Agricultural machine operation path planning method and device based on multi-channel grid perception and reinforcement learning, equipment and medium

Abstract

The application relates to a farm machine operation path planning method and device based on multi-channel grid perception and reinforcement learning, equipment and medium, the method comprising: acquiring a target farmland remote sensing image, identifying passable areas, crops and obstacle areas, and target areas to be operated with poor growth, diseases and insect pests or water and fertilizer shortages through semantic segmentation; constructing a two-dimensional grid map by griding the three types of areas, separating obstacle channels and task channels through semantic decoupling, and generating a multi-channel grid perception matrix; constructing a time sequence feature, extracting grid space features through a convolutional neural network, encoding farm machine motion features through a multilayer perception machine, splicing to generate multi-modal fusion state features; inputting the fusion features into a reinforcement learning strategy network, iteratively optimizing based on a multi-objective reward function, outputting farm machine speed and steering continuous control quantities, and generating an operation path covering the target areas. The application can solve the problems of insufficient farm machine path planning precision and limited operation modes.

Description

Technical Field

[0001] This application relates to the field of agricultural machinery path planning, and in particular to an agricultural machinery operation path planning method, corresponding device, electronic device and computer-readable storage medium based on multi-channel grid perception and reinforcement learning. Background Technology

[0002] With the rapid development of smart agriculture and agricultural automation technology, unmanned agricultural machinery is increasingly widely used in operation scenarios such as sowing, fertilization, plant protection spraying, and farmland inspection. As the core key technology for unmanned agricultural machinery to achieve autonomous operation, the quality of its solution directly determines the operation efficiency, coverage quality and operation safety of the agricultural machinery, and is the core support for achieving precise variable operation in farmland.

[0003] Currently, the mainstream path planning methods in the field of agricultural machinery autonomous driving are mainly divided into two categories: AB line planning method and boundary planning method. Among them, the AB line planning method can quickly generate a single straight line reference path by selecting two reference points A and B within the farmland plot, and the operation threshold is relatively low; however, the path form of this method is fixed, and it is only suitable for regular plots without obstacles and without differentiated operation requirements. It is difficult to adapt to complex farmland operation scenarios with obstacles and the need to target and cover specific areas, and its scenario adaptability is extremely poor.

[0004] Boundary planning methods first obtain the boundary outline of farmland, and then generate reciprocating and other full-coverage operation paths within the boundary according to preset rules. Although this method can achieve full coverage of the plot, it has poor adaptability to irregular farmland plots. In farmland environments with uneven crop growth and complex obstacle distribution, it generally suffers from defects such as high path repetition rate, low operation efficiency, frequent turns, and large wear and tear on agricultural machinery chassis, and cannot meet the needs of high-quality and precise operation.

[0005] More importantly, the existing path planning methods still heavily rely on manually marking coordinate points or recording the work path on-board to establish the work trajectory. Operators need to manually record farmland boundary points, work reference points, or critical path nodes through onboard terminals and GPS devices, and then generate the work path based on the manually recorded coordinates. This work mode has obvious technical shortcomings: the manual marking process is time-consuming and labor-intensive, with a very low degree of automation. This not only significantly increases the preparation cost before the operation, but also the accuracy of manually marked coordinates is easily affected by human operation errors and farmland environmental interference, making it difficult to meet the high precision and high efficiency requirements of unmanned agricultural machinery for autonomous operation. This seriously restricts the large-scale promotion and application of unmanned agricultural machinery in complex farmland scenarios.

[0006] In summary, existing AB line planning methods are ill-suited for complex farmland operation scenarios with obstacles and the need to target specific areas, exhibiting extremely poor scenario adaptability. Furthermore, boundary planning methods suffer from drawbacks such as high path repetition rates, low operational efficiency, frequent turns, and significant wear and tear on agricultural machinery chassis. Therefore, the applicant has made corresponding explorations to address these issues. Summary of the Invention

[0007] The purpose of this application is to solve the above problems by providing a method for agricultural machinery operation path planning based on multi-channel grid perception and reinforcement learning, as well as corresponding devices, electronic devices and computer-readable storage media.

[0008] To achieve the various objectives of this application, the following technical solution is adopted:

[0009] A method for agricultural machinery operation path planning based on multi-channel grid perception and reinforcement learning, proposed to meet one of the purposes of this application, includes:

[0010] Acquire remote sensing images of farmland corresponding to the target farmland, and perform semantic segmentation on the farmland remote sensing images to determine the passable area, crop and obstacle area and target task area in the target farmland. The target task area represents local crop areas with poor growth, pests and diseases, or lack of water and fertilizer.

[0011] The passable area, the crop and obstacle area and the target task area are rasterized to construct a two-dimensional raster map of farmland corresponding to the target farmland. The two-dimensional raster map of farmland is semantically decoupled to separate and generate obstacle passage raster and task passage raster to construct a multi-channel raster perception matrix.

[0012] The motion state vector corresponding to the target agricultural machinery is collected to construct a time-series state feature vector containing historical time-series information. A preset convolutional neural network is used to perform convolutional pooling feature extraction on the multi-channel grid perception matrix to output a grid space feature vector.

[0013] A multilayer perceptron is used to map and encode the temporal state feature vector to output the agricultural machinery motion feature vector. The grid space feature vector and the agricultural machinery motion feature vector are then concatenated and fused to generate multimodal fused state features.

[0014] The multimodal fusion state features are input into a reinforcement learning policy network, and iterative decision-making is performed based on a preset multi-objective reward function. The output is a continuous control action that includes agricultural machinery speed control and agricultural machinery steering control. While avoiding obstacles, an agricultural machinery operation path covering the target task area is generated to complete the agricultural machinery operation path planning based on multi-channel grid perception and reinforcement learning.

[0015] Optionally, the steps of rasterizing the passable area, the crop and obstacle area, and the target task area to construct a two-dimensional raster map of farmland corresponding to the target farmland, semantically decoupling the two-dimensional raster map of farmland, and separating and generating obstacle passage raster and task passage raster to construct a multi-channel raster perception matrix include:

[0016] The passable area of ​​the target farmland, the crop and obstacle area, and the target task area are rasterized to construct a dimension of [missing information]. A two-dimensional raster map of farmland, in which, This indicates the vertical number of raster rows in a two-dimensional raster map of farmland. This indicates the number of horizontal grid columns in the two-dimensional farmland grid map. The value of each grid cell in the two-dimensional farmland grid map is defined as follows: 0 represents a passable area, 0.5 represents a target task area, and 1 represents a crop and obstacle area.

[0017] Semantic decoupling is performed on the farmland 2D grid map to separate and generate obstacle passage grids and task passage grids. The obstacle passage grid retains only the positions with a value of 1 in the farmland 2D grid map, and the task passage grid retains only the positions with a value of 0.5 in the farmland 2D grid map.

[0018] The obstacle channel grid and the task channel grid are stacked to construct a dimension of [dimensional value]. A multi-channel raster sensing matrix.

[0019] Optionally, the steps of collecting the motion state vector corresponding to the target agricultural machinery to construct a temporal state feature vector containing historical time-series information, and using a preset convolutional neural network to perform convolutional pooling feature extraction on the multi-channel raster sensing matrix to output the raster space feature vector include:

[0020] Collect the motion state vector of the target agricultural machinery, including the machinery's position coordinates, speed, heading angle, and work endpoint coordinates;

[0021] Based on a sliding time window of a preset time length, the motion state vectors of multiple consecutive historical moments are spliced ​​together in time sequence to construct a time sequence state feature vector containing historical time sequence information.

[0022] Using a multi-channel grid sensing matrix as input to a convolutional neural network, the multi-channel grid sensing matrix is ​​convolved through a first convolutional layer and then activated by an activation function to obtain a first convolutional feature map.

[0023] After pooling the first convolutional feature map, convolution is performed through the second convolutional layer, and then activated by the activation function to obtain the second convolutional feature map.

[0024] The second convolutional feature map is flattened and then mapped through a fully connected layer to output a raster space feature vector.

[0025] Optionally, the step of mapping and encoding the temporal state feature vector using a multilayer perceptron to output the agricultural machinery motion feature vector, and concatenating and fusing the grid space feature vector with the agricultural machinery motion feature vector to generate multimodal fused state features includes:

[0026] The temporal state feature vector is used as the input of the multilayer perceptron, and its temporal state feature vector is flattened to obtain a one-dimensional flattened feature vector.

[0027] The one-dimensional flattened feature vector is mapped and encoded based on the multilayer perceptron. The one-dimensional flattened feature vector is linearly transformed by the first weight matrix and the first bias vector, and then activated by the activation function to obtain the first encoded feature.

[0028] The first encoded feature is linearly transformed by the second weight matrix and the second bias vector, and then activated by the activation function to obtain the second encoded feature, which is then used as the agricultural machinery motion feature vector.

[0029] The agricultural machinery motion feature vector and the grid space feature vector are concatenated and fused along the channel dimension to generate a multimodal fusion state feature for input to the reinforcement learning strategy network.

[0030] Optionally, the steps of inputting the multimodal fused state features into a reinforcement learning policy network, performing iterative decision-making based on a preset multi-objective reward function, and outputting continuous control actions including agricultural machinery speed control and agricultural machinery steering control include:

[0031] The multimodal fusion state features are subjected to fusion mapping operation. A linear transformation is performed on the fusion weight matrix and the fusion bias vector, and then a nonlinear activation is performed through the activation function to generate a unified state feature vector.

[0032] The unified state feature vector is input into the policy distribution function, and iterative decision-making is performed through a reinforcement learning policy network to output continuous control actions containing agricultural machinery speed control and agricultural machinery steering control. The policy distribution function is used to generate the probability distribution of agricultural machinery speed control and agricultural machinery steering control based on the current unified state feature vector.

[0033] Optionally, the expression for the multi-objective reward function is as follows:

[0034] ,

[0035] in, A reward is given for coverage of the work area when agricultural machinery enters an uncovered work area; As a penalty for repeated operations, a penalty is imposed when agricultural machinery repeatedly visits an already covered area; As a collision penalty, a penalty is imposed when the agricultural machinery collides with an obstacle or the boundary of the farmland; As a path efficiency reward, when a certain amount of the coverage task is completed, the agricultural machinery is given a reward for moving towards the destination; The reward is for completing the task, and an additional reward is given when the agricultural machinery completes the task of covering the entire work area.

[0036] Optionally, the steps of inputting the multimodal fused state features into a reinforcement learning policy network, performing iterative decision-making based on a preset multi-objective reward function, and outputting continuous control actions including agricultural machinery speed control and agricultural machinery steering control include:

[0037] Based on the multi-objective reward function, the reward for each step of the agricultural machinery operation is calculated in real time. When the agricultural machinery visits an uncovered operation area for the first time, a positive reward is given, and a penalty is given when it repeatedly visits an covered area.

[0038] When the agricultural machinery's steering action is closer to straight-line travel, a positive reward is given to reduce drastic control input; when the agricultural machinery goes beyond the map boundary, a boundary collision penalty is given; when the agricultural machinery collides with crops or obstacles, a heavier obstacle collision penalty is given.

[0039] When the coverage rate of the operation exceeds the first preset threshold and the agricultural machinery moves toward the finish line and the distance is shortened, a positive incentive reward is given; a penalty is imposed for each step the agricultural machinery takes to encourage it to complete the task as soon as possible.

[0040] A penalty is imposed when agricultural machinery fails to exit prematurely or reaches the destination but the coverage rate is lower than the second preset threshold; a high coverage rate is awarded when agricultural machinery reaches the destination and the coverage rate is not lower than the second preset threshold.

[0041] Based on the real-time reward value calculated at each step, the reinforcement learning policy network is iteratively optimized to output a continuous control action that includes agricultural machinery speed control and agricultural machinery steering control, which meets the requirements of full operation coverage, obstacle avoidance, and path smoothness.

[0042] A device for agricultural machinery operation path planning based on multi-channel grid perception and reinforcement learning, provided to suit another purpose of this application, includes:

[0043] The semantic segmentation module is configured to acquire remote sensing images of farmland corresponding to the target farmland, and perform semantic segmentation on the farmland remote sensing images to determine the passable area, crop and obstacle area and target task area in the target farmland. The target task area represents local crop areas with poor growth, pests and diseases, or lack of water and fertilizer.

[0044] The grid map construction module is configured to rasterize the passable area, the crop and obstacle area and the target task area to construct a two-dimensional grid map of farmland corresponding to the target farmland, and to semantically decouple the two-dimensional grid map of farmland to separate and generate obstacle passage grid and task passage grid to construct a multi-channel grid perception matrix.

[0045] The dual-modal feature encoding module is configured to collect the motion state vector corresponding to the target agricultural machinery to construct a temporal state feature vector containing historical time series information. A preset convolutional neural network is used to perform convolutional pooling feature extraction on the multi-channel grid perception matrix to output a grid space feature vector.

[0046] The multimodal feature fusion module is configured to use a multilayer perceptron to map and encode the temporal state feature vector to output the agricultural machinery motion feature vector, and to concatenate and fuse the grid space feature vector with the agricultural machinery motion feature vector to generate multimodal fused state features.

[0047] The path planning decision module is configured to input the multimodal fusion state features into the reinforcement learning policy network, perform iterative decision-making based on a preset multi-objective reward function, and output continuous control actions including agricultural machinery speed control and agricultural machinery steering control. While avoiding obstacles, it generates an agricultural machinery operation path covering the target task area to complete the agricultural machinery operation path planning based on multi-channel grid perception and reinforcement learning.

[0048] An electronic device provided for another purpose of this application includes a central processing unit and a memory, the central processing unit being configured to invoke and run a computer program stored in the memory to perform the steps of the agricultural machinery operation path planning method based on multi-channel grid perception and reinforcement learning described in this application.

[0049] A computer-readable storage medium is provided for another purpose of this application, which stores, in the form of computer-readable instructions, a computer program implemented according to the agricultural machinery operation path planning method based on multi-channel grid perception and reinforcement learning, which, when called by a computer, executes the steps included in the corresponding method.

[0050] Compared to existing technologies, this application addresses the shortcomings of existing technologies, such as the AB line planning method's poor adaptability to complex farmland operation scenarios with obstacles and the need for targeted coverage of specific areas, and the high path repetition rate, low operation efficiency, frequent turning, and high wear and tear on agricultural machinery chassis caused by boundary planning methods. This application provides the following beneficial effects, including but not limited to:

[0051] Firstly, this application automatically distinguishes passable areas, crop obstacle areas, and targeted areas for operation in farmland by combining farmland remote sensing images with semantic segmentation and coordinate mapping. This eliminates the need for operators to manually mark boundary points and operation nodes, significantly reducing the workload of preliminary operation preparation, effectively reducing labor costs, and improving the automation level of farmland environmental information acquisition and overall operation preparation efficiency.

[0052] Secondly, after the farmland area is rasterized, semantic decoupling is performed, and obstacle channels and task channels are separated and a multi-channel raster perception matrix is ​​constructed. This can accurately represent the complex farmland terrain, obstacle distribution and target operation area distribution, breaking through the limitations of traditional AB line planning which is only applicable to regular plots and boundary planning which is difficult to adapt to irregular farmland, and has stronger environmental adaptability.

[0053] Thirdly, on the one hand, convolutional neural networks are used to extract the spatial features of the farmland grid environment, and on the other hand, the historical movement time series information of agricultural machinery is combined to construct time series features and complete the encoding fusion. This can take into account both the external farmland environment status and the agricultural machinery's own driving movement status. Compared with a single state input method, the state representation dimensions are richer, providing reliable and complete data support for intelligent decision-making.

[0054] Fourth, based on a multi-objective reward function, the strategy is iteratively optimized from multiple dimensions such as regional coverage, driving stability, obstacle avoidance and protection, operation efficiency, and task completion quality. The output of agricultural machinery speed and steering continuous control quantities can not only accurately avoid obstacles in the field, but also prioritize the completion of targeted crop area coverage operations, effectively reduce invalid driving paths and frequent turning phenomena, reduce agricultural machinery wear and tear, and at the same time significantly improve the rationality of path planning and the quality of precision farmland operations.

[0055] Furthermore, this application realizes a fully automated process from farmland environment perception, feature extraction and fusion to reinforcement learning autonomous decision-making, without the need for human intervention in path planning. It can meet the needs of various unmanned agricultural machinery operations such as plant protection spraying, precision fertilization, and farmland inspection, effectively solving the problems of insufficient accuracy and limited operation modes in traditional planning methods, and facilitating the large-scale deployment and promotion of unmanned agricultural machinery in complex field scenarios. Attached Figure Description

[0056] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0057] Figure 1 This is a flowchart illustrating the agricultural machinery operation path planning method based on multi-channel grid perception and reinforcement learning in the embodiments of this application.

[0058] Figure 2 This is a system architecture diagram of unmanned agricultural machinery path planning based on remote sensing imagery and reinforcement learning in the embodiments of this application;

[0059] Figure 3 This is a block diagram illustrating the environment, agent, and training process of the reinforcement learning policy network in an embodiment of this application.

[0060] Figure 4 The figure shows the simulation results of the coverage process of the unmanned agricultural machinery targeted operation path planning in the embodiments of this application;

[0061] Figure 5 This is a schematic diagram of the agricultural machinery operation path planning device based on multi-channel grid perception and reinforcement learning in the embodiments of this application;

[0062] Figure 6 This is a schematic diagram of the structure of the computer device in the embodiments of this application. Detailed Implementation

[0063] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0064] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0065] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0066] Those skilled in the art will understand that the terms "client," "terminal," and "terminal device" as used herein include both devices that receive wireless signals, devices that only possess wireless signal receiver capabilities without transmission capabilities, and devices with receiving and transmitting hardware, devices that have receiving and transmitting hardware capable of bidirectional communication over a bidirectional communication link. Such devices may include: cellular or other communication devices such as personal computers or tablets, having single-line displays, multi-line displays, or cellular or other communication devices without multi-line displays; PCS (Personal Communications Service) that can combine voice, data processing, fax, and / or data communication capabilities; PDAs (Personal Digital Assistants) that may include radio frequency receivers, pagers, internet / intranet access, web browsers, notebooks, calendars, and / or GPS (Global Positioning System) receivers; and conventional laptops and / or handheld computers or other devices that have and / or include radio frequency receivers. As used herein, "client," "terminal," and "terminal device" can be portable, transportable, installed in a means of transportation (air, sea, and / or land), or suitable and / or configured to operate locally and / or in a distributed manner, operating in any other location on Earth and / or in space. "Client," "terminal," and "terminal device" as used herein can also be a communication terminal, an internet access terminal, or a music / video playback terminal, such as a PDA, a MID (Mobile Internet Device), and / or a mobile phone with music / video playback capabilities, or a smart TV, set-top box, etc.

[0067] The hardware referred to by the names "server," "client," and "service node" in this application is essentially an electronic device with the equivalent capabilities of a personal computer. It is a hardware device with the necessary components revealed by the von Neumann architecture, such as a central processing unit (including an arithmetic logic unit and a control unit), memory, input devices, and output devices. The computer program is stored in its memory, and the central processing unit loads the program stored in the secondary storage into the main memory to run it, execute the instructions in the program, and interact with the input and output devices to complete specific functions.

[0068] It should be noted that the concept of "server" used in this application can also be extended to the case of server clusters. Based on the network deployment principles understood by those skilled in the art, the servers should be logically divided. Physically, these servers can be independent of each other but accessible through interfaces, or they can be integrated into a single physical computer or a computer cluster. Those skilled in the art should understand this flexibility and should not use it to constrain the implementation of the network deployment method in this application.

[0069] One or more of the technical features of this application, unless explicitly specified herein, can be deployed on a server and accessed by a client remotely calling the online service interface provided by the server, or can be directly deployed and run on a client for access.

[0070] Unless otherwise specified, the neural network models referenced or potentially referenced in this application may be deployed on a remote server and invoked remotely on the client, or deployed on a client with the capability to invoke directly. In some embodiments, when running on the client, the corresponding intelligence may be acquired through transfer learning in order to reduce the requirements on the client's hardware resources and avoid excessive consumption of the client's hardware resources.

[0071] Unless otherwise specified, all data involved in this application may be stored remotely on a server or on a local terminal device, as long as it is suitable for use by the technical solution of this application.

[0072] Those skilled in the art will understand that although the various methods in this application are described based on the same concept and thus present commonality among them, they can be performed independently unless otherwise specified. Similarly, the various embodiments disclosed in this application are all based on the same inventive concept; therefore, concepts expressed in the same way, as well as concepts that are appropriately changed for convenience but are expressed differently, should be understood equivalently.

[0073] Unless otherwise expressly stated, the various embodiments disclosed in this application can be combined in a cross-cutting manner to flexibly construct new embodiments, as long as such combination does not depart from the inventive spirit of this application and can meet the needs of the prior art or solve a certain deficiency in the prior art. Those skilled in the art should be aware of such modifications.

[0074] Please see Figures 1 to 4 In one embodiment of the agricultural machinery operation path planning method based on multi-channel grid perception and reinforcement learning of this application, the method includes:

[0075] Step S10: Obtain the farmland remote sensing image corresponding to the target farmland, and perform semantic segmentation on the farmland remote sensing image to determine the passable area, crop and obstacle area and target task area in the target farmland. The target task area represents the local crop area with poor growth, pests and diseases, or lack of water and fertilizer.

[0076] The agricultural machinery operation path planning system in the terminal device can acquire remote sensing images of farmland corresponding to the target farmland, perform semantic segmentation on the farmland remote sensing images, and determine the passable area, crop and obstacle area and target task area in the target farmland. The target task area represents local crop areas with poor growth, pests and diseases, or lack of water and fertilizer. The agricultural machinery is unmanned agricultural machinery.

[0077] Specifically, a YOLO instance segmentation model is trained by constructing a farmland dataset using farmland remote sensing images. This enables the automatic extraction of passable areas, crop and obstacle areas, and target task areas in the target farmland. Furthermore, reinforcement learning algorithms are used to adaptively optimize the path decision-making process of unmanned agricultural machinery in complex environments, thereby achieving autonomous generation of operation paths.

[0078] High-precision semantic perception and coordinate mapping based on deep learning. This application utilizes UAVs to fly low-altitude, full-coverage routes along an "S" shaped path to acquire orthorectified remote sensing images containing rich ground feature information, constructing a dataset. Semantic feature extraction: YOLO models and other deep learning instance segmentation networks are used to transform unstructured image pixels into structured semantic polygons. Through supervised learning on massive offline multi-scene datasets, the model can accurately extract farmland boundaries and remove non-operational areas such as obstacles and roads. After extracting the pixel range of farmland and obstacles, combined with the geographic information inherent in the UAV remote sensing imagery, the two-dimensional pixel coordinates of the image are accurately mapped to real-world geographic coordinates (such as WGS84 latitude and longitude coordinates) through affine transformation, thus providing an absolute position reference for the subsequent actual physical navigation of agricultural machinery.

[0079] This application uses crop growth status as a crucial constraint in path planning, achieving a leap from "blind, full-coverage operation" to "precise, variable-based operation." During the remote sensing image processing stage, monitoring not only identifies the crop range but also assesses crop health by extracting spectral features, texture features, or color indices (such as the visible light vegetation index) from the images. Targeted operation areas are delineated; the system extracts localized crop areas with poor growth, pests and diseases, or water and fertilizer deficiencies as "designated operation areas" (i.e., prescription maps) by setting thresholds.

[0080] Step S20: Rasterize the passable area, the crop and obstacle area and the target task area to construct a two-dimensional raster map of farmland corresponding to the target farmland. Semantically decouple the two-dimensional raster map of farmland and separate and generate obstacle passage raster and task passage raster to construct a multi-channel raster perception matrix.

[0081] Remote sensing images of the target farmland are acquired, and semantic segmentation is performed on the farmland remote sensing images to determine the passable area, crop and obstacle area, and target task area in the target farmland. Then, the passable area, crop and obstacle area, and target task area are rasterized to construct a two-dimensional raster map of the farmland corresponding to the target farmland. Semantic decoupling is performed on the two-dimensional raster map of the farmland to separate and generate obstacle channel raster and task channel raster to construct a multi-channel raster perception matrix.

[0082] In some embodiments, the steps of rasterizing the passable area, the crop and obstacle area, and the target task area to construct a two-dimensional raster map of farmland corresponding to the target farmland, semantically decoupling the two-dimensional raster map of farmland, and separating and generating obstacle passage raster and task passage raster to construct a multi-channel raster perception matrix include:

[0083] Step S201: Rasterize the passable area, crop and obstacle area, and target task area of ​​the target farmland, constructing a dimension of... A two-dimensional raster map of farmland, in which, This indicates the vertical number of raster rows in a two-dimensional raster map of farmland. The value represents the number of horizontal grid columns in the two-dimensional grid map of farmland. Each grid cell in the two-dimensional grid map of farmland is defined as follows: a value of 0 represents a passable area, a value of 0.5 represents a target task area, and a value of 1 represents a crop and obstacle area.

[0084] Step S202: Semantically decouple the farmland two-dimensional grid map and separate and generate obstacle passage grid and task passage grid. The obstacle passage grid retains only the positions with a value of 1 in the farmland two-dimensional grid map, and the task passage grid retains only the positions with a value of 0.5 in the farmland two-dimensional grid map.

[0085] Step S203: Stack the obstacle channel grid and the task channel grid to construct a dimension of A multi-channel raster sensing matrix.

[0086] Specifically, the passable area of ​​the target farmland, the crop and obstacle area, and the target task area are rasterized to construct a dimension of Two-dimensional raster map of farmland ,in, This indicates the vertical number of raster rows in a two-dimensional raster map of farmland. The value represents the number of horizontal grid columns in the two-dimensional grid map of farmland. Each grid cell in the two-dimensional grid map of farmland is defined as follows: a value of 0 represents a passable area, a value of 0.5 represents a target task area, and a value of 1 represents a crop and obstacle area.

[0087] The two-dimensional raster map of the farmland Perform semantic decoupling and split into multiple semantic channels. ,in, Represents an obstacle passage grid. Represents the task channel grid;

[0088] Furthermore, obstacle passage grid The rules for determining the value of are expressed as follows:

[0089]

[0090] That is, only the obstacle positions with a value of 1 in the two-dimensional raster map of farmland are retained, and all other positions are uniformly set to 0.

[0091] Mission Channel Grid The rules for determining the value of are expressed as follows:

[0092]

[0093] That is, only the target task area with a value of 0.5 is retained in the 2D raster map of farmland, and all other locations are uniformly set to 0.

[0094] The obstacle channel grid With the task channel grid By stacking the channels, a structure with dimension [value] is constructed. Multichannel raster sensing matrix .

[0095] Furthermore, after the agricultural machinery covers the target task area, the value of the corresponding grid cell is updated from 0.5 to 0, which is represented as... The task channel grid automatically and synchronously changes (the point disappears) as the grid map is updated, realizing the dynamic update of the input of the convolutional neural network.

[0096] Step S30: Collect the motion state vector corresponding to the target agricultural machinery to construct a time-series state feature vector containing historical time-series information. Use a preset convolutional neural network to perform convolutional pooling feature extraction on the multi-channel grid perception matrix and output the grid space feature vector.

[0097] The passable area, the crop and obstacle area, and the target task area are rasterized to construct a two-dimensional raster map of the target farmland. The two-dimensional raster map of the farmland is semantically decoupled to separate and generate obstacle passage raster and task passage raster. After constructing a multi-channel raster perception matrix, the motion state vector corresponding to the target agricultural machinery is collected to construct a temporal state feature vector containing historical time series information. A preset convolutional neural network is used to perform convolutional pooling feature extraction on the multi-channel raster perception matrix to output the raster space feature vector.

[0098] In some embodiments, the steps of collecting the motion state vector corresponding to the target agricultural machinery to construct a temporal state feature vector containing historical time-series information, and using a preset convolutional neural network to perform convolutional pooling feature extraction on the multi-channel raster perception matrix to output a raster space feature vector include:

[0099] Step S301: Collect the motion state vector of the target agricultural machinery, including the machine's position coordinates, speed, heading angle, and work endpoint coordinates;

[0100] The motion state vector is represented as: ,in, express The motion state vector of the agricultural machinery at any given moment; express The horizontal coordinate of the agricultural machinery's position at any given time; express The vertical coordinate of the agricultural machinery's position at any given time; express The speed of the agricultural machinery at all times; express The heading angle of the agricultural machinery at all times; The x-coordinate of the coordinate system representing the endpoint of the agricultural machinery's operation; The vertical coordinate of the agricultural machinery's endpoint coordinates;

[0101] Step S302: Based on a sliding time window of a preset time length, the motion state vectors of multiple consecutive historical moments are spliced ​​together in time sequence to construct a time sequence state feature vector containing historical time sequence information.

[0102] Based on time length The sliding time window will continuously The motion state vector at each historical moment Perform time-series concatenation to construct a time-series state feature vector containing historical time-series information. .

[0103] Step S303: Using the multi-channel grid sensing matrix as the input of the convolutional neural network, the multi-channel grid sensing matrix is ​​convolved through the first convolutional layer, and then activated by the activation function to obtain the first convolutional feature map.

[0104] Multi-channel raster sensing matrix As input to a convolutional neural network (CNN), it passes through the first convolutional layer. Perform convolution operation on the multi-channel raster sensing matrix to obtain Then activated by the activation function Activation To determine the first convolutional feature map Its expression is .

[0105] Step S304: After performing pooling operation on the first convolutional feature map, convolution operation is performed through the second convolutional layer, and then activated by the activation function to obtain the second convolutional feature map;

[0106] For the first convolutional feature map Perform pooling operations Then, through the second convolutional layer Perform convolution operation to obtain Then activated by the activation function Activate to obtain To determine the second convolutional feature map Its expression is .

[0107] Step S305: Flatten the second convolutional feature map, and then perform feature mapping through a fully connected layer to output a grid space feature vector.

[0108] For the second convolutional feature map Flattening operation to obtain Then through the fully connected layer Feature mapping is performed to obtain In order to determine the raster space feature vector Its expression is .

[0109] Step S40: Use a multilayer perceptron to map and encode the temporal state feature vector to output the agricultural machinery motion feature vector, and then concatenate and fuse the grid space feature vector with the agricultural machinery motion feature vector to generate multimodal fusion state features.

[0110] The motion state vector corresponding to the target agricultural machinery is collected to construct a temporal state feature vector containing historical time-series information. A preset convolutional neural network is used to perform convolutional pooling feature extraction on the multi-channel grid perception matrix. After outputting the grid space feature vector, a multilayer perceptron is used to map and encode the temporal state feature vector to output the agricultural machinery motion feature vector. The grid space feature vector and the agricultural machinery motion feature vector are concatenated and fused to generate multimodal fused state features.

[0111] In some embodiments, the step of mapping and encoding the temporal state feature vector using a multilayer perceptron to output an agricultural machinery motion feature vector, and concatenating and fusing the grid space feature vector with the agricultural machinery motion feature vector to generate a multimodal fused state feature includes:

[0112] Step S401: Use the time-series state feature vector as the input of the multilayer perceptron, and flatten the time-series state feature vector to obtain a one-dimensional flattened feature vector.

[0113] Time-series state feature vectors As input to the multilayer perceptron, its temporal state feature vector is flattened. This yields a one-dimensional flattened eigenvector. .

[0114] Step S402: Based on the multilayer perceptron, the one-dimensional flattened feature vector is mapped and encoded. The one-dimensional flattened feature vector is linearly transformed by the first weight matrix and the first bias vector, and then activated by the activation function to obtain the first encoded feature.

[0115] Based on the multilayer perceptron, the one-dimensional flattened feature vector is processed... Perform mapping encoding using the first weight matrix. With the first bias vector For the one-dimensional flattened feature vector Perform a linear transformation to determine Then activated by the activation function Activate to obtain To determine the first coding feature Its expression is ,in, .

[0116] Step S403: The first encoded feature is linearly transformed by the second weight matrix and the second bias vector, and then activated by the activation function to obtain the second encoded feature. The second encoded feature is used as the agricultural machinery motion feature vector.

[0117] Through the second weight matrix With the second bias vector For the first encoded feature Perform a linear transformation to determine Then activated by the activation function Activate to determine The second coding feature is obtained. The second encoded feature As a feature vector of agricultural machinery motion ,in,

[0118] , ,in, ;

[0119] Step S404: The agricultural machinery motion feature vector and the grid space feature vector are concatenated and fused along the channel dimension to generate a multimodal fusion state feature for input to the reinforcement learning strategy network.

[0120] The agricultural machinery motion feature vector With grid space feature vector Perform channel-dimensional splicing and fusion to determine Generate multimodal fusion state features for input to reinforcement learning policy networks. ,in, .

[0121] Step S50: Input the multimodal fusion state features into the reinforcement learning policy network, perform iterative decision-making based on the preset multi-objective reward function, and output continuous control actions including agricultural machinery speed control and agricultural machinery steering control. While avoiding obstacles, generate an agricultural machinery operation path covering the target task area to complete the agricultural machinery operation path planning based on multi-channel grid perception and reinforcement learning.

[0122] A multilayer perceptron is used to map and encode the temporal state feature vector to output the agricultural machinery motion feature vector. The grid space feature vector and the agricultural machinery motion feature vector are then concatenated and fused to generate a multimodal fused state feature. This multimodal fused state feature is then input into a reinforcement learning policy network. Iterative decision-making is performed based on a preset multi-objective reward function, and continuous control actions including agricultural machinery speed control and agricultural machinery steering control are output. While avoiding obstacles, an agricultural machinery operation path covering the target task area is generated to complete the agricultural machinery operation path planning based on multi-channel grid perception and reinforcement learning.

[0123] In some embodiments, the steps of inputting the multimodal fused state features into a reinforcement learning policy network, performing iterative decision-making based on a preset multi-objective reward function, and outputting continuous control actions including agricultural machinery speed control and agricultural machinery steering control include:

[0124] Step S501: Perform fusion mapping operation on the multimodal fusion state features, and perform linear transformation through the fusion weight matrix and fusion bias vector, and then perform nonlinear activation through the activation function to generate a unified state feature vector.

[0125] For the multimodal fusion state features Perform fusion mapping operations using the fusion weight matrix. With fusion bias vector Perform a linear transformation to determine Then activated by the activation function Perform nonlinear activation to determine Generate a unified state feature vector ,in, .

[0126] Step S502: Input the unified state feature vector into the policy distribution function, perform iterative decision-making through the reinforcement learning policy network, and output continuous control actions including agricultural machinery speed control quantity and agricultural machinery steering control quantity. The policy distribution function is used to generate the probability distribution of agricultural machinery speed control quantity and agricultural machinery steering control quantity based on the current unified state feature vector.

[0127] The unified state feature vector Input to policy distribution function In this process, iterative decision-making is carried out through a reinforcement learning policy network, outputting continuous control actions that include agricultural machinery speed control and agricultural machinery steering control. The strategy distribution function is used to generate the probability distribution of agricultural machinery speed control quantity and agricultural machinery steering control quantity based on the current unified state feature vector.

[0128] In some embodiments, a two-dimensional raster map of farmland can be represented as:

[0129]

[0130] in, Indicates the first Each grid cell.

[0131] The state definition of the two-dimensional raster map of farmland is as follows:

[0132]

[0133] Furthermore, this application abandons traditional heuristic algorithms (such as A* algorithm and Dijkstra's algorithm) and adopts a reinforcement learning mechanism for global path planning. The reinforcement learning environment construction includes a state space and an action space.

[0134] The state space of a reinforcement learning policy network is defined as follows:

[0135]

[0136] in, for The global state space of continuous reinforcement learning is the collection of all the environment and its own state perceived by the agricultural machinery; for The motion state vector of the agricultural machinery at any given moment; A two-dimensional raster map representing farmland is a digital raster model of the real farmland environment, containing semantic information about traversable areas, crop and obstacle areas, and target task areas. This represents the coverage status matrix of the work area, characterizing the work coverage status of each grid location in the farmland.

[0137] Operation area coverage status matrix Represented as:

[0138]

[0139] in, It represents raster coordinates. The corresponding farmland area is the covered area, indicating that the operation has been completed in that area; Represents raster coordinates The corresponding farmland area is an uncovered area, representing an area awaiting or incomplete work.

[0140] Furthermore, the action space is defined, with the agricultural machinery control actions being continuous control variables, represented as follows:

[0141]

[0142] in, express The continuous control actions performed by agricultural machinery at all times This is the speed control variable for agricultural machinery, used to control the speed of the machinery; it is a continuous control variable. This is the steering control variable for agricultural machinery, used to control the direction of travel of the machinery; it is a continuous control variable.

[0143] Furthermore, the motion state of agricultural machinery in a two-dimensional plane can be represented as:

[0144]

[0145] in, express The horizontal coordinate of the agricultural machinery in the farmland plane coordinate system at any given time; express The vertical coordinate of the agricultural machinery in the farmland plane coordinate system at any given time; express The heading angle of the agricultural machinery at any given time is the angle between the direction of travel of the agricultural machinery and the X-axis of the coordinate system. express The speed of agricultural machinery at any given moment, that is, the instantaneous speed of movement of agricultural machinery.

[0146] Furthermore, a kinematic model of the agricultural machinery is constructed. The kinematic model of the agricultural machinery is as follows:

[0147]

[0148] in, express The horizontal coordinate of the agricultural machinery in the farmland plane coordinate system at any given time; express The vertical coordinate of the agricultural machinery in the farmland plane coordinate system at any given time; express The heading angle of the agricultural machinery at all times; The time interval of the agricultural machinery kinematic model is short, that is, the time interval between two adjacent state moments; express The turning angular velocity of the agricultural machinery at all times.

[0149] Furthermore, the expression for the multi-objective reward function is as follows:

[0150] ,

[0151] in, A reward is given for coverage of the work area when agricultural machinery enters an uncovered work area; As a penalty for repeated operations, a penalty is imposed when agricultural machinery repeatedly visits an already covered area; As a collision penalty, a penalty is imposed when the agricultural machinery collides with an obstacle or the boundary of the farmland; As a path efficiency reward, when a certain amount of the coverage task is completed, the agricultural machinery is given a reward for moving towards the destination; The reward is for completing the task, and an additional reward is given when the agricultural machinery completes the task of covering the entire work area.

[0152] Furthermore, reinforcement learning policy networks use a global state space. As input, through the policy distribution function Generate continuous control actions for agricultural machinery This enables the mapping of environmental states to control actions.

[0153] In a further embodiment, the steps of inputting the multimodal fused state features into a reinforcement learning policy network, performing iterative decision-making based on a preset multi-objective reward function, and outputting continuous control actions including agricultural machinery speed control and agricultural machinery steering control include:

[0154] Step S5001: Based on the multi-objective reward function, calculate the reward for each operation action of the agricultural machinery in real time. Give a positive reward when the agricultural machinery visits the uncovered operation area for the first time, and give a penalty when it repeatedly visits the covered area.

[0155] Step S5002: When the agricultural machinery's steering action is closer to straight-line driving, a positive reward is given to reduce drastic control input; when the agricultural machinery exceeds the map boundary, a boundary collision penalty is given; when the agricultural machinery collides with crops or obstacles, a heavier obstacle collision penalty is given.

[0156] Step S5003: When the operation coverage exceeds the first preset threshold and the agricultural machinery moves toward the destination and the distance is shortened, a positive guidance reward is given; a penalty is imposed for each step the agricultural machinery performs to encourage the agricultural machinery to complete the operation task as soon as possible;

[0157] Step S5004: If the agricultural machinery fails to exit early or reaches the destination but the operation coverage rate is lower than the second preset threshold, a penalty will be given; if the agricultural machinery reaches the destination and the operation coverage rate is not lower than the second preset threshold, a high coverage rate positive reward will be given.

[0158] Step S5005: Based on the real-time reward value calculated at each step, iteratively optimize the reinforcement learning policy network and output a continuous control action that includes agricultural machinery speed control and agricultural machinery steering control, which meets the requirements of full operation coverage, obstacle avoidance, and path smoothness.

[0159] Furthermore, after the agricultural machinery executes the control action, it updates the motion state vector for the next moment through a kinematic model. Simultaneously update the work area coverage status matrix. The new global state Feedback is fed into the reinforcement policy network to form a closed-loop training logic of state input → action output → reward calculation → state update → policy iteration, and finally outputs the globally optimal task path that meets the requirements of full task coverage, obstacle avoidance and path smoothness.

[0160] As can be seen from the above embodiments, compared with the prior art, this application addresses the shortcomings of the existing AB line planning method, which is difficult to adapt to complex farmland operation scenarios with obstacles and the need to target and cover specific areas, resulting in extremely poor scenario adaptability, and the boundary planning method, which suffers from high path repetition rate, low operation efficiency, frequent turning, and large wear and tear on the agricultural machinery chassis. This application has, but is not limited to, the following beneficial effects:

[0161] Firstly, this application automatically distinguishes passable areas, crop obstacle areas, and targeted areas for operation in farmland by combining farmland remote sensing images with semantic segmentation and coordinate mapping. This eliminates the need for operators to manually mark boundary points and operation nodes, significantly reducing the workload of preliminary operation preparation, effectively reducing labor costs, and improving the automation level of farmland environmental information acquisition and overall operation preparation efficiency.

[0162] Secondly, after the farmland area is rasterized, semantic decoupling is performed, and obstacle channels and task channels are separated and a multi-channel raster perception matrix is ​​constructed. This can accurately represent the complex farmland terrain, obstacle distribution and target operation area distribution, breaking through the limitations of traditional AB line planning which is only applicable to regular plots and boundary planning which is difficult to adapt to irregular farmland, and has stronger environmental adaptability.

[0163] Thirdly, on the one hand, convolutional neural networks are used to extract the spatial features of the farmland grid environment, and on the other hand, the historical movement time series information of agricultural machinery is combined to construct time series features and complete the encoding fusion. This can take into account both the external farmland environment status and the agricultural machinery's own driving movement status. Compared with a single state input method, the state representation dimensions are richer, providing reliable and complete data support for intelligent decision-making.

[0164] Fourth, based on a multi-objective reward function, the strategy is iteratively optimized from multiple dimensions such as regional coverage, driving stability, obstacle avoidance and protection, operation efficiency, and task completion quality. The output of agricultural machinery speed and steering continuous control quantities can not only accurately avoid obstacles in the field, but also prioritize the completion of targeted crop area coverage operations, effectively reduce invalid driving paths and frequent turning phenomena, reduce agricultural machinery wear and tear, and at the same time significantly improve the rationality of path planning and the quality of precision farmland operations.

[0165] Furthermore, this application realizes a fully automated process from farmland environment perception, feature extraction and fusion to reinforcement learning autonomous decision-making, without the need for human intervention in path planning. It can meet the needs of various unmanned agricultural machinery operations such as plant protection spraying, precision fertilization, and farmland inspection, effectively solving the problems of insufficient accuracy and limited operation modes in traditional planning methods, and facilitating the large-scale deployment and promotion of unmanned agricultural machinery in complex field scenarios.

[0166] Please see Figure 5This application provides an agricultural machinery operation path planning device based on multi-channel grid perception and reinforcement learning, which includes a semantic segmentation module 1100, a grid map construction module 1200, a dual-modal feature encoding module 1300, a multi-modal feature fusion module 1400, and a path planning decision module 1500. The semantic segmentation module 1100 is configured to acquire remote sensing images of the target farmland, perform semantic segmentation on the farmland remote sensing images to determine the passable area, crop and obstacle area, and target task area in the target farmland. The target task area represents localized crop areas with poor growth, pests, diseases, or lack of water and fertilizer. The raster map construction module 1200 is configured to rasterize the passable area, crop and obstacle area, and target task area to construct a two-dimensional raster map of the target farmland. Semantic decoupling is performed on the two-dimensional raster map to separate and generate obstacle passage raster and task passage raster to construct a multi-channel raster perception matrix. The dual-modal feature encoding module 1300 is configured to acquire motion state vectors corresponding to the target agricultural machinery to construct a temporal state containing historical time-series information. The feature vector is extracted by convolutional pooling of the multi-channel grid perception matrix using a preset convolutional neural network, outputting a grid space feature vector. The multimodal feature fusion module 1400 is configured to map and encode the temporal state feature vector using a multilayer perceptron to output an agricultural machinery motion feature vector, and to concatenate and fuse the grid space feature vector with the agricultural machinery motion feature vector to generate a multimodal fused state feature. The path planning and decision module 1500 is configured to input the multimodal fused state feature into a reinforcement learning policy network, perform iterative decision-making based on a preset multi-objective reward function, and output a continuous control action including agricultural machinery speed control and agricultural machinery steering control, thereby generating an agricultural machinery operation path covering the target task area while avoiding obstacles, so as to complete the agricultural machinery operation path planning based on multi-channel grid perception and reinforcement learning.

[0167] Based on any embodiment of this application, please refer to Figure 6 Another embodiment of this application also provides an electronic device, which can be implemented by a computer device, such as... Figure 6The diagram shows the internal structure of a computer device. This computer device includes a processor, a computer-readable storage medium, a memory, and a network interface connected via a system bus. The computer-readable storage medium stores an operating system, a database, and computer-readable instructions. The database stores control information sequences. When executed by the processor, the computer-readable instructions enable the processor to implement an agricultural machinery operation path planning method based on multi-channel grid perception and reinforcement learning. The processor provides computational and control capabilities to support the operation of the entire computer device. The memory stores computer-readable instructions, which, when executed by the processor, enable the processor to execute the agricultural machinery operation path planning method based on multi-channel grid perception and reinforcement learning as described in this application. The network interface of the computer device is used for communication with a terminal. Those skilled in the art will understand that… Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0168] In this embodiment, the processor is used to execute... Figure 5 The specific functions of each module are defined within the device, and the memory stores the program code and various data required to execute these modules. The network interface is used for data transmission between the user terminal and the server. In this embodiment, the memory stores the program code and data required to execute all modules in the agricultural machinery operation path planning device based on multi-channel grid perception and reinforcement learning. The server can call the server's program code and data to execute the functions of all modules.

[0169] This application also provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the agricultural machinery operation path planning method based on multi-channel grid perception and reinforcement learning as described in any embodiment of this application.

[0170] This application also provides a computer program product, including a computer program / instructions that, when executed by one or more processors, implement the steps of the agricultural machinery operation path planning method based on multi-channel grid perception and reinforcement learning as described in any embodiment of this application.

[0171] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. This computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0172] The above description is only a partial embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for agricultural machine operation path planning based on multi-channel grid perception and reinforcement learning, characterized in that, include: Acquire remote sensing images of farmland corresponding to the target farmland, and perform semantic segmentation on the farmland remote sensing images to determine the passable area, crop and obstacle area and target task area in the target farmland. The target task area represents local crop areas with poor growth, pests and diseases, or lack of water and fertilizer. The passable area, the crop and obstacle area and the target task area are rasterized to construct a two-dimensional raster map of farmland corresponding to the target farmland. The two-dimensional raster map of farmland is semantically decoupled to separate and generate obstacle passage raster and task passage raster to construct a multi-channel raster perception matrix. The motion state vector corresponding to the target agricultural machinery is collected to construct a time-series state feature vector containing historical time-series information. A preset convolutional neural network is used to perform convolutional pooling feature extraction on the multi-channel grid perception matrix to output a grid space feature vector. A multilayer perceptron is used to map and encode the temporal state feature vector to output the agricultural machinery motion feature vector. The grid space feature vector and the agricultural machinery motion feature vector are then concatenated and fused to generate multimodal fused state features. The multimodal fusion state features are input into a reinforcement learning policy network, and iterative decision-making is performed based on a preset multi-objective reward function. The output is a continuous control action that includes agricultural machinery speed control and agricultural machinery steering control. While avoiding obstacles, an agricultural machinery operation path covering the target task area is generated to complete the agricultural machinery operation path planning based on multi-channel grid perception and reinforcement learning.

2. The multi-lane grid perception and reinforcement learning based agricultural machine operation path planning method according to claim 1, characterized in that, The steps of rasterizing the passable area, the crop and obstacle area, and the target task area to construct a two-dimensional raster map of the target farmland, semantically decoupling the two-dimensional raster map of the farmland, and separating and generating obstacle passage raster and task passage raster to construct a multi-channel raster perception matrix include: The passable area of ​​the target farmland, the crop and obstacle area, and the target task area are rasterized to construct a dimension of [missing information]. A two-dimensional raster map of farmland, in which, This indicates the vertical number of raster rows in a two-dimensional raster map of farmland. This indicates the number of horizontal grid columns in the two-dimensional farmland grid map. The value of each grid cell in the two-dimensional farmland grid map is defined as follows: 0 represents a passable area, 0.5 represents a target task area, and 1 represents a crop and obstacle area. Semantic decoupling is performed on the farmland 2D grid map to separate and generate obstacle passage grids and task passage grids. The obstacle passage grid retains only the positions with a value of 1 in the farmland 2D grid map, and the task passage grid retains only the positions with a value of 0.5 in the farmland 2D grid map. The obstacle channel grid and the task channel grid are stacked to construct a dimension of [dimensional value]. A multi-channel raster sensing matrix.

3. The agricultural machinery operation path planning method based on multi-channel grid perception and reinforcement learning according to claim 1, characterized in that, The steps of collecting the motion state vector corresponding to the target agricultural machinery to construct a temporal state feature vector containing historical time-series information, and using a preset convolutional neural network to perform convolutional pooling feature extraction on the multi-channel raster sensing matrix to output the raster space feature vector include: Collect the motion state vector of the target agricultural machinery, including the machinery's position coordinates, speed, heading angle, and work endpoint coordinates; Based on a sliding time window of a preset time length, the motion state vectors of multiple consecutive historical moments are spliced ​​together in time sequence to construct a time sequence state feature vector containing historical time sequence information. Using a multi-channel grid sensing matrix as input to a convolutional neural network, the multi-channel grid sensing matrix is ​​convolved through a first convolutional layer and then activated by an activation function to obtain a first convolutional feature map. After pooling the first convolutional feature map, convolution is performed through the second convolutional layer, and then activated by the activation function to obtain the second convolutional feature map. The second convolutional feature map is flattened and then mapped through a fully connected layer to output a raster space feature vector.

4. The agricultural machinery operation path planning method based on multi-channel grid perception and reinforcement learning according to claim 1, characterized in that, The steps of mapping and encoding the temporal state feature vector using a multilayer perceptron to output the agricultural machinery motion feature vector, and concatenating and fusing the grid space feature vector with the agricultural machinery motion feature vector to generate multimodal fused state features include: The temporal state feature vector is used as the input of the multilayer perceptron, and its temporal state feature vector is flattened to obtain a one-dimensional flattened feature vector. The one-dimensional flattened feature vector is mapped and encoded based on the multilayer perceptron. The one-dimensional flattened feature vector is linearly transformed by the first weight matrix and the first bias vector, and then activated by the activation function to obtain the first encoded feature. The first encoded feature is linearly transformed by the second weight matrix and the second bias vector, and then activated by the activation function to obtain the second encoded feature, which is then used as the agricultural machinery motion feature vector. The agricultural machinery motion feature vector and the grid space feature vector are concatenated and fused along the channel dimension to generate a multimodal fusion state feature for input to the reinforcement learning strategy network.

5. The agricultural machinery operation path planning method based on multi-channel grid perception and reinforcement learning according to claim 1, characterized in that, The multimodal fused state features are input into a reinforcement learning policy network, and iterative decisions are made based on a preset multi-objective reward function. The steps for outputting continuous control actions including agricultural machinery speed control and agricultural machinery steering control include: The multimodal fusion state features are subjected to fusion mapping operation. A linear transformation is performed on the fusion weight matrix and the fusion bias vector, and then a nonlinear activation is performed through the activation function to generate a unified state feature vector. The unified state feature vector is input into the policy distribution function, and iterative decision-making is performed through a reinforcement learning policy network to output continuous control actions containing agricultural machinery speed control and agricultural machinery steering control. The policy distribution function is used to generate the probability distribution of agricultural machinery speed control and agricultural machinery steering control based on the current unified state feature vector.

6. The agricultural machinery operation path planning method based on multi-channel grid perception and reinforcement learning according to claim 1, characterized in that, The expression for the multi-objective reward function is as follows: ， in, A reward is given for coverage of the work area when agricultural machinery enters an uncovered work area; As a penalty for repeated operations, a penalty is imposed when agricultural machinery repeatedly visits an already covered area; As a collision penalty, a penalty is imposed when the agricultural machinery collides with an obstacle or the boundary of the farmland; As a path efficiency reward, when a certain amount of the coverage task is completed, the agricultural machinery is given a reward for moving towards the destination; The reward is for completing the task, and an additional reward is given when the agricultural machinery completes the task of covering the entire work area.

7. The agricultural machinery operation path planning method based on multi-channel grid perception and reinforcement learning according to claim 6, characterized in that, The multimodal fused state features are input into a reinforcement learning policy network, and iterative decisions are made based on a preset multi-objective reward function. The steps for outputting continuous control actions including agricultural machinery speed control and agricultural machinery steering control include: Based on the multi-objective reward function, the reward for each step of the agricultural machinery operation is calculated in real time. When the agricultural machinery visits an uncovered operation area for the first time, a positive reward is given, and a penalty is given when it repeatedly visits an covered area. When the agricultural machinery's steering action is closer to straight-line travel, a positive reward is given to reduce drastic control input; when the agricultural machinery goes beyond the map boundary, a boundary collision penalty is given; when the agricultural machinery collides with crops or obstacles, a heavier obstacle collision penalty is given. When the coverage rate of the operation exceeds the first preset threshold and the agricultural machinery moves toward the finish line and the distance is shortened, a positive incentive reward is given; a penalty is imposed for each step the agricultural machinery takes to encourage it to complete the task as soon as possible. A penalty is imposed when agricultural machinery fails to exit prematurely or reaches the destination but the coverage rate is lower than the second preset threshold; a high coverage rate is awarded when agricultural machinery reaches the destination and the coverage rate is not lower than the second preset threshold. Based on the real-time reward value calculated at each step, the reinforcement learning policy network is iteratively optimized to output a continuous control action that includes agricultural machinery speed control and agricultural machinery steering control, which meets the requirements of full operation coverage, obstacle avoidance, and path smoothness.

8. A device for agricultural machinery operation path planning based on multi-channel grid perception and reinforcement learning, characterized in that, include: The semantic segmentation module is configured to acquire remote sensing images of farmland corresponding to the target farmland, and perform semantic segmentation on the farmland remote sensing images to determine the passable area, crop and obstacle area and target task area in the target farmland. The target task area represents local crop areas with poor growth, pests and diseases, or lack of water and fertilizer. The grid map construction module is configured to rasterize the passable area, the crop and obstacle area and the target task area to construct a two-dimensional grid map of farmland corresponding to the target farmland, and to semantically decouple the two-dimensional grid map of farmland to separate and generate obstacle passage grid and task passage grid to construct a multi-channel grid perception matrix. The dual-modal feature encoding module is configured to collect the motion state vector corresponding to the target agricultural machinery to construct a temporal state feature vector containing historical time series information. A preset convolutional neural network is used to perform convolutional pooling feature extraction on the multi-channel grid perception matrix to output a grid space feature vector. The multimodal feature fusion module is configured to use a multilayer perceptron to map and encode the temporal state feature vector to output the agricultural machinery motion feature vector, and to concatenate and fuse the grid space feature vector with the agricultural machinery motion feature vector to generate multimodal fused state features. The path planning decision module is configured to input the multimodal fusion state features into the reinforcement learning policy network, perform iterative decision-making based on a preset multi-objective reward function, and output continuous control actions including agricultural machinery speed control and agricultural machinery steering control. While avoiding obstacles, it generates an agricultural machinery operation path covering the target task area to complete agricultural machinery operation path planning based on multi-channel grid perception and reinforcement learning.

9. An electronic device comprising a central processing unit and a memory, characterized in that, The central processing unit is used to invoke and run a computer program stored in the memory to perform the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores, in the form of computer-readable instructions, a computer program implemented according to any one of claims 1 to 7, which, when invoked by a computer, executes the steps included in the corresponding method.

Images