A method for coordinated operation of multiple sprinkler irrigation machines

Abstract

This invention discloses a method for collaborative operation of multiple sprinkler irrigation machines, including: Step 1, multimodal data acquisition; Step 2, converting the raw multimodal data into a comprehensive feature vector; Step 3, designing soil moisture consistency assessment, terrain adaptability assessment, and crop growth anomaly detection; Step 4, data storage and analysis of operational data; establishing dynamic risk assessment indicators for multi-machine collaborative operation; Step 5, reinforcement learning-based relocation path planning and collaborative optimization. The relocation path planning comprehensively considers path length, soil compaction risk, and number of turns, and uses a near-end strategy optimization algorithm to optimize the path planning online and adjust the path in real time. This invention achieves efficient, safe, and adaptive collaborative operation of multiple sprinkler irrigation machines by fusing multi-source heterogeneous sensing data, constructing a dynamic weighted fusion model, and designing a collaborative operation task scheduling framework.

Description

Technical Field

[0001] This invention belongs to the field of agricultural water-saving irrigation technology, specifically, it relates to a method for coordinated operation of multiple sprinkler irrigation machines. Background Technology

[0002] With the development of large-scale agricultural operations, large sprinkler irrigation machines are widely used in irrigation operations in plain areas. However, the following technical bottlenecks still exist when multiple sprinkler irrigation machines work together: Single sensing modality is insufficient to fully describe the working environment: existing methods mostly rely on a single sensor (such as RTK-GNSS) to obtain location information, lacking the fusion perception of multi-dimensional environmental information such as soil moisture, crop growth, and terrain undulation, resulting in a disconnect between operational decisions and actual conditions.

[0003] Static weight formation control has poor adaptability: The traditional lead-follow mode uses fixed weights to maintain formation, which cannot adapt to dynamic changes such as road width and soil bearing capacity, and is prone to mechanical interference or low traffic efficiency.

[0004] Separation of route planning and job scheduling: Existing solutions handle route planning and job scheduling separately, lacking a unified collaborative optimization framework, resulting in poor coordination between relocation routes and irrigation tasks, and increasing non-operational time.

[0005] Lack of a closed-loop evaluation and adjustment mechanism for multi-machine collaboration: During operation, there is a lack of real-time evaluation and feedback adjustment of the status and environmental changes of each machine, making it impossible to form a closed-loop optimization of perception-decision-control-evaluation. Summary of the Invention

[0006] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a method for coordinated operation of multiple sprinkler irrigation machines.

[0007] To achieve the aforementioned objectives, the technical solution adopted by this invention includes: a method for coordinated operation of multiple sprinkler irrigation machines, comprising sprinkler irrigation machines, specifically including: Step 1: Environmental Sensing and Data Acquisition 1.1 Each sprinkler irrigation machine is equipped with a GNSS dual antenna, an inertial measurement unit, a soil moisture sensor, a multispectral camera, a lidar, and an infrared thermal imager. The GNSS dual antennas are used to acquire the geometric center coordinates and heading angle of the sprinkler; the inertial measurement unit is used for high-frequency attitude compensation; the soil moisture sensor is used to acquire soil moisture content; the multispectral camera is used to calculate the normalized vegetation index (NDVI) to assess crop growth; the lidar is used to construct a three-dimensional point cloud of the road ahead to identify road width, obstacle height, and slope; and the infrared thermal imager is used to identify areas of abnormal soil temperature (to assist in judging soil bearing capacity). Among them, the soil moisture sensor, inertial measurement unit, and GNSS dual antenna reflect the time-series changes in the operating status of the sprinkler irrigation machine and the soil moisture, constituting the time mode; the lidar and infrared thermal imager reflect the terrain undulations, width, and obstacle distribution of the road ahead, constituting the terrain mode; the multispectral camera reflects crop growth, pest and disease conditions, and irrigation uniformity, constituting the crop mode. These data change continuously over time, reflecting the current operating status of the sprinkler irrigation machine and soil conditions, and are the direct basis for judging whether irrigation parameters or operating speed need to be adjusted.

[0008] 1.2 Spatiotemporal alignment of time mode, terrain mode, and crop mode: Time alignment: Based on the inertial measurement unit timestamp, linear interpolation is used for low-frequency data, and sliding window mean aggregation is used for high-frequency data. Let the target time be t, and the sampling time of a certain mode be... and The interpolation formula is: Spatial alignment: All sensing data are uniformly projected onto a local coordinate system with the geometric center of the sprinkler as the origin, the X-axis representing the longitudinal direction of the machine, and the Y-axis representing the lateral direction. After coordinate transformation, the lidar point cloud data generates a terrain elevation map within a 100 m range ahead, with a grid resolution of 0.2 m.

[0009] 1.3. Before input time alignment, real-time quality metrics need to be calculated for multimodal data: Signal-to-noise ratio (SNR): Calculated based on the original sensor signal; an SNR ≥ 20 dB indicates high quality. Packet loss rate (lossRate): Statistics on wireless communication links; a lossRate ≤ 5% is considered normal. Reconstruction error reconErr: Historical data is reconstructed using an autoencoder. If the error exceeds a threshold, it is considered abnormal.

[0010] Step 2: Feature Extraction and Dynamic Weight Fusion 2.1. Transform the raw sensor data (high-dimensional, redundant, and noisy) into compact, semantically rich feature vectors that facilitate subsequent fusion. Design encoders for different multimodal characteristics: A one-dimensional convolutional neural network (1D-CNN) is used, with the input being a sliding time window (window length). Temporal sequences within a stride of 1 second. Network structure: 2 convolutional layers (kernel size 3, number of channels 32-64), batch normalization, ReLU activation, global average pooling, outputting a 128-dimensional temporal feature vector. ; A two-dimensional convolutional neural network (2D-CNN) is used, with a 100 m × 20 m terrain elevation map (resolution 0.2 m, i.e., 500 × 100 pixels) as input. The network structure consists of 3 convolutional layers (kernel size 5, number of channels 16→32→64), max pooling (2×2), and after flattening, a fully connected layer outputs a 128-dimensional terrain feature vector. ; A lightweight Vision Transformer (ViT) is used to divide the image acquired by the multispectral camera into 16×16 pixel patches. After linear projection and position encoding, the patches are input into the encoder of the Transformer (4 layers, 8 attention heads), and the output is a crop feature vector with a dimension of 128. .

[0011] 2.2 Design a three-level integration mechanism: Calculate the original attention score for mode m (m∈{time,terrain,crop}) of sprinkler machine i at time t. The loop variable iterates through all modalities. The values ​​are time, terrain, and crop, respectively. , in , is the weight matrix for the multimodal modes. For trainable vectors shared by all modalities, yes transpose, It is the hyperbolic tangent activation function.

[0012] It takes the dot product of a row vector and a column vector, and outputs a scalar. That is, the unnormalization of the mode at this moment.

[0013] Level 1: Attention weight normalization: ; Calculate the credibility of each modality based on real-time data quality indicators: , Among them, signal-to-noise ratio (SNR), packet loss rate (lossRate), and reconstruction error (reconErr) are... , , As a configurable coefficient, the sigmoid function compresses the output to [0,1]. When At that time, the spatiotemporal alignment data of this modality was marked as untrustworthy and temporarily excluded during fusion.

[0014] Level 3: Weighting and Feature Fusion Multiplying the attention weight by the credibility yields the final fusion weight: The weighted summation yields the comprehensive eigenvector: This comprehensive feature vector The dimension is 128, integrating three modal information: environmental temporal changes, terrain structure, and crop growth.

[0015] Step two differs from traditional methods that rely solely on single location information. This invention integrates multi-source data such as soil, terrain, and crops, and achieves adaptive feature fusion through an attention-reliability dual weighting mechanism, thereby improving the perception robustness in complex farmland environments.

[0016] Step 3: Operational Status Assessment Based on Two-Way Temporal Modeling 3.1. The comprehensive feature vector obtained in step two Inputting time series data into a BiGRU neural network captures forward and backward temporal dependencies. Time window length. (Covering a travel distance of approximately 50 m for the sprinkler), step length 5 s.

[0017] BiGRU neural network structure: 128 hidden layers, 2 layers, forward GRU for forward timing, and backward GRU for backward timing. The output at time step t is: in, In the forward-hidden state, This represents the backward hidden state, where ∥ indicates vector concatenation, and the output dimension is 256.

[0018] 3.2 Based on the output of the BiGRU neural network, three parallel evaluation branches are designed: 3.21 Soil Moisture Consistency Assessment: Determine whether the soil moisture content in the current operating area of ​​the sprinkler irrigation machine matches the preset irrigation target, and output continuous values. ∈[0,1], when it is below 0.6, a supplementary irrigation suggestion is triggered.

[0019] 3.22 Terrain Adaptability Assessment: Determine whether the terrain within 50 m ahead is suitable for the current sprinkler to pass through, and output continuous values. ∈[0,1], when it is below 0.5, formation adjustment or path replanning is triggered.

[0020] 3.23. Crop Growth Anomaly Detection: Determines whether there are abnormal fluctuations in the Normalized Difference Vegetation Index (NDVI) (which may indicate pests, diseases, or uneven irrigation). Outputs a binary value. ∈{0,1}, an abnormal alarm is triggered when the value is 1.

[0021] The three branches share the temporal features extracted by the BiGRU neural network and are jointly optimized using a multi-task learning framework. The total loss function is: in =0.4, =0.3, =0.3, and Mean square error is used. Binary cross-entropy is used.

[0022] Step three combines environmental perception data with time-series modeling to achieve a comprehensive quantitative assessment of soil moisture, terrain adaptability, and crop growth, providing a quantitative basis for subsequent operational decisions.

[0023] Step 4: Multi-machine collaborative task scheduling and path planning 4.1 Construct a three-tier collaborative architecture: End (sprinkler): Responsible for multimodal spatiotemporal aligned data acquisition, local feature extraction, and execution of control commands.

[0024] Edge (Edge Gateway): Deployed in the field management room (one edge gateway is shared by every 2-3 sprinkler irrigation machines), responsible for multi-machine data aggregation, collaborative operation scheduling, and formation reconfiguration decisions. The edge gateway is equipped with an NVIDIA Jetson AGX Orin platform (32 TOPS computing power) and runs a lightweight inference engine.

[0025] Cloud (Control Center): Deployed on a remote server, responsible for global path planning, model training and updating, and job data storage and analysis.

[0026] The edge gateway communicates with the cloud via a 5G network (latency ≤ 20 ms), and communicates with the sprinkler via a 5.8 GHz industrial wireless network (latency ≤ 10 ms).

[0027] 4.2 Constructing a dynamic risk assessment index system for multi-machine collaborative operations: Environmental risks: Average soil bearing capacity, average slope, and obstacle density within 50 m ahead; weighting range: 30±5%. Equipment risks: Remaining water volume of the sprinkler, deviation between current speed and preset speed, and communication signal strength; weighting range: 25±3%. Operational risks: Remaining work area, deviation between current work progress and plan, weighting range: 25±3%; Formation risk: deviation between the spacing of adjacent sprinklers and the theoretical spacing, degree of formation distortion, weight range: 20±2%; Each weight is dynamically adjusted based on real-time threats, using the following formula: + · Where d represents the dimension, d traverses four dimensions: environmental risk, equipment risk, operational risk, and formation risk; Wd is the importance coefficient of the d-th dimension in the comprehensive risk score; α=0.85 is the smoothing factor. The gradient of risk with respect to dimensional features is calculated using an online backpropagation neural network. It is the quantized feature value of the d-th dimension, with its value range normalized to [0,1].

[0028] Based on comprehensive risk score Matching a three-level response mechanism: Low risk (R < 0.3): Maintain current formation and log. For medium risk (0.3 ≤ R < 0.6), reduce speed to 0.3 m / s and increase safety distance to 1.5 times. High risk R ≥ 0.6 triggers formation reconfiguration (column formation or single vehicle passing through), and reports to the cloud; 4.3 The edge gateway adopts a priority queue-based task scheduling algorithm: Task priority is calculated by taking into account the remaining operating time, remaining water volume, and distance from the target point.

[0029] When multiple irrigation machines cross paths, the higher priority irrigation machine is dispatched first, while the lower priority irrigation machine waits in the waiting area (the waiting area is ≥30 m away from the intersection point). When a deadlock (multi-machine loop waiting) is detected, the lowest priority sprinkler is forced to retreat 10 m to break the loop.

[0030] Step four combines a four-dimensional risk quantification assessment of environment, equipment, operation, and formation with adaptive formation adjustment to achieve intelligent scheduling of multi-machine collaborative operations.

[0031] Step 5: Reinforcement Learning-Based Transition Path Planning and Collaborative Optimization 5.1 The relocation route planning comprehensively considers three objectives: route length, soil compaction risk, and number of turns. The cost function is as follows: Where L: total path length (m), weight wL=0.4. S: soil compaction cost, T: number of turns, each turn is calculated based on an equivalent path length of 20 m, weight wT=0.25.

[0032] 5.2 The Proximal Policy Optimization (PPO) algorithm is used for online optimization of path planning. The state space includes: current pose, terrain elevation map 100 m ahead, soil bearing capacity distribution, and target point location. The action space includes: speed adjustment (0.2~0.8 m / s), steering angle adjustment (-0.1~0.1 rad / s), and formation adjustment commands. Reward function design: in This is the path length increment. To increase soil compaction volume, For the increment of the number of turns, This is the collision indication function (-100 upon collision). =0.1, =0.2, =0.1, =100.

[0033] The transition path planning based on deep reinforcement learning is trained offline in the Proximal Policy Optimization (PPO) algorithm (based on historical job data). After training, it is distributed to the edge gateway through federated learning. The edge gateway performs lightweight inference based on the real-time environment to achieve real-time adjustment of the path.

[0034] When a cluster needs to be moved from the current work area to the next work area, the edge gateway executes the following process: 5.31. Obtain the current location, remaining water volume, and operation completion status of all sprinkler irrigation machines; 5.32. Call the relocation path planning function to generate a relocation path for each sprinkler machine; 5.33. Detect path conflicts; if intersections exist, sort them by priority and adjust the time window accordingly. 5.34. Issue the transition command, and the formation will automatically switch to column formation; 5.35 During the relocation process, each sprinkler machine reports its location in real time, and the edge gateway adjusts dynamically.

[0035] Step five incorporates path length, soil compaction risk, and number of turns into a unified cost function, and uses the PPO algorithm to achieve online path optimization, balancing operational efficiency and soil protection.

[0036] Step Six: Closed-Loop Evaluation and Continuous Evolution Mechanism 6.1 After each work block (100 m × 100 m) is completed, the system performs a work quality assessment: Irrigation uniformity: After the operation, soil moisture content was sampled (9 points evenly distributed), and the coefficient of variation CV = standard deviation / mean was calculated. CV ≤ 0.15 is excellent, 0.15~0.25 is acceptable, and >0.25 is unacceptable.

[0037] Work efficiency: The ratio of actual work time to theoretical work time; a ratio ≥0.9 indicates high efficiency.

[0038] Energy efficiency: Electricity consumption per unit area (kWh / acre), compared with historical average.

[0039] The evaluation results are stored on the blockchain (each sprinkler machine acts as a blockchain node) to ensure that the data is immutable and used for subsequent model training and update training and parameter correction.

[0040] 6.2 Constructing a closed-loop optimization system: Edge nodes: Upload anonymized job logs (including perception data, decision records, and evaluation results) weekly. Using a federated learning framework, only model training and gradient updates are uploaded, without disclosing the original data.

[0041] Cloud Center: Monthly aggregation of model training and updates from each edge node, retraining of BiGRU neural network and near-end policy optimization algorithm, and distribution of the optimized BiGRU neural network and near-end policy optimization algorithm model to edge gateway after compression.

[0042] Blockchain-based evidence storage: All model versions, update records, and job evaluation results are stored on the blockchain, forming a traceable knowledge base.

[0043] 6.3 When the system detects the following anomalies, it will automatically switch to degrade mode: Communication interruption: The sprinkler continues to execute according to the last received instruction for 60 seconds. If communication is still not restored, it will automatically decelerate and stop, and activate the audible and visual alarms. It will also use lidar for autonomous obstacle avoidance.

[0044] Positioning failure: Switch to inertial measurement unit + wheel speed meter calculation mode. Positioning error accumulates at 2 m / min. If it is still not recovered after driving 50 m, stop alarm.

[0045] Sensor failure: The faulty mode is automatically masked, and the dynamic weight fusion mechanism automatically reduces the weight of the mode, allowing the system to continue operating in degraded mode.

[0046] Collision risk: If the lidar detects an obstacle entering the restricted area (radius 10 m), the vehicle will brake to a stop and resume driving after the obstacle is removed.

[0047] Step six achieves reliable assessment of job quality and continuous optimization of model parameters with privacy protection, forming a complete closed loop of perception-decision-execution-evaluation-evolution.

[0048] Compared with the prior art, the advantages of the present invention include: (1) The present invention provides a method for collaborative operation of multiple sprinkler irrigation machines, which improves the accuracy of comprehensive environmental perception by integrating multimodal data of soil, terrain and crops, and greatly enhances the adaptability to complex farmland environments; (2) The present invention provides a multi-sprinkler cooperative operation method. The dynamic weight fusion mechanism reduces the response time of formation adjustment and improves the efficiency of narrow road passage. It takes into account path length, soil compaction and number of turns, and balances efficiency and sustainability. (3) The present invention provides a multi-sprinkler cooperative operation method, which optimizes the path to reduce the soil compaction area by 35% and shortens the average length of the transfer path. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a flowchart of steps one to three of a multi-sprinkler cooperative operation method according to the present invention; Figure 2 The flowchart shows steps four to six of a multi-sprinkler cooperative operation method according to the present invention. Detailed Implementation

[0051] In view of the shortcomings of the prior art, the inventors of this invention, through long-term research and extensive practice, have proposed the technical solution of this invention. The technical solution, its implementation process, and principles will be further explained below with reference to the accompanying drawings and specific implementation examples in the embodiments of this application.

[0052] It should be noted that the embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. The described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, the present invention covers any substitutions, modifications, equivalent methods and solutions made on the spirit, principles and scope of the present invention as defined by the claims. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] In the description of this application, the terms "first," "second," "third," and similar terms do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms "one" or "a" do not indicate a quantity limitation, but rather indicate the presence of at least one. Terms such as "include" or "contain" mean that the elements or objects preceding "include" cover the elements or objects listed following "include" or their equivalents, and do not exclude other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect.

[0054] In the description of this application, the terms center, up, down, front, back, left, right, vertical, horizontal, top, bottom, inside, outside, etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application. In addition, when using positional terms such as sides, outer side, top and bottom, etc., it should be understood that they are used only for the convenience of understanding and description, taking into account that the structure may be oriented to other positions.

[0055] In the description of this application, unless otherwise expressly specified and limited, the technical or scientific terms used shall have the ordinary meaning understood by a person with ordinary skills in the art to which this application pertains. Terms such as installation, connection, and linking shall be interpreted broadly, for example, they may refer to fixed connection, detachable connection, mating connection, or integral connection. For a person skilled in the art, the specific meaning of the above terms in this application may be understood according to the specific circumstances.

[0056] The present invention aims to introduce and explain the structural composition of a multi-sprinkler irrigation machine collaborative operation method and the cooperation relationship between the various components. Unless otherwise specified, the dimensions, materials and manufacturing processes of the various components suitable for the multi-sprinkler irrigation machine collaborative operation method in the present invention can be selected according to specific circumstances, and no special limitations or explanations are made here.

[0057] Furthermore, to provide the public with a better understanding of the present invention, certain specific details are described in detail in the following description of the invention. However, those skilled in the art will fully understand the invention even without these detailed descriptions.

[0058] Example Please see Figures 1-2 A method for coordinated operation of multiple sprinkler irrigation machines, including sprinkler irrigation machines, specifically including: Step 1: Environmental Sensing and Data Acquisition 1.1 Each sprinkler irrigation machine is equipped with a GNSS dual antenna, an inertial measurement unit, a soil moisture sensor, a multispectral camera, a lidar, and an infrared thermal imager. The GNSS dual antennas are used to acquire the geometric center coordinates and heading angle of the sprinkler; the inertial measurement unit is used for high-frequency attitude compensation; the soil moisture sensor is used to acquire soil moisture content; the multispectral camera is used to calculate the normalized vegetation index (NDVI) to assess crop growth; the lidar is used to construct a three-dimensional point cloud of the road ahead to identify road width, obstacle height, and slope; and the infrared thermal imager is used to identify areas of abnormal soil temperature (to assist in judging soil bearing capacity). Among them, the soil moisture sensor, inertial measurement unit, and GNSS dual antenna reflect the time-series changes in the operating status of the sprinkler irrigation machine and the soil moisture, constituting the time mode; the lidar and infrared thermal imager reflect the terrain undulations, width, and obstacle distribution of the road ahead, constituting the terrain mode; the multispectral camera reflects crop growth, pest and disease conditions, and irrigation uniformity, constituting the crop mode. These data change continuously over time, reflecting the current operating status of the sprinkler irrigation machine and soil conditions, and are the direct basis for judging whether irrigation parameters or operating speed need to be adjusted.

[0059] 1.2 Spatiotemporal alignment of time mode, terrain mode, and crop mode: Time alignment: Based on the inertial measurement unit timestamp, linear interpolation is used for low-frequency data, and sliding window mean aggregation is used for high-frequency data. Let the target time be t, and the sampling time of a certain mode be... and The interpolation formula is: Spatial alignment: All sensing data are uniformly projected onto a local coordinate system with the geometric center of the sprinkler as the origin, the X-axis representing the longitudinal direction of the machine, and the Y-axis representing the lateral direction. After coordinate transformation, the lidar point cloud data generates a terrain elevation map within a 100 m range ahead, with a grid resolution of 0.2 m.

[0060] Spatiotemporal alignment resolves the sampling frequency differences among multiple sensors (10 Hz, 100 Hz, 1 Hz, 20 Hz) and unifies the time grid; 1.3. Before input time alignment, real-time quality metrics need to be calculated for multimodal data: Signal-to-noise ratio (SNR): Calculated based on the original sensor signal; an SNR ≥ 20 dB indicates high quality. Packet loss rate (lossRate): Statistics on wireless communication links; a lossRate ≤ 5% is considered normal. Reconstruction error reconErr: Historical data is reconstructed using an autoencoder. If the error exceeds a threshold, it is considered abnormal.

[0061] Step 1.3 Data quality assessment: Calculate signal-to-noise ratio, packet loss rate, and reconstruction error to identify sensor faults or communication anomalies.

[0062] Step 2: Feature extraction and dynamic weight fusion, which transforms multi-source heterogeneous data into a unified feature space and dynamically determines the importance of each modality based on the real-time scenario.

[0063] 2.1. Transform the raw sensor data (high-dimensional, redundant, and noisy) into compact, semantically rich feature vectors that facilitate subsequent fusion. Design encoders for different multimodal characteristics: A one-dimensional convolutional neural network (1D-CNN) is used, with the input being a sliding time window (window length). Temporal sequences within a stride of 1 second. Network structure: 2 convolutional layers (kernel size 3, number of channels 32-64), batch normalization, ReLU activation, global average pooling, outputting a 128-dimensional temporal feature vector. ; A two-dimensional convolutional neural network (2D-CNN) is used, with a 100 m × 20 m terrain elevation map (resolution 0.2 m, i.e., 500 × 100 pixels) as input. The network structure consists of 3 convolutional layers (kernel size 5, number of channels 16→32→64), max pooling (2×2), and after flattening, a fully connected layer outputs a 128-dimensional terrain feature vector. ; A lightweight Vision Transformer (ViT) is used to divide the image acquired by the multispectral camera into 16×16 pixel patches. After linear projection and position encoding, the patches are input into the encoder of the Transformer (4 layers, 8 attention heads), and the output is a crop feature vector with a dimension of 128. .

[0064] The time series uses 1D-CNN, the terrain uses 2D-CNN, and the crop uses ViT, all designed to be tailored to the data characteristics, resulting in high extraction efficiency.

[0065] 2.2 Design a three-level integration mechanism: Calculate the original attention score for mode m (m∈{time,terrain,crop}) of sprinkler machine i at time t. The loop variable iterates through all modalities. The values ​​are time, terrain, and crop, respectively. , in , is the weight matrix for the multimodal modes. For trainable vectors shared by all modalities, yes transpose, It is the hyperbolic tangent activation function.

[0066] It takes the dot product of a row vector and a column vector, and outputs a scalar. That is, the unnormalization of the mode at this moment.

[0067] Level 1: Attention weight normalization: ; Calculate the credibility of each modality based on real-time data quality indicators: , Among them, signal-to-noise ratio (SNR), packet loss rate (lossRate), and reconstruction error (reconErr) are... , , As a configurable coefficient, the sigmoid function compresses the output to [0,1]. When At that time, the spatiotemporal alignment data of this modality was marked as untrustworthy and temporarily excluded during fusion.

[0068] Level 3: Weighting and Feature Fusion Multiplying the attention weight by the credibility yields the final fusion weight: The weighted summation yields the comprehensive eigenvector: This comprehensive feature vector The dimension is 128, integrating three modal information: environmental temporal changes, terrain structure, and crop growth.

[0069] Attention weights are automatically adjusted according to scene changes (e.g., terrain weights are increased when there is a narrow road, and crop weights are increased when there are crop abnormalities). The credibility mechanism automatically reduces the weights or excludes low-quality modes. Sensor failures do not affect the overall system operation. Attention weights are visualized, and maintenance personnel can intuitively see what the system is currently focusing on.

[0070] Step two differs from traditional methods that rely solely on single location information. This invention integrates multi-source data such as soil, terrain, and crops, and achieves adaptive feature fusion through an attention-reliability dual weighting mechanism, thereby improving the perception robustness in complex farmland environments.

[0071] Step 3: Based on bidirectional time-series modeling, the operational status of the sprinkler irrigation machine is quantitatively assessed to determine whether soil moisture, terrain adaptability, and crop growth are normal.

[0072] 3.1. The comprehensive feature vector obtained in step two Inputting time series data into a BiGRU neural network captures forward and backward temporal dependencies. Time window length. (Covering a travel distance of approximately 50 m for the sprinkler), step length 5 s.

[0073] BiGRU neural network structure: 128 hidden layers, 2 layers, forward GRU for forward timing, and backward GRU for backward timing. The output at time step t is: in, In the forward-hidden state, This represents the backward hidden state, where ∥ indicates vector concatenation, and the output dimension is 256.

[0074] By utilizing both past and future information, it can identify anomalous patterns more accurately than a one-way RNN.

[0075] 3.2 Based on the output of the BiGRU neural network, three parallel evaluation branches are designed: 3.21 Soil Moisture Consistency Assessment: Determine whether the soil moisture content in the current operating area of ​​the sprinkler irrigation machine matches the preset irrigation target, and output continuous values. ∈[0,1], when it is below 0.6, a supplementary irrigation suggestion is triggered.

[0076] 3.22 Terrain Adaptability Assessment: Determine whether the terrain within 50 m ahead is suitable for the current sprinkler to pass through, and output continuous values. ∈[0,1], when it is below 0.5, formation adjustment or path replanning is triggered.

[0077] 3.23. Crop Growth Anomaly Detection: Determines whether there are abnormal fluctuations in the Normalized Difference Vegetation Index (NDVI) (which may indicate pests, diseases, or uneven irrigation). Outputs a binary value. ∈{0,1}, an abnormal alarm is triggered when the value is 1.

[0078] The three branches share the temporal features extracted by the BiGRU neural network and are jointly optimized using a multi-task learning framework. The total loss function is: in =0.4, =0.3, =0.3, and Mean square error is used. Binary cross-entropy is adopted. The three evaluation branches share underlying features, reducing the amount of computation. At the same time, each task promotes each other, and the output is a continuous value (0~1) instead of a binary judgment, which facilitates the adjustment of subsequent decision thresholds, identifies the trend of soil moisture decline and the trend of terrain change, and issues early warnings before problems occur.

[0079] Step three combines environmental perception data with time-series modeling to achieve a comprehensive quantitative assessment of soil moisture, terrain adaptability, and crop growth, providing a quantitative basis for subsequent operational decisions.

[0080] Step 4: Multi-machine collaborative operation task scheduling and path planning to achieve cloud-edge-device three-level collaboration, automatically adjust the fleet formation according to dynamic risks, and ensure passage efficiency and safety.

[0081] 4.1 Construct a three-tier collaborative architecture: End (sprinkler): Responsible for multimodal spatiotemporal aligned data acquisition, local feature extraction, and execution of control commands.

[0082] Edge (Edge Gateway): Deployed in the field management room (one edge gateway is shared by every 2-3 sprinkler irrigation machines), responsible for multi-machine data aggregation, collaborative operation scheduling, and formation reconfiguration decisions. The edge gateway is equipped with an NVIDIA Jetson AGX Orin platform (32 TOPS computing power) and runs a lightweight inference engine.

[0083] Cloud (Control Center): Deployed on a remote server, responsible for global path planning, model training and updating, and job data storage and analysis.

[0084] The edge gateway communicates with the cloud via a 5G network (latency ≤ 20 ms), and communicates with the sprinkler via a 5.8 GHz industrial wireless network (latency ≤ 10 ms).

[0085] Unified scheduling by edge gateways avoids conflicts caused by multiple devices operating independently, and automatically adjusts according to road conditions (side-by-side on wide roads, longitudinal on narrow roads, and single-vehicle on extremely narrow roads), improving traffic efficiency.

[0086] 4.2 Constructing a dynamic risk assessment index system for multi-machine collaborative operations: Environmental risks: Average soil bearing capacity, average slope, and obstacle density within 50 m ahead; weighting range: 30±5%. Equipment risks: Remaining water volume of the sprinkler, deviation between current speed and preset speed, and communication signal strength; weighting range: 25±3%. Operational risks: Remaining work area, deviation between current work progress and plan, weighting range: 25±3%; Formation risk: deviation between the spacing of adjacent sprinklers and the theoretical spacing, degree of formation distortion, weight range: 20±2%; Each weight is dynamically adjusted based on real-time threats, using the following formula: + · Where d represents the dimension, d traverses four dimensions: environmental risk, equipment risk, operational risk, and formation risk; Wd is the importance coefficient of the d-th dimension in the comprehensive risk score; α=0.85 is the smoothing factor. The gradient of risk with respect to dimensional features is calculated using an online backpropagation neural network. It is the quantized feature value of the d-th dimension, with its value range normalized to [0,1].

[0087] Based on comprehensive risk score Matching a three-level response mechanism: Low risk (R < 0.3): Maintain current formation and log. For medium risk (0.3 ≤ R < 0.6), reduce speed to 0.3 m / s and increase safety distance to 1.5 times. High risk R ≥ 0.6 triggers formation reconfiguration (column formation or single vehicle passing through), and reports to the cloud; The four-dimensional risk assessment (environment, equipment, operation, formation) comprehensively covers influencing factors, and the threat gradient mechanism enables the system to automatically focus on the most dangerous dimension at present.

[0088] 4.3 The edge gateway adopts a priority queue-based task scheduling algorithm: Task priority is calculated by taking into account the remaining operating time, remaining water volume, and distance from the target point.

[0089] When multiple irrigation machines cross paths, the higher priority irrigation machine is dispatched first, while the lower priority irrigation machine waits in the waiting area (the waiting area is ≥30 m away from the intersection point). When a deadlock (multi-machine loop waiting) is detected, the lowest priority sprinkler is forced to retreat 10 m to break the loop.

[0090] Step four combines a four-dimensional risk quantification assessment of environment, equipment, operation, and formation with adaptive formation adjustment to achieve intelligent scheduling of multi-machine collaborative operations.

[0091] Step 5: Based on reinforcement learning, the transfer path planning and collaborative optimization generate the optimal transfer path for multiple sprinkler irrigation machines, balancing the three objectives of path length, soil protection, and number of turns.

[0092] 5.1 The relocation route planning comprehensively considers three objectives: route length, soil compaction risk, and number of turns. The cost function is as follows: Where L represents the total path length (m), with a weight wL = 0.4. S represents the cost of soil compaction, and T represents the number of turns, with each turn calculated based on an equivalent path length of 20 m, with a weight wT = 0.25. The system considers path length, soil compaction, and the number of turns simultaneously, balancing efficiency and sustainability.

[0093] 5.2 The Proximal Policy Optimization (PPO) algorithm is used for online optimization of path planning. The state space includes: current pose, terrain elevation map 100 m ahead, soil bearing capacity distribution, and target point location. The action space includes: speed adjustment (0.2~0.8 m / s), steering angle adjustment (-0.1~0.1 rad / s), and formation adjustment commands. Reward function design: in This is the path length increment. To increase soil compaction volume, For the increment of the number of turns, This is the collision indication function (-100 upon collision). =0.1, =0.2, =0.1, =100.

[0094] The transition path planning based on deep reinforcement learning is trained offline in the Proximal Policy Optimization (PPO) algorithm (based on historical job data). After training, it is distributed to the edge gateway through federated learning. The edge gateway performs lightweight inference based on the real-time environment to achieve real-time adjustment of the path.

[0095] Deep reinforcement learning can handle dynamic environments (movement of other sprinklers, temporary obstacles). When a cluster needs to be moved from the current work area to the next work area, the edge gateway executes the following process: 5.31. Obtain the current location, remaining water volume, and operation completion status of all sprinkler irrigation machines; 5.32. Call the relocation path planning function to generate a relocation path for each sprinkler machine; 5.33. Detect path conflicts; if intersections exist, sort them by priority and adjust the time window accordingly. 5.34. Issue the transition command, and the formation will automatically switch to column formation; 5.35 During the relocation process, each sprinkler machine reports its location in real time, and the edge gateway adjusts dynamically.

[0096] Federated learning aggregates experience from multiple sites, making the model more accurate with use. It only uploads gradients, not raw data, thus protecting privacy.

[0097] Step five incorporates path length, soil compaction risk, and number of turns into a unified cost function, and uses the PPO algorithm to achieve online path optimization, balancing operational efficiency and soil protection.

[0098] Step Six: Closed-loop evaluation and continuous evolution mechanism to form a complete closed loop of perception-decision-control-evaluation-evolution, enabling the system to improve itself.

[0099] 6.1 After each work block (100 m × 100 m) is completed, the system performs a work quality assessment: Irrigation uniformity: After the operation, soil moisture content was sampled (9 points evenly distributed), and the coefficient of variation CV = standard deviation / mean was calculated. CV ≤ 0.15 is excellent, 0.15~0.25 is acceptable, and >0.25 is unacceptable.

[0100] Work efficiency: The ratio of actual work time to theoretical work time; a ratio ≥0.9 indicates high efficiency.

[0101] Energy efficiency: Electricity consumption per unit area (kWh / acre), compared with historical average.

[0102] The evaluation results are stored on the blockchain (each sprinkler machine acts as a blockchain node) to ensure that the data is immutable and used for subsequent model training and parameter correction.

[0103] 6.2 Constructing a closed-loop optimization system: Edge nodes: Upload anonymized job logs (including perception data, decision records, and evaluation results) weekly. Using a federated learning framework, only model gradient updates are uploaded, without disclosing the original data.

[0104] Cloud Center: Monthly aggregation of model updates from each edge node, retraining of BiGRU neural network and near-end policy optimization algorithm, and distribution of the optimized BiGRU neural network and near-end policy optimization algorithm model to edge gateway after compression.

[0105] Blockchain-based evidence storage: All model versions, update records, and job evaluation results are stored on the blockchain, forming a traceable knowledge base.

[0106] The feedback process for assignment quality, from steps two to five, leads to model training, forming a positive cycle of continuous improvement. Blockchain ensures that the evaluation data is tamper-proof and provides real labels for model training. Federated learning enables cross-site knowledge sharing without disclosing the original data.

[0107] 6.3 When the system detects the following anomalies, it will automatically switch to degrade mode: Communication interruption: The sprinkler continues to execute according to the last received instruction for 60 seconds. If communication is still not restored, it will automatically decelerate and stop, and activate the audible and visual alarms. It will also use lidar for autonomous obstacle avoidance.

[0108] Positioning failure: Switch to inertial measurement unit + wheel speed meter calculation mode. Positioning error accumulates at 2 m / min. If it is still not recovered after driving 50 m, stop alarm.

[0109] Sensor failure: The faulty mode is automatically masked, and the dynamic weight fusion mechanism automatically reduces the weight of the mode, allowing the system to continue operating in degraded mode.

[0110] Collision risk: If the lidar detects an obstacle entering the restricted area (radius 10 m), the vehicle will brake to a stop and resume driving after the obstacle is removed.

[0111] The hierarchical fault tolerance mechanism ensures that a single point of failure does not affect the overall operation. All blocking events and model versions are recorded on the blockchain for easy auditing and problem backtracking.

[0112] Step six achieves reliable assessment of job quality and continuous optimization of model parameters with privacy protection, forming a complete closed loop of perception-decision-execution-evaluation-evolution.

[0113] Working principle: Multimodal perception and attention-credibility dual-weighted fusion mechanism, quantitative assessment of job status based on BiGRU bidirectional temporal modeling and multi-task learning, cloud-edge-device collaborative architecture and four-dimensional dynamic risk assessment formation adaptive adjustment, multi-objective transition path planning and federated learning continuous evolution based on PPO deep reinforcement learning, and a blockchain-based job quality closed-loop assessment and continuous evolution mechanism, forming a complete technical closed loop from perception to evolution, realizing the integration of perception-decision-control-assessment-evolution.

[0114] It should be understood that the above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. It should not be considered that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, several simple deductions or substitutions can be made without departing from the concept of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for coordinated operation of multiple sprinkler irrigation machines, characterized in that: Includes the following steps: Step 1, Multimodal Data Acquisition, includes: Time-series modal data: temporal variation patterns of sprinkler operation status and soil moisture; Terrain modality: Data on the terrain undulations, width, and obstacle distribution of the road ahead; Crop modality: Crop growth, pest and disease status, and irrigation uniformity constitute crop modality; Spatiotemporal alignment of time-series modes, terrain modes, and crop modes; Step 2: Convert the raw multimodal data into a comprehensive feature vector: Temporal feature vector, terrain feature vector, crop feature vector; The attention weight is multiplied by the credibility to obtain the final fusion weight, which integrates three modal information: environmental temporal changes, terrain structure, and crop growth. Step 3: Combine the comprehensive feature vector obtained in Step 2. By inputting time series data into a BiGRU neural network, forward and backward temporal dependencies are captured; based on the output of the BiGRU neural network, soil moisture consistency assessment, terrain adaptability assessment, and crop growth anomaly detection are designed. Step four includes: 4.1 Construct a collaborative task scheduling and path planning system for sprinkler irrigation machines, edge gateways, and control centers. The sprinkler irrigation machines are responsible for multimodal spatiotemporal aligned data acquisition, local feature extraction, and execution of control commands. The edge gateways are responsible for machine data aggregation, collaborative task scheduling, and formation reconstruction decisions. The control center is responsible for global path planning, model training and updating, and task data storage and analysis. 4.2 Establish dynamic risk assessment indicators for multi-machine collaborative operations, and match response mechanisms based on comprehensive risk scores. Step 5: Reinforcement learning-based transition path planning and collaborative optimization. The transition path planning comprehensively considers path length, soil compaction risk, and number of turns. The near-end strategy optimization algorithm is used to optimize the path planning online and adjust the path in real time.

2. The method for coordinated operation of multiple sprinkler irrigation machines according to claim 1, characterized in that: The spatiotemporal alignment in step one includes: Time alignment, let the target time be t, and the sampling time of a certain modality be... and The interpolation formula is: Spatial alignment involves projecting all multimodal data onto a local coordinate system with the geometric center of the sprinkler as the origin. The X-axis represents the longitudinal direction of the machine, and the Y-axis represents the lateral direction. After coordinate transformation, the lidar point cloud data generates a terrain elevation map of the foreground.

3. The method for coordinated operation of multiple sprinkler irrigation machines according to claim 1, characterized in that: The comprehensive feature vector Multimodal data fusion is performed using a method that includes the following steps: 2.1 Calculate the original attention score for mode m (m∈{time,terrain,crop}) of sprinkler machine i at time t. The loop variable iterates through all modalities. The values ​​are taken as follows: time (time-series mode), terrain (terrain mode), and crop (crop mode). ， in Let be the weight matrix for the multimodal expressions, and ν be a trainable vector shared by all modalities. It is the transpose of ν. It is the hyperbolic tangent activation function; It takes the dot product of a row vector and a column vector, and outputs a scalar. That is, the unnormalization of the mode at this moment; 2.2 Dynamic Weight Fusion Mechanism Level 1: Attention weight normalization: ; Level 2: Calculate the credibility of each modality based on real-time data quality indicators: ， Where SNR is the signal-to-noise ratio, lossRate is the packet loss rate, reconErr is the reconstruction error, η1=0.5, η2=0.3, η3=0.2 are configurable coefficients, and the sigmoid function compresses the output to [0,1]. Level 3: Weighting and Feature Fusion Multiplying the attention weight by the credibility yields the final fusion weight: The weighted summation yields the comprehensive eigenvector: .

4. The method for coordinated operation of multiple sprinkler irrigation machines according to claim 1, characterized in that: The soil moisture consistency assessment includes determining whether the soil moisture content in the current operating area of ​​the sprinkler irrigation machine matches the preset irrigation target, and outputting continuous values. ∈[0,1], triggering a supplementary watering suggestion when it is below 0.6; The terrain adaptability assessment includes determining whether the terrain within a 50-meter radius ahead is suitable for the current sprinkler to pass through, and outputting continuous values. ∈[0,1], when it is below 0.5, formation adjustment or path replanning is triggered; The crop growth anomaly detection includes determining whether the normalized vegetation index (NVI) shows abnormal fluctuations and outputting a binary value. ∈{0,1}, an abnormal alarm is triggered when the value is 1.

5. The method for coordinated operation of multiple sprinkler irrigation machines according to claim 1, characterized in that: Step 4.2, which involves constructing a dynamic risk assessment index system for multi-machine collaborative operations, includes: Environmental risks: Average soil bearing capacity, average slope, and obstacle density within 50 m ahead; weighting range: 30±5%. Equipment risks: Remaining water volume of the sprinkler, deviation between current speed and preset speed, and communication signal strength; weighting range: 25±3%. Operational risks: Remaining work area, deviation between current work progress and plan, weighting range: 25±3%; Formation risk: deviation between the spacing of adjacent sprinklers and the theoretical spacing, degree of formation distortion, weight range: 20±2%; Each weight is dynamically adjusted based on real-time threats, using the following formula: + · Where d represents the dimension, d traverses four dimensions: environmental risk, equipment risk, operational risk, and formation risk; Wd is the importance coefficient of the d-th dimension in the comprehensive risk score; α=0.85 is the smoothing factor. The gradient of risk with respect to dimensional features. It is the quantized feature value of the d-th dimension, with its value range normalized to [0,1].

6. A method for coordinated operation of multiple sprinkler irrigation machines according to any one of claims 1 or 5, characterized in that: The comprehensive risk score includes It has a three-level response mechanism: Low risk (R < 0.3): Maintain current formation and log. For medium risk (0.3 ≤ R < 0.6), reduce speed to 0.3 m / s and increase safety distance to 1.5 times. High risk R ≥ 0.6 triggers formation reconfiguration and reports to the cloud.

7. A method for coordinated operation of multiple sprinkler irrigation machines according to claim 1, characterized in that: The near-end strategy optimization algorithm performs online optimization of path planning, including: State space optimization: current pose, terrain elevation map 100 m ahead, soil bearing capacity distribution, target point location; Action space optimization: speed adjustment; steering angle adjustment; formation adjustment instructions; design reward function: in This is the path length increment. To increase soil compaction volume, For the increment of the number of turns, For collision indication function, =0.1, =0.2, =0.1, =100; The transition path planning based on deep reinforcement learning is trained offline in the near-end policy optimization algorithm. After training, it is distributed to the edge gateway through federated learning. The edge gateway performs lightweight inference according to the real-time environment to realize real-time adjustment of the path.

8. A method for coordinated operation of multiple sprinkler irrigation machines according to any one of claims 1 or 8, characterized in that: This also includes the following steps performed by the edge gateway when the cluster moves from the current work area to the next work area: 5.

31. Obtain the current location, remaining water volume, and operation completion status of all sprinkler irrigation machines; 5.

33. Call the relocation path planning function to generate a relocation path for each sprinkler machine; 5.

34. Detect path conflicts; if intersections exist, sort by priority and adjust the time window accordingly. 5.

35. Issue the transition command, and the formation will automatically switch to column formation; 5.36 During the relocation process, each sprinkler machine reports its location in real time, and the edge gateway adjusts dynamically.

9. A method for coordinated operation of multiple sprinkler irrigation machines according to any one of claims 1-8, characterized in that: The process also includes step six: after each work block is completed, a work quality assessment is performed, evaluating irrigation uniformity, work efficiency, and energy efficiency. The assessment results are stored on the blockchain to ensure data immutability and are used for subsequent model training, updates, and parameter correction. A closed-loop optimization system is constructed, including uploading anonymized work logs weekly, optimizing the BiGRU neural network and PPO path planning, and storing all model training and update versions, update records, and work assessment results on the blockchain to form a traceable knowledge base. The evaluation results are stored on the blockchain (each sprinkler machine acts as a blockchain node) to ensure that the data is immutable and used for subsequent model training, updates, and parameter correction.

10. A method for coordinated operation of multiple sprinkler irrigation machines according to claim 9, characterized in that: Automatically switch to degrade mode when the following anomalies are detected: Communication interruption: Automatically decelerates and stops, activates audible and visual alarms, and uses lidar for autonomous obstacle avoidance; Location tracking failure: Parking alarm; Sensor failure: The faulty mode is automatically masked, and the dynamic weight fusion mechanism automatically reduces the weight of the mode, so the system continues to operate in degraded mode; Collision risk: The lidar detects an obstacle entering the restricted area, triggers emergency braking to stop the vehicle, and resumes driving after the obstacle is removed.

Images