A distributed electric drive vehicle adaptive drive torque distribution method

Abstract

The present application belongs to the field of four-wheel drive torque distribution technology of distributed electric drive vehicles, and in particular to a kind of adaptive driving torque distribution method of distributed electric drive vehicle.It includes the following steps:S1: establish distributed electric drive vehicle environment model;S2: complete four-wheel drive torque adaptive distribution strategy design based on deep deterministic policy gradient;S3: complete the design of DDPG algorithm neural network structure;S4: design longitudinal driving condition training set;S5: carry out four-wheel drive torque adaptive distribution strategy training based on DDPG;S6: carry out simulation test.The present application enables agent to comprehensively utilize vehicle motion state and road adhesion condition to make driving torque distribution decision, and reduces dependence on road adhesion coefficient estimation information.At the same time, random disturbance and first-order lag link are introduced to road adhesion coefficient observation in training process, training environment with adhesion coefficient estimation error is constructed, and strategy robustness is improved.

Description

Technical Field

[0001] This invention relates to the field of four-wheel drive torque distribution technology for distributed electric drive vehicles, specifically to an adaptive drive torque distribution method for distributed electric drive vehicles. Background Technology

[0002] With the increasing electrification and intelligence of automobiles, vehicle chassis structures are gradually developing towards full drive-by-wire and distributed systems. Compared to centralized electric drive vehicles, distributed electric drive vehicles have independent drive motors for each wheel, allowing for flexible coordination of drive torque based on road surface adhesion conditions. This more fully utilizes the adhesion potential of each wheel, improving vehicle driving efficiency and driving stability. Patent CN202512012829.4 improves driving stability under complex conditions such as wet and slippery roads by acquiring information such as vehicle operating status and road surface adhesion conditions and using a quadratic programming optimization algorithm to adaptively optimize the allocation of four-wheel drive torque. Patent CN202511396975.5 proposes a reinforcement learning-based four-wheel drive torque allocation method for low-adhesion roads, dynamically correcting the control strategy based on the rate of change of the adhesion coefficient, thereby improving the vehicle's stability and adaptability on complex low-adhesion and varying adhesion conditions.

[0003] The methods described above all rely on accurate estimation of the road surface adhesion coefficient to adjust the four-wheel drive torque, adapting to changes in four-wheel adhesion conditions. Existing methods cannot directly measure the road surface adhesion coefficient of the four wheels; they can only obtain it indirectly through relevant estimation methods. However, due to factors such as sensor measurement errors and vehicle model parameter deviations, there will be a certain error between the estimated road surface adhesion coefficient and the actual value, especially under conditions such as sudden changes in road surface adhesion, which may result in significant estimation deviations in a short period. When the drive torque allocation is directly based on the erroneous estimation results, it can easily lead to a mismatch between the allocation result and the actual available adhesion capacity of the wheels. This can cause some wheels to experience excessive slippage due to driving torque exceeding their actual adhesion capacity, or some wheels to fail to fully utilize their adhesion capacity, thus affecting vehicle stability and longitudinal acceleration performance. Therefore, how to achieve adaptive and reasonable allocation of four-wheel drive torque under the condition of errors in road surface adhesion coefficient estimation, and further improve the robustness and environmental adaptability of the strategy, has become an important research topic in the field of longitudinal motion control of distributed electric drive vehicles.

[0004] This invention addresses the problem of mismatch between driving torque allocation decisions and actual road surface adhesion conditions caused by estimation errors in road surface adhesion coefficient observations. It proposes an adaptive driving torque allocation method for distributed electric vehicles based on a deep deterministic policy gradient. During policy training, random perturbations and first-order lag dynamics are introduced into the road surface adhesion coefficient observation information of the four wheels, ensuring that the training environment simultaneously includes adhesion coefficient estimation bias and estimation response lag. This enables the agent to develop adaptability to adhesion estimation errors during the policy learning phase. Summary of the Invention

[0005] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0006] To address the aforementioned technical problems, according to one aspect of the present invention, the present invention provides the following technical solution:

[0007] An adaptive driving torque distribution method for a distributed electric drive vehicle includes the following steps:

[0008] S1: Establish a distributed electric drive vehicle environment model, including a seven-degree-of-freedom vehicle dynamics model, a battery model based on an equivalent circuit model, and a motor model as a torque output unit;

[0009] S2: Complete the design of a four-wheel drive torque adaptive allocation strategy based on deep deterministic policy gradient, including the definition of the state and action space of the DDPG agent and the design of the reward function; the reward function evaluates the current driving torque allocation effect by introducing the actual state response of the vehicle's actual acceleration and four-wheel slip ratio, so that the agent can correct the driving torque allocation decision according to the actual motion state of the vehicle.

[0010] S3: Complete the design of the neural network structure for the DDPG algorithm, which consists of an Actor network and a Critic network; the Actor network is used to directly output specific control actions based on the current state. The Critic network is used to evaluate the expected cumulative reward, i.e., the Q value, for taking a certain action in a specific state.

[0011] S4: Design a longitudinal driving condition training set, which includes vehicle speed information and four-wheel road surface adhesion coefficient information;

[0012] S5: Based on the established vehicle simulation environment and the designed longitudinal driving condition training set, conduct training on the four-wheel drive torque adaptive distribution strategy based on DDPG.

[0013] S6: Conduct simulation tests to verify the robustness of the method to the estimation error of the road adhesion coefficient under complex road adhesion conditions.

[0014] As a preferred embodiment of the adaptive driving torque distribution method for distributed electric drive vehicles described in this invention, in step S1, the seven-degree-of-freedom vehicle dynamics model considers the longitudinal, lateral, yaw, and rotational motions of the entire vehicle, wherein the differential equation for the longitudinal motion of the vehicle is:

[0015]

[0016] The differential equation for the lateral motion of the vehicle is:

[0017]

[0018] The differential equation for the yaw motion of the vehicle is:

[0019]

[0020] In the formula, For the total mass of the vehicle. For longitudinal vehicle speed, Lateral speed, The yaw rate is angular velocity. Let Z be the vehicle's moment of inertia about the Z-axis, and the generalized external forces acting along the vehicle's X and Y axes are respectively... and The generalized external torque about the Z-axis is , The air drag coefficient, The rolling resistance coefficient, The road slope angle;

[0021] Assuming the moment of inertia of each wheel and effective rolling radius If all are the same, then the rotational dynamics equations of the four wheels are:

[0022]

[0023] In the formula, the subscript {Front Left, Front Right, Rear Left, Rear Right} represent different drive motors. To utilize the coefficient of adhesion for the wheels, The angular velocity of the wheel. For the driving torque of the wheels, For the braking torque of the wheels, For the longitudinal force of each tire, This refers to the vertical load on the wheel.

[0024] As a preferred embodiment of the adaptive driving torque distribution method for distributed electric drive vehicles described in this invention, the battery model adopts the Rint equivalent circuit modeling principle, considering the battery as a series circuit composed of a voltage source and a resistor; based on the battery's charging and discharging power input... Open circuit voltage and internal resistance Calculate the operating current of the battery pack based on Kirchhoff's laws. Size, that is:

[0025]

[0026] The motor model is a torque output unit. First, the steady-state torque of the motor drive is calculated, and then a first-order hysteresis element is added to simulate the dynamic response. The steady-state torque of the motor drive... The calculation is as follows:

[0027]

[0028]

[0029] In the formula, This represents the battery's maximum discharge power. The required torque for the motor; This refers to the power required for battery discharge. This refers to the rotor speed of the motor. Constraints on the external characteristic curve of the motor; To improve the working efficiency of the motor.

[0030] As a preferred embodiment of the adaptive driving torque distribution method for distributed electric drive vehicles described in this invention, the specific method of step S2 involves dividing the state space into vehicle actual state information and road environment information; wherein, the vehicle actual state information includes vehicle speed. Four-wheel speed Four-wheel slip ratio and vehicle load This is used to describe the vehicle's current actual dynamic response; to avoid poor longitudinal acceleration performance, the vehicle's actual state information also incorporates the driver's required acceleration. and acceleration tracking error = As a control objective; road environment information includes observation information on the four-wheel road surface adhesion coefficient. and maximum tire adhesion rates of the front and rear axles This is used to characterize the road conditions the vehicle is on and the available adhesion of the tires; the final state space is defined as:

[0031]

[0032] In the torque distribution control architecture based on reinforcement learning for distributed drive vehicles, the action space is defined as the driving torques of each wheel determined by the agent, i.e., the multidimensional continuous vector output by the Actor network. :

[0033]

[0034] in, , , , These are the torque distribution coefficients for the front left, front right, rear left, and rear right wheels, respectively.

[0035] vector The values ​​for each dimension are dimensionless values ​​after Tanh compression to [-1, 1]. The actual wheel driving torque for the interaction between the agent and the vehicle is:

[0036]

[0037] In the formula, This represents the maximum permissible driving torque value for the wheel.

[0038] As a preferred embodiment of the adaptive driving torque allocation method for distributed electric drive vehicles described in this invention, the designed reward function... Vehicle stability penalty Acceleration tracking deviation penalty item Slip ratio constraint penalty item Road surface adhesion constraint penalty item Physical constraint penalty item for drive motor and energy loss penalty The composition is as follows:

[0039]

[0040] In the formula, This is a weighting coefficient used to balance the vehicle's longitudinal acceleration performance, stability, and energy consumption;

[0041] Using the minimum sum of squares of tire load rates as the optimization objective function, a larger safety margin is obtained, while lateral forces and vehicle stability penalties are ignored. Simplified to:

[0042]

[0043] in, For the longitudinal force of each tire, For the vertical load of the wheel, The adhesion coefficient is used for the wheels; the actual vehicle acceleration and slip ratio are introduced, and the acceleration tracking deviation penalty term and slip ratio constraint penalty term are as follows:

[0044]

[0045]

[0046] In the formula, Accelerating driver demand; This refers to the vehicle's actual acceleration. For the first Actual slip ratio of each wheel This represents the maximum slip ratio.

[0047] To prevent the driving torque of each wheel from exceeding the road adhesion limit, the road adhesion constraint penalty term is designed as follows:

[0048]

[0049] Each drive motor actuator has physical limitations; the allocated torque cannot exceed its capacity. The reward function introduces a penalty term for the physical constraints of the drive motor.

[0050]

[0051] In the formula, This represents the maximum torque output of each drive motor; energy loss penalty is considered to reduce energy consumption.

[0052]

[0053] In the formula, This represents the actual output power of each drive motor. ; The power consumption of each drive motor. , For the first The efficiency of the drive motor on each wheel.

[0054] As a preferred embodiment of the adaptive driving torque distribution method for distributed electric drive vehicles described in this invention, in step S4, during the construction of the uniform road surface adhesion coefficient sample, considering the characteristics of overall consistency and local fluctuations in actual road adhesion conditions, the adhesion coefficient of each wheel is generated by combining the basic adhesion coefficient with local disturbances, expressed as follows:

[0055]

[0056] in, Based on the basic road surface adhesion coefficient, ; For the first Local disturbance amount corresponding to each wheel For non-uniform road surfaces, the four-wheel adhesion coefficient vector is defined as follows:

[0057]

[0058] in , , , These represent the front left, front right, rear left, and rear right wheels, respectively; firstly, the adhesion coefficient of the high-adhesion area is... and adhesion coefficient in low-adhesion areas To generate randomly, that is:

[0059]

[0060]

[0061] And the following constraints must be met to ensure a significant difference in adhesion between the high and low adhesion regions:

[0062]

[0063] For split-plane surfaces, it is necessary to simulate extreme working conditions where the adhesion conditions on the left and right sides differ significantly. The corresponding four-wheel adhesion coefficient is: or ;

[0064] For the contact road surface, it is necessary to simulate a sudden change in the adhesion conditions between the front and rear axles, and the corresponding four-wheel adhesion coefficient is: or ;

[0065] Further, a local low-friction road surface scenario is constructed to simulate the complex working condition of a single vehicle wheel entering the low-friction region. One wheel is randomly selected to enter the low-friction region, while the remaining wheels remain in a high-friction state. The corresponding four-wheel adhesion coefficient is expressed as follows: Or other wheel arrangement methods.

[0066] As a preferred embodiment of the adaptive driving torque distribution method for distributed electric drive vehicles described in this invention, in step S5, for each training round, a set of longitudinal driving condition data of the vehicle is randomly selected as the environmental input. The longitudinal driving condition training data includes the target vehicle speed curve and the road surface adhesion coefficient information corresponding to the four wheels of the vehicle.

[0067] As a preferred embodiment of the adaptive driving torque allocation method for distributed electric drive vehicles described in this invention, random perturbations are introduced into the road adhesion coefficient observation information received by the agent during the training process. To simulate the estimation error of road surface adhesion coefficient during actual vehicle movement; at the same time, a first-order lag element is set for the road surface adhesion coefficient observation information after introducing random disturbance, so as to obtain road surface adhesion coefficient estimation information with dynamic change characteristics, and use it as the environmental observation input of reinforcement learning agent.

[0068] A driver model is constructed using a PI controller. The driver's accelerator or brake pedal opening is calculated based on the deviation between the target vehicle speed and the actual vehicle speed. The vehicle speed tracking error between the target speed and the actual speed is defined as:

[0069]

[0070] in, Target speed; The actual vehicle speed; the output of the PI controller is expressed as:

[0071]

[0072] in, For proportional gain; This is the integral gain; The driver's accelerator or brake pedal opening, for Vehicle speed tracking error at any given time.

[0073] As a preferred embodiment of the adaptive driving torque distribution method for distributed electric drive vehicles described in this invention, the accelerator pedal opening, brake pedal opening, and the current vehicle operating state are further input into a pre-established Pedal map. The Pedal map establishes a correspondence between the pedal opening and the target longitudinal acceleration of the vehicle, thereby converting the driver's longitudinal operation intention into the target acceleration of the vehicle, realizing the analysis of the driver's longitudinal driving intention, and driving the vehicle dynamics model to complete the longitudinal motion simulation.

[0074] Subsequently, a state observation sequence was constructed by combining the vehicle's driving status, driver's required acceleration, and the road surface adhesion coefficient after error processing. The Actor network is based on the current state. Output continuous action instructions This refers to the torque distribution command of the four-wheel drive motor. The output action requires the addition of OU noise, as shown in the following formula:

[0075]

[0076] In the formula, The noise is random noise from the Ornstein-Uhlenbeck noise model; Represents an Actor network;

[0077] After this action is applied to the vehicle model, the vehicle's state response is obtained, and the current reward is calculated according to the designed reward function. At the same time, the state at the next moment is obtained. Subsequently, the state transition sequence is... Samples stored in the experience buffer are randomly selected to update the Critic network's value function and the Actor network's policy function, and training stability is improved through the target network's soft update mechanism.

[0078]

[0079]

[0080] In the formula, For target commentator network parameters; For the critic's network parameters; For target actor network parameters; For actor network parameters, This is a smoothing factor.

[0081] Compared with existing technologies, the advantages of this invention are as follows: By introducing vehicle state information related to real road surface adhesion, such as actual longitudinal acceleration and wheel slip ratio, into the agent's state space, this invention enables the agent to comprehensively utilize vehicle motion state and road surface adhesion conditions to make driving torque allocation decisions, reducing dependence on road surface adhesion coefficient estimation information. Simultaneously, by introducing random perturbations and first-order lag elements into the road surface adhesion coefficient observation during training, a training environment with adhesion coefficient estimation error is constructed, improving policy robustness. Under a 15% road surface adhesion coefficient estimation error, the strategy proposed in this invention can still reasonably allocate driving torque according to vehicle motion state and road surface adhesion conditions, effectively suppressing excessive slippage of low-adhesion wheels, maintaining good vehicle longitudinal acceleration performance, and exhibiting good robustness to road surface adhesion coefficient estimation errors. Attached Figure Description

[0082] To more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and detailed embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0083] Figure 1 This is a flowchart of an adaptive driving torque distribution method for a distributed electric drive vehicle according to the present invention;

[0084] Figure 2This is a schematic diagram of the neural network structure of the DDPG algorithm for an adaptive driving torque allocation method for distributed electric drive vehicles according to the present invention.

[0085] Figure 3 This is a schematic diagram of the DDPG agent training process of the adaptive driving torque distribution method for distributed electric drive vehicles according to the present invention.

[0086] Figure 4 This is a comparison of the torque distribution simulation results when the road adhesion coefficient estimation of the adaptive driving torque distribution method for distributed electric drive vehicles of the present invention has a 15% error (dock road acceleration test condition).

[0087] Figure 5 This is a comparison of the simulation results of longitudinal acceleration of a vehicle when there is a 15% error in the estimation of the road adhesion coefficient in the adaptive driving torque distribution method for a distributed electric drive vehicle of the present invention (under the condition of road acceleration test). Detailed Implementation

[0088] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0089] Secondly, the present invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of the present invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not according to the usual scale. Furthermore, the schematic diagrams are merely examples and should not limit the scope of protection of the present invention. In addition, actual fabrication should include three-dimensional spatial dimensions of length, width, and depth.

[0090] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0091] An adaptive driving torque distribution method for distributed electric drive vehicles, such as Figure 1 As shown, it includes the following steps:

[0092] Step 1: Establish a distributed electric drive vehicle environment model, including a seven-degree-of-freedom vehicle dynamics model, a battery model based on an equivalent circuit model, and a motor model as a torque output unit.

[0093] The seven-degree-of-freedom vehicle dynamics model considers the longitudinal, lateral, yaw, and rotational motions of the entire vehicle, including the differential equation of the longitudinal motion of the vehicle as follows:

[0094] (1)

[0095] The differential equation for the lateral motion of the vehicle is:

[0096] (2)

[0097] The differential equation for the yaw motion of the vehicle is:

[0098] (3)

[0099] In the formula, The total mass of the vehicle (including sprung mass and unsprung mass). For longitudinal vehicle speed, Lateral speed, The yaw rate is angular velocity. Let Z be the vehicle's moment of inertia about the Z-axis, and the generalized external forces acting along the vehicle's X and Y axes are respectively... and The generalized external torque about the Z-axis is , The air drag coefficient, The rolling resistance coefficient, This refers to the road slope angle.

[0100] Assuming the moment of inertia of each wheel and effective rolling radius If all are the same, then the rotational dynamics equations of the four wheels are:

[0101] (4)

[0102] In the formula, the subscript {Front Left, Front Right, Rear Left, Rear Right} represent different drive motors. To utilize the coefficient of adhesion for the wheels, The angular velocity of the wheel. For the driving torque of the wheels, For the braking torque of the wheels, For the longitudinal force of each tire, This refers to the vertical load on the wheel.

[0103] The battery model adopts the Rint equivalent circuit modeling principle, considering the battery as a series circuit consisting of a voltage source and a resistor. The charging and discharging power input of the battery is then used. Open circuit voltage and internal resistance The operating current of the battery pack can be calculated based on Kirchhoff's laws. Size, that is:

[0104] (5)

[0105] The motor model is a torque output unit. First, the steady-state driving torque of the motor is calculated, and then a first-order hysteresis element is added to simulate the dynamic response. (Motor driving steady-state torque) The calculation is as follows:

[0106] (6)

[0107] (7)

[0108] In the formula, This represents the battery's maximum discharge power. The required torque for the motor; This refers to the power required for battery discharge. This refers to the rotor speed of the motor. Constraints on the external characteristic curve of the motor; To improve the working efficiency of the motor.

[0109] Step 2: Complete the design of an adaptive four-wheel drive torque allocation strategy based on Deep Deterministic Policy Gradient (DDPG), including the definition of the state and action space of the DDPG agent and the design of the reward function. The reward function evaluates the current driving torque allocation effect by incorporating actual state responses such as vehicle acceleration and four-wheel slip ratio, enabling the agent to correct the driving torque allocation decision based on the actual vehicle motion state, thereby reducing the strategy's dependence on the estimated road adhesion coefficient.

[0110] This invention divides the state space into vehicle actual state information and road environment information. The vehicle actual state information includes vehicle speed. Four-wheel speed Four-wheel slip ratio and vehicle load This is used to describe the vehicle's current actual dynamic response; to avoid poor longitudinal acceleration performance, the vehicle's actual state information also incorporates the driver's required acceleration. and acceleration tracking error = As a control objective, road environmental information includes observation information on the four-wheel road surface adhesion coefficient. and maximum tire adhesion rates of the front and rear axles This is used to characterize the road conditions the vehicle is in and the available adhesion of the tires. The final state space is defined as follows:

[0111] (8)

[0112] In the torque distribution control architecture based on reinforcement learning for distributed drive vehicles, the action space is defined as the driving torques of each wheel determined by the agent, i.e., the multidimensional continuous vector output by the Actor network. :

[0113] (9)

[0114] in, , , , These are the torque distribution coefficients for the front left, front right, rear left, and rear right wheels, respectively.

[0115] vector The values ​​for each dimension are dimensionless values ​​after Tanh compression to [-1, 1]. The actual wheel driving torque for the interaction between the agent and the vehicle is:

[0116] (10)

[0117] In the formula, This represents the maximum permissible driving torque value for the wheel.

[0118] The reward function designed in this invention Vehicle stability penalty Acceleration tracking deviation penalty item Slip ratio constraint penalty item Road surface adhesion constraint penalty item Physical constraint penalty item for drive motor and energy loss penalty The composition is as follows:

[0119] (11)

[0120] In the formula, This is a weighting coefficient used to balance the vehicle's longitudinal acceleration performance, stability, and energy consumption.

[0121] Using the minimum sum of squares of tire load rates as the optimization objective function to obtain a larger safety margin, this invention neglects lateral forces and vehicle stability penalty terms. Simplified to:

[0122] (12)

[0123] To improve strategy robustness and ensure vehicle acceleration performance on surfaces with varying adhesion, this invention introduces real vehicle state information such as actual acceleration and slip ratio. This information is strongly correlated with actual road surface adhesion, enabling the agent to adjust driving torque allocation decisions based on the vehicle's actual state response, thereby reducing the strategy's dependence on road surface adhesion coefficient estimation information. The acceleration tracking deviation penalty term and the slip ratio constraint penalty term are as follows:

[0124] (13)

[0125] (14)

[0126] In the formula, Accelerating driver demand; This refers to the vehicle's actual acceleration. For the first Actual slip ratio of each wheel This represents the maximum slip ratio.

[0127] To prevent the driving torque of each wheel from exceeding the road adhesion limit, the road adhesion constraint penalty term is designed as follows:

[0128] (15)

[0129] Each drive motor actuator has physical limitations; the allocated torque cannot exceed its capacity. The reward function introduces a penalty term for the physical constraints of the drive motor.

[0130] (16)

[0131] In the formula, This represents the maximum torque output of each drive motor.

[0132] On roads with high uniformity and safety, the focus should be on the overall energy efficiency of the vehicle, and energy loss penalties should be considered to reduce energy consumption.

[0133] (17)

[0134] In the formula, This represents the actual output power of each drive motor. ; The power consumption of each drive motor. , For the first The efficiency of the drive motor on each wheel.

[0135] Step 3: Complete the design of the DDPG algorithm neural network structure. The DDPG algorithm constructed in this invention consists of an Actor network and a Critic network, as follows: Figure 2 As shown, the Actor network is used to directly output specific control actions based on the current state. The Critic network is used to evaluate the expected cumulative reward, or Q-value, of taking a certain action in a specific state. Both networks activate their hidden layers using the ReLU function to handle non-linear data, and the number of hidden layer nodes is set to 128 for both.

[0136] Step 4: To meet the training requirements of reinforcement learning agents facing different road surface adhesion conditions, this invention designs a longitudinal driving condition training set, which includes vehicle speed information and four-wheel road surface adhesion coefficient information.

[0137] The vehicle speed training data can be obtained based on standard driving cycle conditions such as the World Light Vehicle Test Cycle (WLTC) and the China Light Vehicle Test Cycle (CLTC-P), or it can be collected from driving simulators under typical driving scenarios such as highways, suburbs, and urban areas. The duration of each vehicle speed training sequence is set to 1200 seconds and includes three typical driving conditions: urban, suburban, and highway, to ensure that the training data covers different vehicle operating states.

[0138] Based on the vehicle speed training data, road surface adhesion coefficient information corresponding to the four wheels of the vehicle is randomly generated, thereby constructing different road surface adhesion environments and expanding the training samples. The road surface adhesion environments include uniform road surface conditions as well as complex road surface conditions such as split roads, joint roads, and locally low-adhesion roads, enabling the agent to be fully trained under different adhesion conditions, thereby improving the adaptability, robustness, and generalization ability of the learned control strategy to complex road surface environments.

[0139] In constructing the uniform road surface adhesion coefficient sample, considering the characteristics of overall consistency and local fluctuations in actual road adhesion conditions, the adhesion coefficient of each wheel is generated by combining the basic adhesion coefficient with local disturbances. The expression form is as follows:

[0140] (18)

[0141] in, Based on the basic road surface adhesion coefficient, ; For the first Local disturbance amount corresponding to each wheel .

[0142] For non-uniform road surfaces, the four-wheel adhesion coefficient vector is defined as:

[0143] (19)

[0144] in , , , These represent the front left, front right, rear left, and rear right wheels, respectively.

[0145] First, the adhesion coefficient of the high-adhesion region was analyzed. and adhesion coefficient in low-adhesion areas To generate randomly, that is:

[0146] (20)

[0147] (twenty one)

[0148] And the following constraints must be met to ensure a significant difference in adhesion between the high and low adhesion regions:

[0149] (twenty two)

[0150] For split-plane surfaces, it is necessary to simulate extreme working conditions where the adhesion conditions on the left and right sides differ significantly. The corresponding four-wheel adhesion coefficient is: or .

[0151] For the contact road surface, it is necessary to simulate a sudden change in the adhesion conditions between the front and rear axles, and the corresponding four-wheel adhesion coefficient is: or .

[0152] This invention further constructs a local low-friction road surface scenario to simulate the complex working condition of a single vehicle wheel entering the low-friction region. A single wheel is randomly selected to enter the low-friction region, while the remaining wheels remain in a high-friction state. The corresponding four-wheel adhesion coefficient is expressed as follows: Or other wheel arrangement methods.

[0153] For a single training round, the four-wheel road adhesion coefficient remains unchanged and is only updated randomly between different training rounds to ensure the continuity of vehicle dynamics characteristics throughout the entire speed sequence.

[0154] Step 5: Based on the established vehicle simulation environment and the designed longitudinal driving condition training set, train the four-wheel drive torque adaptive distribution strategy based on DDPG. Figure 3 The training process of the DDPG agent is demonstrated.

[0155] For each training round, a set of longitudinal driving condition data of the vehicle is randomly selected as the environmental input. The longitudinal driving condition training data includes the target vehicle speed curve and the road surface adhesion coefficient information corresponding to the four wheels of the vehicle.

[0156] To improve the agent's ability to adapt to road surface adhesion coefficient estimation errors, random perturbations are introduced into the road surface adhesion coefficient observation information received by the agent during training. This method simulates the estimation error of the road surface adhesion coefficient during actual vehicle movement. Simultaneously, a first-order lag is introduced into the road surface adhesion coefficient observation information after random perturbation, thereby obtaining road surface adhesion coefficient estimation information with dynamic changing characteristics, which is then used as the environmental observation input for the reinforcement learning agent.

[0157] A driver model is constructed using a PI controller, and the driver's accelerator or brake pedal opening is calculated based on the deviation between the target vehicle speed and the actual vehicle speed. The vehicle speed tracking error between the target speed and the actual vehicle speed is defined as:

[0158] (twenty three)

[0159] in, Target speed; This refers to the actual vehicle speed.

[0160] The output of the PI controller can be expressed as:

[0161] (twenty four)

[0162] in, For proportional gain; This is the integral gain; The driver's accelerator or brake pedal opening, for Vehicle speed tracking error at any given time.

[0163] The accelerator pedal opening, brake pedal opening, and the vehicle's current operating state are further input into a pre-established Pedal map. The Pedal map establishes a correspondence between pedal opening and the vehicle's target longitudinal acceleration, thereby converting the driver's longitudinal operating intention into the vehicle's target acceleration, realizing the analysis of the driver's longitudinal driving intention, and driving the vehicle dynamics model to complete the longitudinal motion simulation.

[0164] The Pedal map can be designed and calibrated independently based on the specific vehicle's powertrain characteristics, overall vehicle dynamics, and driver operating characteristics, or it can be a Pedal map that has been established and validated in existing research.

[0165] Subsequently, a state observation sequence was constructed by combining information such as vehicle driving status, driver demand acceleration, and road adhesion coefficient after error processing. The Actor network is based on the current state. Output continuous action instructions This refers to the torque distribution command of the four-wheel drive motor. To improve the algorithm's exploration ability during training, OU noise needs to be added to the output action, as shown in the following formula:

[0166] (25)

[0167] In the formula, The noise is random noise from the Ornstein-Uhlenbeck noise model; Represents the Actor network.

[0168] After this action is applied to the vehicle model, the vehicle's state response is obtained, and the current reward is calculated according to the designed reward function. At the same time, the state at the next moment is obtained. Then, the state transition sequence... Samples stored in the experience buffer are randomly selected to update the Critic network's value function and the Actor network's policy function, and training stability is improved through the target network's soft update mechanism.

[0169] (26)

[0170] (27)

[0171] In the formula, For target commentator network parameters; For the critic's network parameters; For target actor network parameters; For actor network parameters, This is a smoothing factor.

[0172] As the training rounds iterate, the agent gradually forms a driving torque allocation strategy that can adapt to different road surface adhesion conditions and adhesion coefficient estimation errors through continuous interaction and strategy optimization. The network and other hyperparameter configurations during the training process are shown in Table 1.

[0173] Table 1: Hyperparameter settings for DDPG in distributed drive vehicle torque distribution algorithm

[0174]

[0175] Step 6: Perform simulation tests on the adaptive driving torque allocation method for distributed electric drive vehicles based on deep deterministic strategy gradient described in this embodiment to verify the robustness of the method to the road adhesion coefficient estimation error under complex road adhesion conditions.

[0176] This embodiment sets up an acceleration scenario on a docking road surface. The initial vehicle speed is 10 km / h. The front wheels of the vehicle move from an ice surface with a coefficient of friction of 0.2 onto a dry road surface with a coefficient of friction of 0.85. Accelerator pedal input is then applied, and the four-wheel drive torque distribution, actual longitudinal acceleration, and required acceleration tracking are recorded. A drive torque optimization distribution strategy based on Sequential Quadratic Programming (SQP) is selected as a comparison method, and simulation tests are conducted under the same vehicle parameters and driving conditions.

[0177] To verify the robustness of the proposed method under conditions where there are errors in the estimation of the road surface adhesion coefficient, a 15% estimation error was introduced between the actual road surface adhesion coefficient and the road surface adhesion coefficient obtained by the control strategy. Simulation tests were then conducted using torque distribution strategies based on SQP and DDPG, respectively. Figure 4 and Figure 5 The simulation results show that when the road surface adhesion coefficient estimation has a 15% error, the torque distribution decision given by the SQP-based algorithm does not match the actual road surface adhesion conditions. This causes the driving torque of some wheels to exceed their actual road surface adhesion capacity, resulting in wheel slippage and triggering the vehicle's anti-slip drive control (ASR) system to intervene. The drive motor output torque is frequently adjusted, thus affecting the vehicle's longitudinal acceleration performance. In contrast, the DDPG-based torque distribution strategy can still reasonably distribute torque among the drive motors according to the vehicle's operating state and road surface adhesion conditions. When the front wheels of the vehicle move from a low-adhesion road surface to a high-adhesion road surface, more driving force is distributed to the front wheels on the high-adhesion road surface, thus fully utilizing the adhesion capacity of the high-adhesion road surface. Throughout the acceleration process, the vehicle does not exhibit significant wheel slippage, avoiding the resulting decrease in acceleration performance while maintaining good demand acceleration tracking performance.

[0178] In summary, the simulation results of this embodiment demonstrate that the adaptive driving torque allocation method for distributed electric drive vehicles based on deep deterministic strategy gradients can still achieve reasonable driving torque allocation even with a 15% error in the road adhesion coefficient estimation, effectively reducing wheel slippage and maintaining good vehicle longitudinal acceleration performance. The strategy proposed in this invention has good adaptability to complex road adhesion conditions and good robustness to road adhesion coefficient estimation errors within a certain range.

[0179] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for adaptive driving torque distribution in a distributed electric drive vehicle, characterized in that, Includes the following steps: S1: Establish a distributed electric drive vehicle environment model, including a seven-degree-of-freedom vehicle dynamics model, a battery model based on an equivalent circuit model, and a motor model as a torque output unit; S2: Complete the design of a four-wheel drive torque adaptive allocation strategy based on deep deterministic policy gradient, including the definition of the state and action space of the DDPG agent and the design of the reward function; the reward function evaluates the current driving torque allocation effect by introducing the actual state response of the vehicle's actual acceleration and four-wheel slip ratio, so that the agent can correct the driving torque allocation decision according to the actual motion state of the vehicle. S3: Complete the design of the neural network structure for the DDPG algorithm, which consists of an Actor network and a Critic network; the Actor network is used to directly output specific control actions based on the current state. The Critic network is used to evaluate the expected cumulative reward, i.e., the Q value, for taking a certain action in a specific state. S4: Design a longitudinal driving condition training set, which includes vehicle speed information and four-wheel road surface adhesion coefficient information; S5: Based on the established vehicle simulation environment and the designed longitudinal driving condition training set, conduct training on the four-wheel drive torque adaptive distribution strategy based on DDPG. S6: Conduct simulation tests to verify the robustness of the method to the estimation error of the road adhesion coefficient under complex road adhesion conditions.

2. The adaptive driving torque distribution method for a distributed electric drive vehicle according to claim 1, characterized in that, In S1, the seven-degree-of-freedom vehicle dynamics model considers the longitudinal, lateral, yaw, and rotational motions of the entire vehicle, including the differential equation of the longitudinal motion of the vehicle: The differential equation for the lateral motion of the vehicle is: The differential equation for the yaw motion of the vehicle is: In the formula, For the total mass of the vehicle. For longitudinal vehicle speed, Lateral speed, The yaw rate is angular velocity. Let Z be the vehicle's moment of inertia about the Z-axis, and the generalized external forces acting along the vehicle's X and Y axes are respectively... and The generalized external torque about the Z-axis is , The air drag coefficient, The rolling resistance coefficient, The road slope angle; Assuming the moment of inertia of each wheel and effective rolling radius If all are the same, then the rotational dynamics equations of the four wheels are: In the formula, the subscript {Front Left, Front Right, Rear Left, Rear Right} represent different drive motors. To utilize the coefficient of adhesion for the wheels, The angular velocity of the wheel. For the driving torque of the wheels, For the braking torque of the wheels, For the longitudinal force of each tire, This refers to the vertical load on the wheel.

3. The adaptive driving torque distribution method for a distributed electric drive vehicle according to claim 2, characterized in that, The battery model adopts the Rint equivalent circuit modeling principle, considering the battery as a series circuit composed of a voltage source and a resistor; based on the battery's charging and discharging power input... Open circuit voltage and internal resistance Calculate the operating current of the battery pack based on Kirchhoff's laws. Size, that is: The motor model is a torque output unit. First, the steady-state torque of the motor drive is calculated, and then a first-order hysteresis element is added to simulate the dynamic response. The steady-state torque of the motor drive... The calculation is as follows: In the formula, This represents the battery's maximum discharge power. The required torque for the motor; This refers to the power required for battery discharge. This refers to the rotor speed of the motor. Constraints on the external characteristic curve of the motor; To improve the working efficiency of the motor.

4. The adaptive driving torque distribution method for a distributed electric drive vehicle according to claim 1, characterized in that, The specific method of S2 is to divide the state space into vehicle actual state information and road environment information; wherein, the vehicle actual state information includes vehicle speed. Four-wheel speed Four-wheel slip ratio and vehicle load This is used to describe the vehicle's current actual dynamic response; to avoid poor longitudinal acceleration performance, the vehicle's actual state information also incorporates the driver's required acceleration. and acceleration tracking error = As a control objective; road environment information includes observation information on the four-wheel road surface adhesion coefficient. and maximum tire adhesion rates of the front and rear axles This is used to characterize the road conditions the vehicle is on and the available adhesion of the tires; the final state space is defined as: In the torque distribution control architecture based on reinforcement learning for distributed drive vehicles, the action space is defined as the driving torques of each wheel determined by the agent, i.e., the multidimensional continuous vector output by the Actor network. : in, , , , These are the torque distribution coefficients for the front left, front right, rear left, and rear right wheels, respectively. vector The values ​​for each dimension are dimensionless values ​​compressed to [-1, 1] using Tanh compression. The actual wheel driving torque for the interaction between the agent and the vehicle is: In the formula, This represents the maximum permissible driving torque value for the wheel.

5. The adaptive driving torque distribution method for a distributed electric drive vehicle according to claim 4, characterized in that, Designed reward function Vehicle stability penalty Acceleration tracking deviation penalty item Slip ratio constraint penalty item Road surface adhesion constraint penalty item Physical constraint penalty item for drive motor and energy loss penalty The composition is as follows: In the formula, This is a weighting coefficient used to balance the vehicle's longitudinal acceleration performance, stability, and energy consumption; Using the minimum sum of squares of tire load rates as the optimization objective function, a larger safety margin is obtained, while lateral forces and vehicle stability penalties are ignored. Simplified to: in, For the longitudinal force of each tire, For the vertical load of the wheel, The adhesion coefficient is used for the wheels; the actual vehicle acceleration and slip ratio are introduced, and the acceleration tracking deviation penalty term and slip ratio constraint penalty term are as follows: In the formula, Accelerating driver demand; This refers to the vehicle's actual acceleration. For the first Actual slip ratio of each wheel This represents the maximum slip ratio. To prevent the driving torque of each wheel from exceeding the road adhesion limit, the road adhesion constraint penalty term is designed as follows: Each drive motor actuator has physical limitations; the allocated torque cannot exceed its capacity. The reward function introduces a penalty term for the physical constraints of the drive motor. In the formula, This represents the maximum torque output of each drive motor; energy loss penalties are considered to reduce energy consumption. In the formula, This represents the actual output power of each drive motor. ; The power consumption of each drive motor. , For the first The efficiency of the drive motor on each wheel.

6. The adaptive driving torque distribution method for a distributed electric drive vehicle according to claim 1, characterized in that, In S4, during the construction of the uniform road surface adhesion coefficient sample, considering the characteristics of overall consistency and local fluctuation in actual road adhesion conditions, the adhesion coefficient of each wheel is generated by combining the basic adhesion coefficient with local disturbances. The expression form is as follows: in, Based on the basic road surface adhesion coefficient, ; For the first Local disturbance amount corresponding to each wheel For non-uniform road surfaces, the four-wheel adhesion coefficient vector is defined as follows: in , , , These represent the front left, front right, rear left, and rear right wheels, respectively; firstly, the adhesion coefficient of the high-adhesion area is... and adhesion coefficient in low-adhesion areas To generate randomly, that is: And the following constraints must be met to ensure a significant difference in adhesion between the high and low adhesion regions: For split-plane surfaces, it is necessary to simulate extreme working conditions where the adhesion conditions on the left and right sides differ significantly. The corresponding four-wheel adhesion coefficient is: or ; For the contact road surface, it is necessary to simulate a sudden change in the adhesion conditions between the front and rear axles, and the corresponding four-wheel adhesion coefficient is: or ; To further simulate the complex condition of a single vehicle wheel entering a low-friction area, a local low-friction road surface scenario is constructed. One wheel is randomly selected to enter the low-friction area, while the remaining wheels remain in a high-friction state. The corresponding four-wheel adhesion coefficient is expressed as follows: Or other wheel arrangement methods.

7. The adaptive driving torque distribution method for a distributed electric drive vehicle according to claim 1, characterized in that, In S5, for each training round, a set of longitudinal driving condition data of the vehicle is randomly selected as the environmental input. The longitudinal driving condition training data includes the target vehicle speed curve and the road surface adhesion coefficient information corresponding to the four wheels of the vehicle.

8. The adaptive driving torque distribution method for a distributed electric drive vehicle according to claim 7, characterized in that, During training, random perturbations are introduced into the road adhesion coefficient observation information received by the agent. To simulate the estimation error of road surface adhesion coefficient during actual vehicle movement; at the same time, a first-order lag element is set for the road surface adhesion coefficient observation information after introducing random disturbance, so as to obtain road surface adhesion coefficient estimation information with dynamic change characteristics, and use it as the environmental observation input of reinforcement learning agent. A driver model is constructed using a PI controller. The driver's accelerator or brake pedal opening is calculated based on the deviation between the target vehicle speed and the actual vehicle speed. The vehicle speed tracking error between the target speed and the actual speed is defined as: in, Target speed; The actual vehicle speed; the output of the PI controller is expressed as: in, For proportional gain; This is the integral gain; This determines the opening of the driver's accelerator or brake pedal. for Vehicle speed tracking error at any given time.

9. The adaptive driving torque distribution method for a distributed electric drive vehicle according to claim 8, characterized in that, The accelerator pedal opening, brake pedal opening, and the vehicle's current operating state are further input into a pre-established Pedal map. The Pedal map establishes a correspondence between pedal opening and the vehicle's target longitudinal acceleration, thereby converting the driver's longitudinal operating intention into the vehicle's target acceleration, realizing the analysis of the driver's longitudinal driving intention, and driving the vehicle dynamics model to complete the longitudinal motion simulation. Subsequently, a state observation sequence was constructed by combining the vehicle's driving status, driver's required acceleration, and the road surface adhesion coefficient after error processing. ; The Actor network depends on the current state. Output continuous action instructions This refers to the torque distribution command of the four-wheel drive motor. The output action requires the addition of OU noise, as shown in the following formula: In the formula, The noise is random noise from the Ornstein-Uhlenbeck noise model; Represents an Actor network; After this action is applied to the vehicle model, the vehicle's state response is obtained, and the current reward is calculated according to the designed reward function. At the same time, the state at the next moment is obtained. ; Then, the state transition sequence Samples stored in the experience buffer are randomly selected to update the Critic network's value function and the Actor network's policy function, and training stability is improved through the target network's soft update mechanism. In the formula, For target commentator network parameters; For the critic's network parameters; For target actor network parameters; For actor network parameters; This is a smoothing factor.

Images