Multi-objective coordinated control method for distributed drive electric vehicle

Abstract

The application discloses a multi-target coordinated control method of a distributed drive electric vehicle, and belongs to the technical field of vehicle control.The application solves the technical problem that the prior art still lacks a multi-target coordinated control method which can be used for the distributed drive electric vehicle, and which can consider control real-time performance, steering stability and energy economy of the whole vehicle under the conditions of model uncertainty and random disturbance.The method comprises an upper-layer partition adaptive control step and a lower-layer multi-target coordinated control step;the upper-layer partition adaptive control step divides and distinguishes convergent areas and non-convergent areas according to whether the real-time running state of the vehicle meets Lyapunov asymptotic convergence constraint conditions, and respectively constructs control methods;the lower-layer multi-target coordinated control step takes the vehicle motion parameters calculated by the upper layer as reference quantities, realizes integrated horizontal and longitudinal control by controlling the torques of the four-wheel electric motors, introduces an energy consumption optimization strategy and effectively limits the wheel slip rate, and realizes the coordinated control among the multi-targets.

Description

Technical Field

[0001] This invention belongs to the field of vehicle control technology, specifically relating to a multi-objective coordinated control method for distributed drive electric vehicles. Background Technology

[0002] With the continuous development of electrification and intelligentization technologies in new energy vehicles, distributed drive electric vehicles, due to their independent drive control of each wheel and higher degrees of control freedom, have shown great application potential in vehicle handling stability control and energy optimization control. However, there are significant coupling relationships among the multiple degrees of freedom (such as lateral and longitudinal) in the dynamics of distributed drive electric vehicles, and issues involving multi-actuator coordination and energy efficiency adjustment are also involved, resulting in a complex control system structure, mutual constraints between control objectives, and high controller design difficulty. To ensure rapid vehicle response and stability while minimizing energy consumption, an adaptive control framework is urgently needed to improve the overall performance of distributed drive electric vehicles.

[0003] Distributed drive electric vehicle control technology is divided into two types based on different approaches to handling the coupling relationship between the vehicle's lateral and longitudinal motions: lateral-longitudinal decoupling control and lateral-longitudinal integrated control. Lateral-longitudinal decoupling control optimizes the vehicle's lateral and longitudinal control independently, simplifying the controller's design complexity and implementation process, and is suitable for linear time-invariant systems. Specifically, based on PID algorithms, fuzzy logic, and linear quadratic regulators (LQRs) as lateral-longitudinal decoupling control methods, it establishes a simple linear model and fixed system parameters by independently optimizing the lateral and longitudinal control dimensions, reducing the complexity of the control system and helping to improve the control performance of a single degree of freedom. However, lateral-longitudinal decoupling control ignores the coupling relationship between the vehicle's lateral and longitudinal motions, making it difficult to cope with complex nonlinear dynamics and external disturbances, resulting in poor robustness. In contrast, lateral-longitudinal integrated control considers both the vehicle's lateral and longitudinal motions, as well as the coupling relationship between them. Through an integrated control strategy, it more accurately describes the vehicle's dynamic characteristics under different driving conditions, improving its adaptability to complex operating conditions. In the highly complex control systems of distributed drive electric vehicles, it is necessary not only to balance the dynamic performance of the lateral and longitudinal motions but also to weigh the vehicle's energy consumption performance. This places higher demands on the control system's ability to handle high-dimensional nonlinear constraints, multiple inputs / outputs, and real-time computation and solution, further increasing the difficulty of control system design.

[0004] Model predictive control (MPC), as a model-based optimization control method, possesses the ability to model and coordinate multi-input multi-output systems, and is suitable for handling high-dimensional coupling and nonlinear constraints. In the integrated lateral and longitudinal control system of distributed drive electric vehicles, MPC can simultaneously optimize the dynamic performance of lateral and longitudinal motion, and it is possible to incorporate the vehicle's energy consumption performance into the optimization framework. However, as a model-based optimization algorithm, the accuracy of the model directly affects the accuracy of MPC's prediction of the system's future state and the reliability of control decision optimization. To address the problems of model uncertainty and random disturbances, and to improve MPC's fault tolerance to model uncertainty and random disturbances, it can be divided into Minimax Optimization, Chance-constrained Optimization, and Data-driven Optimization, depending on the optimization strategy used. Among them, Robust Model Predictive Control (RMPC) introduces minimax optimization into the MPC framework, minimizing the maximum loss that the system may incur under extreme conditions, and achieving global robustness to uncertainty. However, RMPC significantly increases solution time in complex distributed drive electric vehicle control systems and sacrifices overall vehicle performance due to its conservative assumptions about the feasible region. Stochastic MPC (SMPC), as a chance-constrained optimization method, relaxes the strict control over system behavior through probabilistic constraints, improving performance while retaining some robustness. However, SMPC heavily relies on accurate modeling of the disturbance probability distribution and struggles to achieve an optimal trade-off between performance and robustness in confidence level selection. Learning MPC (LMPC), combining data-driven modeling and online learning mechanisms, improves adaptability to complex environments through dynamic model correction and disturbance compensation. Despite its greater flexibility and performance potential, it relies on high-quality data, incurs high training costs, and suffers from control instability due to a lack of robustness in extreme or unknown situations.

[0005] The optimization process involving model uncertainty and random disturbances is often accompanied by high computational burden. However, rapid solution and high-frequency control updates are crucial for ensuring the response speed and stability of distributed electric vehicles under complex operating conditions. To improve the computational efficiency and real-time performance of system control, strategies combining offline and online approaches are widely used. Among them, Explicit Model Predictive Control (EMPC) significantly improves real-time control performance by transforming the online optimization problem into an offline pre-computation process and obtaining the optimal control input through a lookup table in real-time control. However, when the system approaches the stability boundary or faces strong nonlinear disturbances, the pre-computation control law provided by EMPC is difficult to adapt to the rapid evolution of the system state, which may exacerbate the risk of instability. Nevertheless, when the system is in a stable operating domain or the state is predictable, control strategies based on online-offline fusion mechanisms remain an effective and feasible approach, significantly improving computational efficiency and real-time performance while ensuring control accuracy. In addition, in the integrated lateral and longitudinal control of distributed electric vehicles, the introduction of energy consumption optimization not only increases the dimensionality of the control objective but also significantly exacerbates the constraint coupling and the complexity of the feasible solution space in the control system. Therefore, achieving effective coordination between the dynamic performance and energy consumption of distributed drive electric vehicles is the key to improving their overall performance under different operating conditions.

[0006] Based on the above analysis and discussion, existing technologies still lack a multi-objective coordinated control method that can balance control real-time performance, handling stability, and overall vehicle energy economy under conditions of model uncertainty and random disturbances for distributed drive electric vehicles. How to construct control mechanisms with differentiated characteristics for different vehicle operating states and dynamically coordinate the weight relationships between various control objectives based on the vehicle's stable state has become a pressing technical problem to be solved in this field. Summary of the Invention

[0007] To address the technical problem that existing technologies still lack a multi-objective coordinated control method for distributed drive electric vehicles that can balance control real-time performance, handling stability, and overall vehicle energy economy under conditions of model uncertainty and random disturbances, this invention provides a multi-objective coordinated control method for distributed drive electric vehicles.

[0008] The specific method is as follows: The method includes an upper-level partitioned adaptive control step and a lower-level multi-objective cooperative control step. The upper-level partition adaptive control step divides and distinguishes the convergent region from the non-convergent region based on whether the real-time operating state of the vehicle satisfies the Lyapunov asymptotic convergence constraint condition. Within the convergence region, a combination of offline and online methods is used: the vehicle control law is established offline, and the control law is searched online based on the actual state of the vehicle. In the non-convergence region, a real-time solution strategy is constructed online based on the vehicle status and the target convergence region; Deviation optimization is performed at the boundary of the convergence region; The lower-level multi-objective cooperative control step adopts a model predictive control strategy based on a seven-degree-of-freedom vehicle model. The vehicle motion parameters calculated by the upper level are used as reference quantities. The lateral and longitudinal integrated control is achieved by controlling the torque of the four-wheel motors. On the basis of ensuring the stability of the lateral and longitudinal integrated control, an energy consumption optimization strategy is introduced and the wheel slip ratio is effectively limited to achieve cooperative control among multiple objectives.

[0009] Furthermore, the convergence region, based on Lyapunov theory, is a defined subset of the state space of a distributed-drive electric vehicle under a specific control law. Within this subset, the system state is guaranteed to asymptotically converge to the target state. ,in, Represents the system state vector. The accuracy parameters of the convergence region are obtained through iterative solutions. The Lyapunov function is expressed in scalar form.

[0010] Furthermore, a gold mining optimization algorithm is adopted to improve accuracy parameters. The accuracy and efficiency of iterative solutions are comprehensively considered from three aspects: the individual's global optimal fitness, the individual's local fitness, and the fitness among individuals, to select the next position for each individual. The optimal accuracy parameters are obtained in the next iteration. .

[0011] Furthermore, within the convergence region, the high-dimensional convergence region is reduced to a two-dimensional sensitive dimension region using an improved Fourier-Motzkin elimination method. The convergence region and control law are established segmented according to vehicle speed intervals. A weighted smoothing switching strategy is adopted in the buffer section of the vehicle speed interval, and the control quantity is output online via a lookup table. Specifically: The improved Fourier-Motzkin elimination method projects the six-dimensional state convergence region onto the two-dimensional sensitive dimensions of lateral error rate and yaw rate error rate. The six-dimensional state convergence region is a feasible region of the state space constructed under Lyapunov theory based on the three-degree-of-freedom dynamic model of a distributed-drive electric vehicle, which can guarantee the asymptotic convergence of the system state to the target state. The improvement in the improved Fourier-Motzkin elimination method is as follows: when a new inequality is generated in each iteration of the Fourier-Motzkin elimination method, the boundary check method is used to check whether the newly generated inequality is redundant, and the redundant inequalities in the solution space are removed. speed Discretize the interval within the feasible region, i.e. , These represent the lower and upper limits of the vehicle speed range, respectively. Corresponding convergence regions and control laws are established for different speed ranges. When the control law changes due to speed range switching, a buffer zone is set at the speed range boundary. In the buffer zone The control law-based weighted transition method is used, as shown in the following formula: ; In the formula, for Internal control laws for The control law, As a weighting factor, For buffer The control law, Indicates the current position is number 1. Speed ​​range.

[0012] Furthermore, within the non-convergence region, a model predictive control state iterative relation is established based on three-degree-of-freedom vehicle dynamics. Random disturbances are introduced into the state iterative relation to construct a predicted state iterative equation with disturbances. State feedback control is designed to converge the cumulative state error caused by random disturbances. In the optimization solution, the initial state variables are used as decision variables, enabling the controller to adjust the state trajectory in real time within the prediction domain to respond to the deviations caused by disturbances. Based on the reference value of the tracking target, an objective function outside the convergence region is established. In the objective function, the final state of the vehicle is kept within the convergence region through the terminal constraints of the control system. The iterative equation for the predicted state with perturbation is as follows:

[0013] in, To consider the iterative equations of the state variables under random perturbation, This means that, considering random perturbations, based on the current discrete time... Predicted future The system state vector of the step; and These represent the system state matrix and the input matrix, respectively. Represented as a state matrix The exponentiation, corresponding to the state in Iteration transition coefficients within a step; Indicates the current step number The system state vector, This represents the amount of random disturbance. middle, For the current discrete time It is the time base for state prediction. The step index is used to predict the time domain and traverse from the current discrete time step. To the future All moments of the step; Indicates the current step number The applied control input vector, Indicates the current step number Predicted future The control input vector applied step by step, To predict the step size; The objective function outside the region of convergence is as follows: ; in, To predict the total step size of future states, Represents the system state vector. Reference values ​​representing the system state vector; This represents the weighted second norm of the state error. This represents the control increment weighted quadratic norm; To control the change in quantity; The weight assigned to the change in the control quantity. For the terminal constraints of the control system.

[0014] Furthermore, the deviation optimization at the boundary of the convergence region specifically involves: A Bayesian neural network is used to make decisions about the state regions of a distributed drive electric vehicle. By introducing a probability distribution into the network weights, the uncertainty of prediction is quantitatively expressed. During training, the prior distribution of the network connection weights is first defined, and the posterior distribution is obtained by updating it through Bayesian inference. Variational inference and maximizing the lower bound of evidence are used to complete the approximate solution. The online decision takes the vehicle state parameters as input and outputs the region selection probability. A control mode switching penalty term is introduced to avoid frequent switching of control modes. The region selection probability is the probability value of the convergent region and the non-convergent region.

[0015] Furthermore, in the lower-level multi-objective cooperative control step, integrated horizontal and vertical control is achieved through horizontal and vertical dynamic tracking performance objective functions. accomplish: ;in, The observations are for the seven-degree-of-freedom vehicle state equations; This serves as a reference value for following the target. The penalty coefficient for changes in the control quantity; Indicates the first The change in control input for each step Indicates the current moment. This indicates the maximum permissible torque variation of the drive motor, which is determined by the drive motor's own performance. Indicates the first Each control step moment.

[0016] Furthermore, in the lower-level multi-objective cooperative control step, the energy consumption optimization strategy is implemented through an energy consumption optimization objective function. accomplish: ; in, The coefficients of the energy consumption function; These represent the front left wheel, front right wheel, rear left wheel, and rear right wheel, respectively. This represents the probability value of the control law corresponding to the selected convergence region in the upper-level control. For adaptive adjustment of the coefficient matrix; Indicates the current moment Predicted future Step, the first The torque control quantity of the drive motor corresponding to each wheel; Indicates different rotational speeds The optimal torque value for efficiency. This indicates the maximum permissible torque of the drive motor, which is determined by the performance of the drive motor itself.

[0017] Furthermore, in the lower-level multi-objective cooperative control step, the wheel slip ratio is effectively limited through the objective function. accomplish: ; ; in, These are the weights of the objective function; For standardization, express The input values ​​of the function. It is a unified expression paradigm that represents the tire adhesion utilization coefficient of a single wheel. Indicates to Input to the function middle, Indicates the first Axis No. Longitudinal tire force of the side wheel, Indicates the first Axis No. Lateral tire force of the side wheel, Indicates the first Axis No. The normal tire force of the side wheel, Indicates the road surface adhesion coefficient. Indicates wheel index: For the front wheels, For the rear wheels; For the left wheel, This is the right-hand wheel.

[0018] The beneficial effects of the method described in this invention are as follows: This invention proposes a hierarchical multi-model coordinated control strategy. The upper-level control determines the overall motion target of the vehicle from a global perspective, and constructs control constraints for the local optimization of the lower level. The lower-level control optimizes the dynamic response of each wheel based on vehicle state information and global constraints, realizing coordinated optimization of global and local control.

[0019] In the upper-level control structure, this invention divides the state region based on the convergence of the controlled system state, and constructs different control mechanisms according to different state regions and following target information, so as to achieve the optimal balance between real-time performance and fault tolerance in the control method.

[0020] Considering the uncertainty of the controlled system model and the deviation in the state region boundary division caused by random disturbances, a deep learning model based on Bayesian theory is constructed to dynamically adjust and optimize the state region boundary, thereby improving the robustness of the control method.

[0021] In the underlying control, this invention constructs a multi-objective collaborative optimization model, introduces energy consumption performance into the integrated horizontal and vertical control optimization, and constructs a multi-objective weight adaptive adjustment mechanism based on the convergence information of the upper-level system state to achieve a coordinated balance between energy consumption performance and dynamic performance. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the overall process of the method described in this embodiment of the invention. Figure 2 This is a flowchart illustrating the design of the upper-level control convergence region control strategy in an embodiment of the present invention. Figure 3 This is a flowchart illustrating the design of the upper-level control convergence region outside the control strategy in this embodiment of the invention. Figure 4 This is a configuration diagram of a distributed drive electric vehicle in an embodiment of the present invention. Detailed Implementation

[0023] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0024] Example 1 Distributed drive electric vehicle configuration diagram as follows Figure 4As shown in the background technology analysis, there are still some problems in key technologies such as the real-time control performance, handling stability, and overall energy economy of distributed drive electric vehicles. These problems can be summarized as follows: The problem of balancing energy consumption and dynamic performance: For a four-wheel independent drive integrated lateral and longitudinal control system, how to incorporate energy consumption performance into the framework of lateral and longitudinal dynamic performance optimization, and dynamically coordinate the conflicts between different objectives under the condition of satisfying high-dimensional nonlinear constraints, so as to achieve a balance between vehicle energy consumption performance and dynamic performance. Impact of model uncertainty and random disturbances: How to enhance the fault tolerance of the control system and improve the operational stability and energy consumption performance of FWID in the face of the problems of difficulty in accurately observing the state of the control system and random disturbances from the external environment; The challenge of balancing real-time performance and fault tolerance in control systems: Fault tolerance requires complex calculations to handle uncertainties and random disturbances, while real-time performance demands rapid response. To address the difficulty of achieving both simultaneously, how can an optimization mechanism be constructed to ensure the control system possesses fault tolerance capabilities while meeting real-time requirements?

[0025] To address the aforementioned issues, this embodiment proposes a hierarchical multi-model cooperative control strategy. This strategy aims to balance the energy consumption performance of the distributed-drive electric vehicle (DEV) while ensuring good lateral and longitudinal dynamic performance, and simultaneously balancing the real-time performance and fault tolerance of the control system. The upper-level controller analyzes the state convergence of the DEV control system under different following targets and divides the convergence region. Based on different state regions and following target information, an optimization mechanism is constructed to adapt to the control model and control quantity solution method under different regions. To address the state region boundary demarcation deviation problem caused by random disturbances, a deep learning model based on Bayesian theory is constructed to improve the discrimination accuracy of the control system under boundary states. Furthermore, the lower-level controller incorporates the vehicle's energy consumption performance into the integrated lateral and longitudinal control optimization framework, constructing a multi-objective weight adaptive adjustment mechanism. Under the condition of satisfying high-dimensional nonlinear constraints, the conflict between different control objectives is dynamically coordinated to achieve a balance between vehicle energy consumption performance and dynamic performance.

[0026] like Figure 1 The diagram shown is the overall framework of the method described in this embodiment. The method will be described in detail below with reference to the accompanying drawings. The explanation will be divided into two parts: the first part introduces the upper-level control strategy, including the establishment of the convergence region and the design of the control strategy, the design of the control strategy for the non-convergence region, and the optimization of the boundary deviation of the convergence region; the second part introduces the lower-level control strategy.

[0027] 1. Upper-level control strategy: Different control designs are implemented in the upper layer based on the convergence regions defined by the overall vehicle status. Figure 2 A flowchart illustrating the design of the upper-level control convergence region control strategy; Figure 3 This is a flowchart illustrating the design of the control strategy outside the convergence region of the upper-level control in this embodiment of the invention. Within the convergence region, the focus is on computational efficiency, employing a combination of offline and online methods. The vehicle control law is established offline, and online, the control law is searched based on the actual state of the vehicle. Outside the convergence region, the focus is on the vehicle's driving stability and the robust stability of the control strategy. A real-time solution strategy is constructed online based on the vehicle state and the target convergence region, aiming to enable the vehicle to quickly enter the convergence region. Furthermore, considering the influence of random disturbances on the system, the state variables exhibit randomness within a certain range, making it difficult to accurately define the convergence region boundary. A Bayesian neural network is used to address the state region boundary partitioning deviation problem, dynamically adjusting and optimizing the state region boundary.

[0028] 1.1 Upper-level control - Establishment of convergence region and design of control strategy: The convergence region is established based on Lyapunov theory and is a defined subset of the state space of the distributed drive electric vehicle under a specific control law. In this subset of states, the system state is guaranteed to converge asymptotically to the target state. Based on the state of the distributed drive electric vehicle and the state of the tracking target, a three-degree-of-freedom state equation is established in the Frenet coordinate system, as shown in equations (1)-(3).

[0029] (1) (2) (3) In the formula, Represents the system state vector. Represents the system output vector. Represents the system control input vector. , , These are the system state matrix, input matrix, and output matrix, respectively. , , , These are the lateral position error between the vehicle and the reference trajectory, the heading error between the vehicle's heading angle and the tangent direction of the reference trajectory, the longitudinal position error of the vehicle along the reference trajectory, and the speed error between the vehicle's longitudinal velocity and the desired velocity. , These are the vehicle's total mass and acceleration, respectively. For steering angle; , These are moment of inertia and vehicle speed, respectively. , These are the distances from the front and rear axles to the center of mass, respectively. , These are the lateral stiffness of the front and rear wheels, respectively.

[0030] Based on the derivation of the above three-degree-of-freedom dynamic model, the state equation is finally formulated into a state error expression. For the parameters related to lateral control, the vehicle's desired yaw rate and sideslip angle can be obtained using the dynamic equations of the three-degree-of-freedom dynamic model, as shown in equations (4)-(6).

[0031] (4) (5) (6) In the formula, It is the desired sideslip angle. It is the expected yaw rate. It is the vehicle's desired front wheel steering angle. It is the road surface adhesion coefficient. It is gravitational acceleration.

[0032] Based on the three-degree-of-freedom dynamic model, a control law matrix is ​​introduced. Construct explicit state-control mappings. , This represents the optimal control vector. Furthermore, based on Lyapunov theory, a Lyapunov function for distributed drive electric vehicles is established. As shown in formula (7).

[0033] (7) In the formula, Represents the Lyapunov function. It is a positive definite symmetric matrix. This is the state matrix of the closed-loop system.

[0034] To further incorporate the performance characteristics of distributed drive electric vehicles in the target tracking process, an objective function is constructed. As shown in formula (8).

[0035] (8) The final variable obtained is the optimal control input sequence in the future prediction time domain, i.e., the magnitude of the driving torque of each wheel. It achieves control through two steps: first, starting from the current vehicle state, predicting the future under different control sequences... The system achieves two objectives: first, by minimizing the error between the predicted state value and the reference trajectory, it ensures lateral and longitudinal tracking performance; second, by penalizing changes in the control input, it limits the amplitude of torque jumps, ensuring system smoothness and safety.

[0036] In the formula, This refers to the total step size for predicting future states in model predictive control. The system's continuous-time output vector. for Discretized expression; The ideal reference value for the state parameters is determined by the target being followed. The current discrete control moment is denoted as step number. , To predict the step index in the time domain; To control the change in quantity; This represents the weighted second norm of the state error. This represents the weighted quadratic norm of the control increment. and Total step size The weight value of each step in the process. The weight of the change in the control quantity in the objective function.

[0037] In solving formula (8), the method of inverse iteration is based on the Riccati equation. The optimal control law obtained by the Riccati equation has good control effect under unconstrained conditions. However, in a four-wheel independent system, there are many state and control constraints due to the limitations of the working state of the vehicle components. Therefore, it is necessary to describe a feasible convergence region to ensure that the single control law converges within this region.

[0038] Solving for the region of convergence while satisfying Lyapunov's theorem. ,in, The accuracy parameters of the convergence region are obtained through iterative solutions. The Lyapunov function, expressed in scalar form, is mentioned above. The simplified representation; "scalar form of Lyapunov function" refers to the Lyapunov function. The output is a scalar value, not that the input variables are scalars. This is the previous text. In simplified mathematical representations, both refer to the exact same Lyapunov function. It is a state vector The abbreviation symbol, its physical meaning is the same as Maintain consistency.

[0039] To obtain more accurate parameters The Gold Rush Optimization (GRO) algorithm is used to improve parameters. The accuracy and efficiency of iterative solutions. GRO simulates the behavior of miners searching for gold in unknown areas and has good optimization capabilities for handling multi-dimensional information. Therefore, GRO is effective in solving... It performs well in six-dimensional state-space problems. The GRO expression is established as shown in equation (9).

[0040] (9) In the formula, Indicates the number of iterations. Indicates the first The current position vector of an individual miner at the next iteration; , , The next position selection of an individual is comprehensively selected from three aspects: the individual's global optimal fitness, the individual's local fitness, and the fitness among individuals. The specific expressions are shown in formulas (10)-(12).

[0041] (10) (11) (12) In the formula, In the t-th iteration, the... The current position vector of each individual miner (optimal solution), corresponding to a candidate solution for the parameters to be optimized. Indicates the first The adaptive weight coefficients of each miner dynamically change with the number of iterations, used to balance the algorithm's global exploration and local exploitation capabilities. This represents the optimal position for a miner in the gold mining optimization algorithm; , and To randomly select the miners' locations; This is the current location of the miner. and Random numbers between 0 and 1; and These represent the maximum number of iterations and the current number of iterations.

[0042] GRO is applied in solving convergence regions. During the process, select the problem to be solved. As a miner's position, i.e. , This is the number of iterations. To obtain the optimal... ,make To reach the maximum range, in terms of fitness calculation, comprehensive consideration is taken into account. and .

[0043] in, Represented as the first The sum of the values ​​representing the locations of individual miners indirectly represents... The size of the range of values. This represents the proportion of initial states that conform to formula (26) at the miner's current position. Specifically, in Randomly generated within the range There are n initial state points, and the number of initial state points that satisfy Lyapunov's theorem and the state constraints is n. ,Right now . The indirectness of the Monte Carlo method represents the solution. The degree of difference from the true convergence region, Indicates the first The accuracy parameters of each candidate convergence region.

[0044] Compared to ordinary iterative solution methods, GRO can effectively handle high-dimensional problems and obtain results more efficiently. The region. However, the solution obtained through GRO iteration. This represents a six-dimensional spatial region. The high-dimensional spatial region brings complex computational complexity to subsequent exploration of boundary stability, making it difficult to establish a control model with real-time capabilities and robustness in online applications.

[0045] By analyzing the sensitivity of the state equation (Equation (1)) of the distributed drive electric vehicle, it can be seen that the control system is sensitive to the second dimension (lateral error rate). ) and the fourth dimension (yaw rate error rate) The disturbances to ) are most sensitive. Therefore, in order to optimize The problem of the curse of dimensionality in six-dimensional regions was addressed using an improved Fourier-Motzkin elimination method. The six-dimensional region is projected onto the second and fourth dimensions to construct a new two-dimensional region. , As shown in formulas (13)-(15).

[0046] (13) (14) (15) In the formula The variables eliminated in the current iteration; To remove Other variables besides; , and They represent the first , No. and the State variables in linear inequality constraints The coefficient; , and They represent the first , No. and the Unsupplied variables in linear inequality constraints The coefficient; , and They represent the first , No. and the The right-hand boundary values ​​of a linear inequality constraint; , and They respectively represent the constraints except for A linear combination of all other state variables.

[0047] The Fourier-Motzkin elimination method transforms a high-dimensional linear inequality system into a low-dimensional problem by gradually eliminating variables, thereby realizing the projection region of a multi-dimensional region onto a specific dimension. Formulas (13)-(15) are transformed by gradually eliminating variables. The new inequality constraint equations are obtained, as shown in equation (16).

[0048] (16) Analysis of the Fourier-Motzkin elimination method reveals that its essential principle is to transform the dimensional information of the elimination into new inequalities that apply to the system. However, each iteration generates new inequalities, increasing computational complexity. Therefore, when the Fourier-Motzkin elimination method generates new inequalities in each iteration, a boundary check is performed to examine whether the newly generated inequalities are redundant, removing redundant inequalities from the solution space and reducing computational complexity.

[0049] During the process of constructing the convergence region and solving the control law, vehicle speed... As important parameters influencing the values ​​of state transition matrices A and B (as in formula (1)), they directly affect the definition of the convergence region and the value of the control law. Therefore, the vehicle speed... Discretize the interval within the feasible region, i.e. , They represent respectively

[0050] For each speed range, a lower and upper limit are defined, and corresponding convergence regions and control laws are established for each speed range. When the control law changes due to speed range switching, a buffer zone is set at the speed range boundary to ensure the continuity or smoothness of the control input. In the buffer area The control law weighted transition method is adopted, as shown in formula (17).

[0051] (17) In the formula, for Internal control laws for The control law, As a weighting factor, For buffer The control law; Indicates the current position is number 1. The vehicle speed is divided into speed ranges, which are used to differentiate control laws for different speed segments. This is because the vehicle speed is divided into... , ,...., These intervals, It is a sequence number used to mark "which speed range the vehicle is currently in".

[0052] In summary, based on the state-space equations of the control system for distributed-drive electric vehicles tracking targets, a six-dimensional state convergence region is constructed using Lyapunov, and the GRO algorithm is employed to optimize the efficiency of iterative solution of the convergence region. To reduce the dimensionality of the convergence region and achieve better real-time deployment, an improved Fourier-Motzkin elimination method is used to reduce the six-dimensional convergence region to a two-dimensional convergence region. The system exhibits strong stability within the state convergence region, thus placing greater emphasis on real-time optimization. By establishing a mapping relationship between the convergence region and the control law offline, and selecting the appropriate control law based on the correspondence between the actual state and the convergence region during online application, the real-time calculation process of online control quantities is transformed into an online table lookup process, improving the real-time performance of the control system.

[0053] 1.2 Upper-level control - Design of control strategy for non-convergent regions In the non-convergence region, due to system constraints, the search-based control law method cannot guarantee state convergence. Therefore, it is necessary to establish an online real-time solution method for the control quantity based on the current state and the following target information. The model predictive control iterative relationship is established based on the three-degree-of-freedom vehicle dynamics, as shown in Equation (18).

[0054] (18) in, The meaning is that at the current discrete time... Predicted future The system state vector of the step, Represented as a state matrix The exponentiation, corresponding to the state in Iterative transition coefficients within a step, The meaning represented is the current discrete time. The applied control input vector, The meaning represented is the current discrete time. The system state vector, The meaning represented is the current discrete time. Predicted future The control input vector applied step by step, To predict the step size. The systems mentioned in the text all refer to distributed drive electric vehicle dynamics systems.

[0055] Considering the presence of random disturbances in the system, the state iteration equation in formula (18) is redefined in a new form, as shown in formula (19): (19) in, For random perturbation quantities, To formulate iterative equations for state variables considering random perturbations; middle, The current discrete control moment serves as the time reference for state prediction. The step index in the time domain is used to traverse from the current time step. To the future All moments in the step. Together, these two factors allow for a complete calculation of the future from the current moment. The cumulative effect of random perturbations on the vehicle state at each step.

[0056] As shown in Equation 19, the random disturbances in the system state equations accumulate errors as the control time domain increases, and the coefficient matrix... This determines the convergence of the accumulated error. Therefore, by designing state feedback control, the state error value generated by random disturbances can be minimized. It exhibits convergence characteristics. The state feedback equation is shown in equation (20).

[0057] (20) in, Indicates the first Step-by-step correction of the control input vector, Indicates the first The actual system state vector at each step.

[0058] When the feedback control strategy is applied to the control system, the cumulative state error value generated by random disturbance is shown in Equation (21).

[0059] (twenty one) By adopting different Value eigenvalues ​​of a matrix ,when Time disturbance error It tends to converge, that is , express of Power of 1.

[0060] Due to the influence of random disturbances, the actual state of the system deviates from the state predicted by the nominal model, thus affecting the prediction accuracy and control performance. MPC predicts state changes based on the initial state, and the initial state error is amplified in the prediction time domain, affecting control accuracy and constraint satisfaction. To suppress system errors caused by disturbances, this invention uses the initial state variable as the decision variable in the optimization solution, enabling the controller to adjust the state trajectory in real time within the prediction domain to respond to the deviations caused by disturbances. The state iteration matrix under the nominal model is shown in formula (22).

[0061] (twenty two) Based on the reference value of the tracking target, an objective function is established outside the convergence region, as shown in formulas (23)-(25).

[0062] (twenty three) (twenty four) (25) In the formula, For a custom weight matrix, and These are the lower and upper limits of the system state variables, determined by the vehicle's own performance. The upper bound of the maximum random disturbance of the system is determined through theoretical analysis and experimental calibration. and These are the lower and upper limits of the control input, respectively, determined by the characteristics of the vehicle's actuators.

[0063] in, This serves as a reference value for the system state vector; and These are the weights assigned to changes in control quantities and terminal states, respectively, and their values ​​are determined based on calibration. As the final constraint of the control system, the final state of the vehicle (the yaw rate error and the center of gravity sideslip angle error of the vehicle) is within the convergence region, and its specific expression is shown in formula (24). This represents the stable convergence region of the system, which is the set of states that satisfy the Lyapunov stability criterion. It is a function that is either equal to 1 or equal to 0, and its value depends on... Does it belong to As shown in Formula 24.

[0064] In summary, state feedback control mitigates the impact of random disturbances on the control system. Specifically, under model predictive control, the optimal control sequence obtained by solving the objective function of formula (23) is applied to the system. Random disturbances cause inconsistencies between the actual vehicle state information and the calculated optimal state information. The rolling optimization strategy in model predictive control improves the stability of the control system, but it does not take corresponding optimization measures for future disturbances. Therefore, when calculating the optimal control sequence of model predictive control, the initial state variables are used as decision variables to solve for the optimal state value at the current moment. This value is then applied to the system through feedback control to adjust the control quantity accordingly for the consistency between the theoretical and actual states in the next step.

[0065] 1.3 Optimization of Boundary Deviation in Convergence Region Because the system is affected by random disturbances, the state variables exhibit randomness within a certain range, making it difficult to accurately determine the control mode at the convergence region boundary. To ensure the stability of the control system under different disturbance conditions, this invention employs a Bayesian Neural Network (BNN) to determine and decide on the state region of a distributed drive electric vehicle. The BNN combines the advantages of neural networks and Bayesian inference, introducing a probability distribution into the network weights to quantify the uncertainty in predictions during model training. BNNs demonstrate excellent performance in handling complex decision-making problems caused by random disturbances.

[0066] Before training in BNN, the prior distribution of network connection weights is defined and treated as random variables. During training, Bayesian inference is performed on them, as shown in Equation (26).

[0067] (26) In the formula BNN weights The probability distribution; Represented as a multivariate normal distribution, It follows a pattern with a mean of 0 and a covariance matrix of . The multivariate normal distribution, The variance parameter represents the variance of a multivariate normal distribution and controls the dispersion of the weight distribution. Represent the identity matrix, ensuring that each weight component is independent. Represents the weight vector The dimension is the total number of weights to be trained in the BNN.

[0068] During training, the prior distribution is updated using new observations to obtain the posterior distribution. This reflects the actual constraints and impact of training data on the weights. BNN weights The posterior distribution update expression is shown in equation (27).

[0069] (27) In the formula The input values ​​for the training data; The target set for the training data; It is the output value of the neural network, representing the weights. Enter below The corresponding predicted value.

[0070] Calculate marginal likelihood Need to consider all possible weights Integrating is difficult to achieve in high-dimensional spaces; therefore, the posterior distribution... Typically, it is not analytical. To overcome this problem, variational inference introduces a tractable distribution. To approximate the posterior distribution This is achieved by maximizing the Evidence Lower Bound (ELBO). Approximate posterior distribution The ELBO expression is shown in formula (28).

[0071] (28) in, Let be the expected log-likelihood, representing the variational distribution. Below, data Given input and weight The expected value of the log-likelihood; Let KL divergence be the expression. and The difference between them, namely .

[0072] In the online application of BNN, parameters such as the rate of change of the vehicle's lateral error, the rate of change of the yaw angle error, and the vehicle speed are collected in real time as inputs to the BNN. After the BNN makes a decision, it outputs the probability values ​​of selecting two state regions. In the decision-making process, a control mode switching penalty term is introduced to avoid frequent switching of control modes, as shown in formula (29).

[0073] (29) In the formula In order to be in Decision outcome at time ( This represents the control law corresponding to the selected convergence region; (Represented as the numerical value of the online calculated control quantity); and They are respectively and The expected return depends on the magnitude of the probability values ​​of the two state regions output by the BNN. This is the penalty coefficient for switching control modes. This indicates the total number of time steps for control decisions.

[0074] In summary, the upper-level control system divides the four-wheel independent drive system into state regions and constructs different control mechanisms based on different state regions and the following target information to achieve an optimal balance between real-time performance and fault tolerance. Reinforcement learning (BNN) is employed to optimize the control mode boundary partitioning deviation caused by random disturbances. Compared to rule-based mode selection strategies, BNN quantifies the decision uncertainty caused by system disturbances during the decision-making process by calculating the posterior distribution, thus enabling a more robust control mechanism within the uncertainty region.

[0075] 2. Lower-level control - Design of control strategy The upper-level controller determines the motion reference of the entire vehicle from a global perspective based on the tracking target and vehicle state. However, the lower-level controller is responsible for the precise control of the drive components. Therefore, the lower-level controller needs to establish a more accurate model to characterize the dynamic changes of the distributed drive electric vehicle. The lower-level controller adopts a model predictive control strategy based on a seven-degree-of-freedom vehicle model, using the yaw rate, acceleration change, sideslip angle, etc., calculated by the upper level as reference quantities, and achieves integrated lateral and longitudinal control by controlling the torque of the four-wheel motors. In addition, an energy consumption optimization strategy is introduced on the basis of ensuring the stability of the integrated lateral and longitudinal control to achieve coordinated control among multiple objectives. The state space equations of the lower-level controller are shown in equations (30)-(31).

[0076] (30) (31) in, This represents the system state vector of the lower-level controller. This represents the system output vector of the lower-level controller. This represents the system control input vector of the lower-level controller. , , These represent the state matrix, input matrix, and output matrix of the lower-level model, respectively. This represents the nonlinear state transition function of a seven-degree-of-freedom vehicle model. Indicates the longitudinal speed of the vehicle. Indicates the vehicle's lateral speed. Indicates the yaw rate of the vehicle. , These are the steering angle and the tire radius, respectively. For the total mass of the vehicle. and These are the distances from the front and rear axles to the center of mass, respectively.

[0077] A multi-objective control framework is established to collaboratively optimize energy consumption and longitudinal and lateral dynamic tracking performance. The objective functions for longitudinal and lateral dynamic tracking performance are shown in Equation (32).

[0078] (32) in, The observations are for the seven-degree-of-freedom vehicle state equations (Formulas 30-31); This serves as a reference value for following the target. The penalty coefficient for changes in the control quantity; and Step size The weight matrix for each step in the process. Indicates the first The change in control input for each step This indicates the maximum permissible torque variation of the drive motor, which is determined by the motor's own performance; here... Indicates the future step number within the prediction time domain. The whole represents the present moment. Beginning, the future's first Each control step moment.

[0079] objective function Through two weighted optimization steps, synergistic optimization of lateral and longitudinal dynamic tracking and control smoothness is achieved within the model predictive control framework: the first step minimizes the error between the vehicle's predicted state and the reference target, ensuring the vehicle's motion tracking performance; the second step normalizes and penalizes changes in the control input, limiting the torque jump amplitude of the drive motor and ensuring the safety and smoothness of the control process. By solving for the control sequence that minimizes the objective function in each control cycle, stable and high-performance control of the distributed drive electric vehicle is ultimately achieved.

[0080] In distributed drive electric vehicle systems, the motor's speed and torque directly affect its efficiency. The motor and wheels are mechanically coupled, and the motor's speed has a fixed proportional relationship with the vehicle speed. Therefore, the value of the motor torque, as a control variable, directly impacts the motor's efficiency. To incorporate the motor's energy consumption characteristics into the integrated lateral and longitudinal control and to construct an objective function solvable by quadratic programming (QP), the motor's operating map is discretized, and a table of optimal torque values ​​for motor efficiency at different speeds is constructed. Therefore, in practical control, the value of the motor torque should follow the optimal torque. The energy consumption optimization objective function is shown in formula (33).

[0081] (33) In the formula, The coefficients of the energy consumption function; They represent the front left wheel, front right wheel, rear left wheel, and rear right wheel, respectively. The probability value of the control law corresponding to the selected convergence region of the BNN output in the upper-level control; Indicates the current moment Predicted future Step, the first The torque control quantity of the drive motor corresponding to each wheel; This indicates the maximum permissible torque of the drive motor, which is determined by the performance of the drive motor itself. The adaptive adjustment coefficient matrix is ​​expressed as shown in equation (34).

[0082] objective function The goal is to reduce overall vehicle energy consumption by ensuring that the actual operating torque of the four-wheel drive motor is as close as possible to the optimal torque at the current speed, while meeting the vehicle's motion control requirements. This is achieved by calculating and squaring the normalized difference between the actual torque of each wheel and the optimal torque, transforming the energy consumption optimization problem into a solvable quadratic programming problem. Simultaneously, an adaptive adjustment coefficient is introduced to dynamically adjust the weight of energy consumption optimization based on the vehicle's current state, achieving coordinated control of energy consumption optimization and vehicle stability.

[0083] (34) In the formula This is the amplitude coefficient of the sine term; The frequency coefficient of the sine term; These are the weighting coefficients for the exponential term; This is the decay coefficient of the exponential term.

[0084] When the state system of a distributed drive electric vehicle is in the convergence region, the vehicle exhibits high stability, thus energy consumption performance should be optimized. However, when the stability of the vehicle's target following decreases, the weight of energy consumption should be reduced, prioritizing the improvement of following stability. To achieve this, this invention establishes a mapping relationship between the probability assessment of the convergence region by the BNN in the upper-level control and the performance weights of the lower-level control. This enables the four-wheel motor to adaptively adjust the weights of energy consumption performance and following dynamic performance, ensuring a coordinated balance between energy consumption and stability in the distributed drive electric vehicle.

[0085] Furthermore, an excessively high wheel slip ratio can lead to a loss of vehicle control, especially during cornering, which can easily cause sideslip and lead to accidents. An excessively high slip ratio can also cause severe tire wear and the risk of tire blowout. Therefore, in order to improve the stability of the vehicle's following control, it is necessary to effectively limit the wheel slip ratio and reduce the impact of sideslip and instability factors. The objective function is established as shown in equations (35)-(36).

[0086] (35) (36) in, These are the weights of the objective function; For standardization, express The input values ​​of the function. It is a unified expression paradigm that represents the tire adhesion utilization coefficient of a single wheel. Indicates to Input to the function middle, Indicates the first Axis No. Longitudinal tire force of the side wheel, Indicates the first Axis No. Lateral tire force of the side wheel, Indicates the first Axis No. The normal tire force of the side wheel, Indicates the road surface adhesion coefficient. Indicates wheel index: For the front wheels, For the rear wheels; For the left wheel, This is the right-hand wheel.

[0087] The objective function aims to improve driving stability by limiting the tire adhesion utilization of each wheel, preventing excessive wheel slip ratio that could lead to vehicle sideslip and loss of control.

[0088] The three objective functions constructed in this invention , , Ultimately, the control objective is achieved by adjusting the independent drive of the four wheels. In the actual solution, the three objective functions are merged into a unified multi-objective optimization problem, which is solved by quadratic programming (QP) to find the four-wheel torque distribution sequence that minimizes the overall objective function.

[0089] : Optimizes lateral and longitudinal dynamic tracking performance to ensure that the vehicle follows the reference trajectory and speed. : Optimize motor energy consumption to bring the torque of each wheel close to the optimal operating point. Tire adhesion utilization and slip ratio limitation prevent wheel slippage and ensure driving stability.

[0090] The three factors are weighted and combined into a single overall objective function. Under the same optimization framework, the torque of the four wheels is solved to achieve coordinated optimal control of "tracking performance, energy efficiency, and driving stability". The separate listing is only for the purpose of explaining the control objectives and constraints of each part.

[0091] As can be seen from the above analysis, the final objective function fully considers the vehicle's dynamic performance, energy consumption performance, changes in control quantity, and tire characteristics, thus realizing multi-objective cooperative control of the four-wheel independent drive system. The overall objective function can be obtained from equations (32), (33), and (35), as shown in equations (37)-(38).

[0092] (37) (38) In the formula, and These represent the maximum and minimum values ​​of the state variables, which are determined by both the vehicle's physical limits and the requirements of its operating conditions. and These represent the maximum and minimum values ​​of the output, respectively, which are determined by the range of the tracking reference value and the control target requirements; and These represent the maximum and minimum values ​​of the control quantity, respectively, which are determined by the physical properties of the drive motor. and These represent the maximum and minimum values ​​of the tire's vertical load, which are determined by the vehicle load and the tire's own characteristics. and These represent the maximum and minimum speeds of the drive motor, respectively, which are determined by the physical properties of the drive motor. and These represent the maximum and minimum output probability values ​​of the BNN, respectively, which are determined by the physical definition of the probability distribution and the control logic.

[0093] In summary, a hierarchical multi-model coordinated control strategy was established to address the multi-objective coordinated control problem between horizontal and vertical integrated control and energy consumption, while ensuring the real-time performance and fault tolerance of the control system. The upper-level control constructed different control mechanisms based on different state regions, achieving an optimal balance between real-time performance and fault tolerance. The lower-level control fully considered the vehicle's dynamic performance, introducing horizontal and vertical integrated control into the energy consumption optimization strategy, and constructing a multi-objective weight adaptive adjustment mechanism based on information from the upper-level control, thus achieving coordinated optimization of energy consumption performance and dynamic performance of the distributed drive electric vehicle.

Claims

1. A multi-objective coordinated control method for a distributed drive electric vehicle, characterized in that, The method includes an upper-level partitioned adaptive control step and a lower-level multi-objective cooperative control step. The upper-level partition adaptive control step divides and distinguishes the convergent region from the non-convergent region based on whether the real-time operating state of the vehicle satisfies the Lyapunov asymptotic convergence constraint condition. Within the convergence region, a combination of offline and online methods is used: the vehicle control law is established offline, and the control law is searched online based on the actual state of the vehicle. In the non-convergence region, a real-time solution strategy is constructed online based on the vehicle status and the target convergence region; Deviation optimization is performed at the boundary of the convergence region; The lower-level multi-objective cooperative control step adopts a model predictive control strategy based on a seven-degree-of-freedom vehicle model. The vehicle motion parameters calculated by the upper level are used as reference quantities. The lateral and longitudinal integrated control is achieved by controlling the torque of the four-wheel motors. On the basis of ensuring the stability of the lateral and longitudinal integrated control, an energy consumption optimization strategy is introduced and the wheel slip ratio is effectively limited to achieve cooperative control among multiple objectives.

2. The multi-objective coordinated control method for a distributed drive electric vehicle according to claim 1, characterized in that, The convergence region, based on Lyapunov theory, is a defined subset of the state space of a distributed electric vehicle under a specific control law. Within this subset, the system state is guaranteed to asymptotically converge to the target state. ,in, Represents the system state vector. The accuracy parameters of the convergence region are obtained through iterative solutions. The Lyapunov function is expressed in scalar form.

3. The multi-objective coordinated control method for a distributed drive electric vehicle according to claim 2, characterized in that, Improve accuracy parameters by using gold mining optimization algorithms The accuracy and efficiency of iterative solutions are comprehensively considered from three aspects: the individual's global optimal fitness, the individual's local fitness, and the fitness among individuals, to select the next position for each individual. The optimal accuracy parameters are obtained in the next iteration. .

4. The multi-objective coordinated control method for a distributed drive electric vehicle according to claim 3, characterized in that, Within the convergence region, the high-dimensional convergence region is reduced to a two-dimensional sensitive dimension region using an improved Fourier-Motzkin elimination method. The convergence region and control law are established segmented according to vehicle speed intervals, and a weighted smooth switching strategy is adopted in the buffer section of the vehicle speed interval. The control quantity is output online through a lookup table method, specifically: The improved Fourier-Motzkin elimination method projects the six-dimensional state convergence region onto the two-dimensional sensitive dimensions of lateral error rate and yaw rate error rate. The six-dimensional state convergence region is a feasible region of the state space constructed under Lyapunov theory based on the three-degree-of-freedom dynamic model of a distributed-drive electric vehicle, which can guarantee the asymptotic convergence of the system state to the target state. The improvement in the improved Fourier-Motzkin elimination method is as follows: when a new inequality is generated in each iteration of the Fourier-Motzkin elimination method, the boundary check method is used to check whether the newly generated inequality is redundant, and the redundant inequalities in the solution space are removed. speed Discretize the interval within the feasible region, i.e. , These represent the lower and upper limits of the vehicle speed range, respectively. Corresponding convergence regions and control laws are established for different speed ranges. When the control law changes due to speed range switching, a buffer zone is set at the speed range boundary. In the buffer zone The control law-based weighted transition method is used, as shown in the following formula: ； In the formula, for Internal control laws for The control law, As a weighting factor, For buffer The control law, Indicates the current position is number 1. Speed ​​range.

5. A multi-objective coordinated control method for a distributed drive electric vehicle according to claim 4, characterized in that, Within the non-convergence region, a model predictive control state iterative relation is established based on three-degree-of-freedom vehicle dynamics. Random disturbances are introduced into the state iterative relation to construct a predicted state iterative equation with disturbances. State feedback control is designed to converge the cumulative state error caused by random disturbances. In the optimization solution, the initial state variables are used as decision variables, enabling the controller to adjust the state trajectory in real time within the prediction domain to respond to the deviations caused by disturbances. Based on the reference value of the tracking target, an objective function outside the convergence region is established. In the objective function, the final state of the vehicle is kept within the convergence region by the terminal constraints of the control system. The iterative equation for the predicted state with perturbation is as follows: in, To consider the iterative equations of the state variables under random perturbation, This means that, considering random perturbations, based on the current discrete time... Predicted future The system state vector of the step; and These represent the system state matrix and the input matrix, respectively. Represented as a state matrix The exponentiation, corresponding to the state in Iteration transition coefficients within a step; Indicates the current step number The system state vector, This represents the amount of random disturbance. middle, For the current discrete time It is the time base for state prediction. The step index is used to predict the time domain and traverse from the current discrete time step. To the future All moments of the step; Indicates the current step number The applied control input vector, Indicates the current step number Predicted future The control input vector applied step by step, To predict the step size; The objective function outside the region of convergence is as follows: ； in, To predict the total step size of future states, Represents the system state vector. Reference values ​​representing the system state vector; This represents the weighted second norm of the state error. This represents the control increment weighted quadratic norm; To control the change in quantity; The weight assigned to the change in the control quantity. For the terminal constraints of the control system.

6. The multi-objective coordinated control method for a distributed drive electric vehicle according to claim 5, characterized in that, The specific steps for bias optimization at the boundary of the convergence region are as follows: A Bayesian neural network is used to make decisions about the state regions of a distributed drive electric vehicle. By introducing a probability distribution into the network weights, the uncertainty of prediction is quantitatively expressed. During training, the prior distribution of the network connection weights is first defined, and the posterior distribution is obtained by updating it through Bayesian inference. Variational inference and maximizing the lower bound of evidence are used to complete the approximate solution. The online decision takes the vehicle state parameters as input and outputs the region selection probability. A control mode switching penalty term is introduced to avoid frequent switching of control modes. The region selection probability is the probability value of the convergent region and the non-convergent region.

7. A multi-objective coordinated control method for a distributed drive electric vehicle according to claim 6, characterized in that, In the lower-level multi-objective cooperative control step, integrated horizontal and vertical control is achieved by dynamically tracking the performance objective function in both directions. accomplish: ;in, The observations are for the seven-degree-of-freedom vehicle state equations; This serves as a reference value for following the target. The penalty coefficient for changes in the control quantity; Indicates the first The change in control input for each step Indicates the current moment. This indicates the maximum permissible torque variation of the drive motor, which is determined by the drive motor's own performance. Indicates the first Each control step moment.

8. A multi-objective coordinated control method for a distributed drive electric vehicle according to claim 7, characterized in that, In the lower-level multi-objective cooperative control step, the energy consumption optimization strategy is implemented through the energy consumption optimization objective function. accomplish: ； in, These are the coefficients of the energy consumption function; These represent the front left wheel, front right wheel, rear left wheel, and rear right wheel, respectively. This represents the probability value of the control law corresponding to the selected convergence region in the upper-level control. For adaptive adjustment of the coefficient matrix; Indicates the current moment Predicted future Step, the first Torque control quantity of the drive motor corresponding to each wheel; Indicates different speeds The optimal torque value for efficiency. This indicates the maximum permissible torque of the drive motor, which is determined by the performance of the drive motor itself.

9. A multi-objective coordinated control method for a distributed drive electric vehicle according to claim 8, characterized in that, In the lower-level multi-objective cooperative control step, the wheel slip ratio is effectively limited through the objective function. accomplish: ； in, These are the weights of the objective function; For standardization, express The input values ​​of the function. It is a unified expression paradigm that represents the tire adhesion utilization coefficient of a single wheel. Indicates to Input to the function middle, Indicates the first Axis No. Longitudinal tire force of the side wheel, Indicates the first Axis No. Lateral tire force of the side wheel, Indicates the first Axis No. The normal tire force of the side wheel, Indicates the road surface adhesion coefficient. Indicates wheel index: For the front wheels, For the rear wheels; For the left wheel, The right wheel.

Images