Search

CN-121411266-B - Multi-agent coupling control-based multi-machine spraying uniformity improving method

CN121411266BCN 121411266 BCN121411266 BCN 121411266BCN-121411266-B

Abstract

The invention discloses a multi-machine spraying uniformity improving method based on multi-agent coupling control, which aims to solve the problems that the spraying deposition amount and coverage rate are not observable on line and can not be verified in the operation process, so that the multi-machine uniformity control can not be closed-loop, the invention generates sediment observation and uncertainty through multi-source data acquisition and space-time alignment and cross-modal fusion, introduces physical constraint neural network and graph neural network combined reconstruction with global mass conservation constraint of Lagrangian pair, wind field graph network short-time prediction and risk sensitive distributed model prediction control, order-preserving calibration and closed-loop updating, and realizes the technical effects of on-line observable and verifiable deposition and coverage, multi-machine collaborative closed-loop control of physical conservation consistency and spray uniformity and coverage improvement.

Inventors

  • LIU QIANLAN
  • DAI ZHENHUA
  • TAN DAOJUN

Assignees

  • 湖南科技学院

Dates

Publication Date
20260508
Application Date
20251031

Claims (8)

  1. 1. A multi-machine spraying uniformity improving method based on multi-agent coupling control is characterized by comprising the following steps: S1, acquiring and synchronizing original observation data and outputting the data; s2, inputting original observation data, and performing cross-modal fusion to obtain a deposition observation set, an observation uncertainty and a working state characteristic set; s3, inputting a sediment observation set, an observation uncertainty and an operation state feature set, executing a physical constraint neural network and graph neural network combined model to reconstruct a sediment field, introducing global mass conservation constraint with Lagrange dual, and outputting a sediment field estimation result, a conservation error index, a sediment gap field and the observation uncertainty; taking a deposit observation set as a graph node observation input, and constructing a multi-agent operation topological graph by using an operation state characteristic set; Explicitly introducing mass conservation constraint of ejection quantity, deposition quantity, evaporation quantity, drift quantity and boundary flux into a loss function of a physical constraint neural network and graph neural network joint model, establishing dual optimization through Lagrangian multipliers to ensure that the ejection quantity is equal to the sum of the deposition quantity, the evaporation quantity, the drift quantity and the boundary flux, and simultaneously taking a convection diffusion equation and an evaporation item as physical residual errors to participate in optimization, and solving to obtain a deposition field estimation result covering an operation area; Calculating conservation error indexes of the global and the subareas based on the deposition field estimation result, and obtaining a deposition notch field according to the difference between the uniform deposition reference of the target and the deposition field estimation result; S4, inputting a deposition field estimation result, a deposition notch field, a conservation error index, an observation uncertainty and an operation state characteristic set, and carrying out wind field map network prediction to form a control preparation set; the wind field graph network specifically comprises: Short-time rolling prediction is performed based on a graph structure constructed by the operation nodes and the spatial adjacent relation, and the mapping relation is written as follows: ; Wherein the method comprises the steps of Is indicated in the position Is the first of (2) A step wind field predictive vector, Network model representing wind field graph and parameters A graph constructed by nodes and adjacent relations in the operation area, Representing short-term prediction step index, Represents the upper limit of the predicted steps, Inputting a feature vector for an operation node corresponding to a position x and a time t under a unified operation coordinate system, wherein the feature vector comprises a feature corresponding to the position x and the time t in an operation state feature set, a deposition estimated value corresponding to the position x and the time t in a deposition field estimated result, a notch value corresponding to the position x and the time t in a deposition notch field, a conservation error index and an observation uncertainty; For the next step of control, wind field prediction and current deposition estimation and deposition gap, global conservation error and observation uncertainty of each future step are arranged into a control preparation set according to coordinates and time stamps: ; Wherein the method comprises the steps of Represents a control preparation set, Representing a uniform time step, A global conservation error index representing statistics within the rolling window, Indicating that the deposition observation is in position And time of day Uncertainty score of (2), Representing a set of discrete sampling nodes under a unified operating coordinate system, the set being organized in a dictionary of coordinates and time stamps to ensure seamless interfacing with a control solution interface, Representing the deposition field estimation results at the location Time of day Is used to estimate the deposition estimate of (a), Indicating the position of the notch field Time of day Is a notch value of (2); S5, inputting a control preparation set, executing risk sensitive distributed model prediction control, and outputting a multi-agent control instruction and a control prediction sequence, wherein the risk sensitive distributed model prediction control refers to mapping an observation uncertainty into space-time risk weights in the distributed model prediction control, and weighting objective function cost items corresponding to a deposition gap field by the risk weights, so that the risk weights are higher when the uncertainty is larger, and a more conservative and robust multi-agent cooperative control instruction is obtained; s6, inputting a multi-agent control instruction and a control prediction sequence, executing according to the multi-agent control instruction and forming job execution observation data and verification pairing data; S7, inputting verification pairing data, performing order-preserving calibration and verification, generating a trigger signal, and outputting a uniformity confidence interval and the trigger signal, wherein the order-preserving calibration refers to that on the basis of the verification pairing data, a monotonic mapping relation is learned, and the uniformity index predicted value is calibrated on the premise of keeping the relative size ordering of the uniformity index predicted value unchanged, so that the calibrated prediction error statistical boundary is consistent with a preset confidence level, and the uniformity confidence interval is constructed and standard-reaching judgment is performed; and S8, performing closed-loop updating according to the trigger signal, and outputting updated original observation data.
  2. 2. The multi-machine spraying uniformity improving method based on multi-agent coupling control according to claim 1, wherein S1 specifically comprises: each intelligent agent respectively collects the flow rate, the pressure of the spray head, the operation speed, the gesture, the height of the relative canopy, the wind speed and the wind direction, the relative positioning data and the land block boundary and the spraying forbidden region information; Performing time synchronization on the acquired data, establishing a unified clock reference and aligning a time stamp, and forming a time sequence of a unified time step by interpolation or resampling on the data with different sampling frequencies; Spatially registering the acquired data, converting the relative positioning data and the gesture into coordinates and gestures under a unified operation coordinate system, and representing land block boundaries and spraying forbidden region information in the unified operation coordinate system; And outputting original observation data, wherein the original observation data are time-ordered and space-marked data sets with coordinates and time stamps.
  3. 3. The multi-machine spraying uniformity improving method based on multi-agent coupling control according to claim 1, wherein S2 is specifically: Inputting the original observation data into a cross-modal fusion model; preprocessing the original observation data, including time stamp alignment, spatial registration correction, noise suppression, scale calibration, missing value filling and outlier processing, to form a preprocessed data set with coordinates and time stamps; Generating a spray deposition observation value at corresponding coordinates and moments by taking the preprocessed data set as input, and organizing the spray deposition observation value as a deposition observation set according to the coordinates and time stamps; outputting uncertainty scores of all spray deposition observation values simultaneously by a cross-modal fusion model, and organizing the uncertainty scores into observation uncertainty according to coordinates and time stamps; extracting and arranging operation speed, attitude, wind speed and wind direction, relative canopy height and relative positioning data based on the preprocessed data set in the fusion process, and organizing the operation speed, attitude, wind speed and wind direction, relative canopy height and relative positioning data into an operation state feature set according to coordinates and time stamps; And outputting a deposit observation set, an observation uncertainty and a job state feature set.
  4. 4. The multi-machine spraying uniformity improving method based on multi-agent coupling control according to claim 1, wherein S4 is specifically: Extracting a deposition field estimation result and a deposition gap field as pneumatic source features according to coordinates and time stamps, and performing space alignment and time alignment on wind speed and wind direction, gesture and operation speed and relative positioning data in an operation state feature set under a unified operation coordinate system to form an input feature set of a wind field graph network; Short-time prediction is carried out in a wind field graph network by taking the input feature set as input, a short-time wind field prediction result in a future rolling window is obtained, and the short-time wind field prediction result is organized according to time step length and coordinates; and arranging the short-time wind field prediction result, the deposition field estimation result, the deposition gap field, the conservation error index and the observation uncertainty into a control preparation set according to coordinates and a time stamp.
  5. 5. The multi-machine spraying uniformity improving method based on multi-agent coupling control according to claim 1, wherein S5 specifically comprises: Inputting a control preparation set into a distributed model predictive control problem; In the distributed model predictive control problem, a short-time wind field predictive result is organized into a disturbance sequence according to time step length, a coupling control target is formed by a deposition field estimation result and a deposition gap field, a conservation error index is used as a hard constraint or weighting penalty term to be added into a constraint set, and risk weight is set according to the uncertainty of observation; Taking a multi-agent operation topological graph as an adjacency relation, establishing a local prediction model and a local optimization problem by each agent, wherein decision variables of the local prediction model are time sequence tracks of operation speed, course, transverse distance and spraying quantity in a prediction window, the local prediction model is updated in a state by using time sequence evolution of operation state characteristics under a unified operation coordinate system, and the disturbance sequences are superposed; The local optimization problem aims at reducing a sediment gap field and considering control stability and working efficiency, wherein the constraints comprise working speed, course change rate, upper and lower limit constraints of transverse spacing and spraying quantity, block boundary and spraying forbidden region constraints, lower limit constraints of relative distance between adjacent working machines and conservation consistency constraints that conservation error indexes do not exceed a preset threshold or are counted in a weighted penalty term mode; And solving under the adjacent coupling term to obtain multi-agent control instructions of the operation speed, the heading, the transverse distance and the spraying quantity of each agent, and forming a control prediction sequence according to the time step length and the predicted value of the deposition and uniformity index in the coordinate organization prediction window.
  6. 6. The multi-machine spraying uniformity improving method based on multi-agent coupling control according to claim 1, wherein S6 is specifically: Issuing a multi-agent control instruction to each agent, executing under a unified operation coordinate system and a unified clock reference, and controlling the operation speed, the course, the transverse interval and the spraying amount according to the time step of the multi-agent control instruction; collecting newly added spray deposition observation values in the execution process, recording corresponding coordinates and time stamps, and organizing the operation execution observation data according to the coordinates and the time stamps; And aligning the operation execution observation data with the control prediction sequence under a unified operation coordinate system and a unified clock reference, and performing one-to-one pairing in a mode that coordinates are identical with time stamps or are regarded as identical within a preset space error limit and a time error limit to form verification pairing data, wherein the verification pairing data comprises each spray deposition observation value, a corresponding predicted deposition value and a uniformity index predicted value of a corresponding space-time position.
  7. 7. The multi-machine spraying uniformity improving method based on multi-agent coupling control according to claim 1, wherein S7 specifically comprises: inputting verification pairing data into a sequence-preserving calibration and verification process; Calculating a calibration score based on the verification pairing data under a unified operation coordinate system and a unified clock reference, wherein the calibration score comprises errors between a uniformity index predicted value of each pairing and a uniformity index actual measurement value calculated by a spray deposition observed value according to a preset uniformity evaluation method, and respectively mapping the predicted deposition value and the spray deposition observed value into a coverage state according to a preset coverage rate judging threshold value and summarizing the errors between the predicted value of the coverage rate and the actual measurement value of the coverage rate; constructing a uniformity confidence interval and a coverage rate confidence interval by using the statistics boundary of the calibration score under a preset confidence level, organizing the uniformity confidence interval according to coordinates and a time stamp, and outputting the uniformity confidence interval; and judging the uniformity confidence interval according to a preset threshold value or judging according to a conservation error index, and generating a trigger signal when the condition of not reaching the standard is met.
  8. 8. The multi-machine spraying uniformity improving method based on multi-agent coupling control according to claim 1, wherein S8 is specifically: Inputting the uniformity confidence interval and the trigger signal into a closed-loop updating process; When the trigger signal indicates that the uniformity confidence interval does not reach the standard, selecting and executing active sampling or model recalibration or risk weight adjustment according to the out-of-range position and the out-of-range amplitude of the uniformity confidence interval; the active sampling is to collect spray deposition observation values according to the out-of-range positions and the out-of-range time periods under a unified operation coordinate system and a unified clock reference, and record coordinates and time stamps to form newly added sampling data; Model recalibration is to recalibrate the parameters of the cross-modal fusion model and the combined model of the physical constraint neural network and the graph neural network and generate a recalibrated data update item; The risk weight is adjusted to update the risk weight parameter in the distributed model prediction control according to the width of the uniformity confidence interval, and the updating is used for the next rolling window in the step S5 without changing the structure of the original observation data; The newly added sampling data and the recalibrated data updating item are combined into original observation data according to coordinates and time stamps under a unified operation coordinate system and a unified clock reference, so that updated original observation data is formed; when the trigger signal indicates that the standard is reached, the original observed data is kept unchanged and the updated original observed data is directly output; the updated raw observation data is output as an input to step S1 to complete the closed loop iteration.

Description

Multi-agent coupling control-based multi-machine spraying uniformity improving method Technical Field The invention relates to the field of unmanned aerial vehicle spraying, in particular to a multi-machine spraying uniformity improving method based on multi-agent coupling control. Background The agricultural plant protection spraying operation has widely adopted a multi-machine cooperation mode, including formation of ground medicine machines and unmanned aerial vehicles and the like. In order to improve coverage rate and uniformity, the prior art has a certain progress in the aspects of path planning, formation control, constant-speed constant-distance spraying, sensing-based self-adaptive parameter adjustment and the like, provides wind speed and wind direction by a commonly used weather station and an onboard anemometer in the aspect of environment modeling, combines a convection diffusion model or an empirical drift model for simulation evaluation, and relies on water-sensitive test paper, tracer sampling and offline laboratory analysis in the aspect of effect evaluation, or adopts sensing means such as images, leaf surface humidity and the like for indirect estimation. In recent years, prediction attempts of environment and deposition by using deep learning and graph structure methods have been made, but the methods have been focused on offline inference or stand-alone scenes. However, the following deficiencies are prevalent in the prior art: 1. The on-line observable and verifiable capability is insufficient, the spray deposition and coverage rate are difficult to obtain and calibrate in real time in the operation process, the off-line sampling and the post evaluation are mainly relied on, the feedback period is long, the space representativeness is limited, and a control closed loop is difficult to form; 2. The method has the defects of strong wind field space distribution and time variability, the existing method mostly adopts a static or simplified model, lacks short-time wind field prediction facing an operation area and coupling propagation among multiple intelligent agents, and the cooperation among multiple machines is always remained on the path and formation level, and lacks coupling optimization and constraint directly facing deposition uniformity; 3. Physical consistency and uncertainty are not utilized, namely mass conservation constraint is not explicitly applied to sediment field estimation generally, mechanisms such as evaporation and drift are not characterized, prediction uncertainty is not utilized in quantification and risk sensitive control, calibration and confidence assessment are also not utilized, and overspray or missing spray is difficult to correct in time. The defects restrict continuous improvement of the multi-machine spraying uniformity and closed-loop guarantee. Therefore, a multi-machine spraying method capable of solving the above-mentioned deficiencies of the prior art is a problem that needs to be solved by those skilled in the art. Disclosure of Invention The invention aims to provide a multi-machine spraying uniformity improving method based on multi-agent coupling control, which aims at solving the problems that in the prior art, the spraying deposition amount and coverage rate are not observable and verifiable on line in the operation process, so that the multi-machine uniformity closed-loop control cannot be realized, the invention provides a technical scheme of acquisition and alignment of multi-source data under unified coordinates and clocks, cross-modal fusion to generate deposition observation and uncertainty, physical constraint neural network and graph neural network combined deposition field reconstruction of global mass conservation constraint with Lagrange dual, wind field network short-time prediction combining deposition gaps and conservation errors, risk sensitive distributed model prediction control taking conservation errors as hard constraint or high weight penalty terms, and on-line verification and triggered closed-loop update based on order preservation calibration. According to the embodiment of the invention, the multi-machine spraying uniformity improving method based on multi-agent coupling control is characterized by comprising the following steps of: S1, acquiring and synchronizing original observation data and outputting the data; s2, inputting original observation data, and performing cross-modal fusion to obtain a deposition observation set, an observation uncertainty and a working state characteristic set; s3, inputting a sediment observation set, an observation uncertainty and an operation state feature set, executing a physical constraint neural network and graph neural network combined model to reconstruct a sediment field, introducing global mass conservation constraint with Lagrange dual, and outputting a sediment field estimation result, a conservation error index, a sediment gap field and the observation uncertainty; S4, in