Tractor double-model driving predictive control method and system based on cultivation scene constraint self-adaption

Abstract

The invention discloses a cultivated land scene constraint self-adaptive tractor double-model driving prediction control method and system, and the method comprises the steps: obtaining the change characteristics of a front wheel rotation angle and cultivation resistance from unit operation data through a deep neural network, and building a controlled quantity unmodeled error data driving model, the output correction model of the data driving model is used for predictively controlling the control quantity solved through rolling optimization, and the corrected steering angle decision quantity and the corrected acceleration decision quantity at the k + 1 moment are used for steering execution and speed adjustment of the tractor. According to the method, the path tracking control stability and scene adaptability of the tractor can be improved.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent agricultural machinery, and in particular relates to a tractor dual-model drive predictive control method with adaptive farmland scene constraints. Background Art

[0002] Path-tracking control is a key technology for autonomous tractor driving. Model Predictive Control (MPC), capable of real-time prediction and optimal control under control input, state, and output constraints, is a commonly used method for high-precision path-tracking control in complex scenarios. In actual tractor farming scenarios, changes in plowing depth cause continuous variations in load and yaw rate, which in turn affects path-tracking stability. Environmental disturbances and changing constraints affect MPC decision-making accuracy, limiting its adaptability to path-tracking control in farming environments. Current model-driven MPC is particularly limited in addressing the continuous variations in load and acceleration caused by changes in plowing depth. Due to the inability to accurately model the system's predictive model, current models have limitations. With the increasing application of unmanned tractors throughout the agricultural production of major grain and oil crops, improving path-tracking control stability and scenario adaptability has become an urgent issue. Summary of the Invention

[0003] In view of the shortcomings of the existing technology, the present invention proposes a tractor dual-model drive predictive control method with adaptive farming scene constraints.

[0004] The present invention achieves the above technical objectives through the following technical means.

[0005] A dual-model driven predictive control method for tractors with adaptive farming scenario constraints:

[0006] A deep neural network is used to obtain the changing characteristics of the front wheel angle and tillage resistance from the unit operation data, and a data-driven model of the unmodeled error of the control variable is established. The output correction model of the data-driven model is used to predict the control variable solved by the rolling optimization. The corrected steering angle decision quantity and acceleration decision quantity at the k+1 moment are used for the steering execution and speed regulation of the tractor.

[0007] Furthermore, the deep neural network is used to obtain the changing characteristics of the front wheel angle and tillage resistance from the unit operation data, and a data-driven model of the unmodeled error of the control quantity is established, specifically:

[0008] Using distributed measurement system to obtain the lateral acceleration of the tillage unit at time k and longitudinal acceleration Tractor traction force F, tractor front wheel angular rate r, tractor heading angle change rate The tractor pitch angle φ and the plow roll angle γ are used to obtain the steering angle decision quantity and acceleration decision quantity of the tractor at time k using the MPC model;

[0009] The tractor front wheel angular rate r and the tractor heading angle change rate at time k are The tractor pitch angle φ and the plow roll angle γ are used as the model inputs, and the angle measurement value of the contact angle sensor at time k is used as the model inputs. Decision-making quantity of tractor steering angle difference To achieve the desired output, train a data-driven model for steering angles.

[0010] The lateral acceleration of the tillage unit at time k is The tractor traction force F, tractor pitch angle φ, and plow roll angle γ are the model inputs, which are composed of the tractor traction force F, the longitudinal acceleration of the tillage unit, and the plow roll angle γ. Corresponding stress F x Calculate the longitudinal resistance force F d , with the longitudinal resistance force F d Corresponding acceleration and the tractor acceleration decision difference To achieve the desired output, train a data-driven model for acceleration;

[0011] Repeat the above steps N times to train the data-driven model of steering angle and acceleration in the farming scenario.

[0012] Furthermore, the control quantity solved by the rolling optimization is predicted by the output correction model of the data-driven model, specifically:

[0013] Collect the tractor front wheel angular rate r and tractor heading angle change rate at time k Tractor pitch angle φ, plow roll angle γ, tractor traction force F, lateral acceleration Longitudinal acceleration An MPC model is established based on the tractor yaw dynamics model to perform rolling optimization solution for the steering angle control variable at time k.

[0014] Collect the lateral acceleration of the tillage unit at time k The tractor traction force F, tractor pitch angle φ, and plow roll angle γ are used to establish an MPC model based on the tractor longitudinal dynamics model, and perform rolling optimization solution for the acceleration control variable at time k;

[0015] The data-driven model of steering angle in farming scenarios is used to predict the steering angle deviation Δδ at time k online. k , and feeds back to the steering angle decision value of the tractor at time k, and obtains the steering angle prediction value of the steering angle sensor at time k+1 Subtract the steering angle measurement value of the steering angle sensor at time k+1 to obtain the residual error of the tractor steering angle decision at time k+1 It is used to correct the tractor steering angle decision-making quantity at time k+1, that is,

[0016] The acceleration change Δa at time k is predicted online using the data-driven model of acceleration in farming scenarios. k , and feeds back to the acceleration decision of the tractor at time k, and obtains the predicted value of the longitudinal resistance force corresponding to the acceleration at time k+1 The acceleration corresponding to the longitudinal resistance force at time k+1 Subtract and get the residual of the tractor acceleration decision quantity at time k+1 It is used to correct the tractor acceleration decision quantity at time k+1, that is,

[0017] Update k=k+1 and repeat the above process to realize the tractor's autonomous plowing path tracking control.

[0018] Furthermore, the tractor yaw dynamics model is:

[0019]

[0020] Where δ is the steering angle of the tractor front wheels, I z is the moment of inertia of the tractor about the vertical axis passing through its center of mass, l f 、l r are the distances between the front and rear axles and the center of mass of the tractor, F xf is the longitudinal force of the tractor front tire, F yf 、F yr are the lateral forces of the front and rear tires of the tractor, m is the total mass of the tillage unit, v x and v y are the longitudinal and lateral velocity components of the tractor, respectively, and ψ is the heading angle of the tractor.

[0021] Furthermore, the longitudinal dynamics model of the tractor is:

[0022]

[0023] Among them, F d =ma d F is the resultant force of the longitudinal resistance of the vehicle, xr is the longitudinal force of the tractor rear tire, a d is the acceleration caused by resistance.

[0024] Furthermore, the deep neural network is a long short-term memory network.

[0025] A tractor dual-model drive predictive control system with adaptive farming scenario constraints, comprising:

[0026] Contact-type rotation angle sensor, installed at the front wheel shaft of the tractor;

[0027] Two single-axis gyroscopes, mounted on the rotating parts of the tractor's left and right front wheels;

[0028] Two MEMS-IMUs, one mounted at the center of the tractor body coordinate system and the other fixed to the plow;

[0029] Dual-antenna RTK receiver, mounted on the tractor;

[0030] The stress sensor is installed at the three-point hitch where the tractor is connected to the plow;

[0031] Signal acquisition and processing module, which collects and synchronizes the output information of MEMS-IMU, contact angle sensor, single-axis gyroscope, dual-antenna RTK receiver, and stress sensor;

[0032] DSP, fuses the synchronized information to obtain the position and position change of the tractor and plow;

[0033] The microcontroller receives the fused information and extracts dynamic features based on the pre-processed inertial information, front wheel angle information, and the deep neural network inference module;

[0034] The on-board computer integrates the dynamic characteristics as soft constraints into the model predictive control constraint equation to realize the tractor model predictive path tracking control with self-adaptation to the tillage scene constraints.

[0035] In the above technical solution, the deep neural network inference module automatically extracts the dynamic change characteristics of the front wheel angle and tillage resistance based on the collected operation data to characterize the dynamic behavior of the tractor during operation.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] 1) A deep neural network is used to extract the changing characteristics of the front wheel steering angle deviation and tillage resistance from the unit operation data, and the front wheel steering angle deviation and tillage resistance compensation control variables solved by MPC rolling optimization are predicted and output online to correct the control decision errors caused by the unmodeled and time-varying parameters of the tractor yaw dynamics model, longitudinal dynamics model and steering execution control system.

[0038] 2) Through multi-sensor data fusion and deep neural network data-driven features, dual-model drive is adopted to match the control decision quantity with the continuous load fluctuation caused by changes in plowing depth, thereby improving the tractor driving control stability in complex plowing scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Schematic diagram of the tractor dual-model drive predictive control system with self-adaptive farming scenario constraints according to the present invention;

[0040] Figure 2 This is a diagram showing the principle of training the data-driven predictive control model of the present invention;

[0041] Figure 3 This is the dual-model driven predictive control flow chart of the present invention. DETAILED DESCRIPTION

[0042] In order to fully and deeply understand the embodiments of the present invention and clearly illustrate the technical solutions of the present invention, the following description will be made with reference to the accompanying drawings. The embodiment mentioned here is only one of many possible embodiments of the present invention and does not include all embodiments. Based on the content disclosed in this specification, ordinary technicians in this field can derive other possible embodiments without involving creative work, and these should be included in the scope of protection of the present invention.

[0043] like Figure 1 As shown, the present invention discloses a tractor dual-model drive prediction control system with self-adaptive farming scene constraints, including a contact angle sensor, two single-axis gyroscopes, two MEMS-IMUs, a dual-antenna RTK receiver, a stress sensor, a power management module, a signal acquisition and processing module, a microcontroller, a deep neural network inference module, a wireless transmission module, a DSP and an on-board computer, wherein the microcontroller adopts an STM32 controller.

[0044] Among them, the contact angle sensor is installed at the front wheel shaft of the tractor to collect the steering angle signal of the tractor's front wheel; two single-axis gyroscopes are respectively installed on the rotating parts of the left and right front wheels of the tractor to collect the angular rates of the left and right front wheels respectively; two MEMS-IMUs, one is installed at the center of the tractor body coordinate system to collect the tractor's lateral and longitudinal accelerations and attitude angles, and the other is fixed to the plow to collect the plow's lateral and longitudinal accelerations and attitude angles; a dual-antenna RTK receiver is installed on the tractor to output the tractor's position, heading and speed; specifically, the output of the tractor MEMS-IMU and the angular rate measurement of the gyroscope and the output of the dual-antenna RTK receiver are sent to the DSP through the signal acquisition and processing module; the output of the plow MEMS-IMU and the output of the dual-antenna RTK receiver are sent to the DSP through the signal acquisition and processing module; the dual-antenna RTK receiver is installed on the tractor; the stress sensor is installed at the three-point suspension connecting the tractor and the plow, and its output is sent to the DSP through the signal acquisition and processing module.

[0045] The signal acquisition and processing module collects and synchronizes the output information of the tractor MEMS-IMU, angle sensor, gyroscope, dual-antenna RTK receiver, stress sensor and plow MEMS-IMU.

[0046] The DSP fuses the synchronized information to obtain the tractor / plow's position and attitude and its change. Specifically, the tractor / plow's position includes its coordinates, and the tractor / plow's attitude includes its speed and heading angle.

[0047] The DSP sends the fused information to the STM32 controller. The STM32 controller extracts the dynamic features based on the pre-processed inertial information, front wheel angle information and deep neural network inference module, and sends the calculation results to the on-board computer via the wireless transmission module.

[0048] The deep neural network inference module automatically extracts the dynamic change characteristics of the front wheel angle and tillage resistance based on the collected operation data to characterize the dynamic behavior pattern during the operation process.

[0049] The on-board computer receives the dynamic features sent by the microcontroller, and further integrates the dynamic features into the model predictive control constraint equation as soft constraints to realize the tractor model predictive path tracking control with adaptive farming scenario constraints, and then sends it to the local or remote terminal via the wireless transmission module.

[0050] The power management module is powered by the tractor battery, and then through the power regulator module to power the dual-antenna RTK receiver, MEMS-IMU, gyroscope, stress sensor, signal acquisition and processing module and DSP.

[0051] The present invention also proposes a dual-model driven predictive control method for a tractor with adaptive tillage scenario constraints. It uses a deep neural network to obtain the changing characteristics of the front wheel angle and tillage resistance from the unit operation data, establishes a data-driven model for the unmodeled error of the control quantity, and uses the output of the data-driven model to correct the control quantity solved by MPC rolling optimization, thereby improving the adaptability of the control decision quantity to the tillage scenario.

[0052] Furthermore, if Figure 2 As shown, the method uses a deep neural network to obtain the changing characteristics of the front wheel angle and tillage resistance from the unit operation data, and establishes a data-driven model for the unmodeled error of the control quantity. The specific steps are:

[0053] A1, using a distributed measurement system to obtain the lateral acceleration of the tillage unit at time k and longitudinal acceleration Tractor traction force F, tractor front wheel angular rate r, tractor heading angle change rate The tractor pitch angle φ and the plow roll angle γ are used to obtain the steering angle decision quantity and acceleration decision quantity of the tractor at time k using the MPC model;

[0054] A2, the tractor front wheel angular rate r and the tractor heading angle change rate at time k The tractor pitch angle φ and the plow roll angle γ are used as the model inputs, and the angle measurement value of the contact angle sensor at time k is used as the model inputs. Decision-making quantity of tractor steering angle difference To achieve the desired output, train a data-driven model for steering angles.

[0055] A3, the lateral acceleration of the tillage unit at time k The tractor traction force F, tractor pitch angle φ, and plow roll angle γ are the model inputs, which are calculated by the tractor traction force F and the longitudinal acceleration of the tillage unit. Corresponding stress F x Calculate the longitudinal resistance force F d , with the longitudinal resistance force F d Corresponding acceleration and the tractor acceleration decision difference To achieve the desired output, train a data-driven model for acceleration;

[0056] A4. Repeat the above steps N times to train a data-driven model for the steering angle and acceleration in the farming scenario. The data-driven model here is the unmodeled error.

[0057] Furthermore, if Figure 3 As shown, the control quantity solved by MPC rolling optimization is corrected by the output of the data-driven model. The specific steps are:

[0058] B1, collect the tractor front wheel angular rate r and tractor heading angle change rate at time k Tractor pitch angle φ, plow roll angle γ, tractor traction force F, lateral acceleration Longitudinal acceleration An MPC model is established based on the tractor yaw dynamics model to perform rolling optimization solution for the steering angle control variable at time k.

[0059] B2, collect the lateral acceleration of the tillage unit at time k The tractor traction force F, tractor pitch angle φ, and plow roll angle γ are used to establish an MPC model based on the tractor longitudinal dynamics model, and perform rolling optimization solution for the acceleration control variable at time k;

[0060] B3, using the data-driven model of steering angle trained in A2 to predict the steering angle deviation Δδ at time k online k, which is fed back to the steering angle decision value of the tractor at time k, and the steering angle prediction value of the steering angle sensor at time k+1 is obtained Subtracting the steering angle measurement value of the steering angle sensor at time k+1, the residual error of the tractor steering angle decision value at time k+1 can be obtained It is used to correct the tractor steering angle decision-making quantity at time k+1, that is,

[0061] B4, using the acceleration data-driven model trained in A3 to online predict the acceleration change Δa at time k k , which is fed back to the acceleration decision of the tractor at time k, and the longitudinal resistance force F at time k+1 is obtained d Corresponding acceleration prediction value The resultant longitudinal resistance force F at time k+1 d Corresponding acceleration Subtracting them, we can get the residual of the tractor acceleration decision quantity at time k+1 It is used to correct the tractor acceleration decision quantity at time k+1, that is,

[0062] B5, use the corrected steering angle and acceleration decision quantities at time k+1 for tractor steering execution and speed adjustment; update k=k+1, repeat B1-B4, and realize autonomous plowing path tracking control of the tractor.

[0063] Furthermore, the tractor yaw dynamics model is:

[0064]

[0065] Where δ is the steering angle of the tractor front wheels, I z is the moment of inertia of the tractor about the vertical axis passing through its center of mass; l f 、l r are the distances between the front and rear wheel axles and the center of mass of the tractor; F xf is the longitudinal force of the tractor front tire; F yf 、F yr are the lateral forces of the front and rear tires of the tractor; m is the total mass of the tillage unit; v x 、v y are the longitudinal and lateral velocity components of the tractor respectively; ψ is the heading angle of the tractor.

[0066] Furthermore, the rolling optimization solution of the steering angle control amount at time k is realized by constructing the following objective function: the yaw dynamics state variable at time k is defined as The control input is the front wheel steering angle u k =δ, let the sampling period be T s , transform the yaw dynamics model into a discrete state space form:

[0067] x k+1 =A d x k +B d u k +C d,k

[0068] in, Define the yaw dynamics prediction state sequence X = [x k+1|k ,x k+2|k ,…,x k+N|k ] T , control sequence U=[u k|k ,u k+1|k ,…,u k+N-1|k ] T , we can get the prediction equation:

[0069] X=Fx k +GU+C d,k

[0070] in, G is a lower triangular matrix, i≥j; construct the minimum v y 、 Tracking error and cost function for minimizing steering angle control amount δ:

[0071]

[0072] Among them, X ref is the expected state sequence, I N is an N×N dimensional unit matrix, is the Kronecker product, where Q is a 2×2 diagonal matrix, R is a scalar, and S is a scalar; Q, R, and S are the equilibrium v y 、 Weight coefficients of tracking performance, control input δ, and control input change Δδ; is the term used to penalize the control increment Δδ, ΔU=UU k-1 , U is the control sequence in the current prediction time domain, U k-1 =[u k-1 ,u k-1 ,…,u k-1 ] T is the control input applied at time k-1.

[0073] Furthermore, the longitudinal dynamics model of the tractor is:

[0074]

[0075] Among them, F d =mad F is the resultant force of the longitudinal resistance of the vehicle, xr is the longitudinal force of the tractor rear tire, a d is the acceleration caused by resistance.

[0076] Furthermore, the rolling optimization solution of the acceleration control amount at time k is realized by constructing the following objective function: the longitudinal dynamic state variable at time k is defined as x k =v x , the control input is acceleration u k =a, let the sampling period be T s , transforming the longitudinal dynamics model into a discrete state space form:

[0077] x k+1 =A vd x k +B vd u k +C vd,k

[0078] Among them, A vd =1, B vd =T s , C vd,k =T s d k ,in is the acceleration disturbance error. Define the longitudinal dynamics prediction state sequence X v =[x k+1|k ,x k+2|k ,…,x k+N|k ] T , control sequence U v =[u k|k ,u k+1|k ,…,u k+N-1|k ] T , we can get the prediction equation:

[0079] X v =F v x k +G v U v +C vd,k

[0080] in, G v is the lower triangular matrix, i≥j; construct the minimum v y Tracking error and cost function for minimizing control amount a:

[0081]

[0082] Among them, X v,refis the expected state sequence, Here Q is a scalar, R is a scalar, and S is a scalar; Q, R, and S are the equilibrium v y Weight coefficients of tracking performance, control input a, and control input change Δa; is the term used to penalize the control increment Δa, ΔU v =U v -U v,k-1 , U v is the control sequence in the current prediction time domain, U v,k-1 =[u k-1 ,u k-1 ,…,u k-1 ] T is the control input applied at time k-1.

[0083] Furthermore, the deep neural network is a long short-term memory network.

[0084] The above description is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A dual-model driven predictive control method for a tractor with adaptive farming scenario constraints, characterized by: A deep neural network is used to obtain the changing characteristics of the front wheel angle and tillage resistance from the unit operation data, and a data-driven model of the unmodeled error of the control variable is established. The output correction model of the data-driven model is used to predict the control variable solved by the rolling optimization. The corrected steering angle decision quantity and acceleration decision quantity at the k+1 moment are used for the steering execution and speed regulation of the tractor.

2. The tractor dual-model drive predictive control method according to claim 1, characterized in that: The deep neural network is used to obtain the changing characteristics of the front wheel angle and tillage resistance from the unit operation data, and a data-driven model of the unmodeled error of the control quantity is established, specifically: Using distributed measurement system to obtain the lateral acceleration of the tillage unit at time k and longitudinal acceleration Tractor traction force F, tractor front wheel angular rate r, tractor heading angle change rate The tractor pitch angle φ and the plow roll angle γ are used to obtain the steering angle decision quantity and acceleration decision quantity of the tractor at time k using the MPC model; The tractor front wheel angular rate r and the tractor heading angle change rate at time k are The tractor pitch angle φ and the plow roll angle γ are used as the model inputs, and the angle measurement value of the contact angle sensor at time k is used as the model inputs. Decision-making quantity of tractor steering angle difference Train a data-driven model of steering angle for the desired output; The lateral acceleration of the tillage unit at time k is The tractor traction force F, tractor pitch angle φ, and plow roll angle γ are the model inputs, which are composed of the tractor traction force F, the longitudinal acceleration of the tillage unit, and the plow roll angle γ. Corresponding stress F x Calculate the longitudinal resistance force F d , with the longitudinal resistance force F d Corresponding acceleration and the tractor acceleration decision difference To achieve the desired output, train a data-driven model for acceleration; Repeat the above steps N times to train the data-driven model of steering angle and acceleration in the farming scenario.

3. The tractor dual-model drive predictive control method according to claim 2, characterized in that: The control quantity solved by the rolling optimization prediction control using the output correction model of the data-driven model is specifically: Collect the tractor front wheel angular rate r and tractor heading angle change rate at time k Tractor pitch angle φ, plow roll angle γ, tractor traction force F, lateral acceleration Longitudinal acceleration An MPC model is established based on the tractor yaw dynamics model to perform rolling optimization solution for the steering angle control variable at time k. Collect the lateral acceleration of the tillage unit at time k The tractor traction force F, tractor pitch angle φ, and plow roll angle γ are used to establish an MPC model based on the tractor longitudinal dynamics model, and perform rolling optimization solution for the acceleration control variable at time k; The data-driven model of steering angle in farming scenarios is used to predict the steering angle deviation Δδ at time k online. k , and feeds back to the steering angle decision value of the tractor at time k, and obtains the steering angle prediction value of the steering angle sensor at time k+1 Subtract the steering angle measurement value of the steering angle sensor at time k+1 to obtain the residual error of the tractor steering angle decision at time k+1 It is used to correct the tractor steering angle decision-making quantity at time k+1, that is, The acceleration change Δa at time k is predicted online using the data-driven model of acceleration in farming scenarios. k , and feeds back to the acceleration decision of the tractor at time k, and obtains the predicted value of the longitudinal resistance force corresponding to the acceleration at time k+1 The acceleration corresponding to the longitudinal resistance force at time k+1 Subtract and get the residual of the tractor acceleration decision quantity at time k+1 It is used to correct the tractor acceleration decision quantity at time k+1, that is, Update k=k+1 and repeat the above process to realize the tractor's autonomous plowing path tracking control.

4. The tractor dual-model drive predictive control method according to claim 3, characterized in that: The tractor yaw dynamics model is: Where δ is the steering angle of the tractor front wheels, I z is the moment of inertia of the tractor about the vertical axis passing through its center of mass, l f 、l r are the distances between the front and rear axles and the center of mass of the tractor, F xf is the longitudinal force of the tractor front tire, F yf 、F yr are the lateral forces of the front and rear tires of the tractor, m is the total mass of the tillage unit, v x and v y are the longitudinal and lateral velocity components of the tractor, respectively, and ψ is the heading angle of the tractor.

5. The tractor dual-model drive predictive control method according to claim 4, characterized in that: The longitudinal dynamics model of the tractor is: Among them, F d =ma d F is the resultant force of the longitudinal resistance of the vehicle, xr is the longitudinal force of the tractor rear tire, a d is the acceleration caused by resistance.

6. The tractor dual-model driving predictive control method according to claim 1, characterized in that: The deep neural network is a long short-term memory network.

7. A system based on the tractor dual-model drive predictive control method according to any one of claims 1 to 6, characterized in that: include: Contact-type rotation angle sensor, installed at the front wheel shaft of the tractor; Two single-axis gyroscopes, mounted on the rotating parts of the tractor's left and right front wheels; Two MEMS-IMUs, one mounted at the center of the tractor body coordinate system and the other fixed to the plow; Dual-antenna RTK receiver, mounted on the tractor; The stress sensor is installed at the three-point hitch where the tractor is connected to the plow; Signal acquisition and processing module, which collects and synchronizes the output information of MEMS-IMU, contact angle sensor, single-axis gyroscope, dual-antenna RTK receiver, and stress sensor; DSP, fuses the synchronized information to obtain the position and position change of the tractor and plow; The microcontroller receives the fused information and extracts dynamic features based on the pre-processed inertial information, front wheel angle information, and the deep neural network inference module; The on-board computer integrates the dynamic characteristics as soft constraints into the model predictive control constraint equation to realize the tractor model predictive path tracking control with self-adaptation to the tillage scene constraints.

8. The system according to claim 7, characterized in that The deep neural network inference module automatically extracts the dynamic change characteristics of the front wheel angle and tillage resistance based on the collected operation data to characterize the dynamic behavior of the tractor during operation.

Images