Combine harvester operation speed control system and method based on multi-operation parameter reward

Abstract

The present invention provides a combine harvester operating speed control system and method based on multiple operating parameter reward systems. The control system includes the following steps: inputting actual operating speed, number of harvested crop supervoxels, grass-to-grain ratio, plant height, stubble height, and swath width information into a Gaussian process regression model of a gravitational search algorithm to obtain a predicted feed rate; establishing a DQN neural network, inputting the predicted feed rate into the DQN neural network, and inputting the trash content, breakage rate, loss rate, and rotation speed as reward functions into the DQN neural network, which outputs a theoretical operating speed; inputting the difference between the actual operating speed and the theoretical operating speed, and the change in this difference, into a fuzzy neural network PID controller, and controlling the forward speed of the combine harvester based on the output of the fuzzy neural network PID controller. This invention improves harvesting quality and efficiency while reducing failure rates and alleviating operator skill requirements and workload.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent agricultural harvesting machinery, and in particular to a combine harvester operating speed control system and method based on multi-operation parameter reward. Background Art

[0002] With the rapid development of modern technology, combine harvesters, as a member of agricultural harvesting machinery, are continuously moving towards intelligent, automated, and multifunctional development. Combine harvesters have complex structures, harsh working environments, and large workload fluctuations. Relying solely on manual control of combine harvester speed without effective feed rate monitoring can easily lead to overload, drum blockage, and excessive loss rates, which will increase combine harvester failures and reduce operating quality. Underloaded operation can also lead to low efficiency, waste of resources, and hindered harvesting.

[0003] Promoting the intelligence and automation of combine harvesters and integrating multiple operating parameters to control operating speed can improve the harvesting quality and efficiency of harvesters while reducing the failure rate and alleviating the technical requirements and work intensity of operators. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a combine harvester operating speed control system and method based on multi-operation parameter reward. Machine vision is used to obtain crop information in advance to calculate the feeding amount. The DQN neural network is used to calculate the optimal operating speed with operating parameters (grain impurity breakage rate, grain loss rate, key part rotation speed) as the reward function and a fuzzy neural network PID controller is used for control. This improves the harvesting quality and efficiency of the harvester while reducing the failure rate and alleviating the technical requirements and work intensity of the operator.

[0005] The present invention achieves the above technical objectives through the following technical means.

[0006] A method for controlling the operating speed of a combine harvester based on multiple operating parameter reward, characterized by comprising the following steps:

[0007] The actual operation speed, number of supervoxels of harvested crops, grass-to-grain ratio, plant height, stubble height and cutting width are input into the Gaussian process regression model GSA-GPR of the universal gravitational search algorithm to obtain the predicted value of feeding amount;

[0008] Establishing a DQN neural network, inputting the predicted value of the feed amount into the DQN neural network, inputting the impurity rate, breakage rate, loss rate, threshing drum speed, conveyor trough speed, and secondary impurity speed into the DQN neural network as reward functions, and the DQN neural network outputs a theoretical operating speed;

[0009] The difference between the actual operating speed and the theoretical operating speed and the variation of the difference are input into the fuzzy neural network PID controller, and the forward speed of the combine harvester is controlled by the output of the fuzzy neural network PID controller.

[0010] Furthermore, establishing the Gaussian process regression model GSA-GPR of the gravitational search algorithm includes the following steps:

[0011] Using sensors to collect field data at time t, the field data includes information on the number of supervoxels of harvested crops, grass-to-grain ratio, plant height, stubble height, and cutting width;

[0012] The number of supervoxels, grass-to-grain ratio, plant height, stubble height, and cutting width of the harvested crops are input into the GPR model as the dataset x. The output of the GPR model is the feed amount f(x). The square exponential kernel function is used: Where x and x′ represent any two sets of data, σ and l are kernel function parameters;

[0013] Set the iterative initial values ​​of the kernel function parameters σ and l and calculate the covariance matrix:

[0014]

[0015] Where x i (i=1, 2, ..., n) represents the i-th data set; n is the maximum number of groups;

[0016] Initialize the population particle number N, maximum number of iterations T and dimension D of the GSA algorithm;

[0017] Generate a GPR hyperparameter population based on the GSA algorithm, use the training set to train the GPR model, and then use the validation set to evaluate the contribution of each particle to the performance of the GPR model. Calculate the fitness value of each particle at time t, and solve the best fitness value best(t) and the worst fitness value worst(t) at time t. Specifically,

[0018]

[0019] Where, fit j (t) represents the fitness value of the jth particle at time t;

[0020] Calculate the mass and acceleration of the particle, specifically:

[0021]

[0022]

[0023] Where: fit i(t) represents the fitness value of the i-th particle at time t; M i (t) represents the mass of the i-th particle at time t; N represents the total number of N particles in the space; represents the sum of the forces exerted on particle i by all other particles in dimension d at time t; rand j Represents a random number between [0,1]; represents the gravitational force exerted by particle i on particle j in d dimensions at time t; represents the acceleration of particle i in dimension d at time t; Kbest represents the K particles with the best adaptability in the population, and the initial value of Kbest is N;

[0024] Update the particle's velocity and position according to the following formula:

[0025]

[0026] Where, represents the velocity of particle i in d dimension at time t, represents the position of particle i in d dimension at time t; represents the velocity of particle i in d dimension at time t+1; represents the position of particle i in d dimension at time t+1;

[0027] Determine whether the maximum number of iterations is met. If not, add 1 to the number of iterations and regenerate the GPR hyperparameter population cycle until the number of iterations is equal to the maximum number of iterations. If it is met, return the optimal hyperparameters and determine the Gaussian process regression model GSA-GPR model f(x)~GP(μ,C) of the gravitational search algorithm, where GP represents Gaussian distribution, μ represents mean, and C represents covariance matrix.

[0028] Furthermore, the DQN neural network training specifically includes the following steps:

[0029] The feeding amount s at time t t The predicted value is input into the DQN neural network, and the impurity rate, breakage rate, loss rate, threshing drum speed, conveyor trough speed and secondary impurity speed are input into the DQN neural network as the reward function;

[0030] Create an experience replay pool, current value network, and target value network;

[0031] Initialize the experience replay pool capacity D', target neural network update frequency T', discount rate γ and number of iterations N'; use random neural network parameters θ to initialize the state-action value function Q;

[0032] According to the feeding amount t, using the ε-greedy strategy, select a random action with probability ε or select the maximum action a of the state-action value function Q at time t with probability (1-ε) t , the maximum action a t is the operating speed of the combine harvester;

[0033] The reward value r at time t is calculated based on the impurity rate, breakage rate, loss rate, threshing drum speed, conveyor trough speed, and secondary impurity speed as the reward function. t and the feed amount s at time t+1 t+1 , calculate the state-action function value Q at time t t =(s t ,a t ,r t ,s t+1 ), the state-action function value Q at time t t Put it into the experience replay pool;

[0034] Uniformly randomly sample a transformation sample data Q from the experience replay pool D' k =(s k ,a k ,r k ,s k+1 ), where s k is the feeding amount at time k; a k is the state-action value function Q at time k k The maximum action of r k is the reward value at time k; s k+1 is the feeding amount at time k+1;

[0035] Will s k Input the current value network and calculate Q(s k ,a k ;θ), where Q(s k ,a k ; θ) represents the Q value at time k, θ is the current random neural network parameter; s k+1 Input the target value network and determine maxQ(s k+1 ,a k+1 ;θ′), where maxQ(s k+1 ,a k+1 ; θ′) represents the Q value at time k+1, θ′ is the target network parameter; calculate the Loss function Δθ in the current value network;

[0036] Update the current random neural network parameters: θ = θ + Δθ;

[0037] Determine whether the interval time reaches n times the target neural network update frequency T'. If the interval time is nT', set θ' = θ, update the target network parameters, and then determine whether the maximum number of iterations is met after updating the target network parameters; if the interval time does not meet the conditions, determine whether the maximum number of iterations is met;

[0038] If the number of iterations N' does not reach the maximum number of iterations, then the number of iterations N' is increased by 1, and the random neural network parameter θ is reused to initialize the state-action value function Q loop until the number of iterations is equal to the maximum number of iterations;

[0039] If the number of iterations N' reaches the maximum number of iterations, the trained DQN neural network model is obtained.

[0040] Furthermore, the Loss function in the current value network is calculated as follows;

[0041] Δθ=[r k +γmaxQ(s k+1 ,a k+1 ;θ′)-Q(s k ,a k ;θ)] 2

[0042] Where r k is the reward value at time k, and γ is the discount rate.

[0043] A combine harvester operating speed control system based on multiple operating parameter reward, comprising an embedded processor, a feed rate module, a grain loss rate sensor, a grain debris and breakage rate sensor, a threshing drum speed sensor, a conveyor trough speed sensor, a secondary debris speed sensor, and an operating speed control module;

[0044] The feeding amount module includes a laser ranging sensor, a binocular vision sensor, a BD / GPS module and a CAN communication module; the laser ranging sensor determines the plant height and stubble height by detecting the round-trip time of the laser contacting the plant height and stubble; the binocular vision sensor obtains a three-dimensional point cloud image of the crop to obtain the number of supervoxels; the BD / GPS module obtains the actual operating speed by obtaining the current position of the combine harvester;

[0045] The feed amount module, the grain loss rate sensor, the grain impurity and breakage rate sensor, the threshing drum speed sensor, the conveyor trough speed sensor, and the secondary impurity speed sensor are communicatively connected to the embedded processor, and the embedded processor is provided with a program for a combine harvester operation speed control method based on multi-operation parameter reward; the operation speed control module performs output control through a D / A conversion module of the embedded processor.

[0046] The beneficial effects of the present invention are:

[0047] 1. The combined harvester operating speed control system and method based on multi-operation parameter reward described in the present invention utilizes a gravitational search algorithm (GSA) to optimize the hyperparameters of a Gaussian process regression model (GPR). This solves the problems of the traditional GPR conjugate gradient method, which is highly dependent on initial values, has a slow convergence rate, and is prone to falling into local optimal solutions during the optimization process. This makes the predicted feed amount obtained by the GSA-GPR model more accurate. The DQN neural network employed incorporates an experience replay pool (Memory pool), a network for updating neural network parameters (MainNet), and a network for updating Q values ​​(TargetNet). The random sampling operation from the Memory pool disrupts the order of the original data, weakening the data's relevance. During the retraining process, the parameters of the MainNet are assigned to the TargetNet at intervals of T, making the algorithm update more stable, thereby improving the accuracy of the obtained optimal operating speed.

[0048] 2. The combine harvester operating speed control system and method based on multi-operation parameter reward described in the present invention adopts a DQN neural network to fuse multiple operating parameters to obtain the optimal operating speed, and then outputs it through fuzzy neural network and network PID control. Compared with traditional manual experience observation and manual control, it is more accurate, effectively improves the harvesting quality and efficiency of the harvester, reduces the failure rate, and reduces the workload of the operator.

[0049] 3. The combine harvester operating speed control system and method based on multi-operation parameter reward described in the present invention can use an embedded processor to collect the combine harvester's operating information in real time, use machine vision technology to obtain feed amount information in advance, collect the rotation speed of multiple key parts, grain impurity and breakage rate, grain loss rate, and operating speed, etc., and monitor and display them on the touch display. If an abnormal situation occurs, an alarm can be issued to prompt the operator. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. The drawings described below are some embodiments of the present invention. For ordinary technicians in this field, it is obvious that other drawings can be obtained based on these drawings without paying any creative work.

[0051] Figure 1 This is a structural block diagram of the combine harvester operating speed control system based on multiple operating parameter reward according to the present invention.

[0052] Figure 2 This is a schematic diagram of the sensor installation position of the combine harvester of the present invention.

[0053] Figure 3 It is a data collection flow chart of the combine harvester of the present invention.

[0054] Figure 4 This is a flow chart of the feed amount prediction according to the present invention.

[0055] Figure 5 This is a flow chart for predicting the optimal operating speed of a combine harvester according to the present invention.

[0056] Figure 6 This is the principle diagram of the fuzzy neural network PID control system described in the present invention.

[0057] In the picture:

[0058] 1-Laser ranging sensor; 2-Binocular vision sensor; 3-Grain impurity and breakage loss rate sensor; 4-Grain loss rate sensor; 5-Conveyor trough speed sensor; 6-Secondary impurity speed sensor; 7-Threshing drum speed sensor. DETAILED DESCRIPTION

[0059] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the protection scope of the present invention is not limited thereto.

[0060] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.

[0061] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "axial", "radial", "vertical", "horizontal", "inside", "outside" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first" and "second" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, "multiple" means two or more, unless otherwise clearly and specifically defined.

[0062] In the present invention, unless otherwise expressly specified or limited, the terms "mounted," "connected," "connect," "fixed," etc. should be understood broadly. For example, they may refer to fixed, detachable, or integral connections; mechanical or electrical connections; direct or indirect connections through an intermediary; or internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0063] The combine harvester operation speed control method based on multiple operation parameter reward according to the present invention comprises the following steps:

[0064] S01: Input the actual operation speed, number of supervoxels of harvested crops, grass-to-grain ratio, plant height, stubble height and cutting width information into the Gaussian process regression model GSA-GPR of the gravitational search algorithm to obtain the predicted value of the feeding amount, such as Figure 4 As shown, specifically:

[0065] S1.1: Use sensors to collect field data at time t, including the number of supervoxels of harvested crops, grass-to-grain ratio, plant height, stubble height, and cutting width information, such as Figure 2 As shown;

[0066] S1.2: The number of supervoxels, grass-to-grain ratio, plant height, stubble height, and swath width of the harvested crops are input into the GPR model as the dataset x. The output of the GPR model is the feed rate f(x). The square exponential kernel function is used: Where x and x′ represent any two sets of data, σ and l are kernel function parameters;

[0067] S1.3: Set the initial values ​​of the kernel function parameters σ and l and calculate the covariance matrix:

[0068]

[0069] Where x i (i=1, 2, ..., n) represents the i-th data set; n is the maximum number of groups;

[0070] S1.4: Initialize the population particle number N, maximum number of iterations T and dimension D of the GSA algorithm;

[0071] S1.5: Generate a GPR hyperparameter population based on the GSA algorithm, train the GPR model using the training set, and then use the validation set to evaluate the contribution of each particle to the GPR model performance. Calculate the fitness value of each particle at time t and solve for the best fitness value best(t) and the worst fitness value worst(t) at time t. Specifically,

[0072]

[0073] Where, fit j (t) represents the fitness value of the jth particle at time t;

[0074] S1.6: Calculate the mass and acceleration of the particle. Specifically:

[0075]

[0076]

[0077] Where: fit i (t) represents the fitness value of the i-th particle at time t; M i (t) represents the mass of the i-th particle at time t; N represents the total number of N particles in the space; represents the sum of the forces exerted on particle i by all other particles in dimension d at time t; rand j Represents a random number between [0,1]; represents the gravitational force exerted by particle i on particle j in d dimensions at time t; represents the acceleration of particle i in dimension d at time t; Kbest represents the K particles with the best adaptability in the population, and the initial value of Kbest is N;

[0078] S1.7: Update the particle's velocity and position according to the following formula:

[0079]

[0080] Where, represents the velocity of particle i in d dimension at time t, represents the position of particle i in d dimension at time t; represents the velocity of particle i in d dimension at time t+1; represents the position of particle i in d dimension at time t+1;

[0081] S1.8: Determine whether the maximum number of iterations is met. If not, add 1 to the number of iterations and return to step S1.5 to regenerate the GPR hyperparameter population loop until the number of iterations is equal to the maximum number of iterations. If it is met, return to the optimal hyperparameters and determine the Gaussian process regression model GSA-GPR model f(x)~GP(μ,C) of the gravitational search algorithm, where GP represents Gaussian distribution, μ represents mean, and C represents covariance matrix.

[0082] S02: Establishing a DQN neural network, inputting the predicted value of the feed amount into the DQN neural network, inputting the impurity rate, breakage rate, loss rate, threshing drum speed, conveyor trough speed, and secondary impurity speed into the DQN neural network as a reward function, and the DQN neural network outputs a theoretical operating speed;

[0083] like Figure 5 As shown, the DQN neural network training specifically includes the following steps:

[0084] S2.1: Set the feeding amount s at time t t The predicted value is input into the DQN neural network, and the impurity rate, breakage rate, loss rate, threshing drum speed, conveyor trough speed and secondary impurity speed are input into the DQN neural network as the reward function;

[0085] S2.2: Create the experience replay pool Memory pool, current value network MainNet and target value network TargetNet;

[0086] S2.3: Initialize the experience replay pool capacity D', the target neural network update frequency T', the discount rate γ, and the number of iterations N'; use the random neural network parameters θ to initialize the state-action value function Q;

[0087] S2.4: According to the feeding amount s t , using the ε-greedy strategy, select a random action with probability ε or select the maximum action a of the state-action value function Q at time t with probability (1-ε) t , the maximum action a t is the operating speed of the combine harvester;

[0088] S2.5: Calculate the reward value r at time t based on the impurity rate, breakage rate, loss rate, threshing drum speed, conveyor trough speed, and secondary impurity speed as the reward function t and the feed amount s at time t+1 t+1 , calculate the state-action function value Q at time t t =(s t ,a t ,r t ,s t+1 ), the state-action function value Q at time t t Put it into the experience replay pool;

[0089] S2.6: Uniformly randomly sample a transformation sample data Q from the experience replay pool D' k =(s k ,a k ,r k ,s k+1 ), where s kis the feeding amount at time k; a k is the state-action value function Q at time k k The maximum action of r k is the reward value at time k; s k+1 is the feeding amount at time k+1;

[0090] S2.7: Change s k Input the current value network and calculate Q(s k ,a k ;θ), where Q(s k ,a k ; θ) represents the Q value at time k, θ is the current random neural network parameter; s k+1 Input the target value network and determine maxQ(s k+1 ,a k+1 ;θ′), where maxQ(s k+1 ,a k+1 ; θ′) represents the Q value at time k+1, θ′ is the target network parameter; calculate the Loss function Δθ in the current value network as follows;

[0091] Δθ=[r k +γmaxQ(s k+1 ,a k+1 ;θ′)-Q(s k ,a k ;θ)] 2

[0092] Where r k is the reward value at time k, and γ is the discount rate.

[0093] S2.8: Update the current random neural network parameters: θ = θ + Δθ;

[0094] S2.9: Determine whether the interval time reaches n times the target neural network update frequency T'. If the interval time is nT', set θ' = θ, update the target network parameters, and after updating the target network parameters, determine whether the maximum number of iterations is met. If the interval time does not meet the condition, determine whether the maximum number of iterations is met.

[0095] S2.10: If the number of iterations N' does not reach the maximum number of iterations, then increase the number of iterations N' by 1, and return to step S2.4 to re-initialize the state-action value function Q using the random neural network parameters θ until the number of iterations reaches the maximum number of iterations;

[0096] S2.11: If the number of iterations N' reaches the maximum number of iterations, the trained DQN neural network model is obtained.

[0097] S03: Inputting the difference between the actual operating speed and the theoretical operating speed and the variation of the difference into the fuzzy neural network PID controller, and controlling the forward speed of the combine harvester through the output of the fuzzy neural network PID controller.

[0098] like Figure 6 As shown, the operating speed control system of the combine harvester adopts fuzzy neural network PID control, collects plant height, stubble height, grass-grain ratio, cutting width, operating speed, and supervoxel number information and inputs it into the GSA-GPR model to obtain the feed amount prediction value, and uses the predicted feed amount value as the input of the DQN neural network. Combined with multiple operating parameters (including impurities, crushing, loss rate, threshing drum, conveyor trough, and secondary impurity speed), the optimal operating speed value is obtained and compared with the current speed value obtained by the speed detection module. The fuzzy neural network PID control system takes the operating speed deviation value e and the deviation value change ec as input, and the output is connected to the operating speed system;

[0099] A fuzzy neural network system is constructed using fuzzy neural networks, which consists of an input layer, a membership function generation layer, a fuzzy reasoning layer, a normalization layer, and an output layer. According to the principle of the PID controller, the three basic parameters of the PID controller are derived from the fuzzy neural network system, namely the gain coefficient Kp, the integral time coefficient Ki, and the differential time coefficient Kd. By adjusting these three coefficient parameters, the fuzzy neural network optimizes the PID parameters and completes the control of the combine harvester operating speed.

[0100] like Figure 1 and Figure 2 As shown, the combine harvester operating speed control system based on multi-operation parameter reward described in the present invention includes an embedded processor, a feeding amount module, a grain loss rate sensor 4, a grain debris and breakage rate sensor 3, a threshing drum speed sensor 7, a conveyor trough speed sensor 5, a secondary debris speed sensor 6, an operating speed control module and a touch display screen; the grain loss rate sensor 4 is used to detect the grain loss rate, the grain debris and breakage rate sensor 3 is used to detect the grain debris and breakage rate, the threshing drum speed sensor 7, the conveyor trough speed sensor 5, and the secondary debris speed sensor 6 are used to respectively detect the threshing drum speed, the conveyor trough speed and the secondary debris speed.

[0101] like Figure 3As shown in the data collection flow chart for a combine harvester, the feed rate module includes a laser range sensor 1, a binocular vision sensor 2, a BD / GPS module, and a CAN communication module. The laser range sensor 1 determines the plant height and stubble height by detecting the round-trip time between the laser contacting the plant height and the stubble. The binocular vision sensor obtains a three-dimensional point cloud image of the crop, divides it into two parts using an octree, and then uses a supervoxel clustering method to cluster and obtain the number of supervoxels. The BD / GPS module obtains the actual operating speed by obtaining the current position of the combine harvester. The CAN communication module receives the grass-to-grain ratio and cutting width information manually input by the touch screen.

[0102] The feeding amount module, the grain loss rate sensor, the grain impurity and breakage rate sensor, the threshing drum speed sensor 7, the conveyor trough speed sensor 5, and the secondary impurity speed sensor 6 are communicatively connected to the embedded processor, and the embedded processor is provided with a program for the combine harvester operation speed control method based on multi-operation parameter reward; the operation speed control module performs output control through the D / A conversion module of the embedded processor.

[0103] Combine harvester feed rate prediction is a nonlinear problem with numerous characteristic parameters. Therefore, a Gaussian process regression model (GSA-GPR) based on the universal gravitational search algorithm (GSA) is used to predict feed rate. Using the GSA algorithm to optimize the hyperparameters of the GPR model, this approach overcomes the traditional GPR conjugate gradient method's strong dependence on initial values, slow convergence, and tendency to fall into local optimal solutions.

[0104] It should be understood that although this specification is described according to various embodiments, not every embodiment contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0105] The series of detailed descriptions listed above are only specific descriptions of feasible embodiments of the present invention. They are not intended to limit the scope of protection of the present invention. Any equivalent embodiments or changes that do not deviate from the technical spirit of the present invention should be included in the scope of protection of the present invention.

Claims

1. A combine harvester operation speed control method based on multiple operation parameter reward, characterized in that: The steps include: The actual operation speed, number of supervoxels of harvested crops, grass-to-grain ratio, plant height, stubble height and cutting width are input into the Gaussian process regression model GSA-GPR of the universal gravitational search algorithm to obtain the predicted value of feeding amount; The Gaussian process regression model GSA-GPR of the gravitational search algorithm is established, which includes the following steps: Using sensors to collect field data at time t, the field data includes information on the number of supervoxels of harvested crops, grass-to-grain ratio, plant height, stubble height, and cutting width; The number of supervoxels, grass-to-grain ratio, plant height, stubble height, and cutting width of the harvested crops are input into the GPR model as the dataset x. The output of the GPR model is the feed amount f(x). The square exponential kernel function is used: Where x and x′ represent any two sets of data, σ and l are kernel function parameters; Set the iterative initial values ​​of the kernel function parameters σ and l and calculate the covariance matrix: Where x i (i=1, 2, ..., n) represents the i-th data set; n is the maximum number of groups; Initialize the population particle number N, maximum number of iterations T and dimension D of the GSA algorithm; Generate a GPR hyperparameter population based on the GSA algorithm, use the training set to train the GPR model, and then use the validation set to evaluate the contribution of each particle to the performance of the GPR model. Calculate the fitness value of each particle at time t, and solve the best fitness value best(t) and the worst fitness value worst(t) at time t. Specifically, Where, fit j (t) represents the fitness value of the jth particle at time t; Calculate the mass and acceleration of the particle, specifically: Where: fit i (t) represents the fitness value of the i-th particle at time t; M i (t) represents the mass of the i-th particle at time t; N represents the total number of N particles in the space; represents the sum of the forces exerted on particle i by all other particles in dimension d at time t; rand j Represents a random number between [0,1]; represents the gravitational force exerted by particle i on particle j in d dimensions at time t; represents the acceleration of particle i in dimension d at time t; Kbest represents the K particles with the best fitness value in the population, and the initial value of Kbest is N; Update the particle's velocity and position according to the following formula: Where, represents the velocity of particle i in d dimension at time t, represents the position of particle i in d dimension at time t; represents the velocity of particle i in d dimension at time t+1; represents the position of particle i in d dimension at time t+1; Determine whether the maximum number of iterations is met. If not, increase the number of iterations by 1 and regenerate the GPR hyperparameter population cycle until the number of iterations is equal to the maximum number of iterations. If it is met, return the optimal hyperparameters and determine the Gaussian process regression model GSA-GPR model f(x)~GP(μ,C) of the gravitational search algorithm, where GP represents the Gaussian distribution, μ represents the mean, and C represents the covariance matrix. Establishing a DQN neural network, inputting the predicted value of the feed amount into the DQN neural network, inputting the impurity rate, breakage rate, loss rate, threshing drum speed, conveyor trough speed, and secondary impurity speed into the DQN neural network as reward functions, and the DQN neural network outputs a theoretical operating speed; The difference between the actual operating speed and the theoretical operating speed and the variation of the difference are input into the fuzzy neural network PID controller, and the forward speed of the combine harvester is controlled by the output of the fuzzy neural network PID controller.

2. The combine harvester operation speed control method based on multiple operation parameter reward according to claim 1, characterized in that: The DQN neural network training specifically includes the following steps: The feeding amount s at time t t The predicted value is input into the DQN neural network, and the impurity rate, breakage rate, loss rate, threshing drum speed, conveyor trough speed and secondary impurity speed are input into the DQN neural network as the reward function; Create an experience replay pool, current value network, and target value network; Initialize the experience replay pool capacity D', target neural network update frequency T', discount rate γ and number of iterations N'; use random neural network parameters θ to initialize the state-action value function Q; According to the feeding amount t , using the ε-greedy strategy, select a random action with probability ε or select the maximum action a of the state-action value function Q at time t with probability (1-ε) t , the maximum action a t is the operating speed of the combine harvester; The reward value r at time t is calculated based on the impurity rate, breakage rate, loss rate, threshing drum speed, conveyor trough speed, and secondary impurity speed as the reward function. t and the feed amount s at time t+1 t+1 , calculate the state-action function value Q at time t t =(s t ,a t ,r t ,s t+1 ), the state-action function value Q at time t t Put it into the experience replay pool; Uniformly randomly sample a transformation sample data Q from the experience replay pool D' k =(S k ,a k ,r k ,s k+1 ), where s k is the feeding amount at time k; a k is the state-action value function Q at time k k The maximum action of r k is the reward value at time k; s k+1 is the feeding amount at time k+1; Will s k Input the current value network and calculate Q(s k ,a k ;θ), where Q(s k ,a k ; θ) represents the Q value at time k, θ is the current random neural network parameter; s k+1 Input the target value network and determine maxQ(s k+1 ,a k+1 ;θ′), where maxQ(s k+1 ,a k+1 ; θ′) represents the Q value at time k+1, θ′ is the target network parameter; calculate the Loss function △θ in the current value network; Update the current random neural network parameters: θ = θ + △θ; Determine whether the interval time reaches n times the target neural network update frequency T'. If the interval time is nT', set θ' = θ, update the target network parameters, and then determine whether the maximum number of iterations is met after updating the target network parameters; if the interval time does not meet the conditions, determine whether the maximum number of iterations is met; If the number of iterations N' does not reach the maximum number of iterations, then the number of iterations N' is increased by 1, and the random neural network parameter θ is reused to initialize the state-action value function Q loop until the number of iterations is equal to the maximum number of iterations; If the number of iterations N' reaches the maximum number of iterations, the trained DQN neural network model is obtained.

3. The combine harvester operation speed control method based on multiple operation parameter reward according to claim 1, characterized in that: Calculate the Loss function in the current value network as follows; △θ=[r k +γmaxQ(s k+1 ,a k+1 ;θ′)-Q(s k ,a k (i)] 2 Where r k is the reward value at time k, and γ is the discount rate.

4. A combine harvester operating speed control system based on multiple operating parameter reward, characterized in that: It includes an embedded processor, a feeding amount module, a grain loss rate sensor (4), a grain impurity and breakage rate sensor (3), a threshing drum speed sensor (7), a conveyor trough speed sensor (5), a secondary impurity speed sensor (6), and an operation speed control module; The feeding amount module comprises a laser ranging sensor (1), a binocular vision sensor (2), and a BD / GPS module; the laser ranging sensor determines the plant height and the stubble height by detecting the round trip time of the laser contacting the plant height and the stubble; the binocular vision sensor obtains a three-dimensional point cloud image of the crop for obtaining the number of supervoxels; the BD / GPS module obtains the actual operating speed by obtaining the current position of the combine harvester; The feed rate module, the grain loss rate sensor, the grain impurity and breakage rate sensor, the threshing drum speed sensor (7), the conveyor trough speed sensor (5), and the secondary impurity speed sensor (6) are communicatively connected to an embedded processor, and the embedded processor is provided with a program of the combine harvester operation speed control method based on multi-operation parameter reward according to any one of claims 1 to 3; the operation speed control module performs output control through a D / A conversion module of the embedded processor.

Images