A vehicle dynamics model incremental updating method based on vehicle-cloud cooperation
By using an incremental update method based on vehicle-cloud collaboration, the problems of aging vehicle dynamics models and limited communication bandwidth were solved, enabling high-precision updates of vehicle dynamics models and improving the real-time performance and robustness of the control system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV
- Filing Date
- 2026-02-27
- Publication Date
- 2026-05-29
AI Technical Summary
Existing technologies in vehicle dynamics models suffer from problems such as model aging, insufficient vehicle-side computing power, and limited communication bandwidth, leading to divergent prediction errors and affecting the robustness of the control system.
An incremental update method based on vehicle-cloud collaboration is adopted. By deploying a lightweight forward inference model on the vehicle and collaborating with a high-performance computing layer on the cloud, incremental training and parameter optimization are achieved. Combined with "surprise" sample selection, hybrid experience playback and physical constraints, a two-way closed-loop update framework is established.
It enables high-precision maintenance of the vehicle dynamics model throughout its entire lifecycle, reduces prediction errors under new operating conditions, and improves the real-time performance and communication efficiency of the control system.
Smart Images
Figure CN122111482A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent connected vehicle control and vehicle networking technology, specifically involving an incremental update method for vehicle dynamics models based on vehicle-cloud collaboration. Background Technology
[0002] Vehicle dynamics models are the core of model predictive control (MPC) in autonomous driving and driver assistance systems. However, the vehicle dynamics system in actual operation is not an ideal linear time-invariant (LTI) system. Its key physical parameters (such as tire lateral stiffness, vehicle mass, and center of gravity position) are affected by the aging of mechanical components, changes in loading tasks, and dynamic switching of external driving environment (such as road surface adhesion coefficient), exhibiting significant time-varying characteristics.
[0003] This phenomenon of "concept drift," caused by the evolution of physical entities, leads to a gradual deviation between the fixed offline training model and the current real-world vehicle state. As mileage accumulates, the prediction error based on fixed parameters will diverge, weakening the robustness of the control system and even causing vehicle instability.
[0004] Existing technologies face a dilemma in addressing this problem: Pure on-vehicle online training is limited by the computing power and storage space of in-vehicle chips, making it difficult to support the enormous overhead of backpropagation in deep neural networks, and is prone to "catastrophic forgetting"; while pure cloud-based centralized training offers sufficient computing power, uploading all high-frequency dynamic data consumes huge communication bandwidth, and network latency prevents the cloud from responding to parameter mutations at the millisecond level. Therefore, there is an urgent need for an incremental update method that can balance real-time performance, communication costs, and model accuracy. Summary of the Invention
[0005] In view of this, the purpose of this invention is to provide an incremental update method for vehicle dynamics models based on vehicle-cloud collaboration, which aims to solve the problems of model aging, insufficient vehicle-side computing power and limited communication bandwidth in the existing technology.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] An incremental update method for vehicle dynamics models based on vehicle-cloud collaboration includes the following steps:
[0008] S1. Construct a vehicle-cloud collaborative closed-loop update framework based on a cyber-physical fusion system;
[0009] The vehicle-cloud collaborative closed-loop update framework includes the following two main lines:
[0010] Vehicle-side edge computing layer: As a physical execution terminal, it deploys a lightweight forward inference model, which is responsible for receiving the current state input and predicting the state at the next moment in real time, providing input to the controller; it also undertakes the tasks of operating condition monitoring and data screening.
[0011] High-performance computing layer in the cloud: As the central brain, it relies on a large-scale GPU cluster to undertake tasks such as data aggregation, incremental model training, parameter optimization and version management;
[0012] The vehicle-side edge computing layer and the cloud-based high-performance computing layer exchange information in a non-real-time manner via 4G / 5G networks;
[0013] S2. Establish an intelligent screening mechanism for high-value, difficult-to-sample cases based on "surprise" at the vehicle-side edge computing layer;
[0014] S3. Establish an incremental learning model update mechanism based on hybrid experience playback and physical constraints in the high-performance computing layer in the cloud;
[0015] S4. Establish a "shadow mode" model validation and parameter distribution process.
[0016] Furthermore, step S2 includes the following sub-steps:
[0017] S2.1 Real-time monitoring and residual calculation;
[0018] The system runs a monitoring thread in the background to collect data on vehicle activity. The true state value at any given moment And obtain the predicted values of the lightweight forward inference model. Calculate the residual norm between the two. ;
[0019] S2.2 Define the data trigger function It is used to determine whether the current working condition has a high degree of "surprise";
[0020]
[0021] In the formula, Represents the state vector of the th The actual values of the dimensional components; Represents the state vector of the th Predicted values of dimensional components; The baseline tolerance threshold is set to three times the root mean square error of the offline model on the test set, following... Guidelines; This is the sensitivity weighting coefficient.
[0022] S2.3 Timing Context Capture and Upload;
[0023] when At that time, it was determined to be a high-value, difficult sample; the system automatically backtracked and extracted samples before and after the trigger time. Step time window sequence Store in a high-priority cache;
[0024] When the cache accumulation reaches the target or the network environment is good, package and upload to the cloud.
[0025] Furthermore, step S3 includes the following sub-steps:
[0026] S3.1 Construct a hybrid experience replay mechanism;
[0027] Maintain a fixed-capacity memory sample library in the cloud. It stores benchmark samples covering typical historical operating conditions;
[0028] When receiving incremental dataset At that time, according to the preset mixing ratio Replay a subset of historical data from the memory bank. Construct a hybrid training set :
[0029]
[0030] S3.2 Introduces physical feasibility constraints;
[0031] Constructing a physical constraint loss function This is used to prevent overfitting or violations of physical principles caused by purely data-driven approaches.
[0032] Lateral acceleration based on model prediction Using Newton's second law to deduce the equivalent lateral force Construct penalty terms:
[0033]
[0034] Where N represents the total number of samples; k represents the sample index; This represents the longitudinal force in the k-th sample. This represents the lateral force in the k-th sample. This represents the vertical load under the k-th sample; This is an estimate of the maximum adhesion coefficient of the road surface;
[0035] S3.3 Parameter Iterative Update;
[0036] Define the overall objective function :
[0037]
[0038] in, Losses due to new missions; Loss of memory due to old tasks; The weighting coefficients for memory loss from previous tasks; For physical constraint loss Weighting coefficients;
[0039] Mini-batch stochastic gradient descent is used to optimize network parameters. Make minor adjustments and updates:
[0040]
[0041] in, This refers to the network parameters before the update; the old parameters. This represents the updated network parameters, the new parameters; This is the learning rate.
[0042] Furthermore, step S4 includes the following sub-steps:
[0043] S4.1 Shadow Mode Verification;
[0044] Updated new model parameters Deployed to a cloud-based high-fidelity simulation platform or run silently in the background of a test vehicle;
[0045] S4.2 Consistency Monitoring;
[0046] The consistency index of the new and old models under normal operating conditions is calculated. The model is validated only when the new model shows a significant improvement in performance under difficult operating conditions and does not experience performance degradation under normal operating conditions, i.e., no severe fluctuations or prediction biases.
[0047] S4.3 OTA deployment;
[0048] Verified parameters are transmitted via over-the-air download technology. The data is then sent to the vehicle to replace the old parameters, completing the online evolution of the model.
[0049] Beneficial effects:
[0050] This invention provides an incremental update method for vehicle dynamics models based on vehicle-cloud collaboration. This method follows the concept of "physical perception—digital evolution—physical feedback," establishing a two-way closed-loop link connecting the vehicle-side entity and the cloud-based digital twin to achieve high-precision maintenance of the dynamics model throughout its entire lifecycle. This method effectively overcomes the catastrophic forgetting defect of deep learning, significantly reducing prediction errors for new operating conditions while maintaining high accuracy for existing conditions.
[0051] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0052] Figure 1 This is a flowchart of an incremental update method for a vehicle dynamics model based on vehicle-cloud collaboration according to the present invention;
[0053] Figure 2 This is a diagram of a model update framework based on vehicle-cloud collaboration.
[0054] Figure 3 This is a schematic diagram of the data filtering logic based on surprise level. Detailed Implementation
[0055] To make the technical solutions, advantages, and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the protection scope of this application.
[0056] like Figures 1-3 As shown, this invention provides an incremental update method for a vehicle dynamics model based on vehicle-cloud collaboration, specifically including the following steps:
[0057] S1. Construct a vehicle-cloud collaborative closed-loop update framework based on Cyber-Physical Systems (CPS).
[0058] 1. Vehicle Edge Layer: As a physical execution terminal, it deploys a lightweight forward inference model (such as DISF-GRU-FNN) to receive the current state input and predict the state at the next moment in real time, providing input to the controller; it also undertakes the tasks of operating condition monitoring and data screening.
[0059] 2. Cloud High-Performance Computing Layer (HPC Layer): Serving as the central brain, it relies on a large-scale GPU cluster to handle tasks such as massive data aggregation, incremental model training, parameter optimization, and version management. The two layers interact non-real-time via 4G / 5G networks.
[0060] S2. At the vehicle-side edge computing layer, perform intelligent screening of high-value, difficult-to-find samples based on "surprise".
[0061] Step 2.1: Real-time monitoring and residual calculation. A monitoring thread runs in the system background, collecting data on vehicle speeds. The true state value at any given moment And obtain the model prediction value. Calculate the residual norm between the two. .
[0062] Step 2.2: Define the data trigger function Used to determine whether the current operating condition has a high degree of "surprise":
[0063]
[0064] in, Represents the first in the state vector Dimensional components (such as yaw rate); The baseline tolerance threshold is set to three times the root mean square error of the offline model on the test set (following...). Guidelines); This is the sensitivity weighting coefficient.
[0065] Step 2.3: Timing Context Extraction and Upload. When At that time, the system determined it to be a high-value, difficult sample. The system automatically backtracked and extracted data before and after the trigger time. Step time window sequence The data is stored in a high-priority cache. When the cache reaches a certain threshold or the network environment is favorable, it is packaged and uploaded to the cloud. This strategy can reduce data upload traffic by more than 90%.
[0066] S3. In the high-performance computing layer in the cloud, perform incremental learning based on hybrid experience playback and physical constraints.
[0067] The cloud receives difficult example samples uploaded by the vehicle and combines them with benchmark samples in the historical memory bank to construct a hybrid training set. The concept of Physical Information Neural Network (PINN) is introduced to add physical feasibility constraints to the loss function and fine-tune the model parameters.
[0068] Step 3.1: Construct a Mixed Experience Replay mechanism. To address "catastrophic forgetting," a fixed-capacity memory sample database is maintained in the cloud. It stores benchmark samples covering typical historical operating conditions (selected based on the principle of maximizing feature coverage).
[0069] When receiving incremental dataset At that time, according to the preset mixing ratio Replay a subset of historical data from the memory bank. Construct a hybrid training set :
[0070]
[0071] Step 3.2: Introduce physical feasibility constraints. To prevent overfitting or violations of physical principles caused by purely data-driven approaches (such as lateral forces exceeding adhesion limits), a physical constraint loss function is constructed. .
[0072] Lateral acceleration based on model prediction Using Newton's second law to deduce the equivalent lateral force Construct penalty terms:
[0073]
[0074] Where N represents the total number of samples; k represents the sample index; This represents the longitudinal force in the k-th sample. This represents the lateral force in the k-th sample. This represents the vertical load under the k-th sample; This is an estimate of the maximum adhesion coefficient of the road surface. The ReLU function guarantees that a penalty gradient is generated only when the resultant force exceeds the physical limit.
[0075] Step 3.3: Parameter Iteration and Update. Define the overall objective function. :
[0076]
[0077] in, Losses due to new missions; Loss of memory due to old tasks; The weighting coefficients for memory loss from previous tasks; For physical constraint loss The weight coefficients are determined. Mini-batch stochastic gradient descent (SGD) is used to adjust the network parameters. Make minor adjustments and updates:
[0078]
[0079] in, This refers to the network parameters before the update; the old parameters. This represents the updated network parameters, the new parameters; A relatively small learning rate.
[0080] S4: Perform "shadow mode" verification and model deployment.
[0081] Step 4.1: Shadow Mode Validation. The updated model parameters... Deployed to a high-fidelity simulation platform in the cloud or running silently in the background of a test vehicle. The new model receives real input but does not participate in control.
[0082] Step 4.2: Consistency Monitoring. Calculate the consistency index of the new and old models under normal operating conditions. The model passes validation only when the new model shows a significant improvement in performance under challenging conditions and does not experience performance degradation (i.e., no severe fluctuations or prediction bias) under normal operating conditions.
[0083] Step 4.3: OTA Deployment. Verified parameters are deployed via Over-the-Air (OTA) technology. The data is then sent to the vehicle to replace the old parameters, completing the online evolution of the model.
[0084] It is hereby declared 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 present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for incremental update of vehicle dynamics model based on vehicle-cloud collaboration, characterized in that, Includes the following steps: S1. Construct a vehicle-cloud collaborative closed-loop update framework based on a cyber-physical fusion system; The vehicle-cloud collaborative closed-loop update framework includes the following two main lines: Vehicle-side edge computing layer: As a physical execution terminal, it deploys a lightweight forward inference model, which is responsible for receiving the current state input and predicting the state at the next moment in real time, providing input to the controller; it also undertakes the tasks of operating condition monitoring and data screening. High-performance computing layer in the cloud: As the central brain, it relies on a large-scale GPU cluster to undertake tasks such as data aggregation, incremental model training, parameter optimization and version management; The vehicle-side edge computing layer and the cloud-based high-performance computing layer exchange information in a non-real-time manner via 4G / 5G networks; S2. Establish an intelligent screening mechanism for high-value, difficult-to-sample cases based on "surprise" at the vehicle-side edge computing layer; S3. Establish an incremental learning model update mechanism based on hybrid experience playback and physical constraints in the high-performance computing layer in the cloud; S4. Establish a "shadow mode" model validation and parameter distribution process.
2. The incremental update method for vehicle dynamics model based on vehicle-cloud collaboration according to claim 1, characterized in that, Step S2 includes the following sub-steps: S2.1 Real-time monitoring and residual calculation; The system runs a monitoring thread in the background to collect data on vehicle activity. The true state value at any given moment And obtain the predicted values of the lightweight forward inference model. Calculate the residual norm between the two. ; S2.2 Define the data trigger function It is used to determine whether the current working condition has a high degree of "surprise". In the formula, Represents the state vector of the th The actual values of the dimensional components; Represents the state vector of the th Predicted values of dimensional components; The baseline tolerance threshold is set to three times the root mean square error of the offline model on the test set, following... Guidelines; This is a sensitivity-weighted coefficient. S2.3 Timing context capture and uploading; when At that time, it was determined to be a high-value, difficult sample; the system automatically backtracked and extracted samples before and after the trigger time. Step time window sequence Store in a high-priority cache; When the cache accumulation reaches the target or the network environment is good, package and upload to the cloud.
3. The incremental update method for a vehicle dynamics model based on vehicle-cloud collaboration according to claim 2, characterized in that, Step S3 includes the following sub-steps: S3.1 Construct a hybrid experience replay mechanism; Maintain a fixed-capacity memory sample library in the cloud. It stores benchmark samples covering typical historical operating conditions; When receiving incremental dataset At that time, according to the preset mixing ratio Replay a subset of historical data from the memory bank. Construct a hybrid training set : S3.2 Introduces physical feasibility constraints; Constructing a physical constraint loss function This is used to prevent overfitting or violations of physical principles caused by purely data-driven approaches. Lateral acceleration based on model prediction Using Newton's second law to deduce the equivalent lateral force Construct penalty terms: Where N represents the total number of samples; k represents the sample index; This represents the longitudinal force in the k-th sample. This represents the lateral force in the k-th sample. This represents the vertical load under the k-th sample; This is an estimate of the maximum adhesion coefficient of the road surface; S3.3 Parameter Iterative Update; Define the overall objective function : in, Losses due to new missions; Loss of memory due to old tasks; The weighting coefficients for memory loss from previous tasks; For physical constraint loss Weighting coefficients; Mini-batch stochastic gradient descent is used to optimize network parameters. Make minor adjustments and updates: in, This refers to the network parameters before the update; the old parameters. This represents the updated network parameters, the new parameters; This is the learning rate.
4. The incremental update method for a vehicle dynamics model based on vehicle-cloud collaboration according to claim 3, characterized in that, Step S4 includes the following sub-steps: S4.1 Shadow Mode Verification; Updated new model parameters Deployed to a cloud-based high-fidelity simulation platform or run silently in the background of a test vehicle; S4.2 Consistency Monitoring; The consistency index of the new and old models under normal operating conditions is calculated. The model is validated only when the new model shows a significant improvement in performance under difficult operating conditions and does not experience performance degradation under normal operating conditions, i.e., there is no severe fluctuation or prediction bias. S4.3 OTA deployment; Verified parameters are transmitted via over-the-air download technology. The data is then sent to the vehicle to replace the old parameters, completing the online evolution of the model.