A path planning method for unmanned agricultural machinery operation based on vehicle status and working environment

Abstract

The present invention discloses a method for planning the operation path of an unmanned agricultural machine based on the vehicle state and the working environment, the method comprising: step 1, specifying the range of a plot according to a satellite remote sensing image 100; step 2, fine-tuning the specified range using image processing to generate an operation surface 101; step 3, standardizing the operation surface using an operation surface preprocessing function 102; step 4, generating a global path according to path rules, motion model rules and process requirements 103; step 5, generating a local path according to a vehicle-mounted sensor 104; step 6, ending the path execution 105. The application of intelligent algorithms can significantly improve operation efficiency, reduce energy consumption, improve operation quality, and enhance the system's adaptive ability and safety.

Description

Technical Field

[0001] The present invention relates to the field of unmanned driving technology, and in particular to a method for perceiving and identifying a geographical environment and then automatically planning a working path for an operating machine. Background Art

[0002] With the continuous development of sensor technology and remote sensing technology and the rapid iteration and update of tensor computing tools, the fully automated production of unmanned driving technology has been implemented in the industrial and agricultural production process. Compared with the new fully automatic operation, traditional manual operation has the disadvantages of low work efficiency, uncontrollable work quality, and low operator quality. In order to meet the automation requirements of industrial upgrading and the increasingly severe employment risks, this method generates unmanned driving routes for unmanned operating machinery by using the global perception provided by remote sensing recognition and the local perception provided by airborne sensors. Together with the operating part, it forms a fully automated unmanned operation system.

[0003] Different from the commonly used Dijkstra algorithm, A* algorithm, and RTT algorithm in path planning, which generally take the shortest or fastest feasible route from the starting point to the end point as the generation goal, the path generation algorithm of the work surface is to ensure that the planned route results can fully cover the entire work surface. On this basis, the path planning principles of full coverage, no duplication, and few turns must be followed to ensure that fuel consumption is reduced and working hours are shortened throughout the operation, thereby improving work efficiency. At the same time, in the process of planning the path covering the work surface, the kinematic characteristics of the vehicle must be fully considered to ensure the feasibility of route execution, and the interaction state with the working environment must be fully considered to ensure the feasibility of route passage. Finally, it is necessary to optimize the tendency of path planning based on the known process requirements provided to meet the generation of work surface trajectories for various demand conditions. Summary of the invention

[0004] In view of the above-mentioned deficiencies of the prior art and in order to meet the new requirements for work surface path planning, the present invention discloses a method for generating an unmanned agricultural machinery work path based on vehicle status and working environment.

[0005] The technical solution of the present invention comprises the following steps:

[0006] Step 1, specifying the range of the plot 100 according to the satellite remote sensing image;

[0007] There are two triggering methods for this process: (1) specifying a location on the remote sensing map; (2) selecting a two-dimensional geometric area on the remote sensing map;

[0008] Step 2, using image processing to fine-tune the specified range to generate a work surface 101;

[0009] Step 3, using the work surface preprocessing function to standardize the work surface 102;

[0010] Step 4, generating a global path 103 according to the path rules, motion model rules and process requirements;

[0011] Step 5, generating a local path according to the vehicle-mounted sensor 104;

[0012] Step 6, path execution ends 105.

[0013] The implementation scheme of the present invention starts with the process of specifying the scope of a plot of land, specifically specifying the scope of the plot of land based on a satellite remote sensing image 100. The process has two triggering modes: (1) specifying a location on the remote sensing map; (2) selecting a two-dimensional geometric area on the remote sensing map.

[0014] After obtaining the specified range result, it is necessary to use image processing methods to fine-tune the specified range to generate a work surface 101. If the range 100 of the plot is specified based on the satellite remote sensing image and the trigger method (1) is used, the system will read the remote sensing image 200 near the location and send it to the image segmentation neural network 201 to segment it into different blocks. The segmented result block at the specified location is sent as a preliminary result to the boundary extraction module 202 to obtain the subdivided block boundary result. Finally, the boundary of the extracted block can be manually adjusted to obtain the specified boundary result.

[0015] If the trigger method (2) is used to specify the range 100 of the plot based on the satellite remote sensing image, the system will use the selected area as the plot range result, and send the plot range result as the preliminary result to the boundary extraction module 202 to obtain the subdivided block boundary result. Finally, the boundary of the block boundary can be manually adjusted to obtain the specified boundary result; if the trigger method (2) is used to specify the range 100 of the plot based on the satellite remote sensing image and it is stated that no detailed processing is required, the boundary of the range result can be manually adjusted to obtain the specified boundary result.

[0016] The convolutional neural network U-Net of the backbone network of the neural network 201 for image segmentation is formed. U-Net is a convolutional neural network architecture for quickly and accurately segmenting images. The backbone network of the neural network for image segmentation provides high accuracy and efficiency in land parcel identification.

[0017] The specified boundary results obtained contain invalid boundaries and non-standard boundary results, and the boundary results need to be further standardized. In the process of standardizing the working surface 102 in the working surface preprocessing function, the working surface simplification module 203 is first used to filter and delete points with less information in the boundary data to extract the skeleton optimized working surface data, and then the working surface segmentation and combination module 204 is used to segment the simplified data into basic triangles and match them into optimal quadrilaterals, and possible concave polygons are segmented into multiple basic quadrilateral units. Finally, the working surface information extraction module 205 is used to give a multi-dimensional standardized description of the characteristics of the quadrilateral.

[0018] After the above steps, the calculation and processing of the input data are completed to obtain the basic input data of the path rule.

[0019] For the path rule function, a prefabricated motion component assembly method is used to ensure fast and reliable path planning, and its components include a linear component 206, a steering component 207 and a fixed process component 208.

[0020] The straight line component 206 includes five elements, namely, a starting point, an end point, a direction, a length, and an adjacency relationship, and four evaluation criteria, namely, proximity relationship, coverage degree, repetition degree, and motion cost, are used as indicators for generating a straight line path.

[0021] The steering component 207 includes seven elements, namely, starting point, end point, direction, length, adjacency, steering method, and steering feasibility, and four evaluation criteria, namely, proximity, coverage, repetition, and motion cost, are used as indicators for generating a steering path.

[0022] The fixed process component 208 includes eight elements, namely, starting point, end point, direction, length, adjacency relationship, process method, process feasibility, and target object operation relationship. Five evaluation criteria, namely, proximity relationship, coverage degree, repetition degree, motion cost, and process necessity, are used as indicators for generating a fixed process path.

[0023] When different components are combined, the component combination module 209 will be responsible for the process of adjacency between different modules to smooth out the defects encountered during splicing and ensure kinematic feasibility.

[0024] When different components encounter conflicts when combined, the conflict may be that the combination module cannot smoothly splice and report an error, or it may be that the evaluation criteria such as proximity, coverage, repetition, motion cost, etc. in the planning process are unqualified. When a conflict occurs, the component conflict handling module 210 needs to be used to compromise, degrade, or even replace the corresponding components.

[0025] After obtaining the optimal quadrilateral after splitting and its standardized description, continue the process of global path generation 103 according to the path rules, motion model rules and process requirements, and send each quadrilateral to the independent work surface path planning 211 for local optimal path planning. The independent work surface path planning 211 uses the path planning neural network 213 to combine the above-mentioned straight line component 206 and fixed process component 208 to obtain the optimal path result within the quadrilateral. If there is only one main quadrilateral, the steering component 207 is spliced ​​and the component combination module 209 and the conflict handling module 210 are called to calculate the global path, and the generation process ends.

[0026] If there are multiple main quadrilaterals, the combined work surface connection method 212 is used to splice the work paths in the two planes using the steering component 207. If the splicing fails, the connection method will return to the independent work surface path planning 211, and the optimal path result will be taken from the second optimal path result until the global path is successfully spliced, and the generation process ends.

[0027] After the generation process is completed, the traffic check 214 will be used to simulate the vehicle passing through the path process, and the difficulty of execution, the degree of coverage of the working surface, the degree of repetition, the movement cost, and the necessity of the process will be evaluated. The evaluation result will be used to execute the path of the subsequent machinery.

[0028] During the execution process, the vehicle sensor 215 continuously returns obstacle information in front of and on the sides of the vehicle. The sensor usually uses a laser radar and a binocular depth camera as hardware, and its data is finally integrated into a spatial obstacle description. When an obstacle appears, the obstacle avoidance component 216 is executed. The component checks the execution component and its adjacent components of the global path planning result at the current time point, and uses a smooth path to modify the current component and its adjacent components to bypass the obstacle.

[0029] When the execution path is completed, the component fine-tuning module 217 records the five evaluation criteria of the actual proximity relationship, coverage, repetition, motion cost, and process necessity of each component in the execution process, and obtains the actual performance of the components under different parameters. The evaluation criteria of different component actions are adjusted according to the learning rate r to overcome the problem that the general model network cannot generate the optimal path due to the conflict of inherent mechanical properties of the vehicle or environmental factors in actual problems.

[0030] When the execution path is completed, the neural network training module 218 inputs the actual results of each vehicle's execution and the fine-tuning results of each component fine-tuning module 217 into the training data set of the path planning neural network 213, and regularly updates the path planning neural network 213 pre-model to optimize the design of the neural network in the actual use environment. The problem that the pre-model is not applicable in the actual environment can make the pre-model perform optimally in the real environment or quickly converge to the optimal range.

[0031] Beneficial effects: The present invention can significantly improve operating efficiency, reduce energy consumption, improve operating quality, and enhance the system's adaptability and safety through the application of intelligent algorithms.

[0032] The U-Net model can efficiently and accurately segment and classify terrain and obstacles, provide high-resolution terrain maps, and make path planning more accurate. By identifying different types of terrain and obstacles, the system can select the most suitable path, avoid complex terrain and obstacles, and improve work efficiency. Reinforcement learning algorithms such as DQN can dynamically adjust path planning based on real-time environmental feedback and select the path with the lowest energy consumption. Intelligent obstacle avoidance strategies can effectively avoid unnecessary stops and turns, reducing operation time and energy consumption. The PSO algorithm uses global search to find the operation path with the highest coverage, lowest energy consumption, and best quality, balancing operation efficiency and quality. The multi-objective optimization algorithm can effectively avoid repeated coverage of operations and improve operation quality. The reinforcement learning algorithm can adapt to different terrains, weather, and operating conditions through continuous learning and adjustment, and improve the system's adaptability. Through real-time data feedback and continuous learning, the system can continuously optimize its own performance and maintain efficient operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is the overall flow chart of the present invention;

[0034] Figure 2 It is a design diagram of the components of the present invention;

[0035] Figure 3 Input image data map for remote sensing;

[0036] Figure 4 This is the segmentation result map of remote sensing data;

[0037] Figure 5 This is a flow chart of the intelligent algorithm for path planning of the present invention. DETAILED DESCRIPTION

[0038] The invention discloses a method for generating an unmanned driving operation path of agricultural machinery based on vehicle status and working environment.

[0039] like Figure 1-2 As shown, the implementation scheme of the present invention starts with the process of specifying the scope of a plot of land, specifically specifying the scope of the plot of land based on a satellite remote sensing image 100. The process has two triggering methods: (1) specifying a location on the remote sensing map; (2) selecting a two-dimensional geometric area on the remote sensing map.

[0040] After obtaining the specified range result, it is necessary to use image processing methods to fine-tune the specified range to generate a work surface 101. If the range 100 of the plot is specified based on the satellite remote sensing image and the trigger method (1) is used, the system will read the remote sensing image 200 near the location and send it to the image segmentation neural network 201 to segment it into different blocks. The segmented result block at the specified location is sent as a preliminary result to the boundary extraction module 202 to obtain the subdivided block boundary result. Finally, the boundary of the extracted block can be manually adjusted to obtain the specified boundary result.

[0041] If the trigger method (2) is used to specify the range 100 of the plot based on the satellite remote sensing image, the system will use the selected area as the plot range result, and send the plot range result as the preliminary result to the boundary extraction module 202 to obtain the subdivided plot boundary result. Finally, the boundary of the block boundary can be manually adjusted to obtain the specified boundary result; if the trigger method (2) is used to specify the range 100 of the plot based on the satellite remote sensing image and it is stated that no detailed processing is required, the boundary of the range result can be manually adjusted to obtain the specified boundary result.

[0042] The convolutional neural network U-Net of the backbone network of the neural network 201 for image segmentation is formed. U-Net is a convolutional neural network architecture for quickly and accurately segmenting images. The backbone network of the neural network for image segmentation provides high accuracy and efficiency in land parcel identification.

[0043] The image segmentation neural network 201 uses a U-Net architecture, and the output of each layer can be expressed as:

[0044] X i+1 =C i (D i (X i ))

[0045] Among them, X i represents the encoder output feature map of the i-th layer, X i+1 represents the output feature map of the i+1th layer, C i represents the convolution operation of the i-th layer, D i represents the downsampling operation of the i-th layer;

[0046] For the decoder part, add upsampling and skip connection feature map merging:

[0047]

[0048] Among them, U i represents the upsampling operation of the i-th layer, C′ i represents the convolution operation of the i-th layer; X′ irepresents the output feature map of the i-th layer of the decoder, X n-i represents the output feature map of the encoder layer ni, X′ i+1 represents the output feature map of the decoder layer i+1, represents the concatenation operation of the feature map, and n is the total number of layers in the encoder;

[0049] For binary image segmentation, the binary cross entropy loss function is used:

[0050]

[0051] Among them, L1 represents the value of the loss function, which is used to measure the difference between the model prediction and the true label. Indicates the summation of all pixels, y i is the true label of the i-th pixel, is the predicted probability of the ith pixel, N is the total number of pixels in the batch, Represents the predicted probability The natural logarithm of Represents the natural logarithm of the complement of the predicted probability.

[0052] The specified boundary results obtained contain invalid boundaries and non-standard boundary results, and the boundary results need to be further standardized. In the process of standardizing the working surface 102 in the working surface preprocessing function, the working surface simplification module 203 is first used to filter and delete points with less information in the boundary data to extract the skeleton optimized working surface data, and then the working surface segmentation and combination module 204 is used to segment the simplified data into basic triangles and match them into optimal quadrilaterals, and possible concave polygons are segmented into multiple basic quadrilateral units. Finally, the working surface information extraction module 205 is used to give a multi-dimensional standardized description of the characteristics of the quadrilateral.

[0053] After the above steps, the calculation and processing of the input data are completed to obtain the basic input data for path planning.

[0054] For the path rule function, a prefabricated motion component assembly method is used to ensure fast and reliable path planning, and its components include a linear component 206, a steering component 207 and a fixed process component 208.

[0055] The straight line component 206 includes five elements, namely, a starting point, an end point, a direction, a length, and an adjacency relationship, and four evaluation criteria, namely, proximity relationship, coverage degree, repetition degree, and motion cost, are used as indicators for generating a straight line path.

[0056] The steering component 207 includes seven elements, namely, starting point, end point, direction, length, adjacency, steering method, and steering feasibility, and four evaluation criteria, namely, proximity, coverage, repetition, and motion cost, are used as indicators for generating a steering path.

[0057] The fixed process component 208 includes eight elements, namely, starting point, end point, direction, length, adjacency relationship, process method, process feasibility, and target object operation relationship. Five evaluation criteria, namely, proximity relationship, coverage degree, repetition degree, motion cost, and process necessity, are used as indicators for generating a fixed process path.

[0058] Explanation of the above elements and evaluation criteria:

[0059] Proximity relationship: describes the proximity between components. It is hoped that components are as close as possible to reduce the cost of switching between components during execution.

[0060] Coverage: describes the coverage of the component to the entire workspace. It is hoped that the component will cover as much of the plot area as possible.

[0061] Degree of repetition: describes the size of the repeated cross-area between components. It is hoped that components will not repeat the areas that have already been executed.

[0062] Movement cost (also referred to as movement model rules): how long it takes and how much fuel it consumes to execute the above components under the vehicle movement model;

[0063] Process necessity: Evaluate whether the fixed process component is necessary;

[0064] When different components are combined, the component combination module 209 will be responsible for the process of adjacency between different modules to smooth out the defects encountered during splicing and ensure kinematic feasibility.

[0065] When different components encounter conflicts when combined, the conflict may be that the combination module cannot smoothly splice and report an error, or it may be that the evaluation criteria such as proximity, coverage, repetition, motion cost, etc. in the planning process are unqualified. When a conflict occurs, the component conflict handling module 210 needs to be used to compromise, degrade, or even replace the corresponding components.

[0066] After obtaining the optimal quadrilateral after splitting and its standardized description, continue the process of global path generation 103 according to the path rules, motion model rules and process requirements, and send each quadrilateral to the independent work surface path planning 211 for local optimal path planning. The independent work surface path planning 211 uses the path planning neural network 213 to combine the above-mentioned straight line component 206 and fixed process component 208 to obtain the optimal path result within the quadrilateral. If there is only one main quadrilateral, the steering component 207 is spliced ​​and the component combination module 209 and the conflict handling module 210 are called to calculate the global path, and the generation process ends.

[0067] If there are multiple main quadrilaterals, the combined work surface connection method 212 is used to splice the work paths in the two planes using the steering component 207. If the splicing fails, the connection method will return to the independent work surface path planning 211, and the optimal path result will be taken from the second optimal path result until the global path is successfully spliced, and the generation process ends.

[0068] After the generation process is completed, the traffic check 214 will be used to simulate the vehicle passing through the path process, and the difficulty of execution, the degree of coverage of the working surface, the degree of repetition, the movement cost, and the necessity of the process will be evaluated. The evaluation result will be used to execute the path of the subsequent machinery.

[0069] During the execution process, the vehicle sensor 215 continuously returns obstacle information in front of and on the sides of the vehicle. The sensor usually uses a laser radar and a binocular depth camera as hardware, and its data is finally integrated into a spatial obstacle description. When an obstacle appears, the obstacle avoidance component 216 is executed. The component checks the execution component and its adjacent components of the global path planning result at the current time point, and uses a smooth path to modify the current component and its adjacent components to bypass the obstacle.

[0070] When the execution path is completed, the component fine-tuning module 217 records the five evaluation criteria of the actual proximity relationship, coverage, repetition, motion cost, and process necessity of each component in the execution process, and obtains the actual performance of the components under different parameters. The evaluation criteria of different component actions are adjusted according to the learning rate r to overcome the problem that the general model network cannot generate the optimal path due to the conflict of inherent mechanical properties of the vehicle or environmental factors in actual problems.

[0071] When the execution path is completed, the neural network training module 218 inputs the actual results of each vehicle's execution and the fine-tuning results of each component fine-tuning module 217 into the training data set of the path planning neural network 213, and regularly updates the path planning neural network 213 pre-model to optimize the design of the neural network in the actual use environment. The problem that the pre-model is not applicable in the actual environment can make the pre-model perform optimally in the real environment or quickly converge to the optimal range.

[0072] The specific execution process of the path planning neural network 213 of the present invention is as follows:

[0073] Step 1: Data collection and variable definition;

[0074] Collect multimodal data: ground sensor data, vehicle mechanical parameter data and historical operation data;

[0075] Define the environment and reward mechanism:

[0076] S represents the state space, including the vehicle's motion state and the state of the relationship between the vehicle and the environment; A represents the action space, steering angle, acceleration; R represents the reward function, the formula is as follows;

[0077] R(S t ,a t )=w1r near (S t ,a t )+w2r corss (S t ,a t )+w3r repeat (S t ,a t )+w4r move (S t ,a t )+w5r process (S t ,a t )

[0078] Among them, R represents the reward function, which refers to the five evaluation criteria of proximity, coverage, repetition, motion cost, and process necessity; s t represents the state at time t; a t represents the action taken at time t; r near Represents a reward reflecting proximity; r cross represents the reward reflecting the coverage level; r repeat indicates a reward that reflects the degree of repetition; r move represents the reward reflecting the cost of movement; r process Indicates a reward that reflects the necessity of the process; w n Indicates the weight of each factor;

[0079] Step 2: Deep Q network;

[0080] Use a deep Q network model to estimate the Q value of each action taken in each state, which represents the long-term cumulative reward of taking this action in the current state;

[0081] Q(s,a;θ)

[0082] Among them, s represents the state, a represents the action, and θ represents the parameters of the network;

[0083] The mean square error (MSE) is used as the loss function to measure the difference between the Q value output by the network and the target Q value.

[0084]

[0085] Where L is the loss function, N is the number of samples, (s, a, r, s′) is the sample in the experience replay buffer, including state s, action a, reward r, next state s′, Q(s, a; θ) is the network model output, which represents the Q value of taking action a in state s, Q′(s′, a′; θ′) is the target Q value, which represents the Q value of taking action a′ in the next state s′, θ represents the network parameters, θ′ represents the parameters of the target network, and γ represents the discount factor;

[0086] Use stochastic gradient descent to minimize the loss function and update the network parameters θ:

[0087]

[0088] Among them, θ′ represents the updated parameter vector; θ represents the current parameter vector; α is the learning rate, which is used to control the step size of parameter update; Represents the gradient of the loss function L with respect to the parameter θ, that is, the rate of change of the loss function with respect to the parameter;

[0089] The output of the deep Q network is integrated with the PSO algorithm to form the input of the operation path planning; the PSO optimization result is used to generate the final operation path, taking into account the terrain, obstacles and multi-objective optimization results; the speed and position update formula of its particles is:

[0090] V i (t+1)=wV i (t)+c1r1(pBest i -P i (t))+c2r2(gBest-P i (t))

[0091] P i (t+1)=P i (t)+V i (t+1)

[0092] Among them, P i (t) represents the position of the i-th particle at the t-th time, P i (t+1) represents the position of particle i at time t+1, V i (t) represents the velocity of the ith particle at the tth time, V i(t+1) represents the velocity of particle i at time t+1, pBest i represents the best historical position of the ith particle, gBest represents the best historical position of the group, w, c1, c2 are weight coefficients, inertia weight, individual learning factor and social learning factor, respectively, r1 and r2 are random numbers in the interval [0,1];

[0093] The objective function of multi-objective optimization can be expressed as a composite function combining operation efficiency, energy consumption and operation quality:

[0094] f(P)=αE(P)+βC(P)+γQ(P)

[0095] Among them, f(P) represents a composite function that combines work efficiency, energy consumption and work quality, which is used to evaluate the overall performance of solution P, E(P) represents work efficiency, C(P) represents energy consumption, Q(P) represents work quality, and α, β and γ are the corresponding weight coefficients respectively.

[0096] Implementation Case 1

[0097] There is a designated plot of land, and it is necessary to plan the path for the unmanned agricultural machinery operation. In order to conduct specific numerical analysis:

[0098] Plot area: 10,000 square meters, machine width: 2 meters, number of obstacles: 5, obstacle positions: [(20,30), (50,50), (75,20), (30,70), (60,80)], the initial range of the plot is provided by satellite remote sensing images, processed using the image segmentation and fine-tuning module

[0099] Step 1: Specify the plot area

[0100] The boundary of the plot is obtained through remote sensing image processing. Assume that the initial boundary is a square of 100 meters by 100 meters. Image segmentation uses the neural network U-Net to segment the plot and refine the boundary. The following are the boundary points after the plot is segmented:

[0101] Boundary points = [(0,0), (0,100), (100,100), (100,0)]

[0102] Step 2: Image processing and boundary fine-tuning

[0103] Using the boundary extraction module and manual adjustment, the subdivided block boundary results are obtained. Assume that the boundary points after fine-tuning are:

[0104] Fine-tuning boundary points = [(1,1), (1,99), (99,99), (99,1)]

[0105] Step 3: Standardization of the work surface

[0106] The obtained boundary results are standardized, the skeleton data of the working surface is extracted, and it is divided into basic quadrilateral units. The standardized working surface data is:

[0107] Working surface quadrilateral element = [((1,1),(1,50),(50,50),(50,1)),…]

[0108] Step 4: Generate basic input data for path planning

[0109] The standardized work surface data is further processed to generate basic input data for path planning.

[0110] Step 5: Path Planning

[0111] Linear component example

[0112] The parameters of a line component are:

[0113] Starting point: (1,1), End point: (1,50), Direction: Vertical, Length: 49 meters, Adjacent relationship: Connected to the next straight line component;

[0114] Evaluation criteria: Coverage: 49m 2 ,Repeat degree: 0m 2 ,Movement cost: 49s / 49ml steering assembly example

[0115] The parameters of a steering component are: starting point: (1,50), end point: (2,50), steering method: 90 degree right turn, steering feasibility: high;

[0116] Evaluation criteria: Coverage: 1m 2 ,Repeat degree: 0m 2 ,Movement cost: 1s / 1ml;

[0117] Example of a fixed process component

[0118] The parameters of a fixed process component are: starting point: (2,50), end point: (2,99), direction: vertical, length: 49 meters, process feasibility: high;

[0119] Evaluation criteria: Coverage: 49m 2 ,Repeat degree: 0m 2 ,Movement cost: 1s / 1ml,Technology necessity: high;

[0120] Step 6: Global path generation

[0121] The components are combined to generate a global path using a neural network through path planning. The generated global path includes the following segments:

[0122] Global path = [(1,1)→(1,50),(1,50)→(2,50),(2,50)→(2,99),…]

[0123] Implementation Case 2

[0124] The unmanned agricultural machinery needs to perform path planning and operation in a simulated farmland. The environmental characteristics of the farmland include: size: 100 meters x 100 meters, terrain type: flat land, slope, obstacles (stones, trees), soil conditions: dry, wet;

[0125] We use a simulator to generate these environmental features and input them into the U-Net model for segmentation and classification.

[0126] U-Net model terrain segmentation, input: 100x100 farmland image;

[0127] U-Net output: Flat land: 60% (6000 pixels); Slope land: 30% (3000 pixels); Obstacles: 10% (1000 pixels);

[0128] Reinforcement Learning Optimization Module

[0129] State space S: terrain type (flat, slope, obstacle), current position, current direction, speed, action space A: forward, turn left, turn right, stop w1, w2 and w3 are 1, -0.5, 2 respectively;

[0130] Training of Deep Q-Network (DQN):

[0131] Generate initial data using the simulator: 100 iterations, 100 steps per iteration; experience replay buffer size: 1000; discount factor γ = 0.9; learning rate α = 0.001

[0132] Initial state: (flat ground, position (0,0), direction 0 degrees, speed 0)

[0133] Initial Action: Straight Line Component

[0134] Reward: Reward function R = proximity + coverage + repetition + motion cost + process necessity

[0135] After 100 iterations, the model performed well in the simulation environment, correctly identifying and avoiding obstacles while optimizing the operation path.

[0136] Multi-objective optimization algorithm (PSO): α = 1, β = -0.55, γ = 2

[0137] Particle swarm initialization: The initial particle positions and velocities are randomly distributed on the farmland. Each particle represents a path planning scheme. After 100 iterations, the optimal path is as follows: total coverage area: 95%, total path length: 500 meters, number of turns: 20 times;

[0138] Results analysis: Through the terrain segmentation of the U-Net model, the path planning of the reinforcement learning optimization module, and the multi-objective optimization algorithm of PSO, we achieved the following goals: efficiently segment and identify farmland terrain, accurately locate obstacles, and use the DQN algorithm to optimize path planning to achieve efficient coverage and obstacle avoidance. The PSO algorithm balances work efficiency, energy consumption, and work quality to generate the optimal work path.

Claims

1. A method for planning the path of an unmanned agricultural machine based on vehicle status and working environment, characterized in that: The following steps are involved: Step 1, specifying the range of the plot 100 according to the satellite remote sensing image; There are two triggering methods for this process: (1) specifying a location on the remote sensing map; (2) selecting a two-dimensional geometric area on the remote sensing map; Step 2, using image processing to fine-tune the specified range to generate a work surface 101; Step 3, using the work surface preprocessing function to standardize the work surface 102; The specific process of step 3 is: In the process of standardizing the working surface 102 in the working surface preprocessing function, the working surface simplification module 203 is first used to filter and delete points with less information in the boundary data to extract the skeleton optimized working surface data, and then the working surface segmentation and combination module 204 is used to segment the simplified data into basic triangles and then match them into optimal quadrilaterals, and possible concave polygons are segmented into multiple basic quadrilateral units. Finally, the working surface information extraction module 205 is used to perform a multi-dimensional standardized description of the features of the quadrilaterals; step 4, global path generation 103 is performed according to the path rules, motion model rules and process requirements; In step 4, the path rules are assembled by using prefabricated motion components to ensure fast and reliable path planning, and the components include a linear component 206, a steering component 207 and a fixed process component 208; The straight line component 206 includes five elements, namely, the starting point, the end point, the direction, the length, and the adjacency relationship, and four evaluation criteria, namely, the proximity relationship, the coverage degree, the repetition degree, and the motion cost, as indicators for generating a straight line path; The steering component 207 includes seven elements, namely, starting point, end point, direction, length, adjacency, steering method, and steering feasibility, and four evaluation criteria, namely, proximity, coverage, repetition, and motion cost, as indicators for generating a steering path; The fixed process component 208 includes eight elements, namely, starting point, end point, direction, length, adjacency, process method, process feasibility, and target object operation relationship. Five evaluation criteria, namely, proximity, coverage, repetition, motion cost, and process necessity, are used as indicators for generating a fixed process path. When different components are combined, the component combination module 209 will be responsible for the process of adjacency between different modules to smooth out the defects encountered during splicing to ensure kinematic feasibility; When different components encounter conflicts when combined, the conflict may be that the combination module reports an error due to inability to smoothly splice or that the proximity relationship, coverage, repetition, and motion cost evaluation standards in the planning process are unqualified. When a conflict occurs, the component conflict processing module 210 needs to be used to compromise, degrade, or even replace the corresponding components; Step 5, generating a local path according to the vehicle-mounted sensor 104; The specific process of step 5 is: During the execution process, the vehicle-mounted sensor 215 will continuously return obstacle information in front of and on the sides of the vehicle. The sensor usually uses a laser radar and a binocular depth-of-field camera as hardware. Its data is finally integrated into a spatial obstacle description. When an obstacle appears, the obstacle avoidance component 216 is executed. The component checks the execution component and its adjacent components of the global path planning result at the current time point, and uses a smooth path to modify the current component and its adjacent components to bypass the obstacle; Step 6, path execution ends 105.

2. The method for planning an unmanned agricultural machinery operation path based on vehicle status and working environment according to claim 1, characterized in that: The specific process of step 1 is: If the trigger method (1) is used according to the range 100 of the plot specified by the satellite remote sensing image, the system will read the remote sensing image 200 near the location, send it to the image segmentation neural network 201, segment it into different blocks, and send the segmented result block at the specified location as the preliminary result to the boundary extraction module 202 to obtain the subdivided block boundary result, and finally manually adjust the boundary of the extracted block to obtain the specified boundary result; If trigger method (2) is used to specify the range 100 of the plot based on the satellite remote sensing image, the system will use the selected area as the plot range result, and send the plot range result as the preliminary result to the boundary extraction module 202 to obtain the subdivided plot boundary result, and finally manually adjust the boundary of the block boundary to obtain the specified boundary result; if trigger method (2) is used to specify the range 100 of the plot based on the satellite remote sensing image and it is stated that no detail processing is required, the boundary of the range result is manually adjusted to obtain the specified boundary result.

3. The method for planning an unmanned agricultural machinery operation path based on vehicle status and working environment according to claim 2, characterized in that: The backbone network of the neural network 201 for image segmentation is formed by a convolutional neural network U-Net.

4. The method for planning an unmanned agricultural machinery operation path based on vehicle status and working environment according to claim 1, characterized in that: The specific process of step 4 is: After obtaining the optimal quadrilateral after splitting and its standardized description, continue to perform the process of global path generation 103 according to the path rules, motion model rules and process requirements, send each quadrilateral to the independent work surface path planning 211 for local optimal path planning, and the independent work surface path planning 211 uses the path planning neural network 213 to combine the above-mentioned straight line component 206 and the fixed process component 208 to obtain the optimal path result in the quadrilateral. If there is only one main quadrilateral, then splice the steering component 207 and call the component combination module 209 and the component conflict processing module 210 to calculate the global path, and the generation process ends; If there are multiple main quadrilaterals, the combined work surface connection method 212 is used to splice the work paths in the two planes using the steering component 207. If the splicing fails, the connection method will return to the independent work surface path planning 211. The optimal path result takes the second optimal path result until the global path is successfully spliced, and the generation process ends; After the generation process is completed, the traffic check 214 will be used to simulate the vehicle passing through the path process, and the difficulty of execution, the degree of coverage of the working surface, the degree of repetition, the movement cost, and the necessity of the process will be evaluated. The evaluation result will be used to execute the path of the subsequent machinery.

5. The method for planning an unmanned agricultural machinery operation path based on vehicle status and working environment according to claim 2, characterized in that: When the execution of the path in step 6 ends 105, The process includes: When the execution path is finished, the component fine-tuning module 217 records the five evaluation criteria of the actual proximity relationship, coverage, repetition, motion cost, and process necessity of each component in the execution process, obtains the actual performance of the component under different parameters, and adjusts the evaluation criteria of different component actions according to the learning rate r, so as to overcome the problem that the general model network cannot generate the optimal path due to the conflict of the inherent mechanical property factors of the vehicle or the environmental factors in actual problems; When the execution path is completed, the neural network training module 218 inputs the actual results of each vehicle's execution and the fine-tuning results of each component fine-tuning module 217 into the training data set of the path planning neural network 213, and regularly updates the path planning neural network 213 pre-model to optimize the design of the neural network in the actual use environment. The problem that the pre-model is not applicable in the actual environment can make the pre-model perform optimally in the real environment or quickly converge to the optimal range.

6. The method for planning the path of an unmanned agricultural machine operation based on vehicle status and working environment according to claim 3, characterized in that: The image segmentation neural network 201 uses a U-Net architecture, and the output of each layer is represented as follows: X i+1 =C i (D i (X i )) Among them, X i represents the encoder output feature map of the i-th layer, X i+1 represents the output feature map of the i+1th layer, C i represents the convolution operation of the i-th layer, D i represents the downsampling operation of the i-th layer; For the decoder part, add upsampling and skip connection feature map merging: Among them, U i represents the upsampling operation of the i-th layer, C i ′ represents the convolution operation of the i-th layer; X i ′ represents the output feature map of the i-th layer of the decoder, X n-i represents the output feature map of the encoder layer ni, X i ' +1 represents the output feature map of the decoder layer i+1, represents the concatenation operation of the feature map, and n is the total number of layers in the encoder; For binary image segmentation, the binary cross entropy loss function is used: Among them, L1 represents the value of the loss function, which is used to measure the difference between the model prediction and the true label. Indicates the summation of all pixels, y i is the true label of the i-th pixel, is the predicted probability of the ith pixel, N is the total number of pixels in the batch, Represents the predicted probability The natural logarithm of Represents the natural logarithm of the complement of the predicted probability.

7. The method for planning the path of an unmanned agricultural machine operation based on vehicle status and working environment according to claim 5, characterized in that: The boundary extraction module 202, the work surface simplification module 203, the work surface segmentation and combination module 204, the work surface information extraction module 205, the component combination module 209, the component conflict handling module 210, the passage inspection 214, the obstacle avoidance component 216, and the component fine-tuning module 217 are all implemented through software programming.

8. The method for planning the path of an unmanned agricultural machine operation based on vehicle status and working environment according to claim 5, characterized in that: The path planning neural network 213 specifically performs the following process: Step a: data collection and variable definition; Collect multimodal data: ground sensor data, vehicle mechanical parameter data and historical operation data; Define the environment and reward mechanism: S represents the state space, including the vehicle's motion state and the state of the relationship between the vehicle and the environment; A represents the action space, steering angle, acceleration; R is the reward function, the formula is as follows; R(S t ,a t )=w1r near (S t ,a t )+w2r corss (S t ,a t )+w3r repeat (S t ,a t )+w4r move (S t ,a t )+w5r process (S t ,a t ) Among them, R represents the reward function, which refers to the five evaluation criteria of proximity, coverage, repetition, motion cost, and process necessity; s t represents the state at time t; a t represents the action taken at time t; r near Represents the reward reflecting the proximity relationship; r cross represents the reward reflecting the coverage level; r repeat indicates a reward that reflects the degree of repetition; r move represents the reward reflecting the cost of movement; r process Indicates a reward that reflects the necessity of the process; w n Indicates the weight of each factor, n = 1, 2, ... 5; Step b: Deep Q network; Use a deep Q network model to estimate the Q value of each action taken in each state, which represents the long-term cumulative reward of taking this action in the current state; Q(s,a;θ) Among them, s represents the state, a represents the action, and θ represents the parameters of the network; The mean square error (MSE) is used as the loss function to measure the difference between the Q value output by the network and the target Q value. Where L is the loss function, N is the number of samples, Q(s,a;θ) is the network model output, which represents the Q value of taking action a in state s, Q′(s′,a′;θ′) is the target Q value, which represents the Q value of taking action a′ in the next state s′, θ represents the network parameter, θ′ represents the parameter of the target network, and γ represents the discount factor; Use stochastic gradient descent to minimize the loss function and update the network parameters θ: Among them, θ′ represents the updated parameter vector; θ represents the current parameter vector; α is the learning rate, which is used to control the step size of parameter update; Represents the gradient of the loss function L with respect to the parameter θ, that is, the rate of change of the loss function with respect to the parameter; The output of the deep Q network is integrated with the PSO algorithm to form the input of the operation path planning; the PSO optimization result is used to generate the final operation path, taking into account the terrain, obstacles and multi-objective optimization results; the speed and position update formula of its particles is: V i (t+1)=wV i (t)+c1r1(pBest i -P i (t))+c2r2(gBest-P i (t)) P i (t+1)=P i (t)+V i (t+1) Among them, P i (t) represents the position of the i-th particle at the t-th time, P i (t+1) represents the position of particle i at time t+1, V i (t) represents the velocity of the ith particle at the tth time, V i (t+1) represents the velocity of particle i at time t+1, pBest i represents the best historical position of the ith particle, gBest represents the best historical position of the group, w, c1, c2 are weight coefficients, inertia weight, individual learning factor and social learning factor, respectively, r1 and r2 are random numbers in the interval [0,1]; The objective function of multi-objective optimization can be expressed as a composite function combining operation efficiency, energy consumption and operation quality: f(P)=αE(P)+βC(P)+γQ(P) Among them, f(P) represents a composite function that combines work efficiency, energy consumption and work quality, which is used to evaluate the overall performance of solution P, E(P) represents work efficiency, C(P) represents energy consumption, Q(P) represents work quality, and α, β and γ are the corresponding weight coefficients respectively.

Images