Distributed independent electric drive vehicle torque distribution control method based on layering

Abstract

The invention discloses a distributed independent electric drive vehicle torque distribution control method based on layering. Comprising the steps that the vehicle state is collected in real time through a vehicle-mounted sensor network, and a layered reinforcement learning controller is constructed and comprises an upper-layer decision module and a lower-layer safety barrier module; the upper layer decision-making module adopts a DDPG algorithm, outputs an original action vector according to a vehicle state, and converts the original action vector into a torque distribution coefficient through a Softmax function; and the lower-layer safety barrier module corrects the torque distribution coefficient. And constructing a dynamic weight reward function, and dynamically adjusting an energy consumption efficiency weight, a tire load rate weight and a torque change rate weight according to working conditions. And deploying an incremental learning mechanism, freezing a part of network parameters and only finely adjusting a key layer when road adhesion sudden change is detected, and quickly adjusting a control strategy. The comprehensive performance of the distributed driving electric automobile in the aspects of safety, stability and energy efficiency is remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving control technology, and particularly relates to a hierarchical distributed independent electric drive vehicle torque distribution control method. Background Technology

[0002] In the field of torque distribution control for distributed drive electric vehicles, existing technical solutions have several limitations. Rule-based control methods require the pre-setting of numerous operating conditions, leading to significant system response delays in complex scenarios (such as sharp turns on icy or snowy roads), making it difficult to prevent vehicle instability caused by tire force saturation. These rule-based control methods typically rely on conditions and logic set by engineers based on experience. While they can provide some control under common operating conditions, their adaptability and response speed are clearly insufficient when faced with complex situations without pre-set conditions. This limitation stems from their static rule system, which cannot be dynamically adjusted to cope with real-time changes in the environment and vehicle state.

[0003] Model predictive control (MPC) is an advanced control strategy that optimizes the future system state through predictive models, theoretically achieving multi-objective optimal control. However, MPC has an extremely heavy computational burden, especially when considering multiple input variables and constraints, resulting in very high time costs for online solution of optimization problems. This makes it difficult for MPC to meet real-time requirements under high-speed dynamic conditions, particularly in emergency situations requiring rapid response, such as high-speed obstacle avoidance or sudden road surface changes. In these situations, the delay in MPC can lead to the failure of vehicle stability control.

[0004] Existing control strategies are mostly based on fixed-parameter models, which assume specific vehicle parameters and environmental conditions during design, such as the coefficient of friction between the tires and the road surface, the vehicle's center of gravity position, and inertial parameters. However, in actual driving, the coefficient of friction may change suddenly due to factors such as weather and road conditions (e.g., driving from a dry asphalt road into a flooded or icy road), and the vehicle load may also change due to variations in the number of passengers or the weight of cargo. These changes can lead to a sharp degradation in the performance of control strategies based on fixed-parameter models, as the models cannot adapt to new environments or vehicle states in a timely manner, thus affecting the vehicle's handling stability and safety.

[0005] While end-to-end reinforcement learning methods possess some adaptive potential, optimizing control strategies through interactive learning with the environment, their training process carries significant safety risks. During the training phase, vehicles may need to experience various hazardous conditions to learn the optimal strategy, which could potentially lead to accidents in real-world driving. Furthermore, end-to-end reinforcement learning methods have limited online adaptability to new conditions, typically requiring substantial data and time for training. Their inability to quickly adapt to new situations is a serious drawback for vehicle control systems that require real-time adjustments.

[0006] Furthermore, torque distribution control requires optimizing multiple objectives simultaneously, including energy efficiency, handling stability, and smoothness of execution. Traditional methods typically employ a fixed-weight trade-off strategy, where the weight coefficients for each objective are pre-set in the control algorithm. However, this fixed-weight strategy cannot dynamically adjust the optimization focus based on real-time operating conditions. For example, when the vehicle is operating at high temperatures, it may be necessary to reduce the weight of energy efficiency and increase the weight of motor protection to prevent overheating and damage. Existing methods struggle to achieve this dynamic adjustment, resulting in suboptimal control performance under different operating conditions. Summary of the Invention

[0007] This invention proposes a hierarchical distributed independent electric drive vehicle torque distribution control method to solve the problems existing in the prior art.

[0008] To achieve the above objectives, the present invention provides a hierarchical distributed independent electric drive vehicle torque distribution control method, comprising the following steps: The vehicle status is collected in real time through an onboard sensor network, including longitudinal speed, lateral speed, yaw rate, four-wheel slip ratio, motor temperature, and estimated road adhesion coefficient. A hierarchical reinforcement learning controller is constructed, comprising an upper-layer decision-making module and a lower-layer safety barrier module. The upper-layer decision-making module outputs the original action vector based on the collected vehicle state and converts it into torque distribution coefficients through the Softmax function. The lower-layer safety barrier module corrects the torque distribution coefficients based on the tire load rate to prevent excessive wheel load. A dynamic weighted reward function is constructed to dynamically adjust the reward weights according to the vehicle status and control objectives. The reward weights include energy efficiency weights, tire load rate weights, and torque change rate weights. When a sudden change in the road surface adhesion coefficient is detected, an incremental learning mechanism is triggered to adjust the key layer, thereby quickly adjusting the control strategy.

[0009] Optionally, obtaining the road surface estimated adhesion coefficient includes: The rate of change of longitudinal force in each tire is calculated using the torque signal from the hub motor. Establish a mapping relationship between the slip ratio and the road surface adhesion coefficient under specific slip ratio conditions; The road surface adhesion coefficient was estimated by normalization.

[0010] Optionally, the expression for obtaining the estimated value of the road surface adhesion coefficient is: In the formula, Indicates the first The longitudinal force of each tire is transmitted through the torque of the hub motor. Wheel rotational inertia Wheel angular acceleration and wheel radius calculate: Indicates the first The slip ratio of each tire, for driving conditions, is defined as: in For the longitudinal speed of the vehicle, This represents the angular velocity of the wheel.

[0011] This represents the partial derivative of the longitudinal force with respect to the slip ratio, in the low slip ratio region ( ) Calculated using real-time differential approximation: The normalization coefficient, obtained through experimental calibration, reflects the proportional relationship between tire longitudinal stiffness and adhesion coefficient under standard road conditions.

[0012] Optionally, the upper-level decision-making module adopts the DDPG algorithm, in which the Actor network takes the state vector as input and outputs the original action vector, which is then converted into torque distribution coefficients through the Softmax layer.

[0013] Optionally, when the load rate of any tire exceeds a preset threshold, the lower-level safety barrier module calculates the maximum allowable torque based on a simplified linear tire model and corrects the distribution coefficient.

[0014] Optionally, when the maximum tire load rate exceeds a preset safety threshold, a segmented penalty function is activated to monitor and penalize the tire load rate in real time.

[0015] Optionally, the piecewise penalty function expression is: ; In the formula, This represents the maximum tire load rate.

[0016] Optionally, the expression for the dynamic weighted reward function is: ; In the formula, , where represents the drive power efficiency term, For wheel torque, The angular velocity of the wheel. This represents the total maximum power of the motor. This represents the maximum tire load rate. , which is the torque change rate, reflects ride comfort. , where is the motor temperature compensation weight, Motor temperature (take the highest value). The rated temperature (e.g., 80℃). This is a scale parameter (e.g., 10℃). The yaw rate compensation weights are as follows: The yaw rate is angular velocity. The threshold is (e.g., 0.5 rad / s). This is the smoothness weighting constant.

[0017] Optionally, the incremental learning mechanism freezes the parameters of the Actor network's convolutional and LSTM layers and adjusts the weights of the fully connected layers when a sudden change in road surface adhesion is detected.

[0018] Optionally, the incremental learning mechanism includes: establishing a mutation detector based on tire force feedback, which determines road characteristic mutations and triggers incremental learning by comparing the actual tire force with the nominal model prediction value in real time.

[0019] Compared with the prior art, the present invention has the following advantages and technical effects: The present invention proposes a hierarchical distributed independent electric drive vehicle torque distribution control method, which has significant advantages and multiple technical effects compared with the prior art: First, this invention achieves rapid response and safe control under complex operating conditions by constructing a hierarchical reinforcement learning controller. The upper-level decision module utilizes the DDPG algorithm to make macroscopic torque distribution decisions based on real-time vehicle status and environmental information; the lower-level safety barrier module is responsible for real-time monitoring and adjustment at the microscopic level, ensuring that the torque output of each wheel is within a safe range and effectively preventing tire force saturation. This hierarchical architecture not only improves the system's response speed but also enhances its safety, ensuring stable vehicle operation under various conditions.

[0020] Secondly, this invention utilizes a dynamic weighted reward function and an incremental learning mechanism to adapt to changes in the environment and vehicle status in real time, effectively improving the system's environmental adaptability and multi-objective collaborative capabilities. The dynamic weighted reward function allows the system to adjust the weights of various control objectives based on real-time operating conditions. For example, it automatically increases the weight of stability when a decrease in road surface adhesion coefficient is detected, or reduces the weight of energy efficiency to protect the motor under high-temperature conditions. The incremental learning mechanism enables the system to quickly adjust its strategy when facing new operating conditions. By freezing some network parameters and fine-tuning only key layers, it significantly reduces online learning time and computational resource requirements, thereby achieving rapid adaptation.

[0021] Furthermore, this invention introduces action space mapping technology and a safety-term optimization mechanism to further optimize the accuracy of torque distribution and vehicle stability. The action space mapping technology uses a Softmax function to convert the original action vector into a probability distribution, ensuring that the torque distribution coefficients meet physical constraints and avoiding total torque overshoot. The safety-term optimization mechanism designs a nonlinear gradient penalty function in the reward function to monitor and penalize tire load rates in real time, forcing the agent to actively avoid tire saturation zones during decision-making, thereby improving vehicle stability and safety.

[0022] Finally, this invention significantly improves the overall performance of distributed drive electric vehicles in terms of safety, stability, and energy efficiency. Through an innovative hierarchical reinforcement learning architecture and multiple optimization mechanisms, this method not only enhances the vehicle's adaptability and response speed under complex operating conditions but also strengthens the system's robustness and multi-objective cooperative capabilities. Specifically, the vehicle's stability control performance under complex conditions such as emergency obstacle avoidance and sudden road surface changes is significantly improved; its adaptability to different environmental conditions is enhanced, enabling rapid response to changes in road adhesion coefficient and vehicle load; and the dynamic adjustment capability of multi-objective optimization allows the vehicle to achieve the optimal balance between energy efficiency, handling stability, and smooth operation under various conditions. These improvements not only enhance the overall performance of the vehicle but also provide drivers with a safer and more stable driving experience, while also offering new directions and ideas for the development of torque distribution control technology for electric vehicles. Attached Figure Description

[0023] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention. Detailed Implementation

[0024] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0026] Example 1 like Figure 1 As shown, this embodiment provides a hierarchical distributed independent electric drive vehicle torque distribution control method, including the following steps: The vehicle status is collected in real time through an onboard sensor network, including longitudinal speed, lateral speed, yaw rate, four-wheel slip ratio, motor temperature, and estimated road adhesion coefficient. A hierarchical reinforcement learning controller is constructed, comprising an upper-layer decision-making module and a lower-layer safety barrier module. The upper-layer decision-making module outputs the original action vector based on the collected vehicle state and converts it into torque distribution coefficients through the Softmax function. The lower-layer safety barrier module corrects the torque distribution coefficients based on the tire load rate to prevent excessive wheel load. A dynamic weighted reward function is constructed to dynamically adjust the reward weights according to the vehicle status and control objectives. The reward weights include energy efficiency weights, tire load rate weights, and torque change rate weights. When a sudden change in the road surface adhesion coefficient is detected, an incremental learning mechanism is triggered to adjust the key layer, thereby quickly adjusting the control strategy.

[0027] Furthermore, vehicle status, including longitudinal speed, is collected in real time through an onboard sensor network. Lateral velocity yaw rate Four-wheel slip ratio Motor temperature Road surface adhesion coefficient estimation This provides a foundation for subsequent control. Among these factors is the road surface adhesion coefficient. The estimation employs a tire stiffness inverse estimation mechanism: the rate of change of longitudinal force of each tire is calculated using the torque signal from the hub motor. Establish a mapping relationship under specific slip ratio conditions: Calculate the rate of change of longitudinal force in each tire using the torque signal from the hub motor: The road surface adhesion coefficient was estimated by normalization. The normalization coefficient, obtained through experimental calibration, reflects the proportional relationship between tire longitudinal stiffness and adhesion coefficient under standard road conditions.

[0028] Torque via hub motor Wheel rotational inertia Wheel angular acceleration and wheel radius calculate: Indicates the first The slip ratio of each tire, for driving conditions, is defined as: in For the longitudinal speed of the vehicle, This represents the angular velocity of the wheel.

[0029] Where the normalization coefficient The range [0.15, 0.25] was obtained through training on a large dataset of 20 different road surfaces and 100 different working conditions, covering a full range of scenarios from icy surfaces to dry asphalt. This estimated value serves as the state vector. The key input directly affects control decisions.

[0030] Furthermore, constructing a hierarchical reinforcement learning controller includes: Upper-level decision module: Employs the DDPG algorithm, with the Actor network inputting the state vector. The original action vector [a1,a2,a3,a4] is output. The original action vector [a1,a2,a3,a4] output by the Actor network is transformed using the Softmax function to form a probability distribution: ; here, This represents the normalized torque distribution coefficient for each wheel. The use of the Softmax function ensures that the sum of all distribution coefficients is 1, and the final torque executed at each wheel consists of two parts: one is the average torque value calculated based on the total driving torque. ; Second, based on safety correction amount After adjustments, the final output torque is expressed as: ; Lower safety barrier module: When any tire load rate The maximum permissible torque is calculated based on a simplified linear tire model. And adjust the distribution coefficient to prevent excessive wheel load. Its expression is: .

[0031] Furthermore, a safety term optimization mechanism for tire force saturation risk is constructed within the reinforcement learning reward function. Its core lies in designing a nonlinear gradient penalty function to dynamically ensure vehicle stability. When the maximum tire load rate... When the threshold of 0.8 is exceeded, a segmented penalty function is activated: ; This function passes through quadratic terms. A gradient relationship is established between the penalty intensity and the tire load rate, forcing the agent to actively avoid tire saturation zones during the decision-making process. This mechanism, together with the lower-level safety barrier module, forms a collaborative protection system: at the decision-making level, the reward function guides the agent to dynamically adjust the torque distribution coefficient. When real-time detection is performed at the execution layer When the value is greater than 0.95, the safety barrier module injects a torque correction amount within 1ms. .

[0032] The expression for the dynamic weighted reward function is: In the formula, For the drive power efficiency term, where For wheel torque, The angular velocity of the wheel. This represents the total maximum power of the motor. This represents the maximum tire load rate. This represents the torque variation rate, reflecting ride comfort. , where is the motor temperature compensation weight, Motor temperature (take the highest value). The rated temperature (e.g., 80℃). This is a scale parameter (e.g., 10℃). The yaw rate compensation weights are as follows: The yaw rate is angular velocity. The threshold is (e.g., 0.5 rad / s). This is the smoothness weighting constant.

[0033] Furthermore, to enable online model evolution, a mutation detector based on tire force feedback is established. This is achieved by comparing the actual tire force with the nominal model prediction in real time. ; in The current average actual adhesion coefficient, nominal adhesion coefficient Update using a sliding window (window duration 10 seconds).

[0034] ; when If the value >0.2 persists for 200ms (e.g., when a vehicle enters a flooded area), it is considered a sudden change in road surface characteristics and incremental learning is triggered. This mechanism significantly improves adaptability to operating conditions and avoids the performance degradation caused by model mismatch in traditional solutions.

[0035] During incremental learning, a hierarchical parameter update strategy is implemented: all weights of the feature extraction layer (including convolutional and LSTM layers) are frozen to preserve general tire dynamics features and avoid the knowledge loss problem caused by retraining in traditional methods. This applies only to fully connected layers. Fine-tuning is performed to significantly reduce the computational load. The update process uses an exponentially decaying learning rate. ; in t represents the initial learning rate, and t represents the number of fine-tuning iterations.

[0036] Furthermore, this embodiment also provides a performance verification method: defining a comprehensive performance index: ; Where the coefficients α=0.5, β=0.3 and γ=0.2, The baseline energy consumption was used. Then, a dual lane change scenario was employed: simulating an emergency lane change maneuver while the vehicle was traveling at a high speed of 80 km / h. The vehicle needed to complete two directional changes within a specified distance, during which the system adjusted the torque of each wheel in real time to ensure stable response to steering commands, verifying the stability and accuracy of the control system under emergency lane change scenarios. Next, emergency braking on a low-friction surface was simulated: simulating an emergency braking maneuver on a slippery surface with a friction coefficient μ=0.3. The system needed to quickly distribute torque to each wheel to prevent wheel lock-up and maintain vehicle stability, verifying the stability and safety of the control system under low-friction conditions.

[0037] The present invention has the following effects: Hierarchical Reinforcement Learning Controller: Traditional control methods, such as rule-based control and model predictive control (MPC), have significant limitations when dealing with complex operating conditions. Rule-based control relies on a large number of pre-set operating conditions, leading to response delays in complex scenarios not covered, and even failing to effectively prevent vehicle instability caused by tire force saturation. While MPC can optimize multiple objectives, its computational burden is heavy, making it difficult to meet the real-time requirements of high-speed dynamic conditions. This method constructs a hierarchical reinforcement learning controller, combining an upper-level decision-making module and a lower-level safety barrier module, to achieve rapid response and safe control under complex operating conditions. This hierarchical architecture not only improves the system's response speed but also enhances its safety, ensuring stable vehicle operation under various conditions.

[0038] Dynamic weighted reward function and incremental learning mechanism: Existing control strategies are mostly based on fixed-parameter models, making it difficult to adapt to conditions such as sudden changes in road adhesion coefficient or vehicle load variations. While end-to-end reinforcement learning has adaptive potential, the training process carries safety risks and has limited online adaptability to new conditions. This method utilizes a dynamic weighted reward function, which can adjust the weights of each control objective according to real-time conditions. For example, it automatically increases the weight of stability when a decrease in road adhesion coefficient is detected, or reduces the weight of energy efficiency under high-temperature conditions to protect the motor. Simultaneously, the incremental learning mechanism enables the system to quickly adapt to new conditions. By freezing some network parameters and fine-tuning only key layers, it significantly reduces online learning time and computational resource requirements, thus achieving rapid adaptation. This dynamic adjustment and rapid learning capability significantly improves the system's environmental adaptability and multi-objective collaborative capability.

[0039] Action space mapping technology and safety-term optimization mechanism: Traditional multi-objective optimization methods employ fixed-weight trade-off strategies, which cannot dynamically adjust the optimization focus based on real-time operating conditions. This method introduces action space mapping technology, using the Softmax function to convert the original action vector into a probability distribution, ensuring that the torque distribution coefficient meets physical constraints and avoiding total torque overshoot. Furthermore, the safety-term optimization mechanism incorporates a nonlinear gradient penalty function in the reward function to monitor and penalize tire load rates in real time, forcing the agent to actively avoid tire saturation zones during decision-making. The application of these technologies further optimizes the accuracy of torque distribution and vehicle stability, ensuring safe driving under various operating conditions.

[0040] Performance Enhancement: Overall, the proposed method significantly improves the comprehensive performance of distributed drive electric vehicles in terms of safety, stability, and energy efficiency. Through an innovative hierarchical reinforcement learning architecture and multiple optimization mechanisms, this method not only improves the vehicle's adaptability and response speed under complex operating conditions but also enhances the system's robustness and multi-objective cooperative capabilities. Specifically, the vehicle's stability control performance under complex conditions such as emergency obstacle avoidance and sudden road surface changes is significantly improved; adaptability to different environmental conditions is enhanced, enabling rapid response to changes in road adhesion coefficient and vehicle load; and the dynamic adjustment capability of multi-objective optimization allows the vehicle to achieve the optimal balance between energy efficiency, handling stability, and smooth operation under various operating conditions. These improvements not only enhance the overall performance of the vehicle but also provide drivers with a safer and more stable driving experience, while also offering new directions and ideas for the development of torque distribution control technology for electric vehicles.

[0041] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A hierarchical distributed independent electric drive vehicle torque distribution control method, characterized in that, Includes the following steps: The vehicle status is collected in real time through an onboard sensor network, including longitudinal speed, lateral speed, yaw rate, four-wheel slip ratio, motor temperature, and estimated road adhesion coefficient. A hierarchical reinforcement learning controller is constructed, comprising an upper-layer decision-making module and a lower-layer safety barrier module. The upper-layer decision-making module outputs the original action vector based on the collected vehicle state and converts it into torque distribution coefficients through the Softmax function. The lower-layer safety barrier module corrects the torque distribution coefficients based on the tire load rate to prevent excessive wheel load. A dynamic weighted reward function is constructed to dynamically adjust the reward weights according to the vehicle status and control objectives. The reward weights include energy efficiency weights, tire load rate weights, and torque change rate weights. When a sudden change in the road surface adhesion coefficient is detected, an incremental learning mechanism is triggered to adjust the key layer, thereby quickly adjusting the control strategy.

2. The method according to claim 1, characterized in that, The acquisition of the road surface estimated adhesion coefficient includes: The rate of change of longitudinal force in each tire is calculated using the torque signal from the hub motor. Establish a mapping relationship between the slip ratio and the road surface adhesion coefficient under specific slip ratio conditions; The road surface adhesion coefficient was estimated by normalization.

3. The method according to claim 1, characterized in that, The expression for obtaining the estimated value of the road surface adhesion coefficient is: In the formula, Indicates the first The longitudinal force of each tire is transmitted through the torque of the hub motor. Wheel rotational inertia Wheel angular acceleration and wheel radius calculate: Indicates the first The slip ratio of each tire, for driving conditions, is defined as: in For the longitudinal speed of the vehicle, The wheel's angular velocity; This represents the partial derivative of the longitudinal force with respect to the slip ratio. This is the normalization coefficient.

4. The method according to claim 1, characterized in that, The upper-level decision-making module adopts the DDPG algorithm, whose Actor network takes the state vector as input and outputs the original action vector, which is then converted into torque distribution coefficients through the Softmax layer.

5. The method according to claim 1, characterized in that, When the load rate of any tire exceeds a preset threshold, the lower-level safety barrier module calculates the maximum allowable torque based on a simplified linear tire model and corrects the distribution coefficient.

6. The method according to claim 1, characterized in that, When the maximum tire load rate exceeds the preset safety threshold, a segmented penalty function is activated to monitor and penalize the tire load rate in real time.

7. The method according to claim 1, characterized in that, The expression for the piecewise penalty function is: ； In the formula, This represents the maximum tire load rate.

8. The method according to claim 1, characterized in that, The expression for the dynamic weighted reward function is: ； In the formula, For the drive power efficiency term, For maximum tire load rate, This represents the rate of change of torque. For motor temperature compensation, For yaw rate compensation, This is the smoothness weighting constant.

9. The method according to claim 1, characterized in that, When the incremental learning mechanism detects a sudden change in road surface adhesion, it freezes the parameters of the Actor network's convolutional and LSTM layers and adjusts the weights of the fully connected layers.

10. The method according to claim 1, characterized in that, The incremental learning mechanism includes: establishing a mutation detector based on tire force feedback, which determines road characteristic mutations and triggers incremental learning by comparing the actual tire force with the nominal model prediction value in real time.

Images