A method and system for scheduling a UAV swarm based on a multi-layer recurrent neural network

Abstract

The application discloses a kind of based on multilayer recurrent neural network's unmanned aerial vehicle group scheduling method and system, it is related to unmanned aerial vehicle scheduling technical field, including the following steps: obtaining the state change sequence of offline unmanned aerial vehicle in the preset time window before offline, obtain initial state vector;When detecting re-online, extract first state vector and carry out vector fitting matching with initial state vector;State fluctuation trajectory is constructed, and the change of state fluctuation trajectory under different task time scales is evaluated, state fluctuation gradient distribution is obtained and is converted into continuous state vector sequence, and with the reference state vector cluster of preset is fitted to fill in truth, and according to the recognition of shock weak area of fill in truth result;The state vector of shock weak area is converted into correction mask, the input state vector of multilayer recurrent neural network's offline unmanned aerial vehicle is corrected and used for unmanned aerial vehicle group scheduling prediction, solve the problem that state information is lagged behind or abnormal after unmanned aerial vehicle offline re-online.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle scheduling, and more particularly to a multi-layer recurrent neural network-based unmanned aerial vehicle group scheduling method and system. BACKGROUND

[0002] Unmanned aerial vehicle group scheduling refers to reasonably arranging and optimizing the flight path, task allocation, cooperative action and resource utilization of each unmanned aerial vehicle in the process of multiple unmanned aerial vehicles cooperating to perform complex tasks, so as to achieve efficient completion and safety guarantee of the overall task. Unmanned aerial vehicle group scheduling not only involves trajectory planning of a single unmanned aerial vehicle, but also needs to comprehensively consider spatial cooperation, obstacle avoidance, communication coordination and dynamic state changes among unmanned aerial vehicles in the cluster.

[0003] However, when a dropped unmanned aerial vehicle appears in the unmanned aerial vehicle group and is reconnected after being dropped, since the state information cannot be continuously updated and synchronized during the drop, the state data returned by the dropped unmanned aerial vehicle is lagging behind or abnormal, and cannot accurately reflect the current real state. The discontinuity and deviation of such state information not only affect the task execution of the unmanned aerial vehicle itself, but also interfere with the cooperative scheduling and overall operation efficiency of the entire unmanned aerial vehicle group, increasing the complexity and uncertainty of system scheduling. SUMMARY

[0004] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide a multi-layer recurrent neural network-based unmanned aerial vehicle group scheduling method and system, which solves the problem of state information lagging behind or being abnormal caused by the reconnection of a dropped unmanned aerial vehicle after being dropped, by supplementing the state fluctuation of the dropped unmanned aerial vehicle and identifying the weak shock area.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical scheme:

[0006] A multi-layer recurrent neural network-based unmanned aerial vehicle group scheduling method, comprising the following steps: detecting a dropped unmanned aerial vehicle in a target unmanned aerial vehicle group, and obtaining a state change sequence of the dropped unmanned aerial vehicle in a preset time window before being dropped, to obtain an initial state vector; when the dropped unmanned aerial vehicle is detected to be reconnected, a first state vector is extracted and vector fitting matching is performed with the initial state vector; a state fluctuation trajectory is constructed according to the matching result, and the change of the state fluctuation trajectory under different task time scales is evaluated to obtain a state fluctuation gradient distribution; the state fluctuation gradient distribution map is converted into a continuous state vector sequence, and vector fitting supplement is performed with a preset reference state vector cluster, and a weak shock area is identified according to the supplement result; the state vector of the weak shock area is converted into a correction mask, and the input state vector of the multi-layer recurrent neural network of the dropped unmanned aerial vehicle is corrected and used for unmanned aerial vehicle group scheduling prediction.

[0007] In a preferred embodiment, the dropped unmanned aerial vehicle in the target unmanned aerial vehicle group is detected, and a state change sequence of the dropped unmanned aerial vehicle in a preset time window before dropping is obtained to obtain an initial state vector. Specifically, the dropped unmanned aerial vehicle in the target unmanned aerial vehicle group is detected, and state data streams in a preset time window before dropping are intercepted, the state data streams including three-dimensional position coordinates, pitch angles, yaw angles, and power system load rates; the state data streams are subjected to time slicing processing to obtain discrete state sequences at equal intervals; the discrete state sequences are screened by using Lyapunov indexes to obtain a plurality of key frame subsets; spatial attitude parameters of the plurality of key frame subsets are obtained and mapped to an SO(3) rotation group to obtain the initial state vector.

[0008] In a preferred embodiment, when the dropped unmanned aerial vehicle is detected to be reconnected, a first state vector is extracted and vector fitting matching is performed with the initial state vector. Specifically, reconnection response data are collected at the moment of reconnection of the dropped unmanned aerial vehicle, and a first state vector is extracted, the reconnection response data including GPS positioning, inertial measurement unit readings, and battery transient response values; the first state vector and the initial state vector are aligned according to a time axis by using a time warping algorithm to construct a DTW distance matrix; a preset time window is divided into a plurality of local windows, and Mahalanobis distances of the local windows are calculated along a main diagonal direction of the dynamic time warping distance matrix to obtain an anti-interference distance layer; a first derivative sequence of the anti-interference distance layer is extracted to construct a curvature feature matrix; the curvature feature matrix is subjected to Hessian transformation 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.

[0009] In a preferred embodiment, the change of the state fluctuation trajectory under different task time scales is evaluated to obtain a state fluctuation gradient distribution. Specifically, under different task time scales, the state fluctuation trajectory is subjected to multi-round Hamiltonian Monte Carlo sampling to obtain a candidate trajectory probability distribution set; the candidate trajectory probability distribution set is iteratively screened by using a greedy algorithm to obtain an optimal sampling path; feature values of preset different task stages are extracted along the optimal sampling path, and the feature values 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.

[0010] In a preferred embodiment, the state fluctuation gradient distribution map is converted into a continuous state vector sequence. Specifically, the state fluctuation gradient distribution map is subjected to Gaussian sampling to obtain a discrete state point set; the discrete point set is reconstructed into a continuous vector field by using a sampling interpolation algorithm, 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 sequence.

[0011] In a preferred embodiment, the vector fitting complementation with the preset reference state vector cluster is specifically: extracting the reference state vector cluster from the historical task library to construct the vector manifold space; calculating the shortest path of the continuous state vector sequence to the reference state vector cluster to obtain the tangent vector field; decomposing the continuous state vector sequence into normal component and tangent component based on the tangent vector field; projecting the normal component into the vector manifold space to construct the residual vector; and performing phase compensation on the residual vector by time axis sliding convolution to obtain the complemented state sequence.

[0012] 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 graph convolution operation on the vertex vector to obtain a topological feature matrix; performing Hadamard product operation on the topological feature matrix and a preset stability basis matrix, and performing singular value decomposition for compression to obtain a binary correction mask matrix.

[0013] In a preferred embodiment, the offline UAV input state vector of the multi-layer recurrent neural network is modified, specifically: performing Hadamard product operation on the binary correction mask matrix and the original state vector of the offline UAV to obtain a modified input state vector; inputting the modified input state vector into the time convolution module of the multi-layer recurrent neural network, and predicting the optimal scheduling path of the cluster through the gated recurrent unit.

[0014] The technical effects and advantages of the UAV cluster scheduling method and system based on a multi-layer recurrent neural network of the present application are:

[0015] 1. The present application realizes high-precision fitting and matching of the reconnection state and the initial state by key frame extraction and spatial pose mapping of the state sequence before dropping, combined with dynamic time warping and second-order optimization path constraint; further utilizes state fluctuation gradient evaluation and frequency domain analysis under multiple time scales to accurately identify and compensate the state oscillation weak area, generates a correction mask by graph convolution and singular value decomposition, dynamically modifies the input state vector, and finally performs intelligent scheduling prediction through a deep recurrent neural network. The overall scheme significantly improves the scheduling stability and task execution efficiency of the UAV cluster in complex environments, enhances the robustness and adaptive ability of the system to the state fluctuation of the offline reconnection, and guarantees the continuous and reliable cooperative operation of the UAV cluster. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The flowchart of the UAV cluster scheduling method based on a multi-layer recurrent neural network of the present application.

[0017] Figure 2 The structure diagram of the UAV cluster scheduling system based on a multi-layer recurrent neural network of the present application. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort belong to the protection scope of the present application.

[0019] Embodiment 1, Figure 1 A UAV group scheduling method based on a multi-layer recurrent neural network is given, comprising the following steps:

[0020] S1, detecting a dropped UAV in a target UAV group, and obtaining a state change sequence of the dropped UAV in a preset time window before dropping, to obtain an initial state vector;

[0021] In this example, a dropped UAV in a target UAV group is detected, and a state change sequence of the dropped UAV in a preset time window before dropping is obtained, to obtain an initial state vector, which is specifically:

[0022] A dropped UAV in a target UAV group is detected, and a state data stream in a preset time window before dropping is intercepted, the state data stream comprising three-dimensional position coordinates, a pitch angle, a yaw angle and a power system load rate;

[0023] The state data stream is subjected to time slicing processing, to obtain discrete state sequences at equal intervals;

[0024] The discrete state sequences are screened by using Lyapunov indexes, to obtain a plurality of key frame subsets;

[0025] Spatial attitude parameters of the plurality of key frame subsets are obtained and mapped to an SO(3) rotation group, to obtain an initial state vector.

[0026] It should be noted that during the operation of the UAV group, the system will monitor the communication state and data transmission of each UAV in real time. Each UAV will periodically report its state data according to the set frequency, including position, attitude and system health parameters. Once a UAV fails to upload data as scheduled for a plurality of continuous periods, or its communication signal is interrupted for more than a preset time threshold, for example, more than one second, the system will determine that the UAV has dropped. At this time, the UAV will be marked as a dropped UAV.

[0027] Further, after detecting the offline UAV, the system traces back the state change data of the offline UAV in a period of time before it goes offline, and divides the data into a series of discrete frame sequences according to fixed time intervals. Then, in order to extract the key frames that are most representative of the dynamic behavior of the UAV, the system introduces Lyapunov index for analysis. Lyapunov index is a mathematical tool for measuring the sensitivity of a system to initial conditions, which can determine whether the system state is unstable or has a chaotic trend. By calculating the state change rate and disturbance amplification between each frame, the system can identify frames with significant state changes that have a greater impact on future states, and retain them as key frames.

[0028] Finally, after screening the key frame subset, the system focuses on extracting the parameters related to spatial position and attitude in these key frames to construct an initial state vector with more physical meaning. The spatial attitude parameters mainly include the three-dimensional position coordinates corresponding to each frame, i.e., the specific position of the UAV in space, and the attitude angle information, usually including pitch angle, yaw angle and roll angle, which together describe the orientation and rotation state of the UAV in space.

[0029] In order to enhance the processing capability of the rotation information, these parameters are mapped to the three-dimensional rotation group space, usually called SO(3) rotation group, which is a mathematical structure for describing arbitrary three-dimensional rotation states. The mapping process essentially converts discrete spatial attitude information into a unified vector format that is more suitable for neural network processing, so that each key frame is represented as a standard rotation state, and these states are then concatenated or fused in time sequence to form a complete initial state vector, laying the foundation for subsequent state matching and group scheduling prediction.

[0030] S2, when the offline UAV is detected to be online again, a first state vector is extracted and vector fitting matching is performed with the initial state vector;

[0031] In this example, when the offline UAV is detected to be online again, a first state vector is extracted and vector fitting matching is performed with the initial state vector, specifically:

[0032] The reconnection response data is collected at the moment of reconnection of the offline UAV, and a first state vector is extracted, wherein the reconnection response data includes GPS positioning, inertial measurement unit readings and battery transient response values;

[0033] A time warping algorithm is used to align the first state vector and the initial state vector along the time axis to construct a DTW distance matrix;

[0034] A preset time window is divided into a plurality of 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;

[0035] The first derivative sequence of the anti-interference distance layer is extracted, and a curvature feature matrix is constructed;

[0036] The Hessian transformation is performed on the curvature feature matrix to generate a second-order optimization path constraint;

[0037] Based on the second-order optimization path constraint, the first state vector is compensated by piecewise interpolation to obtain a fitting matching vector.

[0038] It should be noted that when a certain offline unmanned aerial vehicle (for example, numbered UAV-05) re-establishes connection with the control system, the system will immediately collect sensor data at the re-online instant of the unmanned aerial vehicle to form a first state vector. The state data is derived from multiple sensor channels, mainly including GPS positioning information at the current time (used to obtain the three-dimensional coordinates of the unmanned aerial vehicle), real-time readings of the inertial measurement unit (including acceleration and angular velocity), and instantaneous current, voltage and temperature response values of the battery (reflecting the system load and energy state). These data are combined into a complete six-dimensional or multi-dimensional state vector to reflect the overall operating state of the unmanned aerial vehicle when it reconnects.

[0039] Subsequently, in order to determine whether the current state of the unmanned aerial vehicle is consistent with the operating state before it went offline, the system will align and match the first state vector at the reconnection instant with the initial state vector constructed previously in time. Since the two data often do not have strict time consistency, a dynamic time warping algorithm is used to synchronize and align the two vector sequences in time axis, establish a nonlinear distance matrix, and construct a multi-dimensional dynamic time warping distance map.

[0040] In order to improve the robustness to non-stationary state changes, the system divides the preset time window into several local small windows, and extracts the distance values in the dynamic time warping matrix along the main diagonal direction in each small window, and calculates the Mahalanobis distance as the anti-interference measurement index of matching. In this way, the local similarity of the two state vectors in each time period can be effectively identified.

[0041] Next, the system takes the first derivative of the anti-interference distance result sequence to form a curvature feature matrix reflecting the rate of state change. This matrix is further sent to the Hessian transformation module to generate a second-order optimization constraint structure reflecting the local extreme value and change trend of the path. Through these constraint information, the system can compensate by piecewise interpolation based on the original first state vector, thereby correcting and adjusting the missing or distorted state segment, and finally forming a matching vector highly fitted with the initial state vector.

[0042] S3, constructing a state fluctuation trajectory according to the matching result, and evaluating the change of the state fluctuation trajectory under different task time scales to obtain a state fluctuation gradient distribution;

[0043] In this example, a state fluctuation trajectory is constructed according to the matching result, specifically:

[0044] Taking the fitted matching vector as the starting point, three analysis windows of long, medium and short are divided according to the task time scale;

[0045] Extract the trajectory segment of similar scene in the historical task period in the long window, and construct the background fluctuation baseline;

[0046] Calculate the Mahalanobis distance between the current trajectory and the background fluctuation baseline in the medium window, and mark the deviation threshold interval;

[0047] In the short window, the instantaneous fluctuation trend is predicted by Gaussian process regression, and the spatio-temporal coupled state fluctuation trajectory is generated by fusing the outputs of the three windows.

[0048] In this example, the state fluctuation gradient distribution is obtained by evaluating the change of the state fluctuation trajectory under different task time scales, specifically:

[0049] Under different task time scales, multiple rounds of Hamiltonian Monte Carlo sampling are performed on the state fluctuation trajectory to obtain a set of candidate trajectory probability distributions;

[0050] The optimal sampling path is obtained by iteratively selecting the trajectory probability distribution set using a greedy algorithm;

[0051] The feature values of the preset different task stages are extracted along the optimal sampling path, and the feature values are multi-dimensionally fused to obtain the state fluctuation gradient distribution tensor;

[0052] Project the state fluctuation gradient distribution tensor to construct the state fluctuation gradient distribution map.

[0053] It should be noted that after the initial state vector and the re-online unmanned aerial vehicle state are fitted and matched in the previous stage, the system constructs the state fluctuation trajectory of the unmanned aerial vehicle according to the matching result. The trajectory describes the state change process of the unmanned aerial vehicle from before dropping out, during dropping out to re-online, and reflects the continuous evolution of its spatial attitude, power response, etc.

[0054] In order to evaluate the change trend of the state fluctuation trajectory under different task time scales, first define multiple task-related time scales, such as short-term stable control scale (in seconds), medium-term cooperative flight scale (in tens of seconds) and long-term task planning scale (in minutes). For each time scale, a large number of candidate trajectory sets with physical feasibility are generated by using the Hamiltonian Monte Carlo sampling method to perturb the original trajectory multiple times. Each candidate trajectory reflects the possible state evolution trend in the current time scale in the form of probability.

[0055] A greedy search algorithm is introduced to iteratively filter the candidate trajectory probability distribution set. The algorithm starts from the initial state, and preferentially selects the trajectory node with the highest confidence and optimal state continuity, gradually constructing a sampling trajectory path that is optimal at the current time scale. This path is considered to be the most likely state fluctuation evolution curve at this time scale.

[0056] Further, along the optimal sampling path, the state feature values of each key node are extracted at different task stages (such as takeoff, formation adjustment, target tracking, return, etc.). These feature values include attitude angle rate of change, velocity mutation, load response change, position offset, etc., which reflect the dynamic stability and collaborative adaptability of the UAV at each task stage. Through multi-dimensional fusion technology, these features are integrated into a unified state fluctuation gradient distribution tensor.

[0057] Finally, the tensor is converted into a two-dimensional or three-dimensional state fluctuation gradient distribution map through projection mapping operation. The distribution map clearly shows the fluctuation intensity and evolution direction of the UAV state over time at different task scales, which can be used to identify potential unstable points, high-risk behavior patterns or collaborative imbalance trends.

[0058] S4, convert the state fluctuation gradient distribution map into a continuous state vector sequence, and perform vector fitting and completion with the preset reference state vector cluster, and identify the weak shock area according to the completion result;

[0059] In this example, the state fluctuation gradient distribution map is converted into a continuous state vector sequence, specifically:

[0060] Gaussian sampling is performed on the state fluctuation gradient distribution map to obtain a discrete state point set;

[0061] 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;

[0062] The decomposed components are encoded into a continuous state vector sequence according to the time sequence.

[0063] It should be noted that first, Gaussian sampling is performed on the generated state fluctuation gradient distribution map. In this process, the system performs weighted sampling according to the gradient intensity distribution area, preferentially selecting high gradient areas for dense sampling and low gradient areas for sparse sampling. The sampling results form a set of discrete state points, each representing the fluctuation amplitude and direction of the UAV state at a certain time node and task stage.

[0064] Further, an interpolation algorithm is used to restore the discrete state points to a vector field with time continuity. The interpolation process not only preserves the relative trends between data, but also reasonably completes some missing or sparse areas. The complete process of the state changing continuously along the time axis can be simulated through the interpolated vector field.

[0065] Subsequently, the continuous vector field is decomposed to divide the entire vector field into components according to the change rate, direction characteristics or state dimensions. For example, the vector field can be decomposed into linear velocity components, angular velocity components and power load components, etc., to reflect the dynamic response characteristics of the UAV in various state variables. Finally, the decomposed state components are uniformly coded in time sequence to be converted into a sequence of continuous state vectors with time labels. Each state vector corresponds to a specific time point and contains physical state information of multiple dimensions.

[0066] In this example, vector fitting is performed with a preset reference state vector cluster, specifically:

[0067] Extracting a reference state vector cluster from a historical task library to construct a vector manifold space;

[0068] Calculating the shortest path of the continuous state vector sequence to the reference state vector cluster to obtain a tangent vector field;

[0069] Based on the tangent vector field, the continuous state vector sequence is decomposed into normal components and tangent components;

[0070] Projecting the normal component into the vector manifold space to construct a residual vector;

[0071] Using time axis sliding convolution to compensate the phase of the residual vector to obtain the compensated state sequence.

[0072] It should be noted that first, flight records similar to the current UAV task environment, flight type or execution stage are selected from the historical task library. The state vectors in these records are uniformly preprocessed to form a diversified reference state vector cluster, which is embedded in a high-dimensional state space in the form of a manifold. This manifold space can fully preserve the local geometric relationship and task evolution trend in the state sequence.

[0073] Subsequently, the shortest path trajectory of 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 fitting degree of the current state sequence relative to the reference state evolution direction, and provides a geometric basis for the next step of decomposition and compensation.

[0074] Next, the continuous state vector sequence is decomposed in the tangent vector field to obtain tangent components and normal components, respectively. The tangent components represent the natural evolution of the current state sequence along the reference task trajectory direction, and the normal components represent abnormal disturbances deviating from the reference trajectory. The system further projects the normal components into the manifold space to extract residual vectors with deviations.

[0075] To correct these deviations, a time-axis sliding convolution mechanism is used to perform phase compensation processing on the residual vectors. This mechanism detects state deviation fluctuations within a local time window and gradually corrects the state deviations by smoothing convolution and phase alignment strategies, avoiding transient mutations or physical violations.

[0076] In this example, the weak oscillation region is identified according to the true result, specifically:

[0077] The Fourier descriptor decomposition is performed on the state sequence after the true result, and the frequency energy spectrum is extracted;

[0078] The resonance frequency band and harmonic components exceeding the preset threshold in the energy spectrum are detected;

[0079] The resonance frequency band is reversely mapped to the state vector space-time domain to locate the phase conflict interval;

[0080] The regional oscillation intensity coefficient is calculated based on the amplitude of the harmonic components, and the region with a coefficient greater than the critical value is selected as the weak oscillation region.

[0081] It should be noted that first, the Fourier descriptor decomposition is performed on the state vector sequence after the true result. 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 over time. During the decomposition process, the system regards the state vector as a multi-dimensional time series signal, extracts the frequency spectrum structure of each dimension through Fourier transform, and constructs a complete frequency energy spectrum diagram.

[0082] Further, the energy spectrum is analyzed, especially the frequency band region with prominent energy. By setting an empirical threshold (e.g., twice the average spectral energy), the system can automatically detect the resonance frequency band and the main harmonic components contained therein that exceed the threshold in the energy spectrum. These high-energy frequency bands usually correspond to strong periodic fluctuations or coordination imbalance phenomena in the state signal.

[0083] Subsequently, these identified resonance bands are back-projected onto the time axis and spatial dimensions of the original state vector through frequency-time reverse mapping, thus locating specific state spatiotemporal segments. These regions usually exhibit strong fluctuations in attitude, velocity, or power load within a short time, indicating the presence of potential phase conflicts or system spontaneous oscillations. To quantify the intensity of these abnormal regions, the system further calculates an oscillation intensity coefficient for each located region based on the extracted harmonic component amplitudes. This coefficient takes into account factors such as frequency amplitude, duration, and oscillation concentration. If the oscillation intensity coefficient of a certain spatiotemporal region exceeds a pre-set critical threshold, the system marks this region as an oscillation weak region.

[0084] S5, the state vector of the oscillation weak region is converted into a correction mask, and the input state vector of the multi-layer recurrent neural network of the offline unmanned aerial vehicle is corrected and used for unmanned aerial vehicle group scheduling prediction.

[0085] In this example, the state vector of the oscillation weak region is converted into a correction mask, specifically:

[0086] Extracting the vertex state vector of the oscillation weak region;

[0087] Performing graph convolution operation on the vertex vector to obtain a topological feature matrix;

[0088] Performing Hadamard product operation on the topological feature matrix and the pre-set stability base matrix, and performing singular value decomposition for compression to obtain a binary correction mask matrix.

[0089] It should be noted that first, representative nodes are extracted from each oscillation weak region, i.e., the vertex state vectors in these regions, as high-dimensional representations of local region state characteristics. These vertex state vectors reflect the transient extreme value or trend reversal point of the region when the oscillation occurs, and have structural representativeness.

[0090] Next, based on the graph structure construction mechanism, the similarity relationship between these vertex vectors is organized into an undirected graph structure, and a graph convolution operation is performed. In this process, the system uses a graph convolutional neural network to extract topological dependency features between these state points, generating a topological feature matrix reflecting structural connections and dynamic resonance trends. This matrix reveals the collaborative oscillation patterns and propagation paths between different state dimensions in the oscillation region.

[0091] Subsequently, the generated topological feature matrix is subjected to Hadamard product operation with the stability benchmark matrix trained in advance, i.e., the corresponding elements are multiplied. This operation retains the abnormal features in the oscillation region while incorporating the system stability experience model to suppress false abnormal points. The resulting product matrix is the to-be-compressed correction mask with structural information.

[0092] To reduce the operation complexity of the subsequent model and extract the core disturbance component, singular value decomposition (SVD) processing is performed on the correction matrix. By retaining the first few singular value components, the system effectively compresses the correction information and eliminates the low-weight redundant part. The final output is a binary correction mask matrix, where each element represents whether the state of the corresponding dimension needs to be dynamically corrected, with 1 indicating correction and 0 indicating no adjustment. The mask will directly act on the state input end of the multi-layer recurrent neural network to control the activation, inhibition and correction of the state vector dimension, thereby improving the robustness and convergence of the network when processing unstable state inputs.

[0093] In this example, the input state vector of the offline unmanned aerial vehicle of the multi-layer recurrent neural network is corrected, specifically:

[0094] The binary correction mask matrix is multiplied by the original state vector of the offline unmanned aerial vehicle to obtain the corrected input state vector;

[0095] 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.

[0096] It should be noted that first, the original input state vector of the offline unmanned aerial vehicle after data recovery is processed dimension by dimension and multiplied by the correction mask matrix, i.e. the corresponding elements are multiplied. For each oscillation dimension, the original state value will be masked or replaced by the smoothed value after noise reduction interpolation to eliminate the influence of high-frequency jitter on network judgment; for the dimensions not marked for correction, the state value remains unchanged. The corrected input state vector obtained finally is more stable and has higher predictability.

[0097] In actual operation, the multi-layer recurrent neural network receives state vectors from all unmanned aerial vehicles (including offline unmanned aerial vehicles after data recovery and correction), and through time series modeling and spatial aggregation operations, outputs the optimal scheduling path of the entire unmanned aerial vehicle cluster at the current time, including the position guidance, cooperative target, speed adjustment suggestion, etc. of each unmanned aerial vehicle.

[0098] In this example, the multi-layer recurrent neural network, specifically:

[0099] The multi-layer recurrent neural network includes an input layer, a time convolution layer, an attention mechanism layer, a bidirectional GRU layer, and an output layer, specifically:

[0100] The input layer receives the corrected input state vector;

[0101] The time convolution layer uses an expanded convolution kernel to extract multi-scale time series features;

[0102] The attention mechanism layer weightedly fuses key state parameters;

[0103] The bidirectional GRU layer constructs a state transition probability model.

[0104] The output layer generates a scheduling instruction set containing a heading angle, a speed and a formation parameter.

[0105] Embodiment 2, Figure 2 The application provides a UAV group scheduling system based on a multilayer recurrent neural network, which comprises an offline detection module, a vector matching module, a gradient distribution module, a weak recognition module and a network correction module.

[0106] The offline detection module is used for detecting an offline UAV in a target UAV group and obtaining a state change sequence of the offline UAV in a preset time window before the offline UAV goes offline, so as to obtain an initial state vector.

[0107] The vector matching module is used for extracting the first state vector and performing vector fitting matching with the initial state vector when the offline UAV is detected to be online again.

[0108] The gradient distribution module is used for constructing a state fluctuation trajectory according to a matching result and evaluating changes of the state fluctuation trajectory under different task time scales, so as to obtain a state fluctuation gradient distribution.

[0109] The weak recognition module is used for converting the state fluctuation gradient distribution map into a continuous state vector sequence, performing vector fitting and supplementing true values with a preset reference state vector cluster, and recognizing a weak shock area according to a supplementing true value result.

[0110] The network correction module is used for converting the state vector of the weak shock area into a correction mask, correcting an input state vector of the offline UAV of the multilayer recurrent neural network and being used for UAV group scheduling prediction.

[0111] The above formulas are all dimensionless numerical values, the formulas are obtained by collecting a large amount of data to simulate a formula of a nearest real situation, and preset parameters in the formulas are set by a person skilled in the art according to actual conditions.

[0112] The above embodiments can be realized by software, hardware, firmware or any combination thereof, in whole or in part. When realized by software, the above embodiments can be realized in the form of a computer program product in whole or in part.

[0113] Those skilled in the art can understand that the modules and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0114] In addition, each functional module in each embodiment of the present application can be integrated in one processing module, or each module can exist physically alone, or two or more modules can be integrated in one module.

[0115] The above is merely specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any skilled in the art within the technical scope disclosed in the present application can easily think of changes or replacements, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0116] Finally: the above is only the preferred embodiment of the present application, and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for drone swarm scheduling based on a multi-layer recurrent neural network, characterized in that, Includes the following steps: Detect disconnected drones in the target drone swarm and obtain the state change sequence of the disconnected drones within a preset time window before they went offline, thus obtaining the initial state vector; When a disconnected drone is detected to be back online, the first state vector is extracted and matched with the initial state vector. Based on the matching results, a state fluctuation trajectory is constructed, and the changes of the state fluctuation trajectory under different task time scales are evaluated to obtain the state fluctuation gradient distribution. The gradient distribution map of state fluctuations is transformed into a continuous state vector sequence, and vector fitting is performed to complete the data with a preset reference state vector cluster. Based on the data completion results, weak oscillation regions are identified. The state vector of the weak oscillation region is transformed into a correction mask, which is used to correct the input state vector of the disconnected drone in the multi-layer recurrent neural network and then used for drone swarm scheduling prediction. 2.The multi-layer recurrent neural network based UAV swarm scheduling method of claim 1, wherein, The detection involves identifying disconnected drones within the target drone swarm and obtaining a sequence of state changes of the disconnected drones within a preset time window before they went offline, resulting in an initial state vector. Specifically: Detect disconnected drones in the target drone swarm and capture the status data stream of a preset time window before disconnection. The status data stream includes three-dimensional position coordinates, pitch angle, yaw angle and power system load rate. The state data stream is processed by time slicing to obtain a discrete state sequence with equal intervals; The Lyapunov exponent was used to filter the discrete state sequence to obtain several key frame subsets; The spatial pose parameters of several keyframe subsets are obtained and mapped to the SO(3) rotation group to obtain the initial state vector. 3.The method of claim 2, wherein, When a disconnected drone is detected to have reconnected, the first state vector is extracted and matched with the initial state vector. Specifically: Reconnection response data is collected at the moment the disconnected drone reconnects, and a first state vector is extracted. 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 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 using the dynamic time warp distance matrix to obtain the anti-interference distance layer. Extract the first derivative sequence of the anti-interference distance layer to construct the curvature feature matrix; Perform a Hessian transformation on the curvature feature matrix to generate second-order optimized path constraints; Based on the second-order optimization path constraint, the first state vector is piecewise interpolated and compensated to obtain the fitted matching vector.

4. The multi-layer recurrent neural network based UAV swarm scheduling method of claim 3, wherein, The evaluation of the changes in the state fluctuation trajectory at different task time scales yields the state fluctuation gradient distribution, specifically as follows: At different task time scales, multiple rounds of Hamiltonian Monte Carlo sampling were performed on the state fluctuation trajectory to obtain a set of candidate trajectory probability distributions; A greedy algorithm is used to iteratively filter the probability distribution set of the selected trajectory to obtain the optimal sampling path; Feature values ​​of different task stages are extracted along the optimal sampling path, and the feature values ​​are fused in multiple dimensions 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 multi-layer recurrent neural network based UAV swarm scheduling method of claim 4, wherein, The state fluctuation gradient distribution is transformed into a continuous state vector sequence, specifically as follows: The state fluctuation gradient distribution is Gaussian sampled to obtain a discrete state point set; The discrete point set is reconstructed into a continuous vector field by a sampling interpolation algorithm, and the continuous vector field is decomposed to obtain decomposed components; The decomposed components are encoded into a continuous state vector sequence in time sequence.

6. The multi-layer recurrent neural network based UAV swarm scheduling method of claim 5, wherein, The vector fitting completion is performed with a preset reference state vector cluster, specifically as follows: The reference state vector cluster is extracted from a historical task library to construct a vector manifold space; The shortest path of the continuous state vector sequence to the reference state vector cluster is calculated to obtain a tangent vector field; The continuous state vector sequence is decomposed into a normal component and a tangent component based on the tangent vector field; The normal component is projected into the vector manifold space to construct a residual vector; The residual vector is phase compensated by time axis sliding convolution to obtain a completed state sequence.

7. The multi-layer recurrent neural network based UAV swarm scheduling method of claim 6, wherein, The state vector of the weak oscillation region is transformed into a correction mask, specifically as follows: The vertex state vector of the weak oscillation region is extracted; The vertex vector is subjected to graph convolution operation to obtain a topological feature matrix; The topological feature matrix is subjected to Hadamard product operation with a preset stability basis matrix, and is compressed by singular value decomposition to obtain a binary correction mask matrix.

8. The multi-layer recurrent neural network based UAV swarm scheduling method of claim 7, wherein, The offline UAV input state vector of the multi-layer recurrent neural network is corrected, specifically as follows: The binary correction mask matrix is subjected to Hadamard product operation with the original state vector of the offline UAV 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.

9. A multi-layer recurrent neural network-based UAV swarm scheduling system, applied to the multi-layer recurrent neural network-based UAV swarm scheduling method of any one of claims 1-8, characterized in that, The offline detection module, the vector matching module, the gradient distribution module, the weakness identification module, and the network correction module are included: The offline detection module is used to detect the offline UAV in the target UAV cluster, and to obtain the state change sequence of the offline UAV in a preset time window before the offline, to obtain an initial state vector; The vector matching module is used to extract a first state vector when the offline UAV is detected to be online again, and to perform vector fitting matching with the initial state vector; The gradient distribution module is used to construct a state fluctuation trajectory according to the matching result, and to evaluate the change of the state fluctuation trajectory at different task time scales to obtain a state fluctuation gradient distribution; The weakness identification module is used to transform the state fluctuation gradient distribution into a continuous state vector sequence, to perform vector fitting completion with a preset reference state vector cluster, and to identify a weak oscillation region according to the completion result; The network correction module is used to transform the state vector of the weak oscillation region into a correction mask, to correct the input state vector of the offline UAV of the multi-layer recurrent neural network, and to be used for UAV cluster scheduling prediction.

Images