Weeding machine automatic row control translation system and method based on machine intelligence

Abstract

The invention relates to the technical field of agricultural machinery automation, in particular to a weeding machine automatic row control translation system and method based on machine intelligence, and the method comprises the steps that a multi-mode sensing module obtains environment data and crop information data; the environment map construction module outputs row reference parameters through a crop row feature extraction algorithm, integrates the environment data and the row reference parameters, solves the three-dimensional pose of the weeding machine through a beam adjustment algorithm, and constructs a high-precision map of the working environment; the decision control module obtains a target row translation track based on a near-end strategy optimization algorithm, and outputs a control instruction; the self-adaptive execution module is used for driving the weeding execution mechanism to automatically perform row-to-row translation weeding along crop rows; and the self-learning module is used for establishing a row translation performance evaluation index according to the full-link operation data, and dynamically updating a target row translation track through a continuous learning algorithm. Therefore, the problems of high seedling injury risk, poor terrain adaptability, low operation efficiency and the like in the prior art are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of agricultural machinery automation, in particular to an automatic row alignment and translation system and method of a weeding machine based on machine intelligence. BACKGROUND

[0002] Under the background of accelerating global agricultural intelligence and fine operation, the demand for efficient and accurate inter-row weed management in large-scale farmland planting is increasingly prominent. Combined with the shortage of agricultural labor resources and the deepening of large-scale operation trend, the automation and precision of weeding operation have become the core link to improve the crop growing environment, ensure yield and reduce management cost. Under this situation, it is particularly urgent to develop an automatic row alignment and translation system of a field weeding machine based on machine intelligence. The system can realize real-time visual recognition and positioning of crop rows, and realize automatic and accurate row alignment and translation control of the weeding unit relying on intelligent decision algorithm, so as to meet the core needs of improving weeding operation efficiency, reducing manual operation intensity and avoiding crop damage.

[0003] However, the current traditional weeding machine operation mainly relies on manual driving or simple mechanical guiding device to complete, which has the problems of high risk of damaging seedlings, poor terrain adaptability and low operation efficiency: first, the risk of damaging seedlings increases due to high fatigue and unstable precision of manual operation, which directly aggravates the risk of damaging seedlings; second, the mechanical guiding device is affected by the complex field terrain such as soil softness difference and ridge and furrow fluctuation, which often causes row alignment failure or deviation, highlights the short board of poor terrain adaptability, and makes it difficult to guarantee the weeding precision and operation effect. At the same time, the traditional operation mode is limited by the disconnection between perception control, which makes it difficult to dynamically integrate and closed-loop control the visual perception line position information with the real-time posture, speed and execution mechanism state of the weeding machine, further restricting the operation efficiency. In addition, the simplification of control strategy makes the system form a control blind area in complex scenes such as fuzzy crop seedling line, dense weed shielding or dramatic changes in field illumination conditions, which leads to accumulation of row alignment deviation, lagging or excessive adjustment of correction response, not only aggravates the risk of damaging seedlings, but also reduces the efficiency due to operation interruption and repeated operation, which has a significant negative impact on operation smoothness, weeding thoroughness and crop protection effect. SUMMARY

[0004] The present application provides an automatic row alignment and translation system and method of a weeding machine based on machine intelligence to solve the problems of high risk of damaging seedlings, poor terrain adaptability and low operation efficiency in the prior art.

[0005] The first aspect embodiment of the application provides an automatic row translation system of a weeding machine based on machine intelligence, comprising: a multi-modal perception module, an environment map construction module, a decision control module, an adaptive execution module, and a self-learning module; wherein the multi-modal perception module is configured to obtain environment data and crop information data; the environment map construction module is configured to output row reference parameters based on the crop information data through a crop row feature extraction algorithm, fuse the environment data and the row reference parameters, and calculate a three-dimensional pose of the weeding machine in real time through a beam adjustment algorithm to construct a high-precision map of the working environment; the decision control module is configured to obtain a target row translation trajectory through a near-end strategy optimization algorithm based on the high-precision map of the working environment, and output a control instruction; the adaptive execution module is configured to drive a weeding execution mechanism to automatically translate along the crop row for weeding based on the target row translation trajectory and the control instruction; and the self-learning module is configured to establish a row translation performance evaluation index according to full-link operation data, and dynamically update the target row translation trajectory through a continuous learning algorithm.

[0006] Preferably, the multi-modal perception module comprises a laser radar perception unit and an environment parameter sensing unit, wherein the laser radar perception unit is configured to obtain crop row distribution and weed feature crop information data; and the environment parameter sensing unit is configured to obtain illumination variation, temperature and humidity fluctuation, and complex terrain diversified environment data.

[0007] Preferably, the environment map construction module comprises a reference parameter unit, a data fusion unit, a BA pose optimization unit, and a map construction unit, wherein the reference parameter unit is configured to dynamically position a row reference line through a crop row feature extraction algorithm, and output row reference parameters; the data fusion unit is configured to fuse the environment data and the row reference parameters, and perform feature extraction and coordinate alignment based on a spatial correlation algorithm; the BA pose optimization unit is configured to obtain a three-dimensional pose and a feature point position by using a beam adjustment method according to the feature extraction and coordinate alignment, and correct cumulative errors in combination with loop detection; and the map construction unit is configured to construct a high-precision map of the working environment based on the three-dimensional pose, the feature point position, and the row reference parameters.

[0008] Preferably, the decision control module comprises an intelligent strategy decision maker and a control instruction generation unit, wherein the intelligent strategy decision maker is configured to generate a target row translation trajectory that meets the working constraints by using a near-end strategy optimization algorithm based on the high-precision map of the working environment and the device state; and the control instruction generation unit is configured to convert the target row translation trajectory into an execution mechanism signal, and output a standardized control instruction.

[0009] Preferably, the self-adaptive execution module comprises an instruction receiving and pose adjusting unit and an automatic alignment unit, wherein the instruction receiving and pose adjusting unit is configured to receive the control instruction and adjust the transverse position, working height and angle of the weeding mechanism; and the automatic alignment unit is configured to autonomously identify the row-column structure of the crop growth, accurately align the weeding machine with the crop row and make the weeding machine perform the weeding operation along the row.

[0010] Preferably, the self-learning module comprises a performance evaluation index unit and a continuous learning unit, wherein the performance evaluation index unit is configured to establish a performance evaluation system and key quantitative indicators according to the full-link operation data; and the continuous learning unit is configured to optimize the target alignment and translation trajectory of the weeding machine by using a continuous learning algorithm.

[0011] The second aspect of the present application provides a weeding machine automatic alignment and translation method based on machine intelligence, comprising: obtaining environment data and crop information data; outputting alignment reference parameters by using a crop row feature extraction algorithm on the crop information data; fusing the environment data and the crop information data, performing feature extraction and coordinate alignment based on a spatial correlation algorithm, and solving the three-dimensional pose of the weeding machine by using a beam adjustment algorithm to construct a high-precision map of the working environment; obtaining the target alignment and translation trajectory of the weeding machine by using a near-end strategy optimization algorithm based on the high-precision map of the working environment, the three-dimensional pose of the weeding machine and the alignment reference parameters, converting the target alignment and translation trajectory into an execution mechanism signal, and outputting a control instruction; driving the weeding execution mechanism to automatically perform alignment and translation along the crop row for weeding based on the target alignment and translation trajectory and the control instruction; obtaining full-link operation data by using a time synchronization and calibration algorithm, establishing alignment and translation performance evaluation indicators, evaluating the weeding machine, dynamically updating the alignment and translation performance evaluation indicators of the weeding machine and optimizing the target alignment and translation trajectory by using a continuous learning algorithm.

[0012] The third aspect of the present application provides an electronic device, comprising: a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a weeding machine automatic alignment and translation method based on machine intelligence as described in the above embodiments.

[0013] The fourth aspect of the present application provides a computer readable storage medium having a computer program stored thereon, wherein the program is executed by a processor to implement a weeding machine automatic alignment and translation method based on machine intelligence as described in the above embodiments.

[0014] The fifth aspect of the present application provides a computer program product comprising a computer program or instructions for implementing a weeding machine automatic alignment and translation method based on machine intelligence as described in the above embodiments.

[0015] Thus, the present application includes the following advantages:

[0016] The embodiment of the present application effectively deals with complex terrain and crop shielding interference in the field through the multi-modal environment perception module and crop row feature extraction algorithm, cooperates with the light beam adjustment pose solution and high-precision map construction module to improve the three-dimensional positioning accuracy and operation environment perception ability of the weeding machine; relies on the near-end strategy optimization algorithm and decision control module to accurately plan the row translation trajectory and output control instructions, strengthens the path tracking robustness in the scene of crop row bending, local missing, etc., improves the row translation accuracy, at the same time, the adaptive execution module drives the weeding unit, combines dynamic trajectory following and real-time response mechanism, can automatically translate and weed along the crop row, effectively avoids the crop damage and missing problems caused by trajectory deviation, controls the weeding operation quality; the self-learning module dynamically optimizes the decision strategy and translation trajectory through the performance evaluation index and continuous learning algorithm, improves the adaptive ability and long-term operation efficiency of the system in diversified farmland environment, significantly improves the weeding efficiency, guarantees the safety of crops and reduces the cost of manual intervention. Thus, the problems of high risk of checking seedlings, poor terrain adaptability and low operation efficiency in the prior art are solved.

[0017] Additional aspects and advantages of the present application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0018] The above and / or additional aspects and advantages of the present application will become apparent and be readily appreciated from the following description, including the appended drawings, wherein:

[0019] Figure 1 A structure diagram of an automatic row translation system of a weeding machine based on machine intelligence is provided according to an embodiment of the present application;

[0020] Figure 2 A schematic diagram of a multi-modal perception module is provided according to an embodiment of the present application;

[0021] Figure 3 A schematic diagram of an environment map construction module is provided according to an embodiment of the present application;

[0022] Figure 4 A schematic diagram of a decision control module is provided according to an embodiment of the present application;

[0023] Figure 5 A schematic diagram of an adaptive execution module is provided according to an embodiment of the present application;

[0024] Figure 6 A schematic diagram of a self-learning module is provided according to an embodiment of the present application;

[0025] Figure 7 FIG. 1 is a schematic diagram of a machine intelligence-based automatic row translation system for a weeding machine according to an embodiment of the present application;

[0026] Figure 8 FIG. 2 is a flowchart of a machine intelligence-based automatic row translation method for a weeding machine according to an embodiment of the present application;

[0027] Figure 9 FIG. 3 is a schematic diagram of a machine intelligence-based automatic row translation method for a weeding machine according to an embodiment of the present application;

[0028] Figure 10 FIG. 4 is a structural schematic diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0030] A machine intelligence-based automatic row translation system and method for a weeding machine according to an embodiment of the present application will be described below with reference to the drawings. In view of the low work efficiency mentioned in the background art, the present application provides a machine intelligence-based automatic row translation system for a weeding machine. In the system, through a multi-modal environment perception module and a crop row feature extraction algorithm, complex terrain and crop shielding interference in the field are effectively dealt with, and a light beam adjustment pose solution and a high-precision map construction module are used to improve the three-dimensional positioning accuracy of the weeding machine and the work environment perception ability. Relying on a near-end strategy optimization algorithm and a decision control module, the row translation trajectory is accurately planned and control instructions are output, the path tracking robustness is strengthened in the scenes of curved crop rows and local missing, and the row translation accuracy is improved. At the same time, a self-adaptive execution module drives the weeding unit, combines dynamic trajectory following and real-time response mechanism, and can automatically translate and weed along the crop row, effectively avoiding crop damage and missing problems caused by trajectory deviation, and controlling the weeding work quality. Through performance evaluation indexes and a continuous learning algorithm, the self-learning module dynamically optimizes the decision strategy and the translation trajectory, improves the adaptive ability and long-term work efficiency of the system in diversified farmland environments, significantly improves the weeding efficiency, guarantees the safety of crops, and reduces the cost of manual intervention. Thus, the problems of high risk of damaging seedlings, poor terrain adaptability, and low work efficiency in the prior art are solved.

[0031] Figure 1 FIG. 1 is a structural schematic diagram of a machine intelligence-based automatic row translation system for a weeding machine according to an embodiment of the present application.

[0032] The embodiment of the application provides a weeding machine automatic alignment translation system based on machine intelligence, which comprises a system 10.

[0033] A multi-modal perception module 100, an environment map construction module 200, a decision control module 300, an adaptive execution module 400, and a self-learning module 500.

[0034] The multi-modal perception module 100 is configured to acquire environment data and crop information data; the environment map construction module 200 is configured to output alignment reference parameters based on the crop information data through a crop row feature extraction algorithm, fuse the environment data and the alignment reference parameters, and calculate the three-dimensional pose of the weeding machine in real time through a beam adjustment algorithm to construct a high-precision environment map of the working environment; the decision control module 300 is configured to obtain a target alignment translation trajectory through a near-end strategy optimization algorithm based on the high-precision environment map of the working environment and output a control instruction; the adaptive execution module 400 is configured to drive a weeding execution mechanism to automatically perform alignment translation weeding along the crop row based on the target alignment translation trajectory and the control instruction; and the self-learning module 500 is configured to establish an alignment translation performance evaluation index according to full-link operation data and dynamically update the target alignment translation trajectory through a continuous learning algorithm.

[0035] It can be understood that, in the embodiment of the application, the multi-modal environment perception module and the crop row feature extraction algorithm effectively deal with complex field topography and crop shielding interference, and the beam adjustment pose calculation and high-precision map construction module improves the three-dimensional positioning accuracy of the weeding machine and the working environment perception capability; the near-end strategy optimization algorithm and the decision control module are relied on to accurately plan the alignment translation trajectory and output the control instruction, thereby enhancing the path tracking robustness in the scenes of curved crop rows, local missing, etc., improving the alignment translation accuracy, and at the same time, the adaptive execution module drives the weeding unit, combines the dynamic trajectory following and the real-time response mechanism, and can automatically perform translation weeding along the crop row, effectively avoiding the crop damage and missing problems caused by trajectory deviation, and controlling the weeding operation quality; the self-learning module dynamically optimizes the decision strategy and the translation trajectory through the performance evaluation index and the continuous learning algorithm, improves the adaptive ability and long-term operation efficiency of the system in diversified farmland environments, significantly improves the weeding efficiency, guarantees the safety of crops, and reduces the cost of manual intervention. Thus, the problems of high risk of checking seedlings, poor terrain adaptability, and low operation efficiency in the prior art are solved.

[0036] In the embodiment of the application, the multi-modal perception module 100 further comprises a laser radar perception unit and an environment parameter sensing unit as shown in Figure 2

[0037] The laser radar perception unit is configured to acquire crop row distribution and weed feature crop information data; and the environment parameter sensing unit is configured to acquire light variation, temperature and humidity fluctuation, and complex terrain diversified environment data.​

[0038] It can be understood that the embodiments of the present application can comprehensively obtain the plant row distribution, weed feature crop information, and illumination change, temperature and humidity fluctuation, and complex terrain diversified environmental data by integrating the laser radar perception unit and the environmental parameter sensing unit, to provide multi-dimensional environmental information for subsequent weed removal path planning and adaptive control; the laser radar perception unit can accurately identify the spatial distribution characteristics of crops and weeds, effectively distinguish target plants and weeds, and ensure the high accuracy of weed removal operation; the environmental parameter sensing unit can dynamically adjust the system working parameters according to the temperature and humidity fluctuation and the terrain change, effectively adapt to the complex environmental conditions in the field, and ensure the stability and reliability of the system operation in the diversified environments such as undulating terrain and different temperature and humidity.

[0039] For example, in the intelligent weed removal operation in a wheat field, the self-propelled weeding machine carrying the laser radar perception unit and the environmental parameter sensing unit is operating efficiently: the laser radar perception unit accurately identifies the neat row spacing distribution of wheat (plant row distribution) through real-time scanning, clearly distinguishes the positions, leaf profiles and growth densities of oats, parietaria and other weeds (weed features) mixed therein, and records the plant information such as the plant height and stem spacing of wheat; the environmental parameter sensing unit synchronously monitors the illumination change (such as the influence of strong light at noon on laser reflection), the air temperature and humidity fluctuation (the toughness characteristics of weeds under 28℃ and 70% humidity), and the complex terrain data such as slight slope and low-lying place in the field. These data are fed back to the weeding machine control system in real time, so that the weeding machine can accurately travel along the crop row, start the weeding tool according to the weed position, and adjust the machine height according to the terrain, to efficiently remove weeds and avoid damaging wheat, thereby realizing intelligent and accurate weed removal.

[0040] In the embodiments of the present application, the environmental map construction module 200 includes: as shown in Figure 3 the baseline parameter unit, the data fusion unit, the BA pose optimization unit, and the map construction unit.

[0041] The baseline parameter unit is configured to dynamically position the row baseline by using a crop row feature extraction algorithm, and output the row baseline parameter; the data fusion unit is configured to fuse the environmental data and the row baseline parameter, and perform feature extraction and coordinate alignment based on a spatial correlation algorithm; the BA pose optimization unit is configured to obtain three-dimensional pose and feature point position by using a bundle adjustment method according to the feature extraction and coordinate alignment, and correct the cumulative error by combining loop detection; and the map construction unit is configured to construct a high-precision operation environment map based on the three-dimensional pose and the feature point position, and in combination with the row baseline parameter.

[0042] It is understood that the embodiments of this application can dynamically locate the row baseline through the crop row feature extraction algorithm, improving the baseline ambiguity problem caused by uneven distribution of crop rows in the field, and accurately outputting the row baseline parameters. The data fusion unit integrates environmental data and baseline parameters, achieving coordinate alignment while extracting features, and improving the accuracy of data association. The BA pose optimization unit uses bundle adjustment to obtain the three-dimensional pose and feature point positions, and combines loop closure detection to eliminate pose errors caused by cumulative calculations, ensuring the spatial accuracy of the three-dimensional pose and feature point positions, providing geometrically reliable pose and feature data for subsequent map construction, and improving the accuracy of map construction. The map construction unit integrates multi-dimensional information to construct a high-precision map of the working environment based on the three-dimensional pose, feature point positions, and row baseline parameters. While comprehensively reflecting environmental features, it highlights the row baseline information, providing accurate environmental reference for the automatic row translation of the weeding machine, and improving the reliability of row translation.

[0043] For example, in a cornfield weeding scenario, the environmental map construction module first uses a baseline parameter unit to extract semantic features of corn rows using a binocular camera and LiDAR, and fits a 3D baseline (including parameters such as row spacing and direction). Then, the data fusion unit synchronizes multi-sensor data (camera images, LiDAR point clouds, IMU), and unifies crop features and terrain point clouds into a global coordinate system through cross-modal feature matching and coordinate alignment. Subsequently, the BA pose optimization unit optimizes the pose of each frame based on the bundle adjustment method, combined with baseline constraints and loop closure detection to correct accumulated errors. Finally, the map construction unit integrates and optimizes the pose and environmental data to generate a high-precision map containing a crop row guidance layer, a weed distribution heatmap, and terrain undulations. This supports real-time calibration of row deviations by the weeder (error < 3 cm), enabling precise translation operations, reducing the missed removal rate to 2% and the compaction rate to 1%.

[0044] In this embodiment of the application, the decision control module 300 includes: Figure 4 As shown, there is an intelligent strategy decision-maker and a control command generation unit.

[0045] The intelligent strategy decision-maker generates a target pair translation trajectory that conforms to the operation constraints based on a high-precision map of the operating environment and the equipment status through a near-end strategy optimization algorithm; the control command generation unit is used to convert the target pair translation trajectory into actuator signals and output standardized control commands.

[0046] It can be understood that the intelligent strategy decision maker of the embodiment of the application generates a target row translation trajectory conforming to the work constraint based on the high-precision map of the work environment and the device state through a near-end strategy optimization algorithm; the control instruction generation unit is used to convert the target row translation trajectory into an actuator signal, output a standardized control instruction, improve the rationality of the trajectory generation, the accuracy of the control instruction and the compliance of the work constraint, and provide reliable control support for the automatic row translation of the weeding machine through the structured control logic, thereby reducing the trajectory deviation and the operation mismatch probability.

[0047] It should be noted that converting the target row translation trajectory into an actuator signal is to analyze and discretize the continuous trajectory, extract key parameters such as the translation amount, speed and angle at each time, perform inverse calculation based on the dynamics model of the mechanical structure of the weeding machine, convert the trajectory parameters into motion parameters of the actuator, distribute the control amount according to the mechanism layout, and determine the action instructions of each component; and convert these instructions into physical driving signals such as motor speed and angle to form a standardized control instruction output.

[0048] Dynamics model:

[0049]

[0050] wherein T is the torque required by the actuator; M(q) is an inertia matrix; is a centrifugal force matrix; G(q) is a gravity vector; is the target joint acceleration of trajectory planning; J T is a Jacobian matrix; F ext is an external force received by the end effector.

[0051] For example, when the weeding machine equipped with the system is working in the third row of a corn field with a row spacing of 70 cm, the machine body deviates 10 cm to the right of the target row center line due to uneven ground, the real-time heading angle is 1.5 degrees (deviation to the right), and the forward speed is 0.8 m / s. At this time, the intelligent strategy decision maker makes online decisions based on the real-time constructed high-precision field map and the state of the weeding machine itself using the proximal policy optimization (PPO) algorithm. Under the premise of strictly considering the maximum steering angle limit ± 25 degrees, the maximum lateral acceleration constraint <0.3g, and the crop row safety distance ± 5 cm and other operation constraints, the PPO strategy network generates a target trajectory after <50 ms of calculation: it requires a smooth S-shaped curve to gradually correct the lateral deviation from +10 cm to -2 cm in the next 2.0 seconds, and then fine-tune it back to 0 cm (accurately align the third row center line) in the next 0.5 seconds. Subsequently, the control command generation unit converts this spatial trajectory (a sequence of target lateral displacement at 10 ms intervals) into the control amount of the underlying execution mechanism: it calculates that the target steering wheel angle needs to be linearly changed from the current +3 degrees to -8 degrees in 0.5 seconds, and outputs the standardized CAN2.0B bus command to drive the hydraulic steering valve to execute accurately, achieving efficient and smooth automatic alignment translation.

[0052] In the embodiments of the present application, the adaptive execution module 400 includes, as shown in Figure 5 the instruction receiving and pose adjustment unit, and the automatic alignment unit.

[0053] The instruction receiving and pose adjustment unit is configured to receive the control instruction and adjust the lateral position, working height and angle of the weeding mechanism; and the weeding action driving automatic alignment unit is configured to autonomously identify the row and column structure of crop growth, so that the weeding machine is accurately aligned with the crop row and moves along the row to perform weeding operation.

[0054] It can be understood that, by receiving the control instruction through the instruction receiving and pose adjustment unit, the lateral position, working height and angle of the weeding mechanism are adjusted to ensure accurate response of pose adjustment; the automatic alignment unit autonomously identifies the row and column structure of crop growth, so that the weeding machine is accurately aligned with the crop row and moves along the row to automatically perform adaptive row operation, thereby improving the pose adjustment accuracy of the weeding mechanism, the row operation accuracy and the operation stability, and reducing the row deviation and operation misplacement probability.

[0055] For example, in a corn field with a row spacing of about 60 cm, when the corn grows to a height of 1.2 meters, the weeds in the field begin to grow wildly, at which time the weed control machine based on machine intelligence automatically transverse system is in place. The core adaptive execution module of the system, through the coordinated operation of the two units, carries out precise and efficient weeding operation. The instruction receiving and pose adjustment unit is the pose adjuster of the weed control machine. It first receives the instructions of the control module, which are set based on corn growth data: the transverse needs to deviate from the corn stem by 5 cm to prevent damage, the working height is reduced to 10 cm from the ground to remove the root weeds, and the cutter is tilted by 30 degrees to match the growth direction of the weeds to improve the cutting efficiency. The unit drives the transverse adjustment motor, the height adjustment hydraulic rod and the angle adjustment motor to accurately complete the pose adjustment and prepare for the weeding operation. The weeding action driving automatic transverse unit is the path navigator. During the travel, the visual sensor and laser radar carried by it scan the front corn plants in real time, and analyze the stem arrangement rule through image recognition algorithm. Even if the corn row is slightly curved due to planting errors or terrain undulations, the row direction can be accurately captured. It continuously calculates the deviation of the current travel direction of the weed control machine from the corn row, and constantly adjusts the steering motor of the walking wheel to strictly control the center line deviation within 2 cm, so as to ensure that the machine always travels along the row smoothly. With the cooperation of the two units, the weed control machine can not only adapt to the growth state of crops, but also accurately track the row path, effectively avoiding missing weeds or damaging corn, so that the field weeding operation is more intelligent and efficient.

[0056] In the embodiments of the present application, the self-learning module 500 includes, as shown in Figure 6 the performance evaluation index unit and the continuous learning unit.

[0057] The performance evaluation index unit is configured to establish a performance evaluation system and key quantitative indicators according to full-link operation data; and the continuous learning unit is configured to optimize the target transverse trajectory by using a continuous learning algorithm, to obtain a target transverse trajectory suitable for the weeding operation scenario.

[0058] It can be understood that, the performance evaluation index unit establishes a performance evaluation system and key quantitative indicators based on full-link operation data to clarify the system operation benchmark; and the continuous learning unit optimizes the target transverse trajectory by using a continuous learning algorithm, to obtain a target transverse trajectory suitable for the weeding operation scenario, thereby improving the trajectory optimization adaptability and operation evaluation accuracy, and reducing the transverse deviation and operation adaptation deficiency probability.

[0059] It should be noted that the performance evaluation index unit is used to establish a performance evaluation system and key quantitative indicators based on the full-link operation data. The system must be based on the full-link operation data, clarify the optimization goals and scenario boundaries, collect multi-source heterogeneous data from the perception layer, execution layer, and result layer, and clean and align them; decompose key performance dimensions such as trajectory tracking accuracy, operation efficiency, and environmental adaptability, and select quantifiable indicators such as root mean square error, effective operation speed, and scenario switching success rate; verify the effectiveness of the indicators through historical data comparison or simulation, and allocate weights in combination with business priorities; and finally construct a hierarchical evaluation system.

[0060] For example, in a soybean field on a farm, the intelligent weeding machine encountered problems such as curved crop rows, low light in the morning and evening, and undulating terrain. Its initial straight-line trajectory strategy had shortcomings, including large row offset (peak value 12cm) and high missed weeding rate (5%). Its self-learning module collected full-link data from multiple sensors such as LiDAR and cameras (including crop row outline, light, slope, etc.) through the performance evaluation unit to establish an evaluation system covering accuracy and robustness. Then, the continuous learning unit iteratively optimized the problem scenario. First, it used Bézier curve fitting and attitude compensation to optimize the curved row trajectory, reducing the missed weeding rate to 2.5%. Then, it enhanced low light perception through multimodal fusion (camera + LiDAR) and combined transfer learning to fine-tune the model, so that the standard deviation of the downward offset was ≤4cm and the missed weeding rate was ≤3% in the whole scene, realizing adaptive weeding operation in complex environments.

[0061] This application proposes an automatic row-to-row translation system for weeders based on machine intelligence. Through a multimodal environment perception module and crop row feature extraction algorithm, it effectively addresses complex field terrain and crop shading interference. Combined with a beam adjustment pose calculation and high-precision map construction module, it improves the weeder's 3D positioning accuracy and operational environment perception capabilities. Relying on a near-end strategy optimization algorithm and decision control module, it accurately plans the row-to-row translation trajectory and outputs control commands, enhancing path tracking robustness and improving row-to-row translation accuracy in scenarios such as curved or partially missing crop rows. Simultaneously, an adaptive execution module drives the weeding unit, combining dynamic trajectory following and real-time response mechanisms to automatically perform weeding operations along crop rows, effectively avoiding crop damage and missed weeding due to trajectory deviations, and controlling the quality of weeding operations. A self-learning module dynamically optimizes decision strategies and translation trajectories through performance evaluation indicators and continuous learning algorithms, improving the system's adaptability and long-term operational efficiency in diverse farmland environments, significantly improving weeding efficiency, ensuring crop safety, and reducing manual intervention costs. Therefore, it solves the problems of high risk of damaged seedlings, poor terrain adaptability, and low operational efficiency in existing technologies.

[0062] The following will illustrate an automatic row translation system for a weeder based on machine intelligence through a specific embodiment, such as... Figure 7 As shown, it includes:

[0063] In the main corn-producing areas of the North China Plain, the traditional weeding method has long plagued farmers. A certain planting cooperative in Xingtai City, Hebei Province, once spent 240,000 yuan per year on manual weeding for its 1,200 mu of corn fields. Flood irrigation and chemical weeding also led to soil pesticide residues and a 10% reduction in corn yield. However, the introduction of an automatic row translation system for weeding machines based on machine intelligence in 2024 completely changed the situation. This system, through the coordinated operation of multi-modal perception, environmental map construction, decision control, self-adaptive execution, and self-learning modules, achieved precise removal of weeds between corn rows. Not only did it reduce weeding costs by 60%, but it also increased corn seedling survival rate to 99.5%. The multi-modal perception module, as the nervous system of the system, is composed of multiple types of sensors. The two sets of industrial-level vision cameras on the front of the vehicle can capture color and morphological features of crops with a resolution of 2048x1536 pixels and a frame rate of 30fps. The near-infrared camera can distinguish crops from weeds through spectral differences, and it can work stably even in backlight or cloudy weather. The 16-line laser radar on both sides of the vehicle scans the environment at a density of 100,000 points per second, generating three-dimensional point cloud data to accurately measure corn height, row spacing, and weed distribution density. The vehicle-mounted IMU and Beidou RTK positioning modules record the position, speed, and attitude angle of the weeding machine in real time, with a sampling frequency of 100Hz. In the corn three-leaf stage operation in June 2024, this module successfully identified corn seedling green stems and leaves and distinguished between weeds such as barnyard grass and crabgrass. The laser radar measured an average corn row spacing of 60cm with a maximum deviation of not more than 5cm, laying the foundation for row-by-row operation.

[0064] The environment map construction module first processes the visual image through the crop row feature extraction algorithm, detects the corn seedlings with the YOLOv8 model, and then uses the improved Hough transform algorithm to fit the crop row straight line. The output includes the starting coordinates, row angle (the average is 15 degrees in this field), and row spacing deviation value. Then, the laser radar point cloud, IMU attitude data, and the row alignment reference parameters are fused, and the beam adjustment algorithm is used to solve the three-dimensional pose of the weeding machine in real time, with the positioning error controlled within 3 cm. The final constructed high-precision map of the working environment includes the crop row accurate trajectory layer, the weed distribution heat layer (color marked density, red represents coverage rate over 30%), and the terrain elevation layer (records the ups and downs of the field, the maximum height difference of the field is 8 cm). The map is updated every 0.1 seconds, providing dynamic reference for the decision-making module. The decision control module generates the optimal row translation trajectory based on the high-precision map through the proximal policy optimization algorithm (PPO). As a reinforcement learning method, PPO adjusts the strategy through interaction with the environment, optimizing the three major goals of minimizing row deviation (target ≤ 3 cm), maximizing weeding efficiency (operation speed ≥ 0.8 m / s), and minimizing energy consumption. In actual operation, the trajectory will be dynamically adjusted. For example, when encountering a deviation from 60 cm to 70 cm, the algorithm automatically offsets the trajectory 5 cm towards the row center. When detecting a dense weed area (red area in the heat map), the forward speed is temporarily reduced to 0.5 m / s and the cutter rotation frequency is increased. The output control commands such as lateral translation (range ± 15 cm, accuracy 0.1 cm), forward speed (0.3-1.2 m / s adjustable), and cutter height (set to 8 cm at the three-leaf stage) are transmitted through the CAN bus with a response delay of no more than 50 ms. The adaptive execution module, as the physical executor, is composed of mechanical structure and drive system. The tracked chassis cooperates with the servo motor driven ball screw lateral translation mechanism to realize ± 20 cm lateral fine adjustment. Each of the two symmetrical rotating weeding wheels contains 6 tungsten carbide blades with a spacing of 25 cm according to the corn row spacing. In July 2024, when the corn seedlings were tilted 8 cm due to strong wind, the lateral translation mechanism completed a 5 cm fine adjustment to avoid the seedlings within 0.3 seconds. When encountering a 12 cm high soil block, the cutter was automatically lifted 3 cm to avoid damage. The module also has load adaptive capability, reduces the cutter speed and increases the driving torque to prevent overload when the weed density is too high, and adjusts the tracked grip force parameters to prevent slipping when the soil is wet.The self-learning module optimizes the decision-making model by recording full-link data, collects perception layer original images and point clouds, decision layer trajectory parameters and control instructions, execution layer displacement error, energy consumption and operation effect data (post-camera image shooting calculates weed removal rate and injury rate), establishes evaluation indexes such as row deviation mean (target ≤ 3 cm), weed removal rate (target ≥ 95%), injury rate (target ≤ 1%), unit area energy consumption (target ≤ 0.5 kWh / mu), and dynamically updates PPO parameters through incremental learning algorithm, such as reducing the weight of land edge deviation after the first operation by strengthening the extraction of light scene features, adjusting the track driving torque when continuously working on clay to reduce the slip rate from 8% to 2%, and improving the comprehensive performance by 23% in a growing season.

[0065] On June 15, 2024, before the operation, the technical personnel calibrated the multi-modal perception module. The visual camera was adjusted through the standard calibration board to make the distortion rate ≤1%. The laser radar was calibrated to an error of ±2 cm. The weeder drove along the edge of the plot at 0.5 m / s to construct an initial high-precision map. The northeast dense weed area (coverage rate 42%), the central 3 m low-lying area (maximum water depth 2 cm), and the southwest 5 30 cm stone blocks were identified. The decision-making module planned the sequence of flattening the area first and then the special area. The self-learning module retrieved the planting data from the previous two years to initialize the PPO parameters. The reinforcement learning algorithm optimized the logic of low-lying area passability and stone block obstacle avoidance. On June 16, the formal operation, under the morning light of 20000 lux, the multi-modal perception module clearly identified the boundary between crops and weeds. The environmental map updated the trajectory every 0.1 second. The decision-making module generated real-time instructions. In the flat area, the weeder moved forward at 1.0 m / s, with lateral fine-tuning to stabilize the row deviation within ±2 cm. The weeding wheel rotated at 300 rpm, with a removal rate of 98%. In the northeast dense weed area, the thermal map showed a coverage rate of 42%. The decision-making module reduced the speed to 0.6 m / s, increased the cutter speed to 400 rpm, and increased the cutting depth by 1 cm to ensure the cutting of the root system. At 9 am, the weeder entered the central low-lying area. The laser radar detected a decrease in elevation, and the IMU recorded a 3-degree tilt of the vehicle body. The map module updated the terrain, and the decision-making module activated the anti-skid mode, increasing the track driving torque by 20% and the lateral adjustment frequency from 10 Hz to 15 Hz, making the maximum deviation in this area only 4 cm without slipping. At 2 pm, the southwest detected 30 cm stone blocks. The decision-making module reduced the speed to 0.3 m / s and shifted 10 cm laterally to avoid obstacles. The operation was completed in 3 seconds without damaging the crops. After 3 days of operation, sample detection showed that the average weed removal rate in 100 points was 96.3% (98.5% in flat areas and 92.1% in dense areas), and the injury rate was 0.4%, which was much lower than the traditional mechanical 5%. The average row deviation was 2.1 cm, with a maximum of not more than 5 cm. In terms of economic benefits, 150 mu of operation per day required 3 operators, with a labor cost of 96,000 yuan, a reduction of 144,000 yuan. The cost of pesticides per mu decreased from 30 yuan to 8 yuan, saving 264,000 yuan. It is estimated that the yield per mu will increase by 8% and the income will increase by 150,000 yuan. A growing season will recover 60% of the investment. In the key technology breakthrough, the multi-modal fusion anti-interference technology can distinguish crops and weeds under 5000-100000 lux light. The dynamic trajectory optimization algorithm learns the bending trend of the crop row through PPO and adjusts the deviation within 3 cm, reducing it by 70% compared with traditional machinery. The self-learning performance iteration mechanism records data of different soil, crop stage, and weed species to establish a multi-scenario model. It optimizes the track parameters to improve efficiency by 15% for the characteristics of the cooperative clay, and adjusts the cutter height detection logic to avoid leaf shielding errors.This system realizes three reductions and three increases: herbicide cost reduction of 60%, pesticide use reduction of 73%, and seedling damage rate reduction of 92%, and increases weed removal rate by 28%, operation efficiency by 300%, and yield per mu by 8%. In the future, the system will upgrade multi-crop adaptation models, increase pest and disease identification, and build a regional intelligent agricultural platform. As Li Jianguo, the head of the cooperative, said, the intelligent weeding system not only saves money, but also brings hope for technology-based farming. It changes traditional farmland from relying on weather to making decisions based on weather, and the blueprint for precision agriculture is gradually unfolding in the North China Plain.

[0066] In summary, the embodiments of the present application realize precise weeding along the crop row and dynamic adaptation to complex environments through the coordinated operation of multi-modal perception, dynamic map construction, intelligent decision-making, adaptive execution, and self-learning. This enables the weeding machine to always move steadily along the crop row, only removing weeds between the rows without damaging young seedlings. Through continuous learning and optimization, the system adapts to complex scenarios such as changes in light, terrain undulations, and uneven weed distribution. It can reduce labor costs by 60%, reduce pesticide use by 73%, significantly reduce agricultural inputs, increase operation efficiency by 300%, and achieve a weed removal rate of 96.3% with a seedling damage rate of only 0.4%. This solves the problem of traditional weeding not keeping up with weed growth, ensuring crop survival rates. It also reduces pesticide residues and soil pollution, promoting green agricultural development. The intelligent technology replaces the extensive mode and provides a replicable paradigm for large-scale precision agriculture, helping to transform agriculture from relying on weather to making decisions based on weather. Ultimately, it achieves an 8% increase in yield per mu, providing a solid foundation for food security and farmer income.

[0067] Next, a method for automatically aligning and moving a weeding machine based on machine intelligence is described with reference to the accompanying drawings according to an embodiment of the present application.

[0068] As shown in Figure 8 The method for automatically aligning and moving a weeding machine based on machine intelligence includes the following steps:

[0069] In step S101, environmental data and crop information data are obtained.

[0070] It can be understood that the embodiments of the present application capture field terrain, obstacles, and light environment data to perceive the constraints of the work scene in real time, accurately understand the environment, collect crop row position, row distance, and growth state, anchor the alignment target, clarify the alignment reference, improve alignment accuracy, reduce damage to crops during weeding, enhance adaptability to complex field environments, reduce dependence on manual monitoring, and improve automation and scalability of operations.

[0071] In step S102, the crop information data is output to the alignment reference parameters through a crop row feature extraction algorithm, the environmental data and the crop information data are fused, feature extraction and coordinate alignment are performed based on a spatial correlation algorithm, the three-dimensional pose of the weeding machine is solved through a beam adjustment algorithm, and a high-precision map of the work environment is constructed.

[0072] Wherein, the row reference parameter is a reference standard for rows as a unit, and by specifying the reference scale of each row, the precise evaluation or standardized management of specific units is realized.

[0073] It can be understood that the embodiments of the present application realize coordinate alignment and feature matching of multi-source data through unified reference, make the bundle adjustment algorithm calculate the ground reference of the three-dimensional pose of the device, constrain the positioning deviation by the features such as the trend and spacing of the crop row, improve the pose accuracy, improve the accuracy and stability of the weeding operation, reduce the problem of missing weeds or mistakenly damaging crops, reduce the dependence on high-end sensors to control costs, and provide data support for subsequent intelligent operations such as path planning and parameter optimization.

[0074] It should be noted that the three-dimensional pose is a comprehensive spatial state representation that describes the position (translation) and attitude (rotation) of an object in three-dimensional space.

[0075] The spatial correlation algorithm is an analysis method for data with spatial properties such as position and topology, which analyzes spatial position, distance, adjacency, and other relationships to mine implicit correlation patterns between spatial objects or attributes.

[0076] The bundle adjustment algorithm is a global optimization method used in photogrammetry, computer vision, and other fields to optimize camera parameters and three-dimensional spatial point coordinates to minimize observation errors.

[0077] Crop row feature extraction algorithm:

[0078]

[0079] Wherein, k is the inclination of the straight line in the two-dimensional coordinate system; n is the total number of effective feature points detected in the crop row; x i y i The discrete feature points are the basis data for fitting the straight line, and the overall row center line is deduced from the distribution trend of these points; is the sum of the product of the horizontal and vertical coordinates of all feature points; x i is the horizontal coordinate of the i-th point; y i is the vertical coordinate of the i-th point; is the sum of the horizontal coordinates; is the sum of the vertical coordinates; is the sum of the squares of the horizontal coordinates of all feature points; is the square of the sum of the horizontal coordinates.

[0080] Spatial correlation algorithm:

[0081]

[0082] wherein, R is a rotation matrix; t is a translation vector; N is the number of points in the source point cloud p, i = 1 is the summation of all items from 1 to N; p i is the coordinate in three-dimensional space; q σ(i) is p i is the index of the nearest neighbor point in the target point cloud q.

[0083] Bundle adjustment algorithm:

[0084]

[0085] wherein, P j is a set of three-dimensional poses of the weeding machine at different times; X i is a three-dimensional feature point in the working environment; ij is the summation of all i and j; x ij is a two-dimensional vector; π is the projection calculation of a three-dimensional point to a two-dimensional image.

[0086] For example, in a corn field at the jointing stage, through the data collected by the vehicle-mounted visual sensor and laser radar, the row reference parameters obtained by the crop row feature extraction algorithm are as follows: row spacing 65 cm, guiding the weeding machine to adjust the working width to avoid the plants; row direction angle 3 degrees (north deviation east), used to calibrate the forward direction of the machine body to keep parallel with the crop row; first row reference point coordinates (2.5 m, 1.2 m, 0.3 m), as the spatial reference for the starting positioning of the work; row straightness deviation 5 cm, supporting the dynamic correction of the path of the weeding machine to adapt to the natural bending of the field. These parameters together constitute the spatial ruler for the row operation of the agricultural machine, providing the core basis for precise weeding.

[0087] In step S103, based on the high-precision map of the working environment, combined with the three-dimensional pose of the weeding machine and the row reference parameters, the target row translation trajectory of the weeding machine is obtained through the proximal strategy optimization algorithm, the target row translation trajectory is converted into an actuator signal, and a control instruction is output.

[0088] wherein, the row translation trajectory is a motion path with a corresponding relationship formed by multiple objects in translational motion due to functional requirements.

[0089] It can be understood that the embodiments of the present application can dynamically plan a precise moving path in the direction perpendicular to the crop row according to the real-time three-dimensional pose of the weeding machine and the crop row reference information provided by the high-precision map through the proximal strategy optimization algorithm; perform precise alignment of the weeding actuator, ensure that the weeding operation acts on the weed area between the rows, protect the crops, improve the weeding effect, adapt to the dynamic changes in the field, and improve the working efficiency, reduce resource waste and mechanical loss.

[0090] It should be noted that the actuator signal is a physical signal used by the controller to control the actuator to deliver control instructions and receive state feedback.

[0091] Proximal policy optimization algorithm:

[0092] L CLIP (θ)=E t [min(r t (θ)A t ,clip(r t (θ),1-∈,1+∈)A t ]

[0093] Wherein, L CLIP is the target to be optimized by the clipping algorithm; θ is the neural network model parameter generating the row translation trajectory; E t is the statistical average; r is the probability ratio of the new and old strategies; A t is the benefit of measuring the strategy update, making the strategy more inclined to improve the actions with high advantages; r t (θ) is the probability ratio of the new and old strategies in the foregoing; 1-∈, 1+∈ are the upper and lower limits of clipping.

[0094] For example, in the corn field weeding operation, the high-precision map has marked the third row of corn center line trajectory as a straight line y=2.5m extending along the x-axis, the row distance is 0.6m, and the requirement for the row reference is that the center axis of the weeding machine needs to maintain a 0.2m offset with the center line (i.e. the target trajectory y=2.7m), and the heading angle deviation from the corn row direction does not exceed 0.5 degrees. When the initial pose of the weeding machine is x=5.0m, y=2.8m, and the heading angle is 1.2 degrees (there is a 0.1m right deviation and a 0.7 degree heading deviation), the proximal policy optimization algorithm is based on the future trajectory prediction of the corn row in the high-precision map to optimize the lateral translation in real time with the goal of minimizing the row deviation, and the generated target row translation trajectory shows that as the weeding machine advances along the x-axis, the y-coordinate is smoothly adjusted from the initial 2.8m to 2.7m and remains stable, and the heading angle is simultaneously corrected to within 0.3 degrees, and then the system converts the trajectory into steering and traveling signals of the actuator to output precise control instructions to ensure that the weeding machine operates stably along the target trajectory.

[0095] In step S104, based on the target row translation trajectory and the control instructions, the weeding execution mechanism is driven to automatically perform row translation weeding along the crop row, the full-link operation data is obtained through the time synchronization and calibration algorithm, the row translation performance evaluation index is established, the weeding machine is evaluated, and the row translation performance evaluation index of the weeding machine is dynamically updated and the target row translation trajectory is optimized through the continuous learning algorithm.

[0096] Wherein, the weeding execution mechanism is the core functional component of the weeding device or the automatic system that directly implements the weeding action, responsible for converting control signals or power into specific physical actions to remove weeds.

[0097] It can be understood that the embodiments of the present application carry out weeding by mechanical cutting, energy inactivation and the like, accurately respond to control instructions, execute the terminal of stable movement along the target parallel translation trajectory, and act on the weed area between crop rows to avoid damage, thereby improving the automation efficiency of weeding operation, reducing labor cost, improving operation safety through accurate trajectory following, reducing crop damage, combining with the continuous learning algorithm to promote trajectory dynamic optimization, reducing pesticide use through physical weeding, and reducing environmental and long-term operation and maintenance costs.

[0098] It should be noted that the full-link operation data refers to the sum of data such as operation state, interaction information and performance index covering each link of the complete business link from the starting point to the end point.

[0099] The parallel translation performance evaluation index is a quantitative measurement standard for measuring the speed, accuracy and stability of an object or system when moving in the horizontal direction.

[0100] The time synchronization and calibration algorithm is to eliminate the problems caused by clock deviation, drift or asynchrony by dynamic adjustment and correction, so that the time reference of different devices, systems or nodes tends to be unified.

[0101] Time synchronization and calibration algorithm:

[0102]

[0103] Wherein, T is the time offset; is the optimal time offset; arg min is the parameter value that minimizes; N is the number of sampling points; i is the sampling point index; p ret is the reference trajectory; t i is the i-th sampling time under the control time reference; p meas is the actual measurement trajectory; t i +T is the conversion of control time t i to sensor time.

[0104] Continuous learning algorithm:

[0105]

[0106] Wherein, L is the difference between the model prediction result and the real target; θ is the variable that the model needs to learn; new is the loss corresponding to the latest weeding machine parallel translation data collected at present; λ is a hyperparameter that needs to be set artificially; i is the traversal of all parameters of the model; F is the abbreviation of information matrix; θ i is the i-th component of the model parameters in the current new task training; is the optimal value of the model parameters after the old task training is completed.

[0107] For example, in the corn field, the electric translation type weeding execution mechanism as the core operation component, through the visual sensor to identify the corn row to generate the target translation trajectory, receive the PWM instruction output by the PID controller, drive the knife group to work along the trajectory transversely; During the operation, the real-time position, motor current and other data of the execution mechanism are collected through time synchronization technology, and are linked with perception, control and weeding effect data to support the evaluation of average position deviation, response delay and other translation performance. The first evaluation found that the trajectory deviation exceeded the threshold and the weed removal rate was insufficient due to the lag of visual recognition and high load, and through the continuous learning algorithm, the pre-compensation trajectory was generated at the sharp bend and the control parameters were dynamically adjusted, finally the average position deviation of the execution mechanism was reduced to 1.7 cm, the response delay was shortened to 0.15 s, the weed removal rate was improved to 94%, and the crop damage rate was reduced to 0.6%, realizing the precision and adaptive upgrading of row weeding.

[0108] PID controller:

[0109]

[0110] Wherein, u(t) is the control instruction for driving the motor to move the knife group; K p is the sensitivity to the current deviation; e(t) is the difference between the target trajectory position and the actual position; K i is the influence of the cumulative historical deviation; is the quantification of long-term deviation; K d is the sensitivity to the deviation change speed; is the prediction of future deviation trend.

[0111] According to the automatic row translation method of the weeding machine based on machine intelligence provided in the embodiments of the present application, through the multi-modal environment perception module and the crop row feature extraction algorithm, the complex terrain and crop shielding interference in the field are effectively dealt with, and the three-dimensional positioning accuracy and operation environment perception ability of the weeding machine are improved by cooperating with the light beam adjustment pose solution and high-precision map construction module; Relying on the near-end strategy optimization algorithm and the decision control module, the row translation trajectory is accurately planned and the control instruction is output, the path tracking robustness is strengthened in the scenes such as crop row bending and local missing, and the row translation precision is improved. At the same time, the adaptive execution module drives the weeding unit, combines the dynamic trajectory following and real-time response mechanism, and can automatically translate and weed along the crop row, effectively avoiding the crop damage and missing problems caused by trajectory deviation, and controlling the weeding operation quality; The self-learning module dynamically optimizes the decision strategy and translation trajectory through the performance evaluation index and continuous learning algorithm, improves the adaptive ability and long-term operation efficiency of the system in diversified farmland environment, significantly improves the weeding efficiency, protects the safety of crops and reduces the cost of manual intervention. Thus, the problems of high risk of checking seedlings, poor terrain adaptability and low operation efficiency in the prior art are solved.

[0112] An automatic row-parallel translation method for a machine intelligence-based weeding machine will be described below through a specific embodiment, as shown, comprising: Figure 9

[0113] At a certain 10,000-mu wheat planting base in Chuzhou, Anhui, a tracked intelligent weeding machine integrating multi-dimensional perception and intelligent decision-making system is conducting spring seedling stage weeding operation, and its whole process operation strictly relies on the closed-loop technology architecture of environment perception-data fusion-trajectory planning-execution feedback-dynamic optimization: at 5:30 am, after the operation system starts self-checking, the sensor groups distributed at different positions of the machine body enter the working state synchronously-the 16-line laser radar at the top (scanning frequency 20 Hz, horizontal field of view 360 degrees, vertical field of view 45 degrees, ranging accuracy ±1 cm) performs three-dimensional environment scanning on the operation area at a density of 2 million points per second, focusing on collecting the ridge height (range 5-30 cm) of the wheat field boundary, the concave-convex degree (maximum fluctuation 4 cm) of the soil surface, and the sporadic distribution of stone obstacles (diameter ≥5 cm, i.e. marked); the three-camera array (including RGB camera, near-infrared camera, and thermal imaging camera, with a resolution of 2560×1440 and a frame rate of 30 fps, and a lens focal length of 8 mm) at the front end is installed at a height of 1.2 m from the ground to shoot the wheat seedling image at a 30-degree angle of depression, the RGB camera captures the color characteristics of the plants (wheat seedlings are tender green, with RGB value range R: 120-150, G: 180-220, B: 80-110), the near-infrared camera enhances the gray difference between vegetation and soil (vegetation reflectivity ≥60%, soil ≤30%), and the thermal imaging camera records the temperature distribution of the plants (wheat seedlings have a body temperature of 18-22℃, and weeds are slightly lower by 1-2℃ due to transpiration); the spectrometer (wavelength resolution 1 nm, sampling interval 0.2 s) in the middle of the machine body collects the ground reflectance spectrum in real time through the optical fiber probe, and calculates the NDVI value of wheat (0.6-0.75), the NDVI value of weeds (0.3-0.5), and the spectral difference index between the two; at the same time, the vehicle-mounted GNSS module (supports Beidou+GPS dual-mode RTK, positioning frequency 10 Hz, plane accuracy ±1 cm, height accuracy ±2 cm) and IMU (6-axis, sampling rate 200 Hz, angular velocity measurement range ±2000 degrees / s, acceleration ±16g) combination outputs the real-time coordinates (near North Latitude 32 degrees 15', East Longitude 118 degrees 20') of the weeding machine, the driving speed (0.5-1.2 m / s), and the attitude parameters (heading angle 0-360 degrees, pitch angle -5 degrees to 5 degrees, roll angle -3 degrees to 3 degrees), the environmental sensor group (temperature and humidity sensor, light sensor, soil sensor) respectively collects air temperature (15-18℃), humidity (60%-65%), light intensity (5000-10000 lux), and 0-10 cm deep soil moisture content (18%-22%), soil hardness (1.2-1.8 kg / cm 2 ​), all original data are aggregated through an Ethernet switch, and then filtered and preprocessed (statistical filtering is used for laser radar point cloud denoising, and points outside 3 times the standard deviation of the distance mean are removed; Gaussian filtering with a kernel size of 3*3 is used for camera image denoising) by an edge computing terminal (carrying NVIDIA Jetson AGX Orin, with a computing power of 275 TOPS).

[0114] In the data processing stage, feature extraction of crop information data is the core: firstly, the three-camera images are preprocessed (including white balance correction, distortion correction, and Zhang's calibration method is used for distortion correction with an error of less than or equal to 0.5 pixels), and then the improved Hough transform algorithm (combined with edge detection and morphological operation, the Canny operator is used for edge detection with a threshold of 0.15 and 0.35, and a 3*3 rectangular structure element is used for morphological operation for erosion and expansion) is used to extract the wheat row features, the continuous wheat stem axis is recognized (a feature point is extracted every 10 cm), the least square method is used to fit the straight line equation (y=kx+b) of each row of wheat, and the row reference parameters are calculated, including the average row spacing (18±0.8 cm), the row direction angle (25 degrees±0.5 degrees), the starting point coordinates (x0, y0) and the end point coordinates (x1, y1) of each row, and the dense area of weeds between rows (the area with a pixel ratio of weeds of more than or equal to 20%) is marked. In the data fusion link, the laser radar point cloud and the camera image features are registered based on the spatial correlation algorithm (improved ICP algorithm, iteration times 20, convergence threshold 0.01 m), the image pixel coordinates are converted into camera coordinate system three-dimensional coordinates through the camera external parameter matrix (rotation matrix R, translation vector T), and then unified to the world coordinate system through the coordinate conversion matrix (based on GNSS-IMU calibration results), the coordinate alignment (registration error ≤2 cm) is completed; for abnormal data caused by weed shielding and soil reflection, the RANSAC algorithm (iteration times 50, inlier threshold 0.05 m) is used for removal, and the proportion of abnormal points is about 6%. Subsequently, the Bundle Adjustment algorithm is used to optimize the multi-source data, the camera internal parameters (focal length f, principal point cx, cy) and the laser radar calibration parameters (ranging error model) are used as constraints, the re-projection error equation (error function is Σ||u_i-P(R_iX_j+T_i)||2, where u_i is the image pixel coordinate, and P is the projection matrix) is constructed, and the least square problem is solved through the Levenberg-Marquardt algorithm, and finally the three-dimensional pose (position error ±2.5 cm, attitude angle error ±0.3 degrees) of the weeding machine in the world coordinate system is calculated; combined with the above data, the high-precision map of the working environment is constructed, the map is stored in a raster format with a resolution of 3 cm*3 cm, and includes the accurate position layer of the wheat row, the weed distribution layer (divided into low <20%, medium 20%-40%, and high >40% according to the coverage rate), the terrain elevation layer (elevation accuracy ±1 cm), and the obstacle layer (marking the location of ridges and stones).

[0115] Trajectory planning phase, based on high-precision map, combined with the current three-dimensional pose of weeding machine and the reference parameters of the row, set the core optimization goal: keep the transverse distance between the weeding mechanism (reciprocating weeding knife, knife width 20 cm) and the wheat row stable at 6 ± 0.5 cm to avoid damaging the seedlings. The proximal strategy optimization algorithm (PPO) is used for trajectory generation. The algorithm input state includes: the deviation of the current position and the target row (x deviation, y deviation), the change rate of the row angle (≤0.02 rad / m), soil moisture content (affecting ground adhesion), soil hardness (affecting driving resistance), weed coverage (affecting weeding load); the output is the linear speed and angular speed command of the next step; the reward function is designed as a multi-objective weighted: row deviation reward (deviation ≤1 cm gets 10 points, each increase of 0.5 cm reduces 2 points), trajectory smoothness reward (curvature change ≤0.05 rad / m gets 8 points), work efficiency reward (speed ≥0.8 m / s gets 5 points), energy consumption reward (motor power ≤1.5 kW gets 7 points), through 5000 times of simulation iteration training (each iteration contains 1000 decision steps), the model converges to generate a target row translation trajectory. The trajectory is a multi-segment continuous curve, the straight line segment (middle area of the row) maintains a uniform speed of 0.9 m / s, the turning segment (transition area of row change) uses a quintic polynomial curve planning, the speed is smoothly reduced from 0.9 m / s to 0.4 m / s, and then increased to 0.9 m / s, the turning radius is ≥5 m to ensure stability; after the trajectory is generated, it is converted into execution mechanism control signal through kinematics model: according to the trajectory curvature, the speed difference of the double-track belt is calculated (left track belt speed v1=v-ωL / 2, right track belt v2=v+ωL / 2, where v is the linear speed, ω is the angular speed, L is the track spacing 1.2 m), the track speed command (left track 180-200 r / min, right track 170-190 r / min) is obtained, at the same time, according to the weed coverage, the weeding knife motor PWM signal is adjusted (low coverage duty ratio 60%-70%, medium 70%-80%, high 80%-90%), the control command is output to the execution controller through CANopen protocol.In the execution stage, the weeding machine starts the track driving system and the weeding knife motor according to the instructions. The track moves along the target trajectory through differential motion driven by the hydraulic system, and the weeding knife cuts back and forth at a speed of 2500 r / min. The system synchronously starts the time synchronization and calibration mechanism. Based on the GNSS PPS second pulse signal, all sensor data and actuator feedback signals are time-stamped and aligned (synchronization accuracy ≤0.5 ms). Real-time collection of full-link operation data: actual track speed is obtained through a wheel speed encoder (sampling frequency 100 Hz), post-camera (resolution 1280×720, frame rate 15 fps) is used to take images after operation, the deviation between actual trajectory and target trajectory is calculated (lateral deviation 0.3-2.1 cm, longitudinal deviation 0.2-1.5 cm), current sensor is used to monitor the weeding knife motor current (6-10 A), image recognition is used to calculate the weed removal rate (90%-96%), and GNSS is used to record the operation area (statistics every 5 minutes).

[0116] In the performance evaluation stage, a multi-dimensional evaluation index system is established: row precision (average deviation ≤1.2 cm is excellent, 1.2-2 cm is good), trajectory tracking stability (standard deviation of deviation within 100 m ≤0.6 cm), operation efficiency (≥0.4 hectares / hour), weed removal rate (≥90%), energy consumption index (≤1.0 kWh / hectare), and failure rate (≤0.1 times / hectare). The current operation data shows that the average row deviation is 1.1 cm (excellent), but the instantaneous deviation reaches 2.1 cm in the high weed area (coverage rate 45%), the removal rate is 92% (good), and the operation efficiency is 0.42 hectares / hour (excellent). Based on the evaluation results, the continuous learning algorithm (DDPG model based on deep reinforcement learning) calls the historical data of the last 5 operations (cumulative operation area 3.2 hectares) for model updating. Analysis finds that the trajectory fluctuates in the high weed area due to increased load, so the reward function weight of the proximal policy optimization algorithm is dynamically adjusted, the weed coverage weight is increased from 0.15 to 0.25, and the speed curve of trajectory planning is optimized, with deceleration starting 0.5 m earlier in the high weed area (speed reduced to 0.7 m / s). The PID controller parameters are iteratively optimized, with the proportional coefficient adjusted from 0.7 to 0.85, the integral time shortened from 0.6 s to 0.4 s, and the differential time kept at 0.2 s. The updated system reduces the maximum deviation in the high weed area to 1.6 cm, increases the average removal rate to 94%, and stabilizes the energy consumption at 0.95 kWh / hectare, achieving dynamic optimization of translation performance. During the entire operation process, the system uploads key data to the cloud management platform through the 4G network, displays the operation progress, device status and performance indicators in real time, and the operator can remotely monitor and issue adjustment instructions, ensuring that the intelligent weeding machine maintains high precision and efficiency in the complex wheat field environment, providing reliable technical support for spring wheat seedling management.

[0117] In summary, the embodiments of the present application significantly improve the complex field row accuracy and weeding effect through multi-source sensor fusion and algorithm optimization (1.1 cm average deviation, weed removal rate over 90%, and optimized deviation reduced to 1.6 cm in high weed area), realize accurate operation trajectory by combining high-precision map and three-dimensional pose solution; through continuous learning algorithm and dynamic trajectory planning, ensure stable operation under different soil and weed distribution (trajectory fluctuation standard deviation ≤0.6 cm, speed adaptation response is fast), energy consumption is optimized and efficient; compared with traditional weeding technology, the efficiency is improved by more than 20 times, the pesticide usage is reduced by 60%, the operation energy consumption is reduced to 0.95 kWh / ha, the income benefit is significantly improved under large-scale planting, and the whole link data is automatically recorded, cloud monitoring and dynamic performance optimization are performed, which greatly improves the operation quality, green benefit and intelligent management level, and has strong environmental adaptability and sustainable operation advantages.

[0118] Figure 10 The structure schematic diagram of the electronic device provided by the embodiments of the present application is shown. The electronic device can include:

[0119] The memory 1001, the processor 1002 and the computer program stored in the memory 1001 and executable on the processor 1002.

[0120] The processor 1002 executes the program to realize the automatic row translation method of the weeding machine based on machine intelligence provided in the above embodiments.

[0121] Further, the electronic device further includes:

[0122] The communication interface 1003 is used for communication between the memory 1001 and the processor 1002.

[0123] The memory 1001 is used to store the computer program executable on the processor 1002.

[0124] The memory 1001 can contain a high-speed RAM (Random Access Memory, Random Access Memory) memory, and can also include a non-volatile memory, such as at least one disk memory.

[0125] If the memory 1001, the processor 1002 and the communication interface 1003 are implemented independently, the communication interface 1003, the memory 1001 and the processor 1002 can be connected with each other through a bus and complete communication between each other. The bus can be an ISA (Industry Standard Architecture, Industry Standard Architecture) bus, a PCI (Peripheral Component, Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture, Extended Industry Standard Architecture) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For convenience of representation, Figure 10 Only one thick line is used to represent the bus in the figure, but it does not mean that there is only one bus or only one type of bus.

[0126] Optionally, in a specific implementation, if the memory 1001, the processor 1002 and the communication interface 1003 are integrated on a chip, the memory 1001, the processor 1002 and the communication interface 1003 can complete communication between each other through an internal interface.

[0127] The processor 1002 can be a CPU (Central Processing Unit, Central Processing Unit) or an ASIC (Application Specific Integrated Circuit, Application Specific Integrated Circuit) or one or more integrated circuits configured to implement one or more embodiments of the application.

[0128] The embodiment of the application further provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to implement the automatic alignment translation method of the weeding machine based on machine intelligence.

[0129] In addition, the embodiment of the application further provides a computer program product, which includes a computer program or instructions, and the computer program or instructions are executed to implement the automatic alignment translation method of the weeding machine based on machine intelligence.

[0130] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative expressions of the above terms are not necessarily directed to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. Furthermore, the person skilled in the art can combine and combine the different embodiments or examples described in the present specification and the features of the different embodiments or examples, without contradiction.

[0131] In addition, the terms "first", "second" are used only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Thus, the features defined with "first", "second" can include at least one of the features, explicitly or implicitly. In the description of the present application, the meaning of "a plurality of" is at least two, for example, two, three, etc., unless otherwise explicitly specified.

[0132] Any process or method descriptions in flow charts or described elsewhere herein can be understood as representing one or more steps of a method or process, including a set of executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of the present application includes additional implementation involving other steps, which can be performed at substantially the same time or in reverse order or in other order, depending on the functionality involved, as will be understood by those skilled in the art of the embodiments described herein.

[0133] It should be understood that parts of the present application can be implemented in hardware, software, firmware or a combination thereof. In the above-described embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. As in another embodiment, if implemented in hardware, any of the following technologies known in the art or a combination thereof can be used: discrete logic circuit with logic gate circuit for implementing logic functions on data signals, application specific integrated circuit with suitable combination logic gate circuit, programmable gate array (PGA), field programmable gate array (FPGA) and the like.

[0134] Those skilled in the art of the present technology can understand that all or part of the steps carried out by the above-described embodiment method can be completed by a program instructing the relevant hardware, and the program can be stored in a computer readable storage medium. The program, when executed, includes one or a combination of steps of the method embodiment.

[0135] Although the embodiments of the present application have been shown and described above, it is understood that the above-described embodiments are exemplary and are not to be construed as limiting the present application, and that changes, modifications, substitutions and variations can be made by those skilled in the art without departing from the scope of the present application.

Claims

1. An automatic row-to-row translation system for a weeder based on machine intelligence, characterized in that, include: The module comprises a multimodal perception module, an environment map construction module, a decision control module, an adaptive execution module, and a self-learning module; among which, The multimodal sensing module is used to acquire environmental data and crop information data; The environmental map construction module, based on the crop information data, outputs row reference parameters through a crop row feature extraction algorithm, integrates the environmental data and row reference parameters, and calculates the three-dimensional pose of the weeder in real time through a bundle adjustment algorithm to construct a high-precision map of the working environment. The decision control module, based on the high-precision map of the operating environment, obtains the target's horizontal translation trajectory through a near-end strategy optimization algorithm and outputs control commands. The adaptive execution module drives the weeding execution mechanism to automatically perform row-to-row weeding based on the target row translation trajectory and the control command. The self-learning module is used to establish a performance evaluation index for line translation based on the full-link operation data, and dynamically update the target line translation trajectory through a continuous learning algorithm.

2. The automatic row-to-row translation system for a weeder based on machine intelligence according to claim 1, characterized in that, The multimodal sensing module includes a lidar sensing unit and an environmental parameter sensing unit. The lidar sensing unit is used to acquire data on crop distribution, weed characteristics, and other information. The environmental parameter sensing unit is used to acquire data on changes in light intensity, temperature and humidity fluctuations, and diverse environmental data related to complex terrain.

3. The automatic row-to-row translation system for a weeder based on machine intelligence according to claim 1, characterized in that, The environmental map construction module includes a baseline parameter unit, a data fusion unit, a BA pose optimization unit, and a map construction unit. The baseline parameter unit dynamically locates the row baseline using a crop row feature extraction algorithm and outputs the row baseline parameters. The data fusion unit fuses the environmental data with the row baseline parameters and performs feature extraction and coordinate alignment based on a spatial association algorithm. The BA pose optimization unit obtains the 3D pose and feature point positions using bundle adjustment based on the feature extraction and coordinate alignment, and corrects accumulated errors using loop closure detection. The map construction unit constructs a high-precision map of the working environment based on the 3D pose and feature point positions, combined with the row baseline parameters.

4. The automatic row-to-row translation system for a weeder based on machine intelligence according to claim 1, characterized in that, The decision control module includes an intelligent strategy decision-maker and a control command generation unit. The intelligent strategy decision-maker generates a target pair translation trajectory that conforms to the operation constraints based on a high-precision map of the operating environment and the equipment status through a near-end strategy optimization algorithm. The control command generation unit is used to convert the target pair translation trajectory into actuator signals and output standardized control commands.

5. The automatic row-to-row translation system for a weeder based on machine intelligence according to claim 1, characterized in that, The adaptive execution module includes an instruction receiving and posture adjustment unit and an automatic alignment unit. The instruction receiving and posture adjustment unit is used to receive the control instructions and adjust the lateral position, working height and angle of the weeding mechanism. The weeding action driving automatic alignment unit is used to autonomously identify the row and column structure of crop growth, so that the weeder is precisely aligned with the crop row and moves along the row to carry out weeding operations.

6. The automatic row-to-row translation system for a weeder based on machine intelligence according to claim 1, characterized in that, The self-learning module includes a performance evaluation index unit and a continuous learning unit. The performance evaluation index unit is used to establish a performance evaluation system and key quantitative indicators based on the full-link operation data. The continuous learning unit uses a continuous learning algorithm to optimize the target row translation trajectory to obtain a target row translation trajectory suitable for weeding operation scenarios.

7. A method for applying the automatic row-to-row translation system of a weeder based on machine intelligence as described in any one of claims 1-6, characterized in that, include: Acquire environmental and crop information data; The crop information data is processed by a crop row feature extraction algorithm to output row reference parameters. The environmental data is fused with the crop information data, and feature extraction and coordinate alignment are performed based on a spatial association algorithm. The three-dimensional pose of the weeder is calculated using a bundle adjustment algorithm to construct a high-precision map of the working environment. Based on the high-precision map of the working environment, combined with the three-dimensional pose of the weeder and the row reference parameters, the target row translation trajectory of the weeder is obtained through the near-end strategy optimization algorithm. The target row translation trajectory is converted into an actuator signal and a control command is output. Based on the target row translation trajectory and the control command, the weeding actuator is driven to automatically perform row-to-row weeding along the crop rows. Through time synchronization and calibration algorithms, full-link operation data is obtained, row translation performance evaluation index is established, the weeding machine is evaluated, and the row translation performance evaluation index of the weeding machine is dynamically updated and the target row translation trajectory is optimized through continuous learning algorithms.

8. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, the processor executing the program to implement the automatic row translation method for a weeding machine based on machine intelligence as described in claim 7.

9. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed, they implement the automatic row translation method for a weeding machine based on machine intelligence as described in claim 7.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed, they implement the automatic row translation method for a weeding machine based on machine intelligence as described in claim 7.

Images