive the global tire model parameters and vehicle bus environment information, calculate the confidence interval of the parameter estimates, and generate vehicle stability control signals through evidence theory fusion and output them to the vehicle stability control system. The decision fusion module also outputs vehicle stability state assessment results to the state perception module to dynamically adjust the data sampling frequency; the confidence weights output by the adaptive preprocessing module are used as input weights for the data-driven branch in the model evolution module. The collaborative optimization module adjusts the exploration strategy in the distributed learning framework based on the model prediction error.

2. The vehicle dynamics model identification device according to claim 1, characterized in that, The state awareness module is configured as follows: Calculate the rate of change of the vehicle's yaw rate at the current moment; The rate of change of yaw rate is compared with a preset first threshold. Based on the comparison results, when the rate of change of yaw angular velocity is less than the first threshold, sensor data is collected using a reference sampling frequency; When the rate of change of yaw rate is greater than or equal to the first threshold, sensor data is collected using a high-frequency sampling mode, wherein the frequency of the high-frequency sampling mode is higher than the reference sampling frequency.

3. The vehicle dynamics model identification device according to claim 1, characterized in that, The adaptive preprocessing module is configured as follows: The original sensor data is input into the encoder of the variational autoencoder, which maps the input data to a low-dimensional latent space and outputs a latent feature vector. The latent feature vector is input into the decoder of the variational autoencoder. The decoder integrates an attention mechanism to assign reconstruction weights to different features in the latent feature vector, and decodes the latent feature vector based on the reconstruction weights to output the reconstructed sensor data.

4. The vehicle dynamics model identification device according to claim 3, characterized in that, The adaptive preprocessing module is configured as follows: Calculate the residual between the reconstructed sensor data and the corresponding original sensor data; Obtain the historical error statistics of the sensor from which the original sensor data originates; Based on the residuals, the historical error statistics, and the consistency assessment results between the current data point and the preceding and following time series data, the confidence weights in the range of 0 to 1 are generated.

5. The vehicle dynamics model identification device according to claim 1, characterized in that, The model evolution module includes a gated loop unit, which is configured as follows: The gated loop unit receives the tire slip rate at the current moment; Using the slip ratio as input, a first weight for weighting the physical constraint branch output and a second weight for weighting the data-driven branch output are calculated; The output of the physical constraint branch is multiplied by the first weight to obtain the first weighted result; The output of the data-driven branch is multiplied by the second weight to obtain the second weighted result; The first weighted result is added to the second weighted result to output the estimated value of the tire dynamics parameter.

6. The vehicle dynamics model identification device according to claim 1, characterized in that, The collaborative optimization module uses a distributed learning framework to optimize the local model parameters for each tire, including: Create an independent local agent for each tire, and load the latest global tire model parameters as the initial model for each local agent; Each local agent uses a proximal policy optimization algorithm to update local model parameters within a trust region constraint defined by a physically reasonable range of parameter changes, with the goal of minimizing the difference between local observation data and model predictions.

7. The vehicle dynamics model identification device according to claim 6, characterized in that, The collaborative optimization module, which aggregates local model parameter updates using a federated learning algorithm, includes: The central coordinator periodically collects updated local model parameters from each of the local agents; The central coordinator assigns aggregation weights to each collected local model parameter based on the amount of data corresponding to each local agent. The central coordinator uses the aggregated weights to calculate a weighted average of all local model parameters, generating a new round of global tire model parameters. The newly generated global tire model parameters are distributed to each of the local agents.

8. The vehicle dynamics model identification device according to claim 1, characterized in that, The confidence intervals for calculating parameter estimates in the decision fusion module include: Receive the time series of tire model parameters from the collaborative optimization module; The time series of tire model parameters is smoothed using a Kalman filter to obtain smoothed parameter estimates. Calculate the innovative covariance of the Kalman filter during the smoothing process to quantify the uncertainty measure of the smoothed parameter estimates; The confidence interval for the parameter estimate is determined based on the uncertainty measure.

9. The vehicle dynamics model identification device according to claim 8, characterized in that, The decision fusion module calculates the confidence interval of the parameter estimates, and generates a vehicle stability control signal by fusing it with evidence theory, which is then output to the vehicle stability control system. A recognition framework is constructed, the input of which includes the smoothed parameter estimates and vehicle bus environment information; The DS evidence theory is used to combine and calculate multiple pieces of evidence in the identification framework to obtain a comprehensive trust function; The comprehensive trust function is input into the fuzzy logic rule table, and the vehicle stability control signal is generated through table lookup mapping.

10. The vehicle dynamics model identification device according to claim 1, characterized in that, The initial values ​​of the model parameters in the model evolution module are loaded from a pre-established typical working condition parameter library through a transfer learning strategy, which includes: Identify the similarity between the current vehicle operating condition and the typical operating conditions in the typical operating condition parameter library; Based on the similarity recognition results, the model parameters corresponding to the most similar typical working conditions are selected from the typical working condition parameter library as initial values; Starting from the initial value, the model parameters of the model evolution module are iteratively updated and optimized using the currently collected feature sequence as training data through the gradient descent algorithm.

Images