Adaptive laser weeding method and system

Abstract

The application discloses a self-adaptive laser weeding method and system, and belongs to the technical field of intelligent agricultural equipment. Field images, three-dimensional point clouds and environmental parameters are synchronously collected through a multi-modal sensor to construct a micro-environment dynamic model containing three-dimensional spatial information, geometric characteristics, categories and environmental states of plants. An energy deposition physical model is established based on Fourier heat conduction law to dynamically optimize laser parameters with the goal of killing weeds and minimizing crop temperature rise. A multi-level safety mechanism is used to prevent and control risks in real time during execution. The effect is quantitatively evaluated after operation, and the thermal physical parameters are updated online to realize model self-adaptive calibration. The system comprises a mobile platform, a sensing module, a laser execution module, a shielding structure and an industrial computer. The application realizes a leap from two-dimensional identification to three-dimensional perception, from empirical parameters to physical simulation, and from open-loop execution to closed-loop learning, and significantly improves the precision, safety and environmental adaptability of weeding.

Description

Technical Field

[0001] This invention belongs to the field of intelligent agricultural equipment technology, and relates to an adaptive laser weeding method and system, specifically a precise intelligent weeding method and system that integrates environmental perception, physical modeling and online learning. Background Technology

[0002] Traditional agricultural production relies heavily on chemical herbicides for weed control. Long-term use of chemical herbicides can lead to soil degradation, water pollution, decreased biodiversity, and increased pesticide residues in agricultural products, negatively impacting the ecological environment and human health. Therefore, researching and developing precise, efficient, and environmentally friendly weed control technologies has become an important direction for modern agricultural development.

[0003] In recent years, with the development of smart agriculture technology, physical weeding techniques have gradually attracted attention. Among them, laser weeding technology, due to its non-contact treatment, controllable area of ​​action, and absence of chemical residues, has shown promising application prospects in the field of precision agriculture.

[0004] Existing laser weeding systems typically combine visual recognition with laser action. The basic process is as follows: a visual sensor collects image data of the work area; a deep learning model identifies and locates crops and weeds in the image, obtaining spatial location information of the target weeds; then, according to pre-set laser parameters, laser irradiation is applied to the target location, thereby achieving thermal damage or removal of the weeds.

[0005] However, existing laser weeding technology still has the following problems in practical applications:

[0006] (1) Limited dimensions of perceived information and insufficient basis for decision-making: Existing solutions mostly rely on two-dimensional visual images to identify and locate weeds, lacking effective acquisition and utilization of three-dimensional environmental information in the field, such as plant height, canopy structure, and spatial distance between weeds and crops. When crops and weeds are in close proximity, the system has difficulty accurately assessing the impact of heat diffusion during laser action on surrounding crops, which can easily lead to accidental damage.

[0007] (2) Fixed operating parameters and insufficient environmental adaptability: In existing systems, key parameters such as laser power and irradiation time are usually preset or graded according to weed type, failing to comprehensively consider dynamic factors in the operating environment, including the influence of leaf surface water content, air humidity, soil moisture content, and plant size and physiological state on laser energy absorption and heat conduction. Changes in these factors can lead to unstable laser effects, manifested as insufficient energy utilization or excessive consumption, thereby affecting weeding effectiveness and system operating efficiency.

[0008] (3) Lack of process prediction mechanism and insufficient safety control capability: Before performing laser irradiation, existing systems usually lack modeling and prediction of the interaction process between the laser and the plant and the surrounding environment, and cannot predict the energy deposition and heat diffusion process. Therefore, it is difficult to identify potential risks in a timely manner during actual operation, which may have adverse effects on nearby crops or soil environment, and the safety of the system needs to be improved.

[0009] (4) Lack of feedback and adaptive optimization mechanisms: In existing technologies, most systems have not established a sound mechanism for evaluating and providing feedback on weeding effects, and lack the ability to update models and optimize parameters based on historical operation data. The system is unable to adjust its operation strategy according to different farmland environmental conditions, thus limiting its long-term stability and adaptability.

[0010] In summary, existing laser weeding technologies still have shortcomings in environmental perception, adaptive parameter adjustment, process prediction, and feedback optimization. Therefore, it is necessary to propose a laser weeding method and system that can comprehensively utilize multi-source sensing information, possess process modeling and parameter optimization capabilities, and perform continuous adaptive adjustments to improve the accuracy, safety, and stability of weeding operations. Summary of the Invention

[0011] The purpose of this invention is to address the problems existing in the prior art, such as limited dimensions of sensing information, fixed setting of operation parameters, lack of process prediction mechanism, and lack of adaptive optimization capability. This invention proposes an adaptive laser weeding method and system. This method constructs a field microenvironment model by integrating multi-source sensor data and establishes an energy deposition process model based on physical mechanisms. It analyzes and calculates the laser action process, thereby achieving dynamic optimization of laser weeding operation parameters and improving the accuracy and safety of weeding operations.

[0012] The adaptive laser weeding method provided in this application adopts the following technical solution:

[0013] An adaptive laser weeding method, characterized by comprising the following steps:

[0014] (1) The field environment is synchronously collected by a multimodal sensor system set on a mobile platform. The multimodal sensor system includes: a visual sensor for collecting color image information, a radar sensor for acquiring three-dimensional point cloud data, and a supplementary lighting device for ambient light compensation.

[0015] (2) Based on the multimodal data collected in step (1), a dynamic model of field microenvironment for laser thermal effect prediction is constructed in the industrial control computer. The dynamic model of field microenvironment includes: three-dimensional spatial coordinate information and three-dimensional geometric features of each plant individual in the working area. The three-dimensional geometric features include plant height, canopy projection area and stem diameter, plant category attributes identified based on deep learning model, and current environmental state parameters. The environmental state parameters include at least leaf surface humidity, air temperature and soil humidity.

[0016] (3) When a target weed to be removed is identified, based solely on the three-dimensional structural information of the target weed in the field microenvironment dynamic model, the spatial distance relationship between the target weed and the surrounding crops, and the environmental state parameters, a preset energy deposition physical model based on the Fourier heat conduction law is invoked. With the optimization objective of "making the temperature of the weed growth point reach the lethal threshold while minimizing the temperature rise of the sensitive tissues of the adjacent crops", the unique optimal combination of laser irradiation parameters for the target weed is dynamically solved. The laser irradiation parameters include laser power, irradiation time, and the offset of the focal position in three-dimensional space.

[0017] (4) Based on the optimized laser irradiation parameters, control the laser execution module to irradiate the weed target with laser, and at the same time monitor the safety status of the work area based on real-time sensing data, and trigger the safety response mechanism when the preset risk conditions are detected.

[0018] (5) After the laser irradiation is completed, environmental data is collected again by delay or periodic collection to quantitatively evaluate the weeding effect. The input data of this operation, the laser irradiation parameters and the effect evaluation results are used as training samples to update the physical property parameters such as specific heat capacity and thermal conductivity of plant tissue in the energy deposition physical model online, so as to realize the adaptive calibration of the model for specific farmland environment.

[0019] By adopting the above technical solution, a dynamic model of the field microenvironment is constructed by integrating multimodal data from vision, radar, and environmental sensors. Based on the energy deposition physical model of Fourier's heat conduction law, the laser power, time, and focal position are dynamically optimized to achieve the dual objectives of "weed killing and crop protection." Furthermore, an effect evaluation and online learning closed loop is established, which solves the problems of insufficient dimensions, fixed and rigid parameters, lack of process prediction, and lack of adaptive capabilities in traditional two-dimensional perception. This achieves a leap from experience-based weeding to physical simulation optimization, and from open-loop execution to closed-loop self-evolution.

[0020] Furthermore, in step (1), the data acquisition method of the multimodal sensor system is as follows: the visual sensor continuously acquires RGB image data of the work area at a preset frame rate; the radar sensor synchronously scans within a horizontal range to obtain three-dimensional point cloud data of the work area; wherein, the data frames output by the visual sensor and the radar sensor are assigned the same timestamp and synchronized at the hardware level in the industrial control computer to ensure the sub-pixel level alignment accuracy of subsequent spatial coordinate fusion; the supplementary lighting device is built into the closed-loop control loop of the industrial control computer, the industrial control computer analyzes the brightness and contrast of the image acquired by the visual sensor in real time, and adaptively adjusts the supplementary lighting intensity according to the dynamic changes of ambient light intensity, with the goal of maintaining the minimum image signal-to-noise ratio required for three-dimensional reconstruction; when the ambient light meets the requirements of three-dimensional reconstruction, the supplementary lighting device automatically enters a low-power standby state.

[0021] By adopting the above technical solution, sub-pixel alignment of visual and radar data is achieved through hardware-level timestamp synchronization, and a closed-loop adaptive supplementary lighting system is constructed to dynamically adjust the illumination intensity according to the image brightness. This solves the problems of spatiotemporal asynchrony of multi-source data, unstable illumination, and excessive energy consumption, ensuring high-quality data acquisition and system endurance around the clock.

[0022] Furthermore, in step (2), the specific method for constructing the dynamic model of the field microenvironment includes:

[0023] ① The images collected by the vision sensor are input into a pre-trained deep learning model to identify weeds and crops in the work area and obtain their corresponding two-dimensional pixel coordinates and category information;

[0024] ② By using hand-eye calibration and coordinate transformation, the two-dimensional pixel coordinates are spatially registered and fused with the three-dimensional point cloud data acquired by the radar sensor to determine the three-dimensional spatial position and geometric size parameters of each plant individual in the global coordinate system;

[0025] ③ Combining environmental sensor data or historical data, environmental state parameters are introduced into the field microenvironment dynamic model. The environmental state parameters include one or more of leaf surface humidity, soil humidity, and ambient temperature.

[0026] By adopting the above technical solution, the two-dimensional coordinates of plants are obtained by using a deep learning model to identify plant categories. The global three-dimensional position and geometric size are determined by hand-eye calibration and point cloud registration. Then, environmental sensor data is fused to solve the problems of missing two-dimensional to three-dimensional mapping, difficulty in multi-source data fusion, and unquantified environmental factors, thus generating a complete digital twin model that includes species, location, size, and environment.

[0027] Furthermore, in step (3), the method for dynamically optimizing laser irradiation parameters is as follows: an energy deposition physical model based on Fourier's law of heat conduction is used to describe the absorption, thermal conversion, and diffusion process of laser energy in plant tissues; the lethal temperature is reached at the weed growth point (meristematic tissue). Furthermore, minimizing the thermal impact on surrounding crop tissue is the objective function. By solving this optimization problem, the power P, time t, and focus offset are obtained. The optimal combination of ) can be formally represented as:

[0028]

[0029] in, This represents the simulated temperature at the target weed growth point; Indicates the preset lethal temperature; This represents the simulated temperature of sensitive tissues in the c-th neighboring crop; is the weighting factor; C is the set of neighboring crops.

[0030] By adopting the above technical solution, an energy deposition model is constructed based on Fourier's law of heat conduction. The laser parameter combination is optimized with the dual objective of minimizing the temperature rise of nearby crops while the weed growth point reaches the lethal temperature. This solves the problems of black box in the energy deposition process, blind parameter selection, and unstable weeding effect, making weeding a predictable and calculable precise process instead of trial and error.

[0031] Furthermore, in step (4), the safety response mechanism is a multi-level linkage control, specifically including:

[0032] Level 1 Response: When the radar sensor detects an unexpected moving object approaching the edge of the work area, the control system triggers an audible and visual alarm to provide a warning.

[0033] Level 2 response: When the vision sensor detects damage or abnormal shape of the occlusion structure, it controls the laser execution module to enter a safe working mode and reduces the laser output power;

[0034] Level 3 Response: When severe damage to the obstruction structure and the presence of a living organism entering the core operating area are simultaneously detected, the control system first invokes a pre-trained Bayesian neural network to quantify the uncertainty of the perceived data and calculate the posterior probability distribution of the currently detected target as a "living organism." Simultaneously, based on multi-frame temporal perceived data, the system predicts the target's trajectory and calculates the probability of it entering the laser beam propagation path (optical axis region) within a 200ms time window. The Level 3 response is triggered only when the following conditions are met:

[0035] ① The average probability of the target being identified as a living organism is higher than a preset threshold (e.g., 0.85), and the probability of its predicted trajectory entering the optical axis region is ≥95%;

[0036] ② The visual detection confidence level of the severely damaged occlusion structure is higher than a preset threshold (e.g., 0.9).

[0037] When the above conditions are met simultaneously, the control system immediately cuts off the laser power supply and drives the lifting mechanism to raise the laser beam completely away from its original focal position. If the uncertainty of the Bayesian neural network evaluation is high (such as the target classification probability being distributed across multiple categories, or the trajectory prediction confidence interval being too wide), the system automatically degrades to a level two response (reducing laser power and continuing to observe) instead of directly triggering a level three power outage, in order to avoid accidental shutdowns caused by brief disturbances such as field insects, fallen leaves, and dust.

[0038] By adopting the above technical solution, a three-level progressive safety mechanism is established. The three-level response uses a Bayesian neural network to quantify the uncertainty of biological identification and combines it with 200ms trajectory prediction. The laser power supply is cut off only when strict probability thresholds are met simultaneously, and the level automatically degrades when the uncertainty is high. This solves the problems of insufficient safety risk identification, high false triggering rate and lack of biological protection, and achieves a balance between safety and operational continuity.

[0039] Furthermore, in step (5), the quantitative assessment is achieved by comparing the visual characteristics and three-dimensional morphological changes of the target weeds before and after laser irradiation. The assessment indicators include one or more of the following: plant wilting degree, color change, and structural integrity. The model iterative learning adopts an online learning method, using historical operation data as training samples to continuously update the energy deposition physics model in order to correct the plant tissue thermophysical parameters in the model. The thermophysical parameters include specific heat capacity and thermal conductivity.

[0040] By adopting the above technical solution, data before and after laser irradiation is collected with a delay. The weeding effect is quantitatively evaluated through visual and three-dimensional morphological changes. The operation data is used as training samples to update the thermophysical parameters of plant tissues online. This solves the problems of lack of quantitative standards for weeding effect, deviation between model parameters and actual results, and stagnation of system capabilities, and gives the system the ability to self-evolve and become more accurate with use.

[0041] The adaptive laser weeding system provided in this application adopts the following technical solution:

[0042] An adaptive laser weeding system for implementing the above method is characterized by comprising: a frame, a mobile platform, a lifting mechanism, an industrial control computer, a power module, and a sensing module and an execution module installed on the lifting mechanism;

[0043] The sensing module includes: a visual sensor for acquiring color image data, a radar sensor for acquiring three-dimensional environmental information, and a supplementary lighting device for adjusting the lighting conditions of the work area.

[0044] The execution module includes: a laser for generating a laser beam and a galvanometer system for controlling the laser direction;

[0045] The lower part of the lifting mechanism is provided with a shielding structure to isolate the laser operation area;

[0046] The industrial control computer is electrically or communicatively connected to the sensing module, the execution module, the lifting mechanism, and the mobile platform, respectively, and is used to execute the method steps of any one of claims 1-6.

[0047] By adopting the above technical solution, a hardware system integrating a mobile platform, lifting mechanism, shielding structure, multimodal sensing module, laser execution module, and industrial control computer is established. The industrial control computer coordinates and executes the adaptive weeding method in a unified manner, which solves the problems of software and hardware disconnect, insufficient laser safety protection, and low system modularity, and provides an implementable hardware platform that is fully compatible with the method.

[0048] Furthermore, the mobile platform is a wheeled mobile platform, adopting a dual-wheel differential drive mode, and is controlled by an STM32 series microcontroller through CAN bus communication to achieve precise control of movement, turning and speed.

[0049] By adopting the above technical solution, using a wheeled dual-wheel differential drive structure and having the driver precisely controlled by an STM32 microcontroller via a CAN bus, the problems of poor adaptability to field terrain, insufficient motion control precision, and low communication reliability are solved, enabling flexible movement and precise positioning between narrow crop rows.

[0050] Furthermore, the radar sensor is a 16-line mechanical lidar with a horizontal field of view of 360° and an effective detection distance of not less than 150 meters; the high-definition camera is an industrial-grade RGB camera that supports USB 3.0 high-speed data transmission and a frame rate of not less than 30fps.

[0051] By adopting the above technical solution, configuring a 16-line 360° mechanical lidar (detection distance ≥150 meters) and an industrial-grade RGB camera (USB 3.0 interface, frame rate ≥30fps), the problems of insufficient point cloud density, incomplete field of view coverage, image transmission delay and low frame rate are solved, ensuring complete monitoring of the work area and clear imaging in dynamic scenes.

[0052] Furthermore, the industrial control computer integrates a dual software collaborative platform: the first software platform is developed based on Python and is responsible for running the YOLOv11n-SE deep learning model, processing visual data, constructing micro-environment models, and optimizing laser parameters; the second software platform is developed based on C++ and receives instructions from the first software platform via TCP / IP protocol or RS-485 serial port, and is responsible for the precise triggering control and safety interlocking logic of the laser.

[0053] By adopting the above technical solution, the industrial control computer integrates a first platform developed in Python (responsible for deep learning and parameter optimization) and a second platform developed in C++ (responsible for laser triggering and safety interlocking). The two platforms communicate collaboratively through a standard protocol, which solves the problems of contradiction between algorithm development efficiency and real-time performance, overload of single platform functions, and inconsistent communication protocols. It balances development efficiency and real-time control performance and ensures independent and reliable operation of the safety interlocking.

[0054] In summary, the present invention has at least one of the following beneficial technical effects:

[0055] (1) It has achieved a leap from two-dimensional recognition to three-dimensional environmental perception and modeling:

[0056] Existing technologies are limited to two-dimensional image recognition. This invention integrates visual and radar point cloud data to construct a dynamic digital twin model of the field microenvironment, containing precise three-dimensional coordinates, geometric morphology, and species attributes for each plant. This provides an irreplaceable data foundation for subsequent accurate spatial relationship analysis and physical simulation, and is a prerequisite for the system to achieve high precision and high security.

[0057] (2) A dynamic optimization method for laser parameters based on a physical model is proposed:

[0058] Unlike existing technologies that use preset or simple parameter grading, this invention introduces an energy deposition physics model (based on the law of thermal conduction) into the weeding decision-making process. Before each operation, the system performs a "virtual irradiation" simulation based on the current microenvironment model to quantitatively predict the thermal effects under different laser parameters. With the optimization objective of "effectively killing weeds and minimizing surrounding impacts," the system dynamically solves for the optimal parameter combination. This transforms the weeding process from "empirical" to a precise process that is "calculable, predictable, and optimizable."

[0059] (3) A closed-loop intelligent system covering the entire process of perception, decision-making, execution, and learning has been constructed:

[0060] This invention does not improve the identification or control modules in isolation, but rather designs a complete closed-loop intelligent system architecture. This system comprises the entire chain of "multimodal perception → environmental modeling → physical simulation and optimization → safety control execution → effect evaluation and feedback → online model learning." In particular, the introduction of effect feedback and iterative model learning mechanisms enables the system to accumulate actual field data, continuously optimize its core physical model and decision-making algorithms, and possess a self-evolving capability that becomes increasingly intelligent with use.

[0061] (4) Achieved system-level security control with deep integration of software and hardware:

[0062] This invention deeply integrates mechanical components (sunshade cloth, lifting mechanism), sensing systems (radar, vision), and control systems (industrial computer) through a multi-level safety linkage logic. Safety response is no longer a simple laser switch, but rather a progressive set of measures—including early warning, degradation, and circuit breaking—based on risk levels (object approach, sunshade cloth malfunction, intrusion by a living being). This forms a proactive, comprehensive, and auditable safety protection system, greatly improving system reliability and application security. Attached Figure Description

[0063] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0064] Figure 2 This is a schematic diagram of the dynamic modeling of the field microenvironment and the optimization of laser parameters in this invention.

[0065] Figure 3 This is a flowchart illustrating the multi-level security response mechanism in this invention.

[0066] Figure 4 This is a schematic diagram of the structural composition of the adaptive laser weeding system of the present invention. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with the accompanying drawings and embodiments. The described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0068] This invention provides an adaptive laser weeding method and system based on multimodal perception and microenvironment modeling, which is described below in conjunction with... Figures 1 to 4 The embodiments of the present invention will be described in detail below.

[0069] Example:

[0070] This embodiment uses a laser weeding robot deployed on a wheeled mobile platform as an example to illustrate the specific implementation of the present invention.

[0071] Step S1: Synchronous acquisition of multimodal data.

[0072] After the laser weeding robot is started, its multimodal sensor system begins to work.

[0073] Visual acquisition: Two industrial-grade RGB high-definition cameras (such as the RGB module of Intel RealSense D435) continuously acquire color images of the rows of crops in front at a frame rate of 30fps, and transmit the data to the vehicle-mounted industrial control computer in real time via USB 3.0 interface.

[0074] 3D perception: A 16-line mechanical lidar performs a 360° horizontal rotation scan, generating dense 3D point cloud data of the work area, which is then synchronously transmitted to the industrial control computer.

[0075] Ambient light compensation: The industrial control computer analyzes the average brightness of the image in real time. When the ambient light is insufficient or there are strong shadows, it automatically adjusts the brightness of the supplementary light to ensure the contrast and clarity of the input image are stable, providing a reliable data source for subsequent recognition.

[0076] Step S2: Dynamic modeling of the field microenvironment.

[0077] After receiving the synchronized timestamp-aligned image and point cloud data, the industrial control computer executes the following modeling process (e.g., Figure 2 As shown):

[0078] 1. Two-dimensional recognition: RGB images are input into a YOLOv11n-SE model pre-trained on an agricultural weed dataset for inference. This model incorporates the SEAM attention mechanism, which effectively improves weed recognition accuracy under complex lighting conditions. The model output includes the bounding box, category (e.g., "corn", "foxtail grass"), and confidence score for each detected target in the image.

[0079] 2. Coordinate Fusion and 3D Reconstruction: The spatial transformation matrix between the camera and radar is obtained through pre-completed "hand-eye calibration". Using this matrix, the 2D pixel coordinates of weeds / crops output by YOLO are fused with the depth information of the corresponding spatial location in the radar point cloud to calculate the precise 3D coordinates (x, y, z) of each plant in the robot's global coordinate system, and to estimate its height, approximate canopy volume, and other geometric features.

[0080] 3. Environmental Status Fusion: Combining leaf gloss analysis (inferring surface humidity) with readings from onboard temperature and humidity sensors and possible historical soil moisture data, environmental status parameters such as leaf humidity, air humidity, and soil type are labeled onto the current microenvironment model. Ultimately, a real-time dynamic digital twin model is generated, including information on species, 3D location, geometric dimensions, and the surrounding environment.

[0081] Step S3: Dynamic optimization of laser parameters based on physical simulation.

[0082] Once the model identifies the weeds that need to be removed, the system does not immediately fire a laser, but instead performs virtual surgical planning first:

[0083] 1. Physical Model Invocation: The system incorporates a simplified energy deposition model based on Fourier's law of heat conduction. This model treats plant tissue as a material with specific specific heat capacity, thermal conductivity, and absorption coefficient.

[0084] 2. Set up the simulation scene: Extract the three-dimensional dimensions of the foxtail grass, its precise distance from the nearest corn seedling, and parameters such as the current leaf humidity and air humidity from the microenvironment model as input conditions for the simulation.

[0085] 3. Optimization solution: Using laser power P, irradiation time t, and the position of the laser focus on the weed stem (… ) are the decision variables. Simulation calculations are performed under different (P, t, ... Under the given conditions, the temperature reached at the growing point of foxtail grass. And the temperature rise as heat diffuses to the roots of the adjacent corn seedlings. Solve the following optimization problem:

[0086]

[0087] (Assuming the lethal temperature for weeds is 60°C and the crop protection weighting coefficient is 10).

[0088] 4. Output optimal parameters: A set of optimal parameters is obtained, for example: power = 28W, time = 170ms, focus shifted upwards by 2mm. This set of parameters minimizes the risk of heat damage to crops while ensuring weed killing, and is more energy-efficient than the fixed 30W / 200ms solution.

[0089] Step S4: Safety linkage control and precise execution.

[0090] While preparing to execute the optimized laser command, the system continuously performs safety monitoring (logic as follows). Figure 3 As shown):

[0091] Monitoring work area: The radar continuously scans to confirm that there are no moving objects within the area covered by the blackout cloth; the camera checks whether the blackout cloth is fully laid out through image recognition.

[0092] Tiered Response: When the radar sensor detects an unexpected moving object approaching the edge of the work area, a Level 1 response is triggered, and the control system activates an audible and visual alarm to provide a warning. When the visual sensor detects damage or abnormal shape of the obstruction structure, a Level 2 response is triggered, and the industrial control computer controls the laser execution module to enter a safe working mode and reduce the laser output power. When the system simultaneously detects severe damage to the obstruction structure and the presence of a suspected living object entering the core work area, it first uses a pre-trained Bayesian neural network to perform uncertainty quantification assessment on the perceived data, calculates the posterior probability distribution of the current detected target as a living object, and predicts the target's trajectory based on multi-frame temporal perception data, calculating the probability that it will enter the laser beam propagation path (optical axis region) within the next 200ms time window. Only when the target is a living object is the probability calculated. When three conditions are met simultaneously—the mean probability is higher than a preset threshold (0.85), the probability of the target's predicted trajectory entering the optical axis region is not less than 95%, and the visual detection confidence of severely damaged occlusion structures is higher than a preset threshold (0.9)—a level three response is triggered. The control system immediately cuts off the laser power supply and drives the lifting mechanism to raise the laser beam completely away from its original focal position. If the uncertainty of the Bayesian neural network assessment is high, including situations such as a dispersed target classification probability distribution (high classification entropy) or an excessively wide trajectory prediction confidence interval, the system automatically downgrades to a level two response. It maintains low-power laser output in safe mode and continuously monitors and accumulates more observation data. Once the uncertainty is reduced, the level three response conditions are reassessed, thus effectively avoiding accidental shutdowns caused by temporary environmental interference such as field insects, fallen leaves, and dust.

[0093] Safe execution: After confirming there are no safety risks or the risks have been eliminated, the industrial control computer sends the optimized parameters (28W, 170ms, +2mm) to the laser controller via RS-485 serial port. The laser and galvanometer system work together to complete the precise irradiation of the target weeds within milliseconds.

[0094] Step S5: Performance evaluation and model iterative learning.

[0095] The system's task does not end after laser irradiation:

[0096] 1. Delay assessment: After waiting for about 5 seconds, the control system drives the camera and radar to perform a rapid scan of the original target area again.

[0097] 2. Quantitative Assessment: Compare data before and after irradiation. Analyze images to see if the weeds turn dark green or brown (wilt), and analyze point clouds to see if their height has collapsed and their volume has decreased. Based on these characteristics, a quantitative score for the weed control effect is given (e.g., 0.9 indicates good effect, 0.3 indicates poor effect).

[0098] 3. Model Iteration: The complete data package of this task—input (environmental model parameters), action (laser parameters), and result (effect score)—is used as a new training sample and stored in the database. The background learning process is periodically started, and using the accumulated sample data, algorithms such as gradient descent are employed to fine-tune the thermophysical parameters of plant tissue in the energy deposition model, making the model's predictions increasingly closer to the physical response of real farmland.

[0099] System hardware configuration such as Figure 4 As shown, it includes: a wheeled mobile vehicle frame, an STM32 main control board, a CAN bus driver, a DC servo motor and encoder, a 16-line LiDAR, an RGB camera, a fill light, a 260W fiber laser module, a galvanometer system, a lifting servo motor, a light-blocking cloth, and the core embedded industrial control computer (running the Ubuntu system and the above dual software platforms).

[0100] Through the implementation of the above embodiments, the system of the present invention can stably achieve adaptive and precise weed control. Field trials show that in corn and soybean fields, the system has an accuracy rate of over 94% in identifying common weeds. In extreme scenarios where the distance between weeds and crops is less than 5 mm, the crop damage rate can be controlled below 0.5%. It saves approximately 15%-30% more energy than traditional preset parameter methods. Furthermore, after a week of field learning, the system shows a significant improvement in the success rate of eliminating specific local weeds.

[0101] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An adaptive laser weeding method, characterized in that, Includes the following steps: (1) The field environment is synchronously collected by a multimodal sensor system set on a mobile platform. The multimodal sensor system includes: a visual sensor for collecting color image information, a radar sensor for acquiring three-dimensional point cloud data, and a supplementary lighting device for ambient light compensation. (2) Based on the multimodal data collected in step (1), a dynamic model of field microenvironment for laser thermal effect prediction is constructed in the industrial control computer. The dynamic model of field microenvironment includes: three-dimensional spatial coordinate information and three-dimensional geometric features of each plant individual in the working area. The three-dimensional geometric features include plant height, canopy projection area and stem diameter, plant category attributes identified based on deep learning model, and current environmental state parameters. The environmental state parameters include at least leaf surface humidity, air temperature and soil humidity. (3) When a target weed to be removed is identified, based solely on the three-dimensional structural information of the target weed in the field microenvironment dynamic model, the spatial distance relationship between the target weed and the surrounding crops, and the environmental state parameters, a preset energy deposition physical model based on the Fourier heat conduction law is invoked. With the optimization objective of "making the temperature of the weed growth point reach the lethal threshold while minimizing the temperature rise of the sensitive tissues of the adjacent crops", the unique optimal combination of laser irradiation parameters for the target weed is dynamically solved. The laser irradiation parameters include laser power, irradiation time, and the offset of the focal position in three-dimensional space. (4) Based on the optimized laser irradiation parameters, control the laser execution module to irradiate the weed target with laser, and at the same time monitor the safety status of the work area based on real-time sensing data, and trigger the safety response mechanism when the preset risk conditions are detected. (5) After the laser irradiation is completed, environmental data is collected again by delay or periodic collection to quantitatively evaluate the weeding effect. The input data of this operation, the laser irradiation parameters and the effect evaluation results are used as training samples to update the physical property parameters such as specific heat capacity and thermal conductivity of plant tissue in the energy deposition physical model online, so as to realize the adaptive calibration of the model for specific farmland environment.

2. The adaptive laser weeding method according to claim 1, characterized in that, In step (1), the data acquisition method of the multimodal sensor system is as follows: the vision sensor continuously acquires RGB image data of the work area at a preset frame rate; the radar sensor synchronously scans within a horizontal range to obtain three-dimensional point cloud data of the work area; wherein, the data frames output by the vision sensor and the radar sensor are assigned the same timestamp and synchronized at the hardware level in the industrial control computer to ensure the sub-pixel level alignment accuracy of subsequent spatial coordinate fusion; the supplementary lighting device is built into the closed-loop control loop of the industrial control computer, the industrial control computer analyzes the brightness and contrast of the image acquired by the vision sensor in real time, and adaptively adjusts the supplementary lighting intensity according to the dynamic changes of ambient light intensity, with the goal of maintaining the minimum image signal-to-noise ratio required for three-dimensional reconstruction; when the ambient light meets the requirements of three-dimensional reconstruction, the supplementary lighting device automatically enters a low-power standby state.

3. The adaptive laser weeding method according to claim 1, characterized in that, In step (2), the method for constructing the dynamic model of the field microenvironment includes: ① The images collected by the vision sensor are input into a pre-trained deep learning model to identify weeds and crops in the work area and obtain their corresponding two-dimensional pixel coordinates and category information; ② By using hand-eye calibration and coordinate transformation, the two-dimensional pixel coordinates are spatially registered and fused with the three-dimensional point cloud data acquired by the radar sensor to determine the three-dimensional spatial position and geometric size parameters of each plant individual in the global coordinate system; ③ Combining environmental sensor data or historical data, environmental state parameters are introduced into the field microenvironment dynamic model. The environmental state parameters include one or more of leaf surface humidity, soil humidity, and ambient temperature.

4. The adaptive laser weeding method according to claim 1, characterized in that, In step (3), the method for dynamically optimizing laser irradiation parameters includes: an energy deposition physics model based on Fourier's law of heat conduction to describe the absorption, thermal conversion and diffusion process of laser energy in plant tissues; constructing a corresponding objective function with the optimization objective of minimizing the thermal impact on surrounding crops while ensuring that the temperature at the weed growth point reaches a preset lethal temperature; and obtaining the optimal combination of laser irradiation parameters by solving the optimization problem corresponding to the objective function. The laser irradiation parameters include laser power, irradiation time and focal position offset.

5. The adaptive laser weeding method according to claim 1, characterized in that, In step (4), the safety response mechanism is a multi-level linkage control, including: Level 1 Response: When the radar sensor detects an unexpected moving object approaching the edge of the work area, the control system triggers an audible and visual alarm to provide a warning. Level 2 response: When the vision sensor detects damage or abnormal shape of the occlusion structure, it controls the laser execution module to enter a safe working mode and reduces the laser output power; Level 3 Response: When the obstruction structure is severely damaged and a living organism enters the core operation area, the control system first calls a pre-trained Bayesian neural network to perform uncertainty quantification assessment on the sensing data and calculates the posterior probability distribution of the current detected target as a "living organism". At the same time, based on multi-frame temporal sensing data, the system predicts the target's trajectory and calculates the probability that it will enter the laser beam propagation path (optical axis region) within the next 200ms time window. A level 3 response is triggered only if the following conditions are met: ① The average probability of the target being identified as a living organism is higher than a preset threshold (e.g., 0.85), and the probability of its predicted trajectory entering the optical axis region is ≥95%; ② The visual detection confidence level of the severely damaged occlusion structure is higher than a preset threshold (e.g., 0.9); When the above conditions are met simultaneously, the control system immediately cuts off the laser power supply and drives the lifting mechanism to raise the laser beam so that it completely deviates from the original focal position. If the uncertainty of the Bayesian neural network evaluation is high (such as the target classification probability is distributed across multiple categories, or the trajectory prediction confidence interval is too wide), the system automatically degrades to a level two response (reduces laser power and continues to observe) instead of directly triggering a level three power outage, in order to avoid accidental shutdown caused by temporary interference from field insects, fallen leaves, dust, etc.

6. The adaptive laser weeding method according to claim 1, characterized in that, In step (5), the quantitative assessment is achieved by comparing the visual characteristics and three-dimensional morphological changes of the target weeds before and after laser irradiation. The assessment indicators include one or more of the following: plant wilting degree, color change, and structural integrity. The model iterative learning adopts an online learning method, using historical operation data as training samples to continuously update the energy deposition physics model in order to correct the plant tissue thermophysical parameters in the model. The thermophysical parameters include specific heat capacity and thermal conductivity.

7. An adaptive laser weeding system implementing the method of any one of claims 1-6, characterized in that, include: The vehicle frame, mobile platform, lifting mechanism, industrial control computer, power module, and sensing module and execution module installed on the lifting mechanism; The sensing module includes: a visual sensor for acquiring color image data, a radar sensor for acquiring three-dimensional environmental information, and a supplementary lighting device for adjusting the lighting conditions of the work area. The execution module includes: a laser for generating a laser beam and a galvanometer system for controlling the laser direction; The lower part of the lifting mechanism is provided with a shielding structure to isolate the laser operation area; The industrial control computer is electrically or communicatively connected to the sensing module, the execution module, the lifting mechanism, and the mobile platform, respectively, and is used to execute the method steps of any one of claims 1-6.

8. The adaptive laser weeding system according to claim 7, characterized in that, The mobile platform is a wheeled mobile platform, which adopts a dual-wheel differential drive mode and is controlled by an STM32 series microcontroller through CAN bus communication.

9. The adaptive laser weeding system according to claim 7, characterized in that, The radar sensor is a 16-line mechanical lidar with a horizontal field of view of 360° and an effective detection range of not less than 150 meters; the high-definition camera is an industrial-grade RGB camera that supports USB 3.0 data transmission and a frame rate of not less than 30fps.

10. The adaptive laser weeding system according to claim 7, characterized in that, The industrial control computer integrates a dual software collaborative platform: the first software platform is developed based on Python and is used to run the YOLOv11n-SE deep learning model, process visual data, build micro-environment models, and optimize laser parameters; the second software platform is developed based on C++ and receives instructions from the first software platform via TCP / IP protocol or RS-485 serial port for laser trigger control and safety interlock logic.

Images