Unmanned aerial vehicle group scheduling method and system based on multi-layer recurrent neural network

Abstract

The invention discloses an unmanned aerial vehicle group scheduling method and system based on a multi-layer recurrent neural network, and relates to the technical field of unmanned aerial vehicle scheduling, and the method comprises the following steps: obtaining a state change sequence of offline unmanned aerial vehicles in a preset time window before offline, and obtaining an initial state vector; when re-online is detected, extracting a first state vector and performing vector fitting matching on the first state vector and the initial state vector; constructing a state fluctuation trajectory, evaluating changes of the state fluctuation trajectory under different task time scales, obtaining state fluctuation gradient distribution, converting the state fluctuation gradient distribution into a continuous state vector sequence, performing vector fitting truth compensation on the continuous state vector sequence and a preset reference state vector cluster, and identifying a weak oscillation region according to a truth compensation result; the state vector of the weak oscillation area is converted into a correction mask, the offline unmanned aerial vehicle input state vector of the multi-layer recurrent neural network is corrected and used for unmanned aerial vehicle group scheduling prediction, and the problem that state information is lagged or abnormal due to the fact that the unmanned aerial vehicles are online again after offline is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of drone scheduling, and more specifically, to a drone swarm scheduling method and system based on a multi-layer recurrent neural network. Background Art

[0002] Drone swarm scheduling involves rationally arranging and optimizing each drone's flight path, task allocation, coordinated actions, and resource utilization during the coordinated execution of complex missions by multiple drones, ensuring efficient and safe completion of the overall mission. Drone swarm scheduling not only involves trajectory planning for individual drones but also requires comprehensive consideration of spatial coordination, obstacle avoidance, communication coordination, and dynamic state changes among drones within the swarm.

[0003] However, when a drone in a swarm goes offline and then comes back online, the status information cannot be continuously updated and synchronized during the offline period, resulting in delayed or abnormal status data transmitted back, which cannot accurately reflect its current true state. This discontinuity and deviation in status information not only affects the execution of the drone's own mission, but also interferes with the coordinated scheduling and overall operational efficiency of the entire drone swarm, increasing the complexity and uncertainty of system scheduling. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a drone swarm scheduling method and system based on a multi-layer recurrent neural network. By correcting the state fluctuations of offline drones and identifying weak oscillation areas, the problem of delayed or abnormal status information caused by drones coming back online after being disconnected is solved.

[0005] To achieve the above object, the present invention provides the following technical solutions: A drone swarm scheduling method based on a multi-layer recurrent neural network includes the following steps: detecting offline drones in a target drone swarm, and obtaining a state change sequence of the offline drones in a preset time window before the offline drones are disconnected to obtain an initial state vector; when the offline drone is detected to be back online, extracting a first state vector and performing vector fitting matching with the initial state vector; constructing a state fluctuation trajectory based on the matching result, and evaluating the changes in the state fluctuation trajectory under different task time scales to obtain a state fluctuation gradient distribution; converting the state fluctuation gradient distribution map into a continuous state vector sequence, performing vector fitting and complementing with a preset reference state vector cluster, and identifying oscillation weak areas based on the complementing result; converting the state vector of the oscillation weak area into a correction mask, correcting the input state vector of the offline drone of the multi-layer recurrent neural network, and using it for drone swarm scheduling prediction.

[0006] In a preferred embodiment, the method of detecting offline drones in a target drone group and obtaining a state change sequence of the offline drones in a preset time window before the offline to obtain an initial state vector is specifically as follows: detecting offline drones in a target drone group and intercepting a state data stream in a preset time window before the offline, wherein the state data stream includes three-dimensional position coordinates, pitch angle, yaw angle and power system load rate; performing time slicing processing on the state data stream to obtain a discrete state sequence with equal intervals; using Lyapunov exponents to filter the discrete state sequence to obtain a number of key frame subsets; obtaining spatial posture parameters of the several key frame subsets and mapping them to the SO(3) rotation group to obtain the initial state vector.

[0007] In a preferred embodiment, when it is detected that the offline drone is back online, the first state vector is extracted and vector fitting matching is performed with the initial state vector, specifically: reconnection response data is collected at the moment the offline drone reconnects, and the first state vector is extracted, and the reconnection response data includes GPS positioning, inertial measurement unit readings and battery transient response values; the first state vector and the initial state vector are aligned along the time axis using a time warping algorithm to construct a DTW distance matrix; the preset time window is divided into several local windows, and the Mahalanobis distance of the local window is calculated along the main diagonal direction of the dynamic time warping distance matrix to obtain an anti-interference distance layer; the first-order derivative sequence of the anti-interference distance layer is extracted to construct a curvature feature matrix; the curvature feature matrix is ​​subjected to a Hessian transform to generate a second-order optimization path constraint; based on the second-order optimization path constraint, the first state vector is subjected to piecewise interpolation compensation to obtain a fitting matching vector.

[0008] In a preferred embodiment, the changes in state fluctuation trajectories at different task time scales are evaluated to obtain a state fluctuation gradient distribution, specifically: multiple rounds of Hamiltonian Monte Carlo sampling are performed on the state fluctuation trajectories at different task time scales to obtain a candidate trajectory probability distribution set; a greedy algorithm is used to iteratively screen the selected trajectory probability distribution set to obtain an optimal sampling path; eigenvalues ​​of preset different task stages are extracted along the optimal sampling path, and the eigenvalues ​​are multi-dimensionally fused to obtain a state fluctuation gradient distribution tensor; the state fluctuation gradient distribution tensor is projected to construct a state fluctuation gradient distribution map.

[0009] In a preferred embodiment, the state fluctuation gradient distribution map is converted into a continuous state vector sequence, specifically: Gaussian sampling is performed on the state fluctuation gradient distribution map to obtain a discrete state point set; a sampling interpolation algorithm is used to reconstruct the discrete point set into a continuous vector field, and the continuous vector field is decomposed to obtain decomposed components; the decomposed components are encoded into a continuous state vector sequence in a time series.

[0010] In a preferred embodiment, the vector fitting and correction with the preset reference state vector cluster is specifically as follows: extracting the reference state vector cluster from the historical task library and constructing a vector manifold space; calculating the shortest path from the continuous state vector sequence to the reference state vector cluster to obtain the tangent vector field; based on the tangent vector field, decomposing the continuous state vector sequence into normal components and tangential components; projecting the normal components into the vector manifold space to construct a residual vector; and using time axis sliding convolution to perform phase compensation on the residual vector to obtain a corrected state sequence.

[0011] In a preferred embodiment, the state vector of the oscillation weak area is converted into a correction mask, specifically: extracting the vertex state vector of the oscillation weak area; performing a graph convolution operation on the vertex vector to obtain a topological feature matrix; performing a Hadamard product operation on the topological feature matrix and a preset stability basis matrix, and compressing it using singular value decomposition to obtain a binary correction mask matrix.

[0012] In a preferred embodiment, the input state vector of the offline drone of the multi-layer recurrent neural network is corrected, specifically: a Hadamard product operation is performed on the binary correction mask matrix and the original state vector of the offline drone to obtain a corrected input state vector; the corrected input state vector is input into the time convolution module of the multi-layer recurrent neural network, and the optimal scheduling path of the cluster is predicted through the gated recurrent unit.

[0013] The technical effects and advantages of the drone swarm scheduling method and system based on a multi-layer recurrent neural network of the present invention are as follows: 1. The present invention achieves high-precision fitting and matching of the reconnection state and the initial state by extracting key frames and mapping spatial postures of the state sequence before disconnection, combined with dynamic time warping and second-order optimization path constraints. Furthermore, it uses state fluctuation gradient evaluation and frequency domain analysis under multiple time scales to accurately identify and compensate for weak areas of state oscillation, uses graph convolution and singular value decomposition to generate correction masks, dynamically corrects the input state vector, and finally performs intelligent scheduling prediction through a deep recurrent neural network. The overall solution significantly improves the scheduling stability and task execution efficiency of drone swarms in complex environments, enhances the system's robustness and adaptability to disconnection and reconnection state fluctuations, and ensures the continuous and reliable collaborative operation of drone swarms. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is a flow chart of a drone swarm scheduling method based on a multi-layer recurrent neural network according to the present invention.

[0015] Figure 2 This is a structural schematic diagram of a drone swarm scheduling system based on a multi-layer recurrent neural network in the present invention. DETAILED DESCRIPTION

[0016] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0017] Example 1, Figure 1 The present invention provides a method for scheduling drone swarms based on a multi-layer recurrent neural network, comprising the following steps: S1, detect the offline drones in the target drone group, and obtain the state change sequence of the offline drones in the preset time window before the offline, and obtain the initial state vector; In this example, the offline drones in the target drone group are detected, and the state change sequence of the offline drones in the preset time window before the offline is obtained to obtain the initial state vector, which is specifically: Detecting offline drones in the target drone group and intercepting a state data stream in a preset time window before the disconnection, wherein the state data stream includes three-dimensional position coordinates, pitch angle, yaw angle, and power system load rate; Perform time slicing on the state data stream to obtain a discrete state sequence with equal intervals; The discrete state sequence is screened using Lyapunov exponent to obtain several key frame subsets; The spatial posture parameters of several key frame subsets are obtained and mapped to the SO(3) rotation group to obtain the initial state vector.

[0018] It's important to note that during swarm operation, the system monitors each drone's communication status and data transmission in real time. Each drone regularly reports its status data, including its position, attitude, and system health parameters, at a set frequency. If a drone fails to upload data as scheduled for multiple consecutive cycles, or if its communication signal is interrupted for longer than a preset threshold (e.g., one second), the system deems it offline and marks it as disconnected.

[0019] Furthermore, after detecting a disconnected drone, the system will retrace its state change data for the period before the disconnection and divide this data into a series of discrete frame sequences at fixed time intervals. Subsequently, to extract the key frames that best represent the drone's dynamic behavior, the system introduces Lyapunov exponents for analysis. The Lyapunov exponent is a mathematical tool that measures a system's sensitivity to initial conditions and can determine whether the system's state exhibits unstable or chaotic trends. By calculating the rate of state change and the degree of disturbance amplification between each frame, the system can identify frames with significant state changes and a significant impact on future states, and retain them as key frames.

[0020] Finally, after selecting a subset of keyframes, the system focuses on extracting parameters related to spatial position and attitude within these keyframes to construct a more physically meaningful initial state vector. Spatial attitude parameters primarily include the 3D position coordinates corresponding to each frame—the drone's specific position in space—as well as attitude angles, typically including pitch, yaw, and roll angles, which collectively describe the drone's orientation and rotation in space.

[0021] To enhance the ability to process rotation information, these parameters are uniformly mapped into a three-dimensional rotation group space, commonly referred to as the SO(3) rotation group, a mathematical structure used to describe arbitrary three-dimensional rotational states. The mapping process essentially converts discrete spatial posture information into a unified vector format more suitable for neural network processing, so that each keyframe is represented as a standard rotational state. These states are then spliced ​​or fused in chronological order to ultimately form a complete initial state vector, laying the foundation for subsequent state matching and group scheduling prediction.

[0022] S2, when it is detected that the offline drone is back online, the first state vector is extracted and matched with the initial state vector by vector fitting; In this example, when the offline drone is detected to be back online, the first state vector is extracted and vector fitting is performed with the initial state vector, specifically: Collect reconnection response data and extract the first state vector when the disconnected drone reconnects. The reconnection response data includes GPS positioning, inertial measurement unit readings, and battery transient response values. The time warping algorithm is used to align the first state vector with the initial state vector along the time axis to construct the DTW distance matrix; The preset time window is divided into several local windows, and the Mahalanobis distance of the local windows is calculated along the main diagonal direction of the dynamic time warping distance matrix to obtain the interference-resistant distance layer; Extract the first-order derivative sequence of the anti-interference distance layer and construct the curvature feature matrix; Perform Hessian transformation on the curvature feature matrix to generate second-order optimization path constraints; Based on the second-order optimization path constraint, the first state vector is piecewise interpolated and compensated to obtain a fitting matching vector.

[0023] It's important to note that when a disconnected drone (e.g., UAV-05) reestablishes a connection with the control system, the system immediately collects sensor data from the moment it comes back online, forming a first state vector. This state data, derived from multiple sensor channels, primarily includes the current GPS positioning information (used to obtain the drone's three-dimensional coordinates), real-time readings from the inertial measurement unit (including acceleration and angular velocity), and the battery's instantaneous current, voltage, and temperature response (reflecting system load and energy status). This data is combined into a complete six-dimensional or multi-dimensional state vector, reflecting the drone's overall operational status upon reconnection.

[0024] To determine whether the drone's current state matches its pre-disconnect status, the system then temporally aligns and matches the first state vector at the moment of reconnection with the previously constructed initial state vector. Because two segments of data often lack strict temporal consistency, a dynamic time warping algorithm is employed to synchronize the time axes of the two vector sequences, establishing a nonlinear distance matrix and constructing a multidimensional dynamic time warping distance map.

[0025] To improve robustness against non-stationary state changes, the system divides the preset time window into several smaller local windows. Within each small window, the distance values ​​in the dynamic time warping matrix are extracted along the main diagonal direction, and the Mahalanobis distance is calculated as the interference-resistant metric for matching. This method effectively identifies the local similarity between two state vectors in each time period.

[0026] The system then calculates the first-order derivative of the interference-resistant distance result sequence to form a curvature characteristic matrix reflecting the rate of state change. This matrix is ​​further fed into the Hessian transform module to generate a second-order optimization constraint structure reflecting the local extrema and change trends of the path. Using this constraint information, the system performs piecewise interpolation compensation based on the original first state vector, correcting for missing or distorted state segments and ultimately forming a matching vector that closely matches the initial state vector.

[0027] S3, constructs the state fluctuation trajectory based on the matching results, and evaluates the changes of the state fluctuation trajectory at different task time scales to obtain the state fluctuation gradient distribution; In this example, the state fluctuation trajectory is constructed based on the matching results, specifically: Taking the fitted matching vector as the starting point, the analysis windows are divided into three parts: long, medium and short according to the task time scale; Extract trajectory segments of similar scenes in historical mission cycles within a long window to construct a background fluctuation baseline; Calculate the Mahalanobis distance between the current trajectory and the background fluctuation baseline in the middle window and mark the deviation threshold interval; Gaussian process regression is used to predict the instantaneous fluctuation trend within a short window, and the outputs of three windows are fused to generate the spatiotemporally coupled state fluctuation trajectory.

[0028] In this example, the changes in state fluctuation trajectories at different task time scales are evaluated to obtain the state fluctuation gradient distribution, specifically: At different task time scales, multiple rounds of Hamiltonian Monte Carlo sampling are performed on the state fluctuation trajectory to obtain a probability distribution set of candidate trajectories; A greedy algorithm is used to iteratively screen the probability distribution set of selected trajectories to obtain the optimal sampling path; Extract the eigenvalues ​​of different preset task stages along the optimal sampling path, and perform multi-dimensional fusion of the eigenvalues ​​to obtain the state fluctuation gradient distribution tensor; Project the state fluctuation gradient distribution tensor to construct the state fluctuation gradient distribution map.

[0029] It's important to note that after completing the matching of the initial state vector to the state of the drone after it re-enters the network in the previous stage, the system constructs a state fluctuation trajectory for the drone based on the matching results. This trajectory describes the drone's state changes from before and during the offline period to when it re-enters the network, reflecting the continuous evolution of its spatial attitude, dynamic response, and other aspects.

[0030] To evaluate the evolution of this state fluctuation trajectory at different mission timescales, we first defined multiple mission-related timescales, such as the short-term stability control scale (measured in seconds), the medium-term coordinated flight scale (measured in tens of seconds), and the long-term mission planning scale (measured in minutes). For each timescale, we applied Hamiltonian Monte Carlo sampling to perform multiple rounds of perturbation sampling on the original trajectory, generating a large set of physically feasible candidate trajectories. Each candidate trajectory probabilistically reflects the state evolution trend that may occur within the current timescale.

[0031] A greedy search algorithm is introduced to iteratively screen candidate trajectory probability distributions. Starting from the initial state, the algorithm prioritizes trajectory nodes with the highest confidence and optimal state continuity, gradually constructing an optimal sampling trajectory path at the current time scale. This path is considered to be the most likely state fluctuation evolution curve at that time scale.

[0032] Furthermore, along the optimal sampling path, state feature values ​​of key nodes are extracted at different mission stages (such as takeoff, formation adjustment, target tracking, and return). These feature values ​​include indicators such as attitude angle change rate, velocity mutation, load response change, and position offset, reflecting the UAV's dynamic stability and collaborative adaptability during each mission stage. Using multi-dimensional fusion technology, these features are integrated into a unified state fluctuation gradient distribution tensor.

[0033] Finally, this tensor is converted into a two- or three-dimensional state fluctuation gradient distribution map through a projection mapping operation. This distribution map clearly shows the intensity and evolution direction of the UAV state fluctuations over time at different mission scales, which can be used to identify potential unstable points, high-risk behavior patterns, or coordination disorders.

[0034] S4, converting the state fluctuation gradient distribution map into a continuous state vector sequence, and performing vector fitting and correction with the preset reference state vector cluster, and identifying the oscillation weak area based on the correction result; In this example, the state fluctuation gradient distribution map is converted into a continuous state vector sequence, specifically: Perform Gaussian sampling on the state fluctuation gradient distribution map to obtain a discrete state point set; The sampling interpolation algorithm reconstructs the discrete point set into a continuous vector field and decomposes the continuous vector field to obtain the decomposed components; The decomposed components are encoded into a continuous state vector sequence in time series.

[0035] It should be noted that Gaussian sampling is first performed on the generated state fluctuation gradient distribution map. During this process, the system performs weighted sampling based on the gradient intensity distribution area, prioritizing high-gradient areas for dense sampling and sparse sampling in low-gradient areas. The sampling results form a set of discrete state points, each representing the amplitude and direction of the drone's state fluctuation at a specific time point and mission phase.

[0036] Furthermore, an interpolation reconstruction algorithm is used to restore these discrete state point sets to a time-continuous vector field. This interpolation process not only maintains the relative trends between the data but also reasonably fills in some missing or sparse regions. The interpolated vector field can simulate the complete process of continuous state change on the time axis.

[0037] This continuous vector field is then decomposed, dividing the entire vector field into components based on rate of change, directional characteristics, or state dimensions. For example, the vector field can be decomposed into linear velocity components, angular velocity components, and power load components, each reflecting the dynamic response characteristics of the drone based on various state variables. Finally, these decomposed state components are uniformly encoded in chronological order, converting them into a sequence of continuous state vectors with time labels. Each state vector corresponds to a specific point in time and contains multiple dimensions of physical state information.

[0038] In this example, vector fitting and correction are performed with the preset reference state vector cluster, specifically: Extract reference state vector clusters from the historical task library and construct a vector manifold space; Calculate the shortest path from the continuous state vector sequence to the reference state vector cluster to obtain the tangent vector field; Based on the tangent vector field, the continuous state vector sequence is decomposed into normal components and tangential components; Project the normal component into the vector manifold space to construct the residual vector; The time axis sliding convolution is used to perform phase compensation on the residual vector to obtain the compensated state sequence.

[0039] It should be noted that flight records with similar environments, flight types, or execution phases to the current UAV mission are first selected from the historical mission library. The state vectors in these records are uniformly preprocessed to form a diverse cluster of reference state vectors, which are then embedded in a high-dimensional state space in a manifold form. This manifold space fully preserves the local geometric relationships in the state sequence and the mission evolution trends.

[0040] Subsequently, the shortest path from the current continuous state vector sequence to each reference state point in the manifold space is calculated, thereby constructing a tangent vector field composed of tangent vectors. This vector field reflects the local fit of the current state sequence relative to the evolution direction of the reference state and provides a geometric basis for the next step of decomposition and correction.

[0041] Next, the continuous state vector sequence is decomposed in the tangent vector field to obtain tangential and normal components. The tangential component represents the natural evolution of the current state sequence along the reference task trajectory, while the normal component represents the abnormal perturbation from the reference trajectory. The system further projects the normal component into the manifold space and extracts the deviated residual vector.

[0042] To correct these deviations, a time-axis sliding convolution mechanism is used to perform phase compensation on the residual vector. This mechanism detects fluctuations in state deviations within a local time window and gradually corrects them through smoothing convolution and phase alignment strategies, avoiding transient mutations or physical violations.

[0043] In this example, the weak oscillation areas are identified based on the correction results, specifically: Perform Fourier descriptor decomposition on the padded state sequence and extract the frequency domain energy spectrum; Detecting resonant frequency bands and harmonic components exceeding a preset threshold in the energy spectrum; Reversely map the resonance frequency band to the spatiotemporal domain of the state vector to locate the phase conflict interval; The regional oscillation intensity coefficient is calculated based on the amplitude of the harmonic component, and the areas with a screening coefficient greater than the critical value are considered to be oscillation weak areas.

[0044] It should be noted that the padded state vector sequence is first subjected to Fourier descriptor decomposition. This processing method not only extracts the main components of the original state sequence in the frequency domain, but also preserves the periodic characteristics of the state changes over time. During the decomposition process, the system treats the state vector as a multidimensional time series signal, extracts the spectral structure of each dimension through Fourier transform, and constructs a complete frequency domain energy spectrum.

[0045] The energy spectrum is then analyzed, with particular attention paid to frequency bands with prominent energy. By setting an empirical threshold (e.g., twice the average spectral energy), the system automatically detects resonant frequency bands in the energy spectrum that exceed this threshold, along with their primary harmonic components. These high-energy frequency bands often correspond to strong periodic fluctuations or coordinated imbalances in the state signal.

[0046] Subsequently, these identified resonant frequency bands are reprojected onto the time axis and spatial dimension of the original state vector through frequency-time inverse mapping, thereby locating the specific state space-time segment. These areas usually show strong fluctuations in posture, speed or power load in a short period of time, indicating the presence of potential phase conflicts or spontaneous oscillations of the system. To quantify the intensity of these abnormal areas, the system further calculates an oscillation intensity coefficient for each located area based on the amplitude of the extracted harmonic components. This coefficient takes into account factors such as frequency amplitude, duration and oscillation concentration. If the oscillation intensity coefficient of a certain space-time area exceeds the preset critical threshold, the system marks the area as an oscillation weak area.

[0047] S5, converts the state vector of the weak oscillation area into a correction mask, corrects the input state vector of the offline drone of the multi-layer recurrent neural network and uses it for drone swarm scheduling prediction.

[0048] In this example, the state vector of the weak oscillation area is converted into a correction mask, specifically: Extract the vertex state vector of the weak oscillation region; Perform graph convolution operation on vertex vectors to obtain topological feature matrix; The topological feature matrix is ​​subjected to Hadamard product operation with the preset stability basis matrix, and is compressed using singular value decomposition to obtain a binary modified mask matrix.

[0049] It should be noted that, first, representative nodes are extracted from each weak region of oscillation. These vertex state vectors serve as a high-dimensional representation of the local region's state characteristics. These vertex state vectors reflect the region's transient extreme values ​​or trend reversal points when the oscillation occurs, and are structurally representative.

[0050] Next, based on a graph structure construction mechanism, the similarity relationships between these vertex vectors are organized into an undirected graph structure, and a graph convolution operation is performed. During this process, the system uses a graph convolutional neural network to extract the topological dependency features between these state points, generating a topological feature matrix that reflects the structural connections and dynamic resonance trends. This matrix reveals the coordinated oscillation patterns and propagation paths between different state dimensions in the oscillation region.

[0051] The resulting topological feature matrix is ​​then subjected to a Hadamard product operation (element-wise multiplication) with the previously trained stability benchmark matrix. This operation preserves the anomaly characteristics in the oscillation region while incorporating the empirical model of system stability to suppress false anomalies. The resulting product matrix is ​​the compression correction mask containing structural information.

[0052] To reduce the computational complexity of subsequent models and extract the core perturbation components, a singular value decomposition (SVD) process is performed on the correction matrix. By retaining the first few singular value components, the system effectively compresses the correction information and removes low-weight redundant components. The final output is a binary correction mask matrix, where each element represents whether the state of the corresponding dimension requires dynamic correction (1 indicates correction is required, 0 indicates no adjustment). This mask will be directly applied to the state input of the multi-layer recurrent neural network to control the activation, inhibition, and correction of the state vector dimensions, thereby improving the network's robustness and convergence when processing unstable state inputs.

[0053] In this example, the input state vector of the offline drone of the multi-layer recurrent neural network is corrected as follows: Perform a Hadamard product operation on the binary correction mask matrix and the original state vector of the offline drone to obtain the corrected input state vector; The corrected input state vector is input into the temporal convolution module of the multi-layer recurrent neural network, and the optimal scheduling path of the cluster is predicted through the gated recurrent unit.

[0054] It should be noted that the original input state vector, after incorporating the offline drone, is first processed dimension by dimension and then subjected to a Hadamard product (element-wise multiplication) with the correction mask matrix. For each oscillating dimension, the original state value is masked or replaced with a smoothed value after noise reduction and interpolation to eliminate the impact of high-frequency jitter on network judgment. The state values ​​of uncorrected dimensions remain unchanged. The resulting corrected input state vector is more stable and predictable.

[0055] During actual operation, the multi-layer recurrent neural network receives state vectors from all drones (including offline drones after correction), and through time series modeling and spatial aggregation operations, outputs the optimal scheduling path for the entire human-machine cluster at the current moment, including position guidance, collaborative targets, speed adjustment suggestions, etc. for each drone.

[0056] In this example, the multi-layer recurrent neural network is: The multi-layer recursive neural network includes an input layer, a temporal convolution layer, an attention mechanism layer, a bidirectional GRU layer and an output layer, specifically: The input layer receives the corrected input state vector; The temporal convolution layer uses dilated convolution kernels to extract multi-scale temporal features; Attention mechanism layer weighted fusion of key state parameters; The bidirectional GRU layer builds a state transition probability model; The output layer generates a scheduling instruction set containing heading angle, speed and formation parameters.

[0057] Example 2, Figure 2 The present invention provides a drone swarm scheduling system based on a multi-layer recursive neural network, which includes a disconnection detection module, a vector matching module, a gradient distribution module, a weak identification module, and a network correction module: The offline detection module is used to detect offline drones in the target drone group and obtain the state change sequence of the offline drones in the preset time window before the offline to obtain the initial state vector; The vector matching module is used to extract the first state vector and perform vector fitting matching on the initial state vector when detecting that the offline drone is back online; The gradient distribution module is used to construct the state fluctuation trajectory based on the matching results, and evaluate the changes in the state fluctuation trajectory at different task time scales to obtain the state fluctuation gradient distribution; Weakness identification module, used to convert the state fluctuation gradient distribution map into a continuous state vector sequence, perform vector fitting and correction with the preset reference state vector cluster, and identify the oscillation weak area based on the correction result; The network correction module is used to convert the state vector of the weak shock area into a correction mask, correct the input state vector of the offline drone of the multi-layer recurrent neural network, and use it for drone group scheduling prediction.

[0058] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0059] The above embodiments may be implemented in whole or in part through software, hardware, firmware or any other combination. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product.

[0060] Those skilled in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0061] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0062] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0063] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A drone swarm scheduling method based on a multi-layer recurrent neural network, characterized in that: The following steps are involved: Detect offline drones in the target drone group and obtain the state change sequence of the offline drones in the preset time window before the offline drones are disconnected to obtain the initial state vector; When it is detected that the offline drone is back online, the first state vector is extracted and matched with the initial state vector by vector fitting; The state fluctuation trajectory is constructed based on the matching results, and the changes of the state fluctuation trajectory at different task time scales are evaluated to obtain the state fluctuation gradient distribution; The state fluctuation gradient distribution map is converted into a continuous state vector sequence, and vector fitting and correction are performed with the preset reference state vector cluster, and the weak oscillation area is identified based on the correction results; The state vector of the weak oscillation area is converted into a correction mask, and the input state vector of the offline drone of the multi-layer recurrent neural network is corrected and used for drone swarm scheduling prediction.

2. The method for scheduling drone swarms based on a multi-layer recurrent neural network according to claim 1, characterized in that: The method detects offline drones in the target drone group and obtains the state change sequence of the offline drones in the preset time window before the offline, and obtains the initial state vector, which is specifically: Detecting offline drones in the target drone group and intercepting a state data stream in a preset time window before the disconnection, wherein the state data stream includes three-dimensional position coordinates, pitch angle, yaw angle, and power system load rate; Perform time slicing on the state data stream to obtain a discrete state sequence with equal intervals; The discrete state sequence is screened using Lyapunov exponent to obtain several key frame subsets; The spatial posture parameters of several key frame subsets are obtained and mapped to the SO(3) rotation group to obtain the initial state vector.

3. The method for scheduling drone swarms based on a multi-layer recurrent neural network according to claim 2, characterized in that: When it is detected that the offline drone is back online, the first state vector is extracted and vector fitting matching is performed on the initial state vector, specifically: Collect reconnection response data and extract the first state vector when the disconnected drone reconnects. The reconnection response data includes GPS positioning, inertial measurement unit readings, and battery transient response values. The time warping algorithm is used to align the first state vector with the initial state vector along the time axis to construct the DTW distance matrix; The preset time window is divided into several local windows, and the Mahalanobis distance of the local windows is calculated along the main diagonal direction of the dynamic time warping distance matrix to obtain the interference-resistant distance layer; Extract the first-order derivative sequence of the anti-interference distance layer and construct the curvature feature matrix; Perform Hessian transformation on the curvature feature matrix to generate second-order optimization path constraints; Based on the second-order optimization path constraint, the first state vector is piecewise interpolated and compensated to obtain a fitting matching vector.

4. The method for dispatching drone swarms based on a multi-layer recurrent neural network according to claim 3, characterized in that: The changes in the state fluctuation trajectory at different task time scales are evaluated to obtain the state fluctuation gradient distribution, specifically: At different task time scales, multiple rounds of Hamiltonian Monte Carlo sampling are performed on the state fluctuation trajectory to obtain a probability distribution set of candidate trajectories; A greedy algorithm is used to iteratively screen the probability distribution set of selected trajectories to obtain the optimal sampling path; Extract the eigenvalues ​​of different preset task stages along the optimal sampling path, and perform multi-dimensional fusion of the eigenvalues ​​to obtain the state fluctuation gradient distribution tensor; Project the state fluctuation gradient distribution tensor to construct the state fluctuation gradient distribution map.

5. The method for scheduling drone swarms based on a multi-layer recurrent neural network according to claim 4, characterized in that: The state fluctuation gradient distribution map is converted into a continuous state vector sequence, specifically: Perform Gaussian sampling on the state fluctuation gradient distribution map to obtain a discrete state point set; The sampling interpolation algorithm reconstructs the discrete point set into a continuous vector field and decomposes the continuous vector field to obtain the decomposed components; The decomposed components are encoded into a continuous state vector sequence in time series.

6. The method for dispatching drone swarms based on a multi-layer recurrent neural network according to claim 5, characterized in that: The vector fitting and correction with the preset reference state vector cluster is specifically as follows: Extract reference state vector clusters from the historical task library and construct a vector manifold space; Calculate the shortest path from the continuous state vector sequence to the reference state vector cluster to obtain the tangent vector field; Based on the tangent vector field, the continuous state vector sequence is decomposed into normal components and tangential components; Project the normal component into the vector manifold space to construct the residual vector; The time axis sliding convolution is used to perform phase compensation on the residual vector to obtain the compensated state sequence.

7. The method for dispatching drone swarms based on a multi-layer recurrent neural network according to claim 6, characterized in that: The state vector of the weak shock area is converted into a correction mask, specifically: Extract the vertex state vector of the weak oscillation region; Perform graph convolution operation on vertex vectors to obtain topological feature matrix; The topological feature matrix is ​​subjected to Hadamard product operation with the preset stability basis matrix, and is compressed using singular value decomposition to obtain a binary modified mask matrix.

8. The method for dispatching drone swarms based on a multi-layer recurrent neural network according to claim 7, characterized in that: The correction of the offline drone input state vector of the multi-layer recurrent neural network is specifically as follows: Perform a Hadamard product operation on the binary correction mask matrix and the original state vector of the offline drone to obtain the corrected input state vector; The corrected input state vector is input into the temporal convolution module of the multi-layer recurrent neural network, and the optimal scheduling path of the cluster is predicted through the gated recurrent unit.

9. A drone swarm scheduling system based on a multi-layer recurrent neural network, applied to a drone swarm scheduling method based on a multi-layer recurrent neural network according to any one of claims 1 to 8, characterized in that: It includes offline detection module, vector matching module, gradient distribution module, weak identification module and network correction module: The offline detection module is used to detect offline drones in the target drone group and obtain the state change sequence of the offline drones in the preset time window before the offline to obtain the initial state vector; The vector matching module is used to extract the first state vector and perform vector fitting matching on the initial state vector when detecting that the offline drone is back online; The gradient distribution module is used to construct the state fluctuation trajectory based on the matching results, and evaluate the changes in the state fluctuation trajectory at different task time scales to obtain the state fluctuation gradient distribution; Weakness identification module, used to convert the state fluctuation gradient distribution map into a continuous state vector sequence, perform vector fitting and correction with the preset reference state vector cluster, and identify the oscillation weak area based on the correction result; The network correction module is used to convert the state vector of the weak shock area into a correction mask, correct the input state vector of the offline drone of the multi-layer recurrent neural network, and use it for drone group scheduling prediction.

Images