Motor intelligent control method based on digital twinning and deep reinforcement learning

Abstract

The application relates to a motor intelligent control method based on digital twinning and deep reinforcement learning, a deep reinforcement learning training environment is defined on the basis of a constructed multi-physical field digital twinning body, the training environment is dependent on the multi-physical field digital twinning body, after a strategy network is trained, a preliminary control action output by the strategy network is used as a reference track of a model predictive controller, a constrained finite time domain optimization problem is solved through the model predictive controller, and finally, a control instruction of an electric agricultural machine is generated, the training environment of the deep reinforcement learning comprises motor states, vehicle poses and environment parameters, the action is torque or current instructions of the motors, and a reward function is a multi-objective function of comprehensive energy consumption, operation quality and equipment safety. The process combines long-term global adaptive optimization capability of the deep reinforcement learning and short-term accurate constraint processing capability of the model predictive control, and has excellent control effects of simultaneously considering safety and controllable cost.

Description

Technical Field

[0001] This invention relates to the field of new energy vehicles and intelligent control technology, and more specifically, to an intelligent motor control method based on digital twins and deep reinforcement learning. Background Technology

[0002] With the development of agricultural electrification and intelligentization, new energy electric agricultural machinery, represented by independent or coupled multi-motor drives, has become an important direction. These electric agricultural machines are usually equipped with multiple motors responsible for walking drive, power output (PTO), and hydraulic assistance, respectively. The core challenge lies in achieving efficient, precise, and robust coordination of multiple motors under complex and time-varying farmland conditions, so as to simultaneously optimize operation quality, energy economy, and system reliability.

[0003] Traditional control methods for electric agricultural machinery mainly rely on fixed-rule-based energy management or classic proportional-integral-derivative (PID) cascade control. These methods exhibit significant limitations when facing highly nonlinear soil-implementation interactions, severe load disturbances, and multi-objective optimization requirements (such as efficiency, energy saving, and protection). For example, fixed-rule-based control methods have poor adaptability and cannot cope with undefined operating conditions; proportional-integral-derivative control methods are difficult to tune and lack look-ahead optimization capabilities.

[0004] While model predictive control (MMCC) can handle multiple constraints and perform short-term optimization, its performance heavily relies on accurate agricultural machinery-environment dynamics models, which are costly to identify and maintain online. Deep reinforcement learning offers a novel approach to solving complex sequential decision-making problems and has the potential for model-free adaptive optimization; however, directly applying it to expensive physical agricultural machinery for online training presents safety risks during the exploration process, low data sampling efficiency, and unacceptable training costs. Therefore, current technologies lack a comprehensive solution that can safely and efficiently combine long-term adaptive optimization with short-term precise constraint control for multi-motor systems in electric agricultural machinery under unknown dynamic environments. Summary of the Invention

[0005] The technical problem to be solved by this invention is how to overcome the technical defects of existing electric agricultural machinery systems, such as safety risks, low data sampling efficiency, and unacceptable training costs. In order to overcome the above defects of the prior art, this invention provides a motor intelligent control method based on digital twins and deep reinforcement learning.

[0006] This invention provides a method for intelligent motor control based on digital twins and deep reinforcement learning, comprising the following steps: S1: Construct a multi-physics digital twin of an electric agricultural machine, and define a deep reinforcement learning training environment based on the multi-physics digital twin; in, The multiphysics digital twin includes the multibody dynamics model of the electric agricultural machine, the electromagnetic model of the motor contained in the electric agricultural machine, and the soil-implementation interaction model corresponding to the electric agricultural machine. The deep reinforcement learning training environment includes the state, actions, and reward function of the electric agricultural machinery; The state includes motor state, pose, and environmental parameters; The action is a torque or current command from the motor; The reward function is a multi-objective function that integrates energy consumption, work quality, and equipment safety. S2: In the deep reinforcement learning training environment, a deep reinforcement learning algorithm is used to train a policy network so that the policy network outputs a preliminary control action according to the state of the electric agricultural machinery. S3: The initial control action output by the policy network is used as the reference trajectory of the model predictive controller. The model predictive controller solves the constrained finite-time optimization problem to generate the control command of the electric agricultural machine. S4: When the electric agricultural machinery is running according to the control command, the model parameters of the multiphysics digital twin are calibrated according to the running data, and the model parameters of the strategy network are fine-tuned. S5: Adjust the constraint set of the model predictive controller and / or the weight of the reward function based on the real-time motor condition monitoring results.

[0007] This invention discloses a motor intelligent control method based on digital twins and deep reinforcement learning. Addressing the technical problems of this invention, it defines a deep reinforcement learning training environment based on a constructed multi-physics digital twin, making the training environment dependent on the multi-physics digital twin. After training the policy network, the initial control actions output by the policy network are used as the reference trajectory for the model predictive controller. The model predictive controller solves a constrained finite-time optimization problem, ultimately generating control commands for the electric agricultural machinery. The deep reinforcement learning training environment includes the states of each motor, vehicle pose, and environmental parameters; the actions are the torque or current commands of each motor; and the reward function is a multi-objective function integrating energy consumption, operational quality, and equipment safety. This process creatively combines the long-term global adaptive optimization capability of deep reinforcement learning with the short-term precise constraint handling capability of model predictive control. Deep reinforcement learning acts as an "intelligent navigator," providing optimization direction at an acceptable training cost, while the model predictive controller ensures safe execution, achieving excellent control performance that simultaneously balances safety and cost control. Furthermore, when the electric agricultural machinery operates according to the control commands, a real-time data closed loop from the physical entity to the digital virtual entity is established through steps S4 and S5. This ensures easy data acquisition and enables the control system to continuously learn and self-optimize from actual operating experience, further improving system safety, increasing work efficiency, and adapting to constantly changing environments and tasks. Simultaneously, this offline training deep reinforcement learning strategy reduces the online computational burden, and lightweight model predictive control ensures real-time performance, enhancing the feasibility of deploying the entire system on actual agricultural machinery.

[0008] In one possible implementation, the mathematical representation of the state is: , In the formula, For state, This is the motor speed vector. This is the motor torque vector. The motor current vector, The power supply is in a charged state. For speed, For location, For heading angle, For roll angle, The pitch angle; This state vector takes into account the electrode rotation angular velocity, motor torque, current, power supply capacity, speed, position, and multiple angles of the electric agricultural machinery. Combined with rewards and actions, it makes the entire deep reinforcement learning model more capable of long-term global adaptive optimization.

[0009] In one possible implementation, the electric agricultural machine is an agricultural machine that includes a left traction motor, a right traction motor, and a power output motor.

[0010] In one possible implementation, during the training process of step S2, the policy network is instructed to allocate the motor torque under different soil hardness, slope and battery conditions to maximize the reward; thereby increasing the traction torque to prevent slippage, and also adopting a more energy-efficient torque allocation strategy when the battery is low.

[0011] In one possible implementation, step S3 includes the following steps: S31: Input the status data of the electric agricultural machinery into the strategy network so that the strategy network can output the expected preliminary control action in the next few seconds; S32: A predictive model comprising a simplified kinematic model and a motor dynamic model is established through the model predictive controller; the simplified kinematic model includes an electronic differential steering formula; the motor dynamic model is constructed based on a simplified torque formula. S33: Using the model obtained in step S32 and the preliminary control action, the model predicts the controller to solve a constrained finite-time domain optimization problem, so as to obtain the optimal instantaneous torque command and / or current command from the preliminary control action. S34: The model predictive controller sends the instruction obtained in step S33 to the current loop controller of each motor. The current loop controller uses a proportional-integral control algorithm with feedforward to convert the instruction into the control instruction.

[0012] The technical solution corresponding to the above steps inputs the state data of the electric agricultural machinery into a strategy network and combines it with a model predictive controller to construct a predictive model that includes a simplified kinematic model (containing the electronic differential steering formula) and a motor dynamic model (based on the simplified torque formula). This achieves efficient prediction and optimization of future control actions. Within the finite time domain, by solving a constrained optimization problem, the optimal instantaneous torque or current command is extracted from the initial control action and precisely executed by a PI current loop controller with feedforward. This significantly improves the response speed, control accuracy, and energy efficiency of the electric agricultural machinery under complex operating conditions. Simultaneously, the introduction of the electronic differential steering model enhances the flexibility and stability of steering control, avoiding the wear and energy consumption associated with traditional mechanical differentials. Overall, this achieves a safer, smarter, and more energy-efficient automatic control effect for agricultural machinery, suitable for the urgent needs of precision agriculture for high dynamic and high-reliability control.

[0013] In one possible implementation, the electronic differential steering formula is as follows: , , In the formula, Let be the desired angular velocity of the left traction motor of the electric agricultural machinery. Let be the desired angular velocity of the right traction motor of the electric agricultural machinery. For the target speed, The turning radius is This refers to the track gauge.

[0014] In one possible implementation, the simplified torque formula is as follows: , In the formula, Where is the motor torque, and p is the number of motor pole pairs. It is a permanent magnet flux linkage. , For direct-axis and quadrature-axis inductors, , This represents the corresponding stator current component.

[0015] In one possible implementation, the constrained finite-time optimization problem is formulated as: minimizing a cost function of the form: [equation missing - likely a specific function or requirement] while satisfying given constraints. , in, The cost function value, The torque of the power output motor in the initial control action is [missing information]. This is the vector of motor torque changes. Let be the angular velocity of the left traction motor of the electric agricultural machinery. Let be the angular velocity of the right traction motor of the electric agricultural machinery. The torque of the electric motor that outputs power to the electric agricultural machine is denoted as .

[0016] It solves the above-mentioned constrained optimization problems (such as motor torque limit and battery power limit) in its prediction time domain, and its cost function is to track the reference trajectory and minimize the control change rate, and can output accurate instantaneous control commands that satisfy all safety constraints to each motor driver.

[0017] In one possible implementation, in step S34, for the traction motor, the operating formula for its corresponding current loop controller is: , ; For a power output motor, the operating formula for its corresponding current loop controller is: ; In the formula, This is the reference value for the q-axis current. These are the q-axis current observations. For the desired rotational speed, For the actual measured speed, and For proportional-integral controller parameters, For the estimated load torque, Let J be the motor torque constant, J be the moment of inertia, and B be the coefficient of viscous friction. To control the period, k represents the current time. For time sequence, Based on the base torque, The adaptive gain is negative. To estimate the load, This is the preset load threshold.

[0018] when Exceed At that time, the algorithm actively reduces To coordinate traction and workload, and prevent system stalling.

[0019] In one possible implementation, in step S5, the power output motor is allowed to start or maintain operation only if the following condition is met: , In the formula, For the power output motor to be allowed to start or maintain operating conditions, , This refers to the actual speed of the left and right traction motors. The actual vehicle speed measured by satellite navigation. and The preset minimum safety threshold, This is an AND operation. The safety interlock condition enabled by this power output motor further ensures the safety of starting or maintaining the operation of the power output motor, thus enhancing the overall system safety. Attached Figure Description

[0020] Figure 1 This is a flowchart of a motor intelligent control method based on digital twin and deep reinforcement learning disclosed in an embodiment of this application; Figure 2 This is a diagram of the three-layer intelligent control system architecture disclosed in the embodiments of this application; Figure 3 This is a side view of the rotary tiller disclosed in the embodiments of this application. Detailed Implementation

[0021] First, those skilled in the art should understand that these embodiments are merely used to explain the technical principles of the embodiments of this application and are not intended to limit the scope of protection of the embodiments of this application. Those skilled in the art can make adjustments as needed to adapt to specific application scenarios.

[0022] In the description of the embodiments of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "electrical connection" and "establishing an electrical connection relationship" should be interpreted broadly, that is, it should be understood that two or more parties have an electrical relationship, which can be achieved through a wire, a radio connection, or a combination of both; it can be a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in the embodiments of this application based on the specific circumstances.

[0023] The present application will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0024] See Figure 1 , Figure 2 and Figure 3 This application discloses a method for intelligent motor control based on digital twins and deep reinforcement learning. Figure 1 This is a flowchart illustrating the method. This embodiment uses a dual-track independently driven electric rotary tiller as the electric agricultural machinery. Figure 2 and Figure 3 As shown, in this embodiment, the walking system of the electric rotary tiller is driven independently by two 18kW permanent magnet synchronous motors (PMSMs) for the left and right tracks, respectively. The rotary tillage system is directly driven by a 22kW PMSM for the rotary tiller shaft. The entire machine is powered by a 96V / 150Ah lithium iron phosphate battery pack, which supplies power to each motor via a DC / AC inverter. Figure 2 As shown.

[0025] See Figure 1 The method includes the following steps: S1: Construct a multiphysics digital twin of the electric agricultural machine and define a deep reinforcement learning training environment based on the multiphysics digital twin. In this embodiment, the multiphysics digital twin includes a multibody dynamics model of the electric agricultural machine, an electromagnetic model of the motor contained in the electric agricultural machine, and a soil-implant interaction model corresponding to the electric agricultural machine. This embodiment utilizes multibody dynamics software, a motor simulation library, and a soil mechanics model to construct the digital twin of the electric rotary tiller. The dynamics model can simulate the movement of the vehicle on slopes and rugged terrain; the motor model uses the dq-axis equation to accurately simulate its electromagnetic and thermal characteristics; the soil-rotary tillage blade model is used to generate realistic nonlinear operating resistance.

[0026] Based on a multiphysics digital twin, the deep reinforcement learning training environment defined in this embodiment includes the state, actions, and reward function of the electric agricultural machinery; the state includes the motor state, pose, and environmental parameters, which are mathematically expressed as follows: , In the formula, For state, This is the motor speed vector. This is the motor torque vector. The motor current vector, The power supply is in a charged state. For speed, For location, For heading angle, For roll angle, The pitch angle is given. Since this embodiment uses an electric rotary tiller as the electric agricultural machinery, the mathematical representation of the state is as follows: ,in, The left motor speed, This refers to the speed of the right motor. The speed of the power output motor (PTO) The torque of the left motor. The torque of the right motor. To provide power output motor torque, The current of the left motor is... This is the current of the right motor. For power output motor current, The x and y coordinates of the corresponding positions.

[0027] Correspondingly, the action is a torque or current command for the motor. In this embodiment, the action space consists of torque commands for three motors: that is, The reward function is a multi-objective function that considers energy consumption, operational quality, and equipment safety. In this embodiment, the reward function is expressed as follows: The reward function is designed as follows: ,in This represents the area of ​​work completed within the simulation step. Total power consumption, For the progress of the work, For path tracking error, Penalty for exceeding temperature limits, ~ The weights are adjustable.

[0028] S2: In the deep reinforcement learning training environment, a deep reinforcement learning algorithm is used to train the policy network, enabling the network to output initial control actions based on the state of the electric agricultural machinery. In the digital twin environment, the SAC (Soft Actor-Critic) algorithm is used for offline training. During training, the policy network... The simulation allocates torque to the three motors under varying soil hardness, slope, and battery conditions to maximize rewards. For example, it learns to appropriately reduce PTO torque and increase traction torque to prevent slippage in hard soil, and to adopt a more energy-efficient torque allocation strategy when the battery is low. After training, the parameters of the policy network are stored and downloaded to the onboard VCU.

[0029] S3: The initial control action output by the policy network is used as the reference trajectory of the model predictive controller. The model predictive controller solves the constrained finite-time optimization problem to generate control commands for the electric agricultural machinery.

[0030] In this embodiment, step S3 includes the following steps: S31: Input the status data of the electric agricultural machinery into the strategy network so that the strategy network can output the expected initial control action in the next few seconds.

[0031] S32: A predictive model incorporating a simplified kinematic model and a motor dynamic model is established using a model predictive controller. The simplified kinematic model includes the electronic differential steering formula, which is as follows: , , In the formula, Let be the desired angular velocity of the left motor of the electric rotary tiller. Let be the desired angular velocity of the right motor of the electric rotary tiller. For the target speed, The turning radius is This refers to the track gauge.

[0032] The motor dynamic model is constructed based on the simplified torque formula; the simplified torque formula is as follows: , In the formula, Where is the motor torque, and p is the number of motor pole pairs. It is a permanent magnet flux linkage. , For direct-axis and quadrature-axis inductors, , These represent the corresponding stator current components. Since this embodiment uses a surface-mount motor, the model is simplified to... .

[0033] S33: Using the model obtained in step S32 and the preliminary control actions obtained in step S31, solve the constrained finite-time optimization problem through model prediction controller to obtain the optimal instantaneous torque command and / or current command from the preliminary control actions. The constrained finite-time optimization problem is formulated as: minimizing a cost function of the following form under given constraints: , in, The cost function value, To initially control the torque of the power output motor during the action, This is the vector of motor torque changes. Let be the angular velocity of the left traction motor of the electric agricultural machinery. Let be the angular velocity of the right traction motor of the electric agricultural machinery. The torque of the electric motor that outputs power to the electric agricultural machine is given. The constraint condition in this embodiment is designed as the range of torque values.

[0034] S34: The model predictive controller sends the instructions obtained in step S33 to the current loop controllers of each motor. The current loop controllers use a proportional-integral control algorithm with feedforward to convert the instructions into the control instructions. In step S34, for the traction motor, the operating formula of its corresponding current loop controller is: , ; For a power output motor, the operating formula for its corresponding current loop controller is: ; In the formula, This is the reference value for the q-axis current. These are the q-axis current observations. For the desired rotational speed, For the actual measured speed, and For proportional-integral controller parameters, For the estimated load torque, Let J be the motor torque constant, J be the moment of inertia, and B be the coefficient of viscous friction. To control the period, k represents the current time. For time sequence, Based on the base torque, The adaptive gain is negative. To estimate the load, This is the preset load threshold.

[0035] S4: When the electric agricultural machinery operates according to control commands, the model parameters of the multiphysics digital twin are calibrated based on the operational data, and the model parameters of the policy network are fine-tuned. Specifically, during the operation of the electric rotary tiller, all operational data is uploaded to the cloud via a 4G / 5G network. The cloud system uses this data, such as actual motor response and energy consumption, to automatically calibrate the motor efficiency MAP and soil resistance parameters in the digital twin. Simultaneously, new (state, action, reward) data is added to the playback buffer, and the policy network is periodically updated. Fine-tuning training is performed to better adapt the strategy to this specific field and the current state of the agricultural machinery. After the updated strategy parameters are tested and verified, they can be updated to the vehicle's VCU via OTA (over-the-air).

[0036] S5: Adjust the weights of the constraint set and / or reward function of the model predictive controller based on real-time motor condition monitoring results. Specifically, this involves continuously monitoring the motor condition. For example, if the right traction motor temperature sensor indicates overheating, the diagnostic module will trigger fault-tolerant logic. At this time, the constraints in the model predictive controller optimizer will immediately be adjusted. By temporarily lowering the torque level, the model predictive controller will automatically allocate more torque to the left motor while still meeting the vehicle's basic driving needs. Simultaneously, this event can serve as negative feedback, modifying the weight of the temperature penalty term in the cloud-based deep reinforcement learning training reward function to enable the next-generation strategy to proactively prevent the right motor from overheating under similar operating conditions.

[0037] To ensure safe operation, in step S5, the power output motor is only allowed to start or continue operating if the following conditions are met: , In the formula, For the power output motor to be allowed to start or maintain operating conditions, , This refers to the actual speed of the left and right traction motors. The actual vehicle speed measured by satellite navigation. and The preset minimum safety threshold, The AND operation.

[0038] The intelligent motor control method based on digital twins and deep reinforcement learning disclosed in this embodiment defines a deep reinforcement learning training environment on the basis of a constructed multi-physics digital twin. This training environment is dependent on the multi-physics digital twin. After training the policy network, the initial control actions output by the policy network are used as the reference trajectory for the model predictive controller. The model predictive controller solves a constrained finite-time optimization problem, ultimately generating control commands for the electric agricultural machinery. The deep reinforcement learning training environment includes the states of each motor, vehicle pose, and environmental parameters; the actions are the torque or current commands of each motor; and the reward function is a multi-objective function that integrates energy consumption, operational quality, and equipment safety. This process creatively combines the long-term global adaptive optimization capability of deep reinforcement learning with the short-term precise constraint handling capability of model predictive control. Deep reinforcement learning acts as an "intelligent navigator," providing optimization direction at an acceptable training cost, while the model predictive controller ensures safe execution, achieving excellent control performance that simultaneously balances safety and cost control. Furthermore, when the electric agricultural machinery operates according to the control commands, a real-time data closed loop from the physical entity to the digital virtual entity is established through steps S4 and S5. This ensures easy data acquisition and enables the control system to continuously learn and self-optimize from actual operating experience, further improving system safety, increasing work efficiency, and adapting to constantly changing environments and tasks. Simultaneously, this offline training deep reinforcement learning strategy reduces the online computational burden, and lightweight model predictive control ensures real-time performance, enhancing the feasibility of deploying the entire system on actual agricultural machinery.

[0039] In the description of the embodiments of this application, it should be noted that the terms "inner" and "outer" and other terms indicating direction or positional relationship are based on the direction or positional relationship shown in the drawings. This is only for the convenience of description and does not indicate or imply that the device or component must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this application.

[0040] In the description of this application, the references to terms such as "an embodiment," "some embodiments," "in this embodiment," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

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

Claims

1. A method for intelligent motor control based on digital twins and deep reinforcement learning, characterized in that, Includes the following steps: S1: Construct a multi-physics digital twin of an electric agricultural machine, and define a deep reinforcement learning training environment based on the multi-physics digital twin; in, The multiphysics digital twin includes the multibody dynamics model of the electric agricultural machine, the electromagnetic model of the motor contained in the electric agricultural machine, and the soil-implementation interaction model corresponding to the electric agricultural machine. The deep reinforcement learning training environment includes the state, actions, and reward function of the electric agricultural machinery; The state includes motor state, pose, and environmental parameters; The action is a torque or current command from the motor; The reward function is a multi-objective function that integrates energy consumption, work quality, and equipment safety. S2: In the deep reinforcement learning training environment, a deep reinforcement learning algorithm is used to train a policy network so that the policy network outputs a preliminary control action according to the state of the electric agricultural machinery. S3: The initial control action output by the policy network is used as the reference trajectory of the model predictive controller. The model predictive controller solves the constrained finite-time optimization problem to generate the control command of the electric agricultural machine. S4: When the electric agricultural machinery is running according to the control command, the model parameters of the multiphysics digital twin are calibrated according to the running data, and the model parameters of the strategy network are fine-tuned. S5: Adjust the constraint set of the model predictive controller and / or the weight of the reward function based on the real-time motor condition monitoring results.

2. The intelligent motor control method based on digital twin and deep reinforcement learning according to claim 1, characterized in that, The mathematical expression for the state is: ， In the formula, For state, This is the motor speed vector. This is the motor torque vector. The motor current vector, The power supply is in a charged state. For speed, For location, For heading angle, For roll angle, The pitch angle.

3. The intelligent motor control method based on digital twin and deep reinforcement learning according to claim 2, characterized in that, The electric agricultural machinery refers to agricultural machinery that includes a left traction motor, a right traction motor, and a power output motor.

4. The intelligent motor control method based on digital twin and deep reinforcement learning according to claim 3, characterized in that, During the training process in step S2, the policy network is instructed to allocate the motor torque under different soil hardness, slope, and battery conditions to maximize the reward.

5. The intelligent motor control method based on digital twins and deep reinforcement learning according to claim 3 or 4, characterized in that, Step S3 includes the following steps: S31: Input the status data of the electric agricultural machinery into the strategy network so that the strategy network can output the expected preliminary control action in the next few seconds; S32: A predictive model comprising a simplified kinematic model and a motor dynamic model is established through the model predictive controller; the simplified kinematic model includes an electronic differential steering formula; the motor dynamic model is constructed based on a simplified torque formula. S33: Using the model obtained in step S32 and the preliminary control action, the model predicts the controller to solve a constrained finite-time domain optimization problem, so as to obtain the optimal instantaneous torque command and / or current command from the preliminary control action. S34: The model predictive controller sends the instruction obtained in step S33 to the current loop controller of each motor. The current loop controller uses a proportional-integral control algorithm with feedforward to convert the instruction into the control instruction.

6. The intelligent motor control method based on digital twin and deep reinforcement learning according to claim 5, characterized in that, The electronic differential steering formula is as follows: ， ， In the formula, Let be the desired angular velocity of the left traction motor of the electric agricultural machinery. Let be the desired angular velocity of the right traction motor of the electric agricultural machinery. For the target speed, The turning radius is This refers to the track gauge.

7. The intelligent motor control method based on digital twin and deep reinforcement learning according to claim 6, characterized in that, The simplified torque formula is as follows: ， In the formula, Where is the motor torque, and p is the number of motor pole pairs. It is a permanent magnet flux linkage. , For direct-axis and quadrature-axis inductors, , This represents the corresponding stator current component.

8. The intelligent motor control method based on digital twins and deep reinforcement learning according to claim 6 or 7, characterized in that, The constrained finite-time optimization problem is formulated as follows: minimizing a cost function of the form: [The original text contains several inconsistencies and unclear formatting issues. A more accurate translation would require the full context.] ， in, The cost function value, The torque of the power output motor in the initial control action is [missing information]. This is the vector of motor torque changes. Let be the angular velocity of the left traction motor of the electric agricultural machinery. Let be the angular velocity of the right traction motor of the electric agricultural machinery. The torque of the electric motor that outputs power to the electric agricultural machine is denoted as .

9. The intelligent motor control method based on digital twin and deep reinforcement learning according to claim 8, characterized in that, In step S34, for the traction motor, the operating formula for its corresponding current loop controller is: ， ； For a power output motor, the operating formula for its corresponding current loop controller is: ； In the formula, This is the reference value for the q-axis current. These are the q-axis current observations. For the desired rotational speed, For the actual measured speed, and For proportional-integral controller parameters, For the estimated load torque, Let J be the motor torque constant, J be the moment of inertia, and B be the coefficient of viscous friction. To control the period, k represents the current time. For time sequence, Based on the base torque, The adaptive gain is negative. To estimate the load, This is the preset load threshold.

10. The intelligent motor control method based on digital twin and deep reinforcement learning according to claim 8, characterized in that, In step S5, the power output motor is allowed to start or continue operating only if the following conditions are met: ， In the formula, For the power output motor to be allowed to start or maintain operating conditions, , This refers to the actual speed of the left and right traction motors. The actual vehicle speed measured by satellite navigation. and The preset minimum safety threshold, The AND operation.

Images