An unmanned aerial vehicle cluster management and control system and method based on an artificial intelligence model

Abstract

The application discloses an unmanned aerial vehicle cluster management and control system and method based on an artificial intelligence model, relates to the technical field of artificial intelligence, and comprises the following steps: collecting unmanned aerial vehicle multi-source sensor data according to a preset period, performing self-adaptive preprocessing through an artificial intelligence model to obtain preprocessed data and screened feature data; performing IMU dead calculation based on the preprocessed data to obtain an initial pose sequence, combining residual prediction and weighted least square optimization to obtain a local pose estimation value and a covariance matrix; detecting stable environment features based on the screened feature data and generating hash codes, generating temporary anchor points through multi-machine matching and triangulation optimization; performing global effective anchor point screening and fusion weight calculation based on the local pose estimation value and temporary anchor point information, realizing distributed consistency fusion, and obtaining a cluster reference consistent fusion pose estimation value; and calculating uncertainty based on the fusion pose estimation value, and executing a preset control strategy according to a preset threshold interval.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a drone swarm management system and method based on an artificial intelligence model. Background Technology

[0002] Drone swarm management is a technical field that achieves precise positioning and task execution through multi-drone collaboration, relying primarily on multi-source sensor data fusion and pose estimation technologies. Specifically, the IMU (Inertial Measurement Unit) collects motion attitude data, the UWB / RSSI module provides relative distance and communication signal information, and the image sensor captures environmental features. These three components work together to provide the foundational data for pose calculation. Distributed collaboration is the mainstream mode of swarm management, achieving global task scheduling through data interaction between drones. Complex environments (such as urban canyons, electromagnetic interference zones, and scenes with variable lighting) can easily interfere with sensor data, representing a typical application scenario that swarm management must address. The core technical requirements for this are improving data reliability and collaborative consistency.

[0003] Existing UAV swarm management technologies have significant shortcomings in complex environments. The core issue is the lack of a unified spatial reference, which easily leads to reference drift and inconsistencies in cooperative localization. In traditional distributed solutions, each UAV independently estimates its pose based on its own sensors, without a globally unified reference anchor point. In complex environments, sensor data is affected by noise, obstruction, and electromagnetic interference, making it impossible to verify the reliability of environmental characteristics observed by a single UAV, and easily causing local reference shifts. After long-term operation, the pose references of each UAV diverge and drift, leading to swarm formation collapse and failure of cooperative tasks (such as joint inspections and formation flights), severely limiting the reliability and stability of UAV swarm applications in complex real-world scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide a drone swarm management system and method based on an artificial intelligence model to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for managing and controlling a drone swarm based on an artificial intelligence model, the method comprising the following steps:

[0006] Step 1: Collect multi-source sensor data from the UAV according to a preset cycle, and use an artificial intelligence model for adaptive preprocessing to obtain preprocessed data and filtered feature data;

[0007] Step 2: Based on the preprocessed data, perform IMU calculations to obtain the initial pose sequence. Combine residual prediction and weighted least squares optimization to obtain the local pose estimate and covariance matrix.

[0008] Step 3: Based on the filtered feature data, detect stable environment features and generate hash codes, and generate temporary anchor points through multi-machine matching and triangulation optimization;

[0009] Step 4: Based on the local pose estimation value and temporary anchor point information, perform global effective anchor point screening and fusion weight calculation to achieve distributed consistent fusion and obtain a fused pose estimation value with cluster benchmark consistency.

[0010] Step 5: Calculate the uncertainty based on the fused pose estimation value, and execute the preset control strategy according to the preset threshold range.

[0011] In step 1, noise, outliers and invalid frames in the sensor data are filtered out by preprocessing to output a high-quality data foundation, providing reliable input for subsequent pose estimation and anchor point generation, and avoiding the impact of poor data on control accuracy.

[0012] The multi-source sensor data includes IMU data, image frames, and UWB / RSSI data;

[0013] The artificial intelligence model includes a noise suppression model and a feature extraction model;

[0014] The noise suppression model can adopt a lightweight CNN-LSTM hybrid model. Its core function is to dynamically identify types of noise such as Gaussian noise and impulse noise in IMU data. By adaptively adjusting the filtering parameters (such as the process noise variance Q and observation noise variance R of Kalman filtering), it can achieve accurate noise suppression and avoid the insufficient adaptation of fixed filtering parameters to complex noise.

[0015] The feature extraction model can use the SIFT algorithm, SURF algorithm, or lightweight CNN model (such as MobileNet). Its core function is to efficiently extract distinctive features such as edges and corners from image frames. At the same time, it filters low-quality frames caused by motion blur and poor lighting by using sharpness evaluation (such as variance calculation based on gray level gradient) to ensure the effectiveness of the output feature data.

[0016] The adaptive preprocessing specifically includes:

[0017] The noise suppression model is called to process the IMU data sequence, and the filtering parameters are dynamically adjusted to output the denoised IMU data, denoised as I. The filtering parameter adjustment logic is as follows: the model uses a sliding window (e.g., window size of 20 frames) to statistically analyze the variance, peak factor and other features of the IMU data. After identifying the noise type, the model increases the observation noise variance weight for Gaussian noise and uses median filtering for impulse noise preprocessing before performing Kalman filtering.

[0018] Abnormal jump values ​​in UWB / RSSI data are filtered based on the filtering parameters to output stable UWB / RSSI data, denoted as U. The filtering parameters reuse the noise statistical characteristics output by the noise suppression model, such as the noise variance σ of the IMU data. When the deviation of a certain frame value in UWB / RSSI data from the mean of the sliding window exceeds 3σ, it is determined to be an abnormal jump value, and the abnormal value is replaced by linear interpolation.

[0019] The abnormal jump value refers to data whose deviation from the data mean exceeds a preset deviation threshold.

[0020] The preprocessed data includes data I and U;

[0021] The feature extraction model is invoked to extract the edge and corner features of the image frame, and frames with a sharpness lower than a preset threshold are filtered to obtain the filtered feature data.

[0022] In step 2, the short-term pose is quickly obtained by IMU calculation, and the accumulated error of IMU is corrected by residual prediction. Combined with weighted least squares to fuse multi-source preprocessed data, the accuracy and stability of local pose estimation are improved.

[0023] The IMU dead calculation refers to the short-term calculation performed based on the denoised IMU data using the extended Kalman filter algorithm to output an initial pose sequence; the initial pose sequence contains three-dimensional position and three-dimensional attitude angle parameters;

[0024] The residual prediction is achieved through a time-series prediction model, which takes continuous frame data of the initial pose sequence as input and outputs the cumulative error prediction value (Δx,Δy,Δz,Δα,Δβ,Δγ) calculated by the IMU.

[0025] LSTM or GRU models can be used. The input is the initial pose sequence of consecutive frames, and the output is the cumulative error prediction value of the corresponding frame.

[0026] Where Δx, Δy, and Δz represent the residuals in the three-dimensional position direction, and Δα, Δβ, and Δγ represent the residuals in the three-dimensional attitude angle direction, respectively.

[0027] The weighted least squares optimization uses a sliding window of preset length as the processing unit to construct the objective function: minΣ i=1 K [w I,i (X i -X I,i ) 2 +w U,i (X i -X U,i ) 2 ];

[0028] Where K represents the number of data frames within the sliding window; i∈{1,2,…,K}; wI,i The weight w represents the weight of the data I in the i-th frame. I,i =1 / σ I,i 2 +|ΔX pred,i |,w U,i The weight w represents the weight of the data U in the i-th frame. U,i =1 / σ U,i 2 ;σ I,i σ U,i Let ΔX represent the noise variance of the I and U data, respectively. pred,i X represents the residual prediction value of the i-th frame. i Let X represent the optimized pose of the i-th frame. I,i X U,i These represent the pose estimates corresponding to IMU calculations and UWB measurements, respectively.

[0029] w I,i Introducing residual prediction value |ΔX pred,i | is to reduce the weighting percentage when the accumulated error of IMU calculation is large, σ I,i σ U,i A larger value indicates greater data noise and a smaller weight.

[0030] The local pose estimate X and the corresponding covariance matrix P are obtained by solving. P is a 6×6 matrix, and the diagonal elements correspond to the variances of the three-dimensional position (x,y,z) and the three-dimensional attitude angle (α,β,γ), respectively. The smaller the variance value, the more reliable the pose estimate in that dimension.

[0031] In step 3, stable features in the environment are extracted and transformed into spatial reference anchors to provide a unified spatial benchmark for the cluster. The reliability of the anchors is verified through multi-machine collaboration to avoid benchmark drift caused by the observation error of a single UAV.

[0032] The specific features for detecting stable environments include: for the filtered feature data, a stability score is calculated based on the spatiotemporal continuity and grayscale contrast of the features, and features with scores greater than a preset stability threshold are retained as candidate stable features.

[0033] The stability scoring logic is as follows: Spatiotemporal continuity score = 1 - feature position offset of 10 consecutive frames / feature scale (score is 1 when offset ≤ 2 pixels), grayscale contrast score = maximum grayscale value of feature region - minimum grayscale value (score is 1 when ≥ 100), the comprehensive score is the product of the two, the preset stability threshold is set to 0.8, and features with a comprehensive score ≥ 0.8 are retained;

[0034] The hash encoding generation specifically includes: inputting candidate stable features using a hash generation method, combining local pose estimation values, calculating the observation angle and relative distance of the candidate stable features, and outputting a triplet containing: hash value, observation angle, and relative distance;

[0035] Example of hash generation method: Using the LSH algorithm, the 128-dimensional feature descriptor is mapped to a 64-bit binary hash value, reducing data transmission and matching overhead;

[0036] Observation angle calculation: Based on the transformation relationship between the UAV body coordinate system and the world coordinate system, combined with the pixel coordinates of the feature in the image frame and camera intrinsic parameters (such as focal length f, principal point coordinates (u0, v0)), the azimuth and pitch angles of the feature relative to the UAV are solved; Relative distance calculation: Based on the local pose estimation value and the image ranging model of the feature (such as the triangulation principle), the preliminary distance is obtained, and then corrected by combining UWB measurement data before output;

[0037] The process of generating temporary anchor points through multi-machine matching and triangulation optimization specifically includes:

[0038] The drone broadcasts triples to neighboring drones within its communication range via wireless communication. Simultaneously, it receives triples broadcast by neighboring drones and performs hash matching on the received triples. When the Hamming distance is less than or equal to a preset matching threshold, it is considered a match. The number N of drones that match the same hash value is counted.

[0039] When N is greater than or equal to a preset threshold, the three-dimensional coordinates of the features are solved by a triangulation algorithm based on the local pose estimates, observation angles and relative distances of multiple UAVs; the consistency scores of the three-dimensional coordinates are scored, and features with scores greater than or equal to a preset consistency threshold are determined as temporary anchor points.

[0040] Triangulation algorithm: Based on the principle of multi-view geometry, the local pose of each UAV is taken as the observation viewpoint. Combined with the observation angle and distance of the features relative to each viewpoint, an overdetermined set of equations is constructed, and the three-dimensional coordinates of the features in the world coordinate system are solved by least squares.

[0041] Consistency score: Calculate the standard deviation of the three-dimensional coordinates obtained by multiple UAVs. For example, the score is 1 when the standard deviation is ≤0.2m, and the score decreases by 0.2 for every 0.1m increase in standard deviation. The preset consistency threshold can be set to 0.8.

[0042] Output the three-dimensional coordinates of the temporary anchor point and the corresponding covariance matrix P1; P1 is a 3×3 matrix, and the diagonal elements correspond to the variance of the three-dimensional coordinates (x,y,z), reflecting the uncertainty of the spatial position of the anchor point.

[0043] In step 4, the reliability of the reference benchmark is ensured by filtering global effective anchor points, and the trust level of local and neighboring machine data is balanced by reasonably allocating fusion weights, so as to achieve the unification of the cluster pose benchmark and solve the problem of benchmark inconsistency in the distributed system.

[0044] The global valid anchor point screening process is as follows: after each UAV receives the temporary anchor point information broadcast by its neighboring UAV, it counts the number of UAVs observing the same hash value anchor point and calculates the coordinate deviation of different UAVs observing the anchor point. When the number of observations is greater than or equal to a preset threshold and the coordinate deviation is less than or equal to a preset deviation threshold, it is determined to be a global valid anchor point.

[0045] The temporary anchor point information refers to the temporary anchor point hash value received by the UAV from a neighboring UAV and the three-dimensional coordinates of the anchor point observed by the UAV itself.

[0046] The calculation of the fusion weights specifically includes: inputting the covariance matrix trace of the neighboring machine, the UWB / RSSI signal strength, and the number of globally effective anchor points observed by the neighboring machine, and allocating the neighboring machine fusion weights w through weighted normalization. neigh ; Satisfying Σw neigh +w local =1; where, w local The local weight is determined by the trace of the local covariance matrix. The trace of the covariance matrix is ​​the sum of the diagonal elements of the matrix and is a simplified index for quantifying the uncertainty of pose estimation. The smaller the trace value, the lower the uncertainty.

[0047] The distributed consensus fusion specifically includes: employing a distributed Kalman consensus algorithm, combined with coordinate constraints and fusion weights of globally valid anchor points, to fuse the poses of the local machine and its neighboring machines: X fuse =w local ·X+Σw neigh ·X neigh The three-dimensional coordinates of the globally effective anchor point are used as a fixed reference. During the fusion process, the pose estimation values ​​of the local and neighboring drones are forced to satisfy the "observation equation based on the anchor point", that is, the world coordinates of the anchor point = drone pose + feature coordinates relative to the drone. This constraint corrects pose drift and ensures that the pose reference of the cluster is unified.

[0048] Among them, X fuse X represents the fused pose estimate. neigh This represents the local pose estimate of the neighboring machines; the fused pose estimate X, consistent with the cluster baseline, is obtained. fuse After fusion, X fuse The trace of the covariance matrix is ​​usually smaller than the trace of the covariance matrix between the local machine and all its neighbors;

[0049] In step 5, the reliability of pose estimation is quantified by uncertainty, and the control strategy is dynamically adjusted to ensure positioning accuracy while taking into account cruise efficiency and energy consumption control.

[0050] The calculation of uncertainty based on the fused pose estimate specifically includes:

[0051] Based on the fused pose estimate, Monte Carlo inference is performed (with a preset sampling number of 1000 times, and a covariance matrix P is added to the fused pose estimate at each sampling). fuse (Ignoring random noise, we obtain 1000 sampled poses, statistically analyze the standard deviation of sampled values ​​in each dimension, output probability distribution, and calculate uncertainty S: S = ((σ x 2 +σ y 2 +σ z 2 )+(σ α 2 +σ β 2 +σ γ 2 )) 1 / 2 ;

[0052] Preset thresholds include conservative mode threshold S th1 With recovery mode threshold S th2 S th1 >S th2 A preset control strategy is executed based on the threshold range in which the uncertainty S is located.

[0053] For example: when S≥S th1 At this time: reduce the cruising speed (take the preset conservative speed range), increase the shooting frame rate (take the preset high frequency range), and prioritize the execution of temporary anchor point triangulation tasks;

[0054] When S th1 >S>S th2 Time: Dynamically adjust the shooting frame rate and speed to the preset middle range, taking into account both anchor point detection and cruising efficiency;

[0055] When S≤S th2 Furthermore, when the preset time period is continuously maintained: the normal cruise mode is restored (preset normal speed, low frequency shooting), and the cruise path is executed first;

[0056] It outputs speed control commands, shooting control commands, and task scheduling commands for the UAV, enabling dynamic management and control based on pose reliability.

[0057] A drone swarm management and control system based on an artificial intelligence model, comprising a data preprocessing module, a local pose estimation module, a feature anchor point generation module, and a fusion decision control module;

[0058] The data preprocessing module is used to collect multi-source sensor data from the UAV at a preset period, and obtain preprocessed data and filtered feature data through adaptive preprocessing by an artificial intelligence model. The local pose estimation module is used to perform IMU calculation based on the preprocessed data to obtain the initial pose sequence, and obtain the local pose estimate and covariance matrix by combining residual prediction and weighted least squares optimization. The feature anchor point generation module is used to detect stable environment features and generate hash codes based on the filtered feature data, and generate temporary anchor points through multi-machine matching and triangulation optimization. The fusion decision control module is used to perform global effective anchor point screening and fusion weight calculation based on the local pose estimate and temporary anchor point information to achieve distributed consistent fusion and obtain a fused pose estimate with a consistent cluster baseline. The uncertainty is calculated based on the fused pose estimate, and a preset control strategy is executed according to a preset threshold range.

[0059] The data preprocessing module includes a multi-source acquisition unit and an adaptive processing unit;

[0060] The multi-source acquisition unit is used to acquire multi-source sensor data of the UAV at a preset cycle; the adaptive processing unit is used to call an artificial intelligence model to perform adaptive preprocessing on the acquired data and output denoised data and filtered feature data.

[0061] The local pose estimation module includes a calculation unit, a residual prediction unit, and an optimization unit.

[0062] The calculation unit is used to obtain the initial pose sequence based on the denoised IMU data through IMU brute-force calculation; the residual prediction unit is used to predict the cumulative error residual of the IMU brute-force calculation through a time series prediction model; the optimization unit is used to combine the residual prediction and obtain the local pose estimate and covariance matrix through weighted least squares optimization.

[0063] The feature anchor point generation module includes a feature detection unit, an encoding generation unit, and an anchor point generation unit;

[0064] The feature detection unit is used to detect stable environment features from the filtered feature data; the encoding generation unit is used to generate triple codes containing hash value, observation angle and relative distance for stable features; the anchor point generation unit is used to generate temporary anchor points and their coordinates through hash matching and triangulation optimization between multiple machines.

[0065] The fusion decision control module includes an anchor point screening unit, a weight allocation unit, and a fusion decision unit.

[0066] The anchor point screening unit is used to select globally effective anchor points from temporary anchor points based on multi-machine observation information; the weight allocation unit is used to calculate and allocate the weights required for consistency fusion based on neighboring machine status information; the fusion decision unit is used to perform distributed consistency fusion to obtain the fusion pose consistent with the cluster baseline, and execute a preset control strategy based on its uncertainty.

[0067] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention effectively improves data quality by adaptively preprocessing multi-source sensor data through an artificial intelligence model; it improves the accuracy of local pose estimation by combining residual prediction and weighted least squares optimization; it provides a unified spatial reference for the cluster by generating temporary anchor points and global effective anchor points through multi-machine collaboration; it achieves cluster pose reference consistency based on distributed consistency fusion, reducing the risk of reference drift; and it balances positioning accuracy, cruise efficiency, and energy consumption control through dynamic adjustment of uncertainty control strategies. Attached Figure Description

[0068] Figure 1 This is a flowchart illustrating a drone swarm management system based on an artificial intelligence model according to the present invention. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] Example: Figure 1 As shown, the present invention provides a technical solution, a method for managing and controlling drone swarms based on an artificial intelligence model, the method comprising the following steps:

[0071] Step 1: Collect multi-source sensor data from the UAV according to a preset cycle, and use an artificial intelligence model for adaptive preprocessing to obtain preprocessed data and filtered feature data;

[0072] Step 2: Based on the preprocessed data, perform IMU calculations to obtain the initial pose sequence. Combine residual prediction and weighted least squares optimization to obtain the local pose estimate and covariance matrix.

[0073] Step 3: Based on the filtered feature data, detect stable environment features and generate hash codes, and generate temporary anchor points through multi-machine matching and triangulation optimization;

[0074] Step 4: Based on the local pose estimation value and temporary anchor point information, perform global effective anchor point screening and fusion weight calculation to achieve distributed consistent fusion and obtain a fused pose estimation value with cluster benchmark consistency.

[0075] Step 5: Calculate the uncertainty based on the fused pose estimation value, and execute the preset control strategy according to the preset threshold range.

[0076] In step 1, noise, outliers and invalid frames in the sensor data are filtered out by preprocessing to output a high-quality data foundation, providing reliable input for subsequent pose estimation and anchor point generation, and avoiding the impact of poor data on control accuracy.

[0077] The multi-source sensor data includes IMU data, image frames, and UWB / RSSI data;

[0078] The artificial intelligence model includes a noise suppression model and a feature extraction model;

[0079] The noise suppression model can adopt a lightweight CNN-LSTM hybrid model. Its core function is to dynamically identify types of noise such as Gaussian noise and impulse noise in IMU data. By adaptively adjusting the filtering parameters (such as the process noise variance Q and observation noise variance R of Kalman filtering), it can achieve accurate noise suppression and avoid the insufficient adaptation of fixed filtering parameters to complex noise.

[0080] The feature extraction model can use the SIFT algorithm, SURF algorithm, or lightweight CNN model (such as MobileNet). Its core function is to efficiently extract distinctive features such as edges and corners from image frames. At the same time, it filters low-quality frames caused by motion blur and poor lighting by using sharpness evaluation (such as variance calculation based on gray level gradient) to ensure the effectiveness of the output feature data.

[0081] The adaptive preprocessing specifically includes:

[0082] The noise suppression model is called to process the IMU data sequence, and the filtering parameters are dynamically adjusted to output the denoised IMU data, denoised as I. The filtering parameter adjustment logic is as follows: the model uses a sliding window (e.g., window size of 20 frames) to statistically analyze the variance, peak factor and other features of the IMU data. After identifying the noise type, the model increases the observation noise variance weight for Gaussian noise and uses median filtering for impulse noise preprocessing before performing Kalman filtering.

[0083] Abnormal jump values ​​in UWB / RSSI data are filtered based on the filtering parameters to output stable UWB / RSSI data, denoted as U. The filtering parameters reuse the noise statistical characteristics output by the noise suppression model, such as the noise variance σ of the IMU data. When the deviation of a certain frame value in UWB / RSSI data from the mean of the sliding window exceeds 3σ, it is determined to be an abnormal jump value, and the abnormal value is replaced by linear interpolation.

[0084] The abnormal jump value refers to data whose deviation from the data mean exceeds a preset deviation threshold.

[0085] The preprocessed data includes data I and U;

[0086] The feature extraction model is invoked to extract the edge and corner features of the image frame, and frames with a sharpness lower than a preset threshold are filtered to obtain the filtered feature data.

[0087] In step 2, the short-term pose is quickly obtained by IMU calculation, and the accumulated error of IMU is corrected by residual prediction. Combined with weighted least squares to fuse multi-source preprocessed data, the accuracy and stability of local pose estimation are improved.

[0088] The IMU dead calculation refers to the short-term calculation performed based on the denoised IMU data using the extended Kalman filter algorithm to output an initial pose sequence; the initial pose sequence contains three-dimensional position and three-dimensional attitude angle parameters;

[0089] The residual prediction is achieved through a time-series prediction model, which takes continuous frame data of the initial pose sequence as input and outputs the cumulative error prediction value (Δx,Δy,Δz,Δα,Δβ,Δγ) calculated by the IMU.

[0090] LSTM or GRU models can be used. The input is the initial pose sequence of consecutive frames, and the output is the cumulative error prediction value of the corresponding frame.

[0091] Where Δx, Δy, and Δz represent the residuals in the three-dimensional position direction, and Δα, Δβ, and Δγ represent the residuals in the three-dimensional attitude angle direction, respectively.

[0092] The weighted least squares optimization uses a sliding window of preset length as the processing unit to construct the objective function: minΣ i=1 K [w I,i (X i -X I,i ) 2 +w U,i (X i -X U,i ) 2 ];

[0093] Where K represents the number of data frames within the sliding window; i∈{1,2,…,K}; w I,i The weight w represents the weight of the data I in the i-th frame. I,i =1 / σ I,i 2 +|ΔX pred,i |,w U,i The weight w represents the weight of the data U in the i-th frame. U,i =1 / σU,i 2 ;σ I,i σ U,i Let ΔX represent the noise variance of the I and U data, respectively. pred,i X represents the residual prediction value of the i-th frame. i Let X represent the optimized pose of the i-th frame. I,i X U,i These represent the pose estimates corresponding to IMU calculations and UWB measurements, respectively.

[0094] w I,i Introducing residual prediction value |ΔX pred,i | is to reduce the weighting percentage when the accumulated error of IMU calculation is large, σ I,i σ U,i A larger value indicates greater data noise and a smaller weight.

[0095] The local pose estimate X and the corresponding covariance matrix P are obtained by solving. P is a 6×6 matrix, and the diagonal elements correspond to the variances of the three-dimensional position (x,y,z) and the three-dimensional attitude angle (α,β,γ), respectively. The smaller the variance value, the more reliable the pose estimate in that dimension.

[0096] In step 3, stable features in the environment are extracted and transformed into spatial reference anchors to provide a unified spatial benchmark for the cluster. The reliability of the anchors is verified through multi-machine collaboration to avoid benchmark drift caused by the observation error of a single UAV.

[0097] The specific features for detecting stable environments include: for the filtered feature data, a stability score is calculated based on the spatiotemporal continuity and grayscale contrast of the features, and features with scores greater than a preset stability threshold are retained as candidate stable features.

[0098] The stability scoring logic is as follows: Spatiotemporal continuity score = 1 - feature position offset of 10 consecutive frames / feature scale (score is 1 when offset ≤ 2 pixels), grayscale contrast score = maximum grayscale value of feature region - minimum grayscale value (score is 1 when ≥ 100), the comprehensive score is the product of the two, the preset stability threshold is set to 0.8, and features with a comprehensive score ≥ 0.8 are retained;

[0099] The hash encoding generation specifically includes: inputting candidate stable features using a hash generation method, combining local pose estimation values, calculating the observation angle and relative distance of the candidate stable features, and outputting a triplet containing: hash value, observation angle, and relative distance;

[0100] Example of hash generation method: Using the LSH algorithm, the 128-dimensional feature descriptor is mapped to a 64-bit binary hash value, reducing data transmission and matching overhead;

[0101] Observation angle calculation: Based on the transformation relationship between the UAV body coordinate system and the world coordinate system, combined with the pixel coordinates of the feature in the image frame and camera intrinsic parameters (such as focal length f, principal point coordinates (u0, v0)), the azimuth and pitch angles of the feature relative to the UAV are solved; Relative distance calculation: Based on the local pose estimation value and the image ranging model of the feature (such as the triangulation principle), the preliminary distance is obtained, and then corrected by combining UWB measurement data before output;

[0102] The process of generating temporary anchor points through multi-machine matching and triangulation optimization specifically includes:

[0103] The drone broadcasts triples to neighboring drones within its communication range via wireless communication. Simultaneously, it receives triples broadcast by neighboring drones and performs hash matching on the received triples. When the Hamming distance is less than or equal to a preset matching threshold, it is considered a match. The number N of drones that match the same hash value is counted.

[0104] When N is greater than or equal to a preset threshold, the three-dimensional coordinates of the features are solved by a triangulation algorithm based on the local pose estimates, observation angles and relative distances of multiple UAVs; the consistency scores of the three-dimensional coordinates are scored, and features with scores greater than or equal to a preset consistency threshold are determined as temporary anchor points.

[0105] Triangulation algorithm: Based on the principle of multi-view geometry, the local pose of each UAV is taken as the observation viewpoint. Combined with the observation angle and distance of the features relative to each viewpoint, an overdetermined set of equations is constructed, and the three-dimensional coordinates of the features in the world coordinate system are solved by least squares.

[0106] Consistency score: Calculate the standard deviation of the three-dimensional coordinates obtained by multiple UAVs. For example, the score is 1 when the standard deviation is ≤0.2m, and the score decreases by 0.2 for every 0.1m increase in standard deviation. The preset consistency threshold can be set to 0.8.

[0107] Output the three-dimensional coordinates of the temporary anchor point and the corresponding covariance matrix P1; P1 is a 3×3 matrix, and the diagonal elements correspond to the variance of the three-dimensional coordinates (x,y,z), reflecting the uncertainty of the spatial position of the anchor point.

[0108] In step 4, the reliability of the reference benchmark is ensured by filtering global effective anchor points, and the trust level of local and neighboring machine data is balanced by reasonably allocating fusion weights, so as to achieve the unification of the cluster pose benchmark and solve the problem of benchmark inconsistency in the distributed system.

[0109] The global valid anchor point screening process is as follows: after each UAV receives the temporary anchor point information broadcast by its neighboring UAV, it counts the number of UAVs observing the same hash value anchor point and calculates the coordinate deviation of different UAVs observing the anchor point. When the number of observations is greater than or equal to a preset threshold and the coordinate deviation is less than or equal to a preset deviation threshold, it is determined to be a global valid anchor point.

[0110] The temporary anchor point information refers to the temporary anchor point hash value received by the UAV from a neighboring UAV and the three-dimensional coordinates of the anchor point observed by the UAV itself.

[0111] The calculation of the fusion weights specifically includes: inputting the covariance matrix trace of the neighboring machine, the UWB / RSSI signal strength, and the number of globally effective anchor points observed by the neighboring machine, and allocating the neighboring machine fusion weights w through weighted normalization. neigh ; Satisfying Σw neigh +w local =1; where, w local The local weight is determined by the trace of the local covariance matrix. The trace of the covariance matrix is ​​the sum of the diagonal elements of the matrix and is a simplified index for quantifying the uncertainty of pose estimation. The smaller the trace value, the lower the uncertainty.

[0112] The distributed consensus fusion specifically includes: employing a distributed Kalman consensus algorithm, combined with coordinate constraints and fusion weights of globally valid anchor points, to fuse the poses of the local machine and its neighboring machines: X fuse =w local ·X+Σw neigh ·X neigh The three-dimensional coordinates of the globally effective anchor point are used as a fixed reference. During the fusion process, the pose estimation values ​​of the local and neighboring drones are forced to satisfy the "observation equation based on the anchor point", that is, the world coordinates of the anchor point = drone pose + feature coordinates relative to the drone. This constraint corrects pose drift and ensures that the pose reference of the cluster is unified.

[0113] Among them, X fuse X represents the fused pose estimate. neigh This represents the local pose estimate of the neighboring machines; the fused pose estimate X, consistent with the cluster baseline, is obtained. fuse After fusion, X fuse The trace of the covariance matrix is ​​usually smaller than the trace of the covariance matrix between the local machine and all its neighbors;

[0114] In step 5, the reliability of pose estimation is quantified by uncertainty, and the control strategy is dynamically adjusted to ensure positioning accuracy while taking into account cruise efficiency and energy consumption control.

[0115] The calculation of uncertainty based on the fused pose estimate specifically includes:

[0116] Based on the fused pose estimate, Monte Carlo inference is performed (with a preset sampling number of 1000 times, and a covariance matrix P is added to the fused pose estimate at each sampling). fuse (Ignoring random noise, we obtain 1000 sampled poses, statistically analyze the standard deviation of sampled values ​​in each dimension, output probability distribution, and calculate uncertainty S: S = ((σ x 2 +σ y2 +σ z 2 )+(σ α 2 +σ β 2 +σ γ 2 )) 1 / 2 ;

[0117] Preset thresholds include conservative mode threshold S th1 With recovery mode threshold S th2 S th1 >S th2 A preset control strategy is executed based on the threshold range in which the uncertainty S is located.

[0118] For example: when S≥S th1 At this time: reduce the cruising speed (take the preset conservative speed range), increase the shooting frame rate (take the preset high frequency range), and prioritize the execution of temporary anchor point triangulation tasks;

[0119] When S th1 >S>S th2 Time: Dynamically adjust the shooting frame rate and speed to the preset middle range, taking into account both anchor point detection and cruising efficiency;

[0120] When S≤S th2 Furthermore, when the preset time period is continuously maintained: the normal cruise mode is restored (preset normal speed, low frequency shooting), and the cruise path is executed first;

[0121] It outputs speed control commands, shooting control commands, and task scheduling commands for the UAV, enabling dynamic management and control based on pose reliability.

[0122] A drone swarm management and control system based on an artificial intelligence model, comprising a data preprocessing module, a local pose estimation module, a feature anchor point generation module, and a fusion decision control module;

[0123] The data preprocessing module is used to collect multi-source sensor data from the UAV at a preset period, and obtain preprocessed data and filtered feature data through adaptive preprocessing by an artificial intelligence model. The local pose estimation module is used to perform IMU calculation based on the preprocessed data to obtain the initial pose sequence, and obtain the local pose estimate and covariance matrix by combining residual prediction and weighted least squares optimization. The feature anchor point generation module is used to detect stable environment features and generate hash codes based on the filtered feature data, and generate temporary anchor points through multi-machine matching and triangulation optimization. The fusion decision control module is used to perform global effective anchor point screening and fusion weight calculation based on the local pose estimate and temporary anchor point information to achieve distributed consistent fusion and obtain a fused pose estimate with a consistent cluster baseline. The uncertainty is calculated based on the fused pose estimate, and a preset control strategy is executed according to a preset threshold range.

[0124] The data preprocessing module includes a multi-source acquisition unit and an adaptive processing unit;

[0125] The multi-source acquisition unit is used to acquire multi-source sensor data of the UAV at a preset cycle; the adaptive processing unit is used to call an artificial intelligence model to perform adaptive preprocessing on the acquired data and output denoised data and filtered feature data.

[0126] The local pose estimation module includes a calculation unit, a residual prediction unit, and an optimization unit.

[0127] The calculation unit is used to obtain the initial pose sequence based on the denoised IMU data through IMU brute-force calculation; the residual prediction unit is used to predict the cumulative error residual of the IMU brute-force calculation through a time series prediction model; the optimization unit is used to combine the residual prediction and obtain the local pose estimate and covariance matrix through weighted least squares optimization.

[0128] The feature anchor point generation module includes a feature detection unit, an encoding generation unit, and an anchor point generation unit;

[0129] The feature detection unit is used to detect stable environment features from the filtered feature data; the encoding generation unit is used to generate triple codes containing hash value, observation angle and relative distance for stable features; the anchor point generation unit is used to generate temporary anchor points and their coordinates through hash matching and triangulation optimization between multiple machines.

[0130] The fusion decision control module includes an anchor point screening unit, a weight allocation unit, and a fusion decision unit.

[0131] The anchor point screening unit is used to select globally effective anchor points from temporary anchor points based on multi-machine observation information; the weight allocation unit is used to calculate and allocate the weights required for consistency fusion based on neighboring machine status information; the fusion decision unit is used to perform distributed consistency fusion to obtain the fusion pose consistent with the cluster baseline, and execute a preset control strategy based on its uncertainty.

[0132] In this embodiment, in a complex urban canyon environment (with interference from tall buildings, electromagnetic interference, and sudden changes in lighting), three drones form a cluster to perform road inspection tasks, and the method described in this invention is used to achieve control.

[0133] Data Acquisition and Preprocessing: Three UAVs collect IMU data, image frames, and UWB / RSSI data at preset intervals; the noise suppression model in the artificial intelligence model is called to dynamically identify and filter electromagnetic interference noise in the IMU data, and at the same time, abnormal jump values ​​in UWB / RSSI data are filtered based on the filtering parameters; the feature extraction model is called to extract edge and corner features of image frames, filter blurred frames caused by sudden changes in illumination, and output high-quality preprocessed data and filtered feature data.

[0134] Local pose estimation: Based on the preprocessed IMU data, IMU calculation is performed to obtain the initial pose sequence. The residual prediction value is output through the time series prediction model. The weighted least squares optimization is combined to correct the calculation error and obtain the local pose estimation value and covariance matrix of each mechanism, thus initially reducing the pose drift trend of single mechanism.

[0135] Temporary anchor point generation: Based on the filtered feature data, three drones detect stable corner points (with strong spatiotemporal continuity and high contrast) on the walls of urban buildings and generate hash codes. They calculate the observation angle and distance of the features by combining their local poses and broadcast the triplet of "hash value, observation angle, and distance". The same corner point on the wall is identified by hash matching. After the number of matching is ≥3 (preset threshold), temporary anchor points are generated through triangulation optimization.

[0136] Global fusion and benchmark unification: Each machine receives temporary anchor point information from neighboring machines, counts the number of observations of anchor points with the same hash value (all 3 machines) and the coordinate deviation is ≤ preset threshold, and determines them as global valid anchor points (i.e., spatial benchmarks jointly recognized by the 3 machines); calculates fusion weights based on covariance matrix trace, UWB signal strength and number of valid anchor points, and fuses the poses of the local machine and neighboring machines through distributed Kalman consensus, outputting a fused pose with consistent benchmarks for the 3 machines, eliminating local benchmark drift differences among the machines.

[0137] Dynamic control: The uncertainty of the fused pose is quantified through Monte Carlo inference. If the uncertainty exceeds the conservative mode threshold at a certain moment due to urban canyon obstruction, the system automatically reduces the cruise speed of each aircraft, increases the image acquisition frame rate, and prioritizes the addition of new global effective anchor points. When the uncertainty drops to the normal threshold, normal cruise efficiency is restored.

[0138] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for managing and controlling a drone cluster based on an artificial intelligence model, characterized in that: The method comprises the following steps: Step 1, collecting unmanned aerial vehicle multi-source sensor data at a preset period, and obtaining pretreatment data and screened feature data through artificial intelligence model adaptive preprocessing; Step 2, performing IMU dead reckoning based on the pretreatment data to obtain an initial pose sequence, combining residual prediction and weighted least squares optimization to obtain a local pose estimate value and a covariance matrix; Step 3, based on the screened feature data, detecting stable environment features and generating hash codes, and generating temporary anchor points through multi-machine matching and triangulation optimization; Step 4, based on the local pose estimate value and the temporary anchor point information, performing global effective anchor point screening and fusion weight calculation, realizing distributed consistency fusion, and obtaining a fusion pose estimate value consistent with the cluster reference; Step 5, calculating the uncertainty based on the fusion pose estimate value, and executing a preset control strategy according to a preset threshold interval; In step 3, the detection of stable environment features specifically comprises: for the screened feature data, performing stability scoring based on the spatiotemporal continuity and gray scale contrast of the features, and retaining features with a score greater than a preset stability threshold as candidate stable features; The hash code generation specifically comprises: inputting the candidate stable features into a hash generation method, combining the local pose estimate value, calculating the observation angle and relative distance of the candidate stable features, and outputting a triple including: hash value, observation angle, and relative distance; The generation of temporary anchor points through multi-machine matching and triangulation optimization specifically comprises: The unmanned aerial vehicle broadcasts the triple to neighboring machines within the communication range through wireless communication, simultaneously receives the triples broadcast by the neighboring machines, performs hash matching on the received triples, and determines that the triples are matched when the Hamming distance is less than or equal to a preset matching threshold, and counts the number N of unmanned aerial vehicles matching the same hash value; When N is greater than or equal to a preset number threshold, the three-dimensional coordinates of the features are solved based on the local pose estimate values, observation angles, and relative distances of multiple unmanned aerial vehicles through a triangulation algorithm; features with a consistency score greater than or equal to a preset consistency threshold are determined as temporary anchor points; The three-dimensional coordinates and corresponding covariance matrix of the temporary anchor points are output. 2.The UAV cluster management and control method based on an artificial intelligence model according to claim 1, characterized in that: In step 1, the multi-source sensor data includes IMU data, image frames, UWB / RSSI data; The artificial intelligence model includes a noise suppression model and a feature extraction model; The adaptive preprocessing specifically comprises: The noise suppression model is called to process the IMU data sequence, the filter parameter is dynamically adjusted, and the denoised IMU data I is output; The filter parameter is used to filter abnormal jump values in the UWB / RSSI data, and stable UWB / RSSI data U is output; The abnormal jump value refers to data deviating from the data mean by more than a preset deviation threshold; The pretreatment data includes data I and U; The feature extraction model is called to extract edge and corner features of the image frames, frames with a clarity lower than a preset threshold are filtered, and screened feature data is obtained. 3.The UAV cluster management and control method based on an artificial intelligence model according to claim 2, characterized in that: In step 2, the IMU dead reckoning refers to performing short-term prediction based on the denoised IMU data using an extended Kalman filter algorithm to output an initial pose sequence; the initial pose sequence contains three-dimensional position and three-dimensional attitude angle parameters; The residual prediction is realized by a timing prediction model, and the residual prediction takes continuous frame data of the initial pose sequence as input and outputs predicted values (Δx, Δy, Δz, Δα, Δβ, Δγ) of the cumulative error of the IMU dead reckoning; wherein Δx, Δy and Δz respectively represent residuals in three-dimensional position directions, and Δα, Δβ and Δγ respectively represent residuals in three-dimensional attitude angle directions. The weighted least square optimization takes a preset length of sliding window as a processing unit, and constructs an objective function: min∑ i=1 K [w I,i (X i -X I,i ) 2 +w U,i (X i -X U,i ) 2 ] where K represents the number of data frames in the sliding window; i ∈ {1, 2, …, K}; w I,i represents the weight w of the i-th frame data I I,i = 1 / σ I,i 2 + |ΔX pred,i |, w U,i represents the weight w of the i-th frame data U U,i = 1 / σ U,i 2 ; σ I,i , σ U,i respectively represent the noise variance of I, U data, ΔX pred,i represents the residual prediction value of the i-th frame, X i represents the optimized pose of the i-th frame, X I,i , X U,i respectively represent the pose estimation value corresponding to the IMU dead reckoning and UWB measurement; The local pose estimation value X and the corresponding covariance matrix P are obtained by solving.

4. The method of claim 3, wherein: In step 4, the global effective anchor point screening is specifically: after each unmanned aerial vehicle receives the temporary anchor point information broadcast by the adjacent unmanned aerial vehicle, the number of observation unmanned aerial vehicles of the anchor point with the same hash value is counted, and the coordinate deviation of different unmanned aerial vehicles observing the anchor point is calculated; When the number of observations is greater than or equal to a preset threshold and the coordinate deviation is less than or equal to a preset deviation threshold, the anchor point is determined as a global effective anchor point; The temporary anchor point information refers to the hash value of the temporary anchor point broadcast by the adjacent unmanned aerial vehicle and the three-dimensional coordinates of the anchor point observed by the unmanned aerial vehicle. The fusion weight calculation specifically comprises: inputting a covariance matrix trace of a neighboring machine, UWB / RSSI signal strength, and a global effective anchor point quantity observed by the neighboring machine, and assigning a neighboring machine fusion weight w through weighted normalization neigh ; meets Σw neigh +w local =1; wherein w local represents a local machine weight, which is determined by a local machine covariance matrix trace; The distributed consistency fusion specifically includes: adopting a distributed Kalman consistency algorithm, combining the coordinate constraints of global effective anchor points and fusion weights, and fusing the local and neighbor machine poses: X fuse = w local · X +∑w neigh · X neigh ; wherein X fuse represents the fused pose estimate, X neigh represents the local pose estimate of the neighboring machine; obtaining a cluster reference consistent fused pose estimate X fuse ; In step 5, the calculation of the uncertainty based on the fused pose estimation value specifically includes: Based on the fused pose estimate, a probability distribution is output by Monte Carlo inference, calculating the uncertainty S: S = ((σ x 2 + σ y 2 + σ z 2 + σ α 2 + σ β 2 + σ γ 2 ) 1 / 2 ; The preset threshold value includes a conservative mode threshold value S th1 and a recovery mode threshold value S th2 , S th1 > S th2 ; a preset control strategy is executed based on the threshold interval in which the uncertainty S is located.

5. An unmanned aerial vehicle cluster management and control system based on an artificial intelligence model, applied to the unmanned aerial vehicle cluster management and control method based on an artificial intelligence model in any one of claims 1-4, characterized in that: The system comprises a data preprocessing module, a local pose estimation module, a feature anchor point generation module and a fusion decision control module. The data preprocessing module is configured to collect multi-source sensor data of the unmanned aerial vehicle at a preset period, and to obtain preprocessed data and screened feature data by adaptive preprocessing through an artificial intelligence model. The local pose estimation module is configured to perform IMU dead reckoning based on the preprocessed data to obtain an initial pose sequence, and to obtain a local pose estimation value and a covariance matrix by combining residual prediction and weighted least squares optimization; the feature anchor point generation module is configured to detect stable environmental features and generate hash codes based on the screened feature data, and to generate temporary anchor points by multi-machine matching and triangulation optimization; and the fusion decision control module is configured to perform global effective anchor point screening and fusion weight calculation based on the local pose estimation value and the temporary anchor point information, to realize distributed consistency fusion, and to obtain a fused pose estimation value consistent with the cluster reference. An uncertainty is calculated based on the fused pose estimation value, and a preset control strategy is executed according to a preset threshold interval. The feature anchor point generation module comprises a feature detection unit, an encoding generation unit and an anchor point generation unit. The feature detection unit is configured to detect stable environmental features from the screened feature data; the encoding generation unit is configured to generate a three-tuple encoding comprising a hash value, an observation angle and a relative distance for the stable features; and the anchor point generation unit is configured to generate temporary anchor points and their coordinates by hash matching and triangulation optimization among multiple machines. 6.The UAV cluster management and control system based on an artificial intelligence model according to claim 5, characterized in that: The data preprocessing module comprises a multi-source collection unit and an adaptive processing unit. The multi-source collection unit is configured to collect multi-source sensor data of the unmanned aerial vehicle at a preset period; and the adaptive processing unit is configured to call an artificial intelligence model to perform adaptive preprocessing on the collected data, and to output denoised data and screened feature data.

7. The unmanned aerial vehicle cluster management and control system based on an artificial intelligence model according to claim 6, characterized in that: The local pose estimation module comprises a calculation unit, a residual prediction unit and an optimization unit. The estimation unit is configured to obtain an initial pose sequence by IMU dead reckoning based on the denoised IMU data; the residual prediction unit is configured to predict a cumulative error residual of the IMU dead reckoning by a time series prediction model; and the optimization unit is configured to obtain a local pose estimation value and a covariance matrix by weighted least squares optimization in combination with residual prediction. 8.The UAV cluster management and control system based on an artificial intelligence model according to claim 7, characterized in that: The fusion decision control module comprises an anchor point screening unit, a weight distribution unit and a fusion decision unit. The anchor point screening unit is configured to screen global effective anchor points from temporary anchor points based on multi-machine observation information; the weight distribution unit is configured to calculate and distribute weights required for consistency fusion according to neighbor machine state information; and the fusion decision unit is configured to perform distributed consistency fusion to obtain a fusion pose consistent with a cluster reference, and to perform a preset control strategy based on an uncertainty thereof.

Images