Unmanned aerial vehicle group space structure regularity evaluation method based on brain-like calculation

Abstract

The invention discloses an unmanned aerial vehicle group space structure regularity evaluation method based on brain-like calculation, and belongs to the technical field of unmanned aerial vehicle group space structure evaluation, and the method comprises the steps: calculating the optimal pose estimation of an unmanned aerial vehicle group under a global coordinate system, and extracting a global position set at a selected moment; constructing three variant matrixes of various distance deviations; calculating a correlation coefficient matrix and an autocorrelation value of each variant matrix; mapping the self-correlation value into structural regularity, and calculating confidence in combination with a consistency index; constructing a heterogeneous integrated learning framework based on a spiking neural network and a gradient boosting decision tree model, inputting the self-correlation value into the spiking neural network, inputting the self-correlation value and the structure regularity into the gradient boosting decision tree model, and executing a heterogeneous model fusion strategy to obtain an integrated evaluation result; through system performance verification and evaluation, a core output parameter set of the integrated system is generated, and high-precision global unified positioning in a denial environment is realized.

Description

Technical Field

[0001] This invention belongs to the field of UAV swarm spatial structure evaluation technology, specifically involving a method for evaluating the regularity of UAV swarm spatial structure based on brain-like computing. Background Technology

[0002] In complex environments, drone swarms, with their distributed collaboration advantages and flexible formation characteristics, have become a core force for target acquisition and situational awareness. However, existing drone swarm collaboration technologies still face the following significant shortcomings in practical applications: First, there is a lack of stable global spatial reference in GNSS-denied environments. Traditional models rely heavily on absolute position information provided by the Global Positioning System (GPS / GNSS). In urban canyons, underground spaces, or environments with strong electronic interference, the lack of satellite signals prevents UAV swarms from obtaining accurate coordinates. Cooperative control strategies fail due to the loss of a unified spatial reference, severely restricting the sharing of contextual awareness and consistency of decision-making in large-scale swarms.

[0003] Second, existing technologies lack objective means to assess the regularity of the spatial structure of drone swarms. Currently, the assessment of the regularity, regularity, and repeatability of drone swarm configurations mostly relies on manual visual judgment or simple geometric centroid calculation. This approach is highly subjective, has a single dimension, and lacks scale, rotation, and translation invariance, making it difficult to quantitatively describe the spatial arrangement quality of the swarm during complex maneuvers. This results in a lack of effective closed-loop feedback indicators in the control system.

[0004] Third, the accuracy and robustness of multi-source observation data fusion are insufficient. When dealing with high-speed, non-cooperative targets, single-sensor observations are susceptible to occlusion and noise interference. Traditional algorithms often face spatiotemporal misalignment and observation heterogeneity problems when dealing with highly nonlinear motion, resulting in large deviations in target state estimation and making it difficult to support high-precision cooperative acquisition tasks.

[0005] Fourth, there is an imbalance between real-time performance and security in decision-making algorithms under complex constraints. Traditional methods typically employ step-by-step optimization or rule bases. When faced with multiple conflicting constraints such as "rapid encirclement, obstacle avoidance, collision avoidance, and target visibility," the computational load increases exponentially with the cluster size, which can easily lead to the UAV getting stuck in a decision-making deadlock of "oscillation and hesitation" or "path blocking," making it impossible to achieve millisecond-level real-time response on embedded terminals. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, this invention provides a brain-inspired computing-based method for evaluating the spatial structure regularity of UAV swarms, solving the following technical problems faced by existing UAV swarm collaboration technologies in practical applications: (1) There is a lack of stable global space reference in GNSS denied environments.

[0007] (2) There is a lack of objective means to evaluate the regularity of the spatial structure of clusters.

[0008] (3) The accuracy and robustness of multi-source observation data fusion are insufficient.

[0009] (4) The decision-making algorithm under complex constraints is out of balance between real-time performance and security.

[0010] To achieve the aforementioned objectives, the technical solution adopted by this invention is as follows: a method for evaluating the spatial structure regularity of unmanned aerial vehicle (UAV) swarms based on neuromorphic computing, comprising the following steps: S1. Based on the multi-source sensor data fusion and factor graph optimization model, calculate the optimal pose estimate of the UAV cluster in the global coordinate system and extract the global position set at the selected time. S2. Based on the global position set, by calculating Euclidean distance, removing the mean and performing difference operations, three variant matrices with various distance deviations are constructed to represent the spatial structure of the UAV swarm from multiple perspectives. S3. Define the Pearson correlation coefficient standard and calculate the correlation coefficient matrix and autocorrelation value of each variant matrix; S4. The autocorrelation value is mapped to structural regularity by the improved Chaddock scaling function, and the confidence level is calculated by combining the consistency index. The core parameter set and detailed data set for determining structural regularity are output. S5. Construct a heterogeneous ensemble learning framework based on a spiking neural network and a gradient boosting decision tree model, collaboratively optimize the training framework, input the autocorrelation value into the spiking neural network, input the autocorrelation value and structural regularity into the gradient boosting decision tree model, and execute the heterogeneous model fusion strategy to obtain the ensemble evaluation results. S6. Based on the integration evaluation results, through system performance verification and evaluation, generate the core output parameter set of the integrated system.

[0011] The beneficial effects of this invention are as follows: (1) Achieved high-precision global unified positioning in GNSS denied environments: This invention introduces a factor graph optimization mechanism to deeply integrate IMU odometry, RTK global factors, and UWB relative observation factors, effectively solving the problem of accumulated error in GNSS denied environments. Even when global absolute coordinates are unavailable, a globally consistent spatial coordinate system can still be constructed, laying a solid underlying data foundation for evaluating the collaborative regularity of the swarm and solving the problem of the lack of stable global spatial reference in GNSS denied environments faced by existing UAV swarm collaboration technologies in practical applications.

[0012] (2) An objective and quantitative evaluation system for spatial structure regularity has been established: This invention innovatively proposes an evaluation model based on multivariable autocorrelation analysis. Through the calculation of the relative position matrix (MOP) and Pearson correlation coefficient, the scale, rotation, and translation invariance of the evaluation results is achieved. This method can automatically and quickly quantify the cluster configuration into a clear regularity level, replacing subjective manual judgment. It provides scientific and standardized performance indicators for the automated feedback control of UAV swarms and solves the problem of the lack of objective evaluation methods for the spatial structure regularity of swarms in the practical application of existing UAV swarm collaborative technologies.

[0013] (3) Significantly improves the estimation accuracy and trajectory smoothness of dynamic targets: In trajectory prediction based on multi-source observations, this invention organically combines unscented Kalman filtering (UKF) with a Bézier curve generation algorithm. UKF effectively handles the strong nonlinearity problem in the acquisition process, while the Bézier curve ensures the dynamic feasibility and spatiotemporal continuity of the generated trajectory, significantly improving the accuracy of the system's state estimation of high-speed maneuvering targets and solving the problem of insufficient accuracy and robustness of multi-source observation data fusion in the practical application of existing UAV swarm collaborative technology.

[0014] (4) Efficient and intelligent decision evaluation is achieved through heterogeneous ensemble learning: This invention combines the spatiotemporal pattern recognition capability of brain-like computing with the statistical processing capability of CatBoost ensemble learning. By using SNN to simulate the intuitive perception of spatial configuration by the biological brain and using CatBoost for accurate logical classification, this heterogeneous fusion architecture can greatly reduce the inference latency of the algorithm (<50ms) while ensuring an evaluation accuracy of over 94.7%. This ensures the real-time safe response and rapid encirclement capability of UAV swarms in complex and highly dynamic environments, and solves the imbalance between real-time performance and security of decision algorithms under complex constraints faced by existing UAV swarm collaboration technology in practical applications. Attached Figure Description

[0015] Figure 1 This is a flowchart of a method for evaluating the spatial structure regularity of unmanned aerial vehicle (UAV) swarms based on brain-like computing, according to the present invention.

[0016] Figure 2 Three variant matrices characterize spatial structure features from different perspectives.

[0017] Figure 3 This is the architecture and decision graph for a heterogeneous integrated learning system. Detailed Implementation

[0018] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0019] like Figure 1 As shown, in one embodiment of the present invention, a method for evaluating the spatial structure regularity of unmanned aerial vehicle (UAV) swarms based on neuromorphic computing includes the following steps: S1. Based on the multi-source sensor data fusion and factor graph optimization model, calculate the optimal pose estimate of the UAV cluster in the global coordinate system and extract the global position set at the selected time. S2. Based on the global position set, by calculating Euclidean distance, removing the mean and performing difference operations, three variant matrices with various distance deviations are constructed to represent the spatial structure of the UAV swarm from multiple perspectives. S3. Define the Pearson correlation coefficient standard and calculate the correlation coefficient matrix and autocorrelation value of each variant matrix; S4. The autocorrelation value is mapped to structural regularity by the improved Chaddock scaling function, and the confidence level is calculated by combining the consistency index. The core parameter set and detailed data set for determining structural regularity are output. S5. Construct a heterogeneous ensemble learning framework based on a spiking neural network (SNN) and a gradient boosting decision tree (CatBoost) model, collaboratively optimize the training framework, input autocorrelation values ​​into the spiking neural network, input autocorrelation values ​​and structural regularity into the gradient boosting decision tree model, and execute a heterogeneous model fusion strategy to obtain ensemble evaluation results. S6. Based on the integration evaluation results, the core output parameter set of the integrated system is generated through system performance verification and evaluation.

[0020] S1 includes the following steps: S11. Set the UAV swarm size and observation time series, define the single-unit state vector and the system full state vector, and integrate multi-source sensor data, including odometry observation, RTK global positioning observation and relative observation; S12. Construct objective functions for odometry observations, RTK global positioning observations, and relative observations, and establish a factor graph model by integrating the objective functions. S13. Utilize the sparse structure of the factor graph model to linearize the error function at the current estimation point, and use an incremental optimization algorithm to solve the normal equation to obtain the optimal pose estimate of all UAVs at all times. S14. Based on the optimal pose estimates of all UAVs at all times, extract the global position set at the selected time.

[0021] In S11, the single-machine state vector is defined. and the system's total state vector The specific expression is: In the formula, For drones i At any moment t x-axis coordinates For drones i At any moment t The vertical axis coordinate, For drones i At any moment t The vertical axis coordinates, For drones i At any moment t Roll angle, For drones i At any moment t pitch angle, For drones i At any moment t Yaw angle, It is the transpose symbol. For drone swarm scale, For observation time; Based on the definition of state variables, multi-source sensor data is integrated, specifically including: odometry observations. , indicating drone From time At the time t Relative motion observations are obtained from IMU, visual odometry, or laser odometry, and RTK global positioning observations. , indicating drone At any moment t The absolute global position coordinates, , For matrices, relative to observation , indicating drone With drones j At any moment t The relative position vector is obtained through inter-machine UWB ranging, etc. ; In S12, the factor graph model consists of the odometry factor, the RTK factor, and the relative observation factor. The specific expression of the factor graph model is as follows: In the formula, To estimate the optimal pose for all UAVs at all times, The objective function for odometer observations is... The objective function for RTK global positioning observations is... The objective function is relative to the observation. The error function is relative to the observation. Let be the error function of RTK global positioning observations. Let be the error function of the odometer observation. , and Represents Mahalanobis distance, Let be the covariance matrix of the odometer observation noise. Here is the covariance matrix of the RTK observation noise. Let be the covariance matrix of the relative observation noise. For drones There exists a time set of RTK observations. Let be the set of UAV pairs that have relative observations at time t; In the formula, This is the relative transformation matrix between adjacent time points predicted from the state; in this embodiment, this relative transformation matrix is... Mapping to the Lie algebra space and comparing it with odometer observations yields the odometer error function, which in turn provides the error function of the odometer observations. .

[0022] In the formula, For the reason The transformation matrix of the transformation. For the reason The transformation matrix of the transformation; In the formula, For the reason The rotation matrix formed, For drones The three-dimensional position coordinates, , It is a special three-dimensional Euclidean group, representing the group consisting of all rigid body transformation matrices in three-dimensional space; In this embodiment, the present invention introduces an RTK factor, which provides an absolute position reference to effectively suppress odometry accumulation errors and aligns all UAVs to a unified global coordinate system. The error function of RTK global positioning observation... The specific expression is: To further enhance system robustness, a relative observation factor is introduced. This factor directly constrains the relative geometric relationships between UAVs to improve the internal consistency of the overall configuration, and the error function of the relative observation is used. The specific expression is: In the formula, For drones The three-dimensional position coordinates; In S13, the expression for the normal equation is as follows: In the formula, For Jacobian matrices, For the parameter increment, The block diagonal noise covariance matrix is... , This is a block diagonal matrix constructor, used to arrange the input matrix parameters along the main diagonal to form a block diagonal matrix. In order to be in At this point, the error vector is calculated based on odometer readings, RTK, and relative observations. This is the current estimate; The specific method for constructing the normal equation is as follows: First, at the current estimate Linearize the error function at this point: In the formula, For addition operations on a manifold, SE(3) is typically implemented using an exponential mapping. Jacobian matrix It has a sparse structure: In the formula, e This is the total error vector. For state error components, The observation error component is represented by the odometer factor Jacobian: and It is only related to the pose of adjacent time steps; RTK factor Jacobi: It is only related to the current pose; relative to the observation factor Jacobian: and It is only related to the pose of the two relevant drones; Based on the linearization result, the linearized error is substituted into the objective function: In the formula, For the first The residual vector of each observed factor, Let Jacobian matrix be the residual function with respect to state variables. This is the noise covariance matrix of the corresponding sensor. This represents the squared Mahalanobis distance operation based on the weights of the inverse covariance matrix. By taking the derivative and setting it to zero, we obtain the normal equation.

[0023] The method for solving the normal equation using an incremental optimization algorithm is as follows: S131. Construct the Hessian matrix based on the sparse structure of the factor graph model. ; In the formula, Corresponding drones and drones State variables, For drones Its own state variables, By drone It consists of its own odometry observations and RTK global positioning observations. Only when drones and drones It is non-zero only when there is a relative observation, where, Only when drones The state is non-zero only when there is a relative observation between j and . In the time dimension, the state of each UAV forms a band-like structure.

[0024] S132. To enhance numerical stability, the Levenberg-Marquardt algorithm is used to establish the first linear equation. The specific expression of the first linear equation is as follows: In the formula, To construct a diagonal matrix by taking the elements of the main diagonal of the matrix, It is the damping factor; S133. Solve the first linear equation to obtain the increment. According to the following formula based on the increment Perform iterative updates to obtain the optimal pose estimates for all UAVs at all times. ; In the formula, For addition operations on manifolds, For the first The estimated value of the system's full state vector at the next iteration. For the first The estimated value of the system's full state vector at the next iteration; S14 specifically involves: after solving the above optimization problem, obtaining the optimal pose estimates for all UAVs at all times. ,based on Extract the selected time. global location set The specific expression is: In the formula, For drones At any moment The three-dimensional position coordinates, , For drones i At any moment x-axis coordinates For drones i At any moment The vertical axis coordinate, For drones i At any moment The vertical axis coordinates. This unified and precise global location set. It will serve as a direct input for constructing the relative position matrix, providing a reliable data foundation for assessing the regularity of spatial structures.

[0025] S2 includes the following steps: S21. Calculate the Euclidean distance based on the global location set and extract the ordered neighbor distances to construct the first variant matrix; S22. Calculate the row mean of the first variant matrix, and construct the second variant matrix that measures the distance deviation by removing the mean; S23. Perform differential processing on the sorted neighbor distances to construct a third variant matrix that enhances the sensitivity to local dense structures.

[0026] S21 specifically refers to: S211. Calculate the Euclidean distance between any UAVs based on the global position set, and construct a symmetric distance matrix. R ; In the formula, For drones k With drones l The Euclidean distance between them , It is the L2 norm. For drones The three-dimensional position coordinates, For drones The three-dimensional position coordinates; S212. For any UAV, based on the symmetric distance matrix, obtain the distance set from any UAV to all other UAVs. Sort the elements in the distance set in ascending order through the ordered neighbor distance extraction operation, and select... m The number of values ​​generates a set of neighbor distances. m This is an adjustable parameter representing the number of nearby drones, typically set according to the mission (e.g., based on communication range or sensing capabilities). In this embodiment, it is generally recommended... m The value is between 6 and 10, where the drone is obtained through ordered neighbor distance extraction. k Neighbor distance set The specific expression is: In the formula, For drones k The set of distances to all other drones. , For ordered neighbor distance extraction operation, For drones k The distance to its first nearest neighbor, For drones k The distance to its second nearest neighbor, For drones k To its first The distance to close neighbors, , The upper limit is m , To obtain the minimum value; Vector construction and padding: If drones k The actual number of neighbors is less than m (e.g., in sparse formations or) If the relative position vector is zero, then fill it with zeros until the length is 1. m .

[0027] In summary, drones k relative position vector The complete expression is: in, for dimensional vector, The former The elements are arranged in ascending order, then... elements (if) The value is zero.

[0028] S213. Construct the first variant matrix based on the neighbor distance set. ; In the formula, For drones k The relative position vector, For drones N To its first Distance to nearest neighbors; this embodiment uses a column vector to represent the relative positions of all drones. Combining, building The basic relative position matrix B.

[0029] In S22, to measure the deviation of each neighbor's distance from the "ideal" average distance, a [structure / method] is constructed. Second variant matrix, second variant matrix The specific expression is: In the formula, For the first variant matrix The mean of the first row, For the first variant matrix The mean in the second row, For the first variant matrix The Row mean, first variant matrix The row mean The specific expression is: In the formula, For the first variant matrix The line, number Column elements, i.e. , For drones To its first The distance to the nearest neighbor; Second variant matrix Its elements From the first variant matrix Corresponding element Subtracting the mean of that row gives the result. elements It measures the deviation of each neighbor's distance from the "ideal" average distance.

[0030] In S23, to enhance the sensitivity to locally dense structures (such as clusters), a third variant matrix is ​​constructed. The specific expression is: In the formula, For the first variant matrix The Difference vector of column sort distance; third variant matrix The construction process is as follows: obtain the first variant matrix. For each drone k (i.e., the kth column B[k] of matrix B), calculate the first variant matrix. The Difference vector of column sort distance The derivation of this vector is as follows: the first element of the vector... Defined as The j-th element (j>1) of the vector is defined as and The difference.

[0031] In the formula, For the first variant matrix The The difference between the rows, For drones To its first The distance to close neighbors, For drones To its first Distances to nearest neighbors. Use these difference vectors. As a column, construct Third variant matrix .

[0032] like Figure 2 As shown, in this embodiment, by constructing the first variant matrix to the third variant matrix, the present invention obtains feature representations that characterize the spatial structure of UAV clusters from different perspectives, providing rich input features for subsequent autocorrelation value calculation and regularity level determination.

[0033] S3 includes the following steps: S31. Based on the three variant matrices constructed, calculate the correlation coefficient matrix corresponding to each variant matrix according to the Pearson correlation coefficient to quantify the pairwise correlation within the cluster. S32. Aggregate the correlation coefficient matrix, calculate the autocorrelation value of each variant matrix, and obtain macroscopic regularity indicators; S33. Analyze the characteristics of autocorrelation values ​​and output the set of autocorrelation values ​​and the set of correlation coefficient matrices; In S31, the correlation coefficient matrix corresponding to each variant matrix is ​​calculated, and the elements of the correlation coefficient matrix are... The specific expression is: In the formula, For the first The variant matrix of the first i Line 1 k Column elements, For the first The variant matrix of the first i Line 1 l Column elements, For the first The variant matrix of the first k The mean of the column vectors, For the first The variant matrix of the first l The mean of the column vectors, m Number of neighbors; In this embodiment, the elements of the correlation coefficient matrix It has the following properties: the range of values ​​is ,in Indicates a perfect positive correlation. Indicates a completely negative correlation. This indicates that there is no linear correlation; and the coefficient is symmetric, i.e. = .

[0034] In this embodiment, the first correlation coefficient matrix corresponding to the first variant matrix is ​​calculated. , Elements of the first correlation coefficient matrix The specific expression is: In the formula, For the first variant matrix, the first variant is the... i Line 1 k Column elements, For the first variant matrix, the first variant is the... i Line 1 l Column elements, For the first variant matrix, the first variant is the... k The mean of the column vectors, For the first variant matrix, the first variant is the... l The mean of the column vectors, m Number of neighbors; Calculate the second correlation coefficient matrix corresponding to the second variant matrix. , Elements of the second correlation coefficient matrix The specific expression is: In the formula, For the second variant matrix, the first i Line 1 k Column elements, For the second variant matrix, the first i Line 1 l Column elements, For the second variant matrix, the first k The mean of the column vectors, For the second variant matrix, the first l The mean of the column vectors; Calculate the third correlation coefficient matrix corresponding to the third variant matrix. , Elements of the third correlation coefficient matrix The specific expression is: In the formula, For the third variant matrix, the first i Line 1 k Column elements, For the third variant matrix i Line 1 l Column elements, For the third variant matrix, the first k The mean of the column vectors, For the third variant matrix, the first l The mean of the column vectors; In S32, the autocorrelation values ​​of each variant matrix are calculated. The specific expression is: In the formula, Used for normalization to ensure the comparability of autocorrelation values.

[0035] In this embodiment, the first autocorrelation value of the first variant matrix is ​​calculated. The specific expression is: Calculate the second autocorrelation value of the second variant matrix. The specific expression is: Calculate the third autocorrelation value of the third variant matrix. The specific expression is: In S33, the range of autocorrelation values ​​is: Its physical meaning is defined as follows: The autocorrelation value is approximately 1, which indicates a highly regular spatial structure, meaning that the drones are evenly and orderly distributed. The autocorrelation value is approximately 0, indicating a random distribution with no obvious regularity; An autocorrelation value of approximately -1 indicates a highly irregular distribution, suggesting a clear repulsion or clustering phenomenon. In this embodiment, the stability analysis characteristics of the autocorrelation value calculation method are defined. The method has the following features: Scale invariance: The drone swarm is invariant to overall scaling; Rotation invariance: The drone swarm is invariant to overall rotation; Translation invariance: The drone swarm is invariant to overall translation.

[0036] Next, output the set of autocorrelation values: obtain the set of autocorrelation values ​​based on the autocorrelation values ​​of each variant matrix. Output the set of correlation coefficient matrices: obtain the set of correlation coefficient matrices based on the correlation coefficient matrices corresponding to each variant matrix.

[0037] The set of autocorrelation values ​​and the set of correlation coefficient matrices serve as inputs for determining the regularity level, providing core indicators for the quantitative assessment of the spatial structure regularity of UAV swarms.

[0038] S4 includes the following steps: S41. Calculate the structural regularity of each variant matrix based on the improved Chaddock scaling function, and fuse them to obtain a comprehensive structural regularity and consistency index. S42. Determine the cluster structure level based on the comprehensive structural regularity, calculate the confidence level based on the comprehensive structural regularity and consistency index, and output the core parameter set and detailed data set for structural regularity determination; S41 specifically refers to: First, we define the traditional Chaddock scale as a reference, which is used to assess the correlation between variables. The basic mapping relationship is as follows: It is a very weak correlation. It is a weak correlation. It is moderately relevant. Strong correlation, It is extremely strongly correlated.

[0039] Secondly, considering the spatial structure characteristics of UAV swarms, an improved Chaddock scaling function is defined. Its input is the autocorrelation value. The structural regularity of each variant matrix was calculated based on the improved Chaddock scaling function. The specific expression is: In the formula, For the improved Chaddock scaling function, The function is continuously differentiable and strictly monotonically increasing within its domain, ensuring .

[0040] Calculate the regularity of weighted composite structure The specific expression is: In the formula, For the first variant matrix structure regularity, , For the second variant matrix structure regularity, , For the regularity of the third variant matrix structure, , Based on the structural regularity of the basic distance, To correct the regularity of distance structure weights, The neighbor distance difference structure regularity weights satisfy the following conditions: In this embodiment, the weight allocation is as follows: , , .

[0041] Calculate the consistency index of the overall structural regularity The specific expression is: In the formula, The mean of the structural regularity of the variants. ; In S42, the cluster structure level is determined based on the comprehensive structural regularity. The specific expression is: Among them, level The corresponding drones are distributed in a precise grid, level It corresponds to a completely random distribution.

[0042] Confidence level is calculated based on comprehensive structural regularity and consistency index. The specific expression is: Confidence levels are determined based on confidence scores: For high confidence, At medium confidence level, The confidence level is low.

[0043] Secondly, implement a data validation mechanism to ensure the validity of the output data: range validation: verification. Consistency verification: [This likely refers to a specific type of verification process.] Weight verification: Validation .

[0044] Generate and output the core parameter set and detailed data set for structural regularity determination. and detailed dataset The specific expression is: In the formula, For timestamps, This is the first autocorrelation value. This is the second autocorrelation value. This is the third autocorrelation value.

[0045] In this embodiment, the present invention introduces a spiking neural network to perform deep feature extraction and evaluation of the spatial structural features of the UAV swarm. This process achieves intelligent fusion of high-dimensional features by simulating the information processing mechanism of the biological nervous system.

[0046] In S5, a spiking neural network based on small-world topology is constructed, and a neuronal dynamics model of the spiking neural network is defined to simulate the firing characteristics of biological neurons. The spiking neural network consists of an input layer, a hidden layer, and an output layer connected in sequence. The input layer has 3 neurons, which are used to input the first autocorrelation value to the third autocorrelation value. The hidden layer consists of a first hidden layer and a second hidden layer connected in sequence. The first hidden layer has 128 LIF neurons, and the second hidden layer has 64 LIF neurons. The output layer has 5 neurons, which are used to output the core parameter set of brain-like computing and use softmax pulse counting decoding. The hidden layer uses a small-world network connection pattern with a connection probability of [missing information]. The specific expression is: In the formula, For neurons and neurons Functional distance between It is the attenuation constant. For small-world parameters, The average connectivity; in this embodiment, the parameter is set as: attenuation constant. Small world parameters Average connectivity ; The neuronal dynamics model is specifically expressed as follows: In the formula, This refers to the membrane potential. To adapt to the current, For film capacitors, For leakage conductivity, Slope factor This is the leakage reversal potential. Threshold potential, It is a synaptic current. To adapt to the time constant of the current, To accommodate the current sensitivity coefficient, The increment of the adaptive current after pulse triggering. The pulse firing time, The Dirac function is used to represent the function in... A pulse of time; The specific workflow of a spiking neural network is as follows: A1. Input the autocorrelation value into the spiking neural network and execute the spatiotemporal coding mechanism to convert the continuous autocorrelation value into a discrete pulse sequence pattern; A2. Employ the multi-factor impulse temporal dependent plasticity (STDP) learning rule to dynamically update synaptic weights in order to extract deep features; A3. Combining multi-timescale integration and attention mechanisms for brain-like decision-making, and mapping the decision output vector to a class probability distribution using the Softmax function. And the multi-scale feature vector is defined as brain-like feature. Simultaneously calculate the impulse activity matrix and multi-index confidence; A4. Optimize network performance based on energy efficiency constraints and noise robustness, and output a set of core evaluation parameters for neuromorphic computing; In A1, the specific method for implementing the spatiotemporal coding mechanism is as follows: Three parallel coding strategies are employed to encode the autocorrelation values: rate coding, time coding, and phase coding. A distributed population coding strategy is then used to integrate the encoded features, outputting a pulse sequence pattern. ; In the formula, For the nonlinear mapping function from features to impulse modes, The number of encoding neurons for each feature. For background noise, The weights of the encoded neurons, The input is the autocorrelation value; In A2, the method for dynamically updating synaptic weights using multi-factor impulse temporally dependent plasticity learning rules is as follows: Calculate the synaptic weight increment, and update the synaptic weights using the synaptic weight increment. The specific expression is: In the formula, The global learning rate, The fusion coefficient is the pulse timing-dependent plasticity (STDP) factor. The fusion coefficient is the steady-state plasticity factor. The fusion coefficient of the reward modulation factor, For STDP weights, As steady-state plasticity weights, Adjust the weights for rewards; In the formula, To enhance the amplitude of STDP, For STDP suppression amplitude, To enhance the time constant, To suppress the time constant, This represents the upper threshold of synaptic weights. For the synaptic weights before the update, The time difference between the pulses of the preceding and following neurons; In the formula, This is the steady-state plasticity adjustment coefficient. For neurons emission rate, Parameters for maintaining the target emissivity; in this embodiment, the steady-state plasticity weight is used to maintain the target emissivity. ; In the formula, This is a dynamic reward signal based on classification accuracy; A3 specifically refers to: (1) Perform multi-timescale integration on the spiking activity of the output layer neurons to obtain the brain-like decision output vector. ; In the formula, and These are the short-term and long-term integration windows, respectively. For short-term integral weights, For long-term integral weights, For pulse firing function, For integration variables; (2) Use the Softmax function to process the output vector Perform normalization processing and output the class probability distribution. ; In the formula, This represents the class probability distribution of the output of a spiking neural network. Represents the normalized exponential function, For the Softmax function, For the output layer The integral output of each neuron For the output layer The integral output of each neuron This represents the number of neurons in the output layer. The maximum value indexing method is used to index the output vector. Decoding is performed to obtain the structural hierarchy of the spiking neural network output. ; In the formula, To extract the index at which the function reaches its maximum value. j ; (3) Using a multi-scale integration strategy, the spiking activity of hidden layer neurons is analyzed. Time windows of different lengths Perform pulse counting and integration separately to obtain The integral results at different time scales, the th k Integration results on time scale The specific expression is: In the formula, In the first k On a time scale, the corresponding hidden layer neurons at time 10:00 The pulse firing state vector, , For the first neuron in the hidden layer at time t... The pulse firing state vector, For the second neuron in the hidden layer at time t... The pulse firing state vector, For the H-th neuron in the hidden layer at time t... The pulse firing state vector, For the first The width of the integration window corresponding to each time scale; Concatenate all integration results to obtain multi-scale feature vectors. ; In the formula, This is the integral result for the first time scale. This is the integral result for the second time scale. For the first Integration results over time; This multi-scale feature vector is used as the brain-like feature output, i.e. This is used for subsequent heterogeneous model fusion; (4) Dynamically construct a spur activity matrix through attention neural modulation to achieve selective enhancement of specific feature channels. The specific expression is: In the formula, For the first One neuron in The modulation intensity of the pulse activity at time t, For the first One neuron in The modulation intensity of the pulse activity at time t, For the first One neuron in The modulation intensity of the pulse activity at time t, For the first One neuron in The modulation intensity of the pulse activity at time t, For diagonalization operation, The Sigmoid activation function maps the calculation result to the (0, 1) interval, representing the modulation intensity. Here is the learnable weight matrix in the attention subnetwork. For learnable biases in attention subnetworks, For the first The average firing rate of each neuron within a recent time window. In the formula, , For the hidden layer i Neurons at time The pulse firing state vector; (5) Calculate the multi-index confidence score, which includes synchronicity, entropy, and stability. The specific expression is: In the formula, As a synchronicity indicator, As an entropy index, For stability indicators; In the formula, The number of neurons in the output layer. For reference signal, For output signal, The standard deviation of the output signal. The standard deviation of the reference signal; In the formula, To output the Shannon entropy of the distribution, This is the category probability distribution vector; In the formula, Let be the output probability distribution at time t. This represents the output probability distribution at the previous time step; In A4, the specific method for optimizing network performance based on energy efficiency constraints and noise robustness is as follows: An optimization objective function is constructed based on energy efficiency constraints to improve network performance. The specific expression is: In the formula, Total energy consumption, Energy consumption per pulse For neurons emission rate, Leakage current energy consumption rate This is the energy consumption coefficient for synaptic transmission. For synaptic weights, For preneurons Emission rate; Noise robustness processing is introduced to enhance the stability of the network under uncertain environments. The specific expression is: In the formula, For the original input, It is Gaussian noise. To distribute noise evenly, For uniform noise amplitude, The standard deviation of Gaussian noise; Output the core evaluation parameter set for neuromorphic computing The specific expression is: In the formula, The structure level output by the SNN. For multi-scale feature vectors, For pulse activity matrix; In S5, the specific workflow of the gradient boosting decision tree model is as follows: B1. Set of autocorrelation values Parameters for determining structural regularity Perform dimension alignment and concatenation to construct the input feature vector of the gradient boosting decision tree model. Its specific expression is: B2. By constructing a decision tree sequentially, each new tree is specifically used to correct the residual of the previous tree, and its optimization objective function is... The specific expression is: In the formula, For drones The predicted loss, For the first The regularization term for the trees, For ensemble models (i.e., the SNN-CatBoost heterogeneous ensemble learning framework) for UAVs The predicted output, For drones The true label, K For the total number of decision trees, For drones The input feature vector; The degree-boosting decision tree model employs an ordered boosting strategy to avoid target leakage. Its gradient estimation formula is as follows: In the formula, For drones The predicted loss, To integrate models for drones The predicted output, For drones The input feature vector, For drones The true label, For gradient operators, and The random arrangement of the training samples; In one specific parameter configuration, the model parameters are set as follows: tree depth of 6, learning rate of 0.05, number of iterations of 200, and L2 regularization coefficient of 3.0.

[0047] B3. Input feature vector Input to Decision Tree In the ensemble model, for each decision tree The mapping function is defined using its internal splitting criterion and topological structure. This maps the continuous high-dimensional feature space to a discrete leaf node index space. ,in For the first The total number of leaf nodes in the tree, and the input feature vector. Index of active leaf nodes in each tree It is uniquely determined by the following formula: B4. Generate the corresponding index for each leaf node output by each decision tree. One-Hot Sparse Encoding Vector The first vector element The rule for determining the value of is defined as follows: if and only if hour, ,otherwise Through this discrete coding mechanism, the original numerical features are... Transformed into a structured encoding vector representing the combination of decision paths; B5, will One-hot encoded vectors generated by decision trees The final statistical features are synthesized by concatenating the vertical vectors according to the generation order of the decision tree. ; In the formula, For the first The one-hot encoding vector of each decision tree. For the first The one-hot encoding vector of each decision tree. For the first The one-hot encoding vector of each decision tree.

[0048] In this embodiment, the final statistical characteristics Total dimensions The sum of the number of leaf nodes in all decision trees, and the positional distribution of its non-zero elements, fully characterize the nonlinear mapping path of the input features within the gradient boosting decision tree model, providing a high-order feature representation for subsequent heterogeneous feature fusion.

[0049] like Figure 3As shown in S5, a heterogeneous model fusion strategy is executed through a heterogeneous ensemble learning framework. For the brain-like features output by the spiking neural network and the statistical features output by the gradient boosting decision tree model, the brain-like features and statistical features are concatenated at the feature level to generate a fused feature vector. At the decision level, a decision-level fusion mechanism is executed to perform weighted fusion of the output probabilities of the spiking neural network and the gradient boosting decision tree to obtain the final probability distribution. The fused feature vector and the final probability distribution are used as the ensemble evaluation results. Among them, the fused feature vector The specific expression is: In the formula, It has brain-like characteristics. Statistical characteristics; Subsequently, a decision-level fusion mechanism is implemented to weight and fuse the output probabilities of the two models, resulting in the final probability distribution. The specific expression is: In the formula, This represents the class probability distribution output by the spiking neural network. To improve the class probability distribution output by the gradient boosting decision tree model, For dynamic fusion weights; In the formula, To adjust the parameters, The classification accuracy of the spiking neural network on the validation set. To improve the classification accuracy of the decision tree model on the validation set using gradient descent; In S5, the specific method for collaboratively optimizing the training framework is as follows: (1) A two-stage training strategy is adopted: Independent pre-training stage: the spiking neural network is pre-trained in an unsupervised manner using the STDP rule, and the gradient boosting decision tree model is trained in a supervised manner on traditional features; Joint fine-tuning stage: the parameters of the spiking neural network are fixed, the gradient boosting decision tree model is trained using fused features, and the fused parameters are adjusted based on the performance of the validation set. (2) Define the overall loss function for collaborative optimization training. ; In the formula, To improve the prediction classification loss of the gradient boosting decision tree model, The composite loss of a spiking neural network (SNN) includes classification error loss and burst rate constraint loss. A consistency penalty term is added to the output probability distribution of the heterogeneous model to constrain the consistency of the decision logic between the SNN branch and the CatBoost branch. The adjustment weights for the SNN loss. The adjustment weight for consistency loss.

[0050] S6 includes the following steps: S61. Based on confidence level and final probability distribution Calculate the final confidence level. The specific expression is: In the formula, The set weighting coefficients, , To obtain the maximum probability value in the vector; In this embodiment, the final confidence score linear synthesis derivation employs a linear weighted fusion algorithm to synthesize the aforementioned components in order to balance the regularity indicators of spatial physical characteristics with the statistical judgment results of the deep learning model. In one specific implementation, the weighting coefficients... The value is 0.6, which is used to strengthen the weight of the model prediction results in the confidence assessment.

[0051] S62. Calculate the final cluster structure level based on the final probability distribution. The specific expression is: In the formula, For the final probability distribution The corresponding number in the middle The probability components of each structural level, The function for the maximum element index. For discrete hierarchical space, It indicates a high degree of structure. This indicates good structure. Indicates medium structure. Indicates weak structuring. This indicates no structure; in this embodiment, the invention is based on the maximum a posteriori probability criterion (MAP) in discrete hierarchy space. Find the index value that maximizes the fusion probability.

[0052] S63. Based on the final cluster structure level, the overall accuracy is calculated by comparing the truth labels to verify the system accuracy. The specific expression is: In the formula, The total number of test samples, For test samples The final cluster structure level, For test samples The comparison of truth labels, The value is the Kronecker delta, which takes the value 1 if and only if the predicted value matches the actual value, and 0 otherwise.

[0053] S64. To quantify the degree of certainty of the evaluation results, the final probability distribution is... Perform Shannon entropy calculation to determine the Shannon entropy value. The specific expression is: If entropy value The closer the entropy value is to 0, the higher the certainty and reliability of the evaluation result; if the entropy value is close to its theoretical maximum value... If so, the system is determined to be in an uncertain state; S65. Introduce variance into the input data. Gaussian noise Observe the fluctuation of the final confidence level to determine its robustness and stability index. The specific expression is: In the formula, The final confidence level of the integrated system after introducing Gaussian noise. The final confidence level of the integrated system based on the original input data; S66, When the overall accuracy Shannon entropy and robustness and stability index If the value exceeds the preset threshold, the performance verification is deemed successful, and the core output parameter set of the integrated system is then generated. ; In this embodiment, the present invention performs robustness testing by calculating overall accuracy, Shannon entropy, and robustness stability indices, introducing different levels of noise into the input data. The system's accuracy is maintained under different noise levels. In addition, the generalization ability is verified by evaluating the model's performance on the same distribution test set, cross-distribution test set, and extreme scenario test set.

[0054] In the description of this invention, the above are merely preferred embodiments and are not intended to limit the scope of protection of this invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for evaluating the regularity of spatial structure of a UAV swarm based on brain-like computing, characterized in that, The method comprises the following steps: S1, calculating optimal pose estimation of the UAV cluster in a global coordinate system according to multi-source sensor data fusion and a factor graph optimization model, and extracting a global position set at a selected time; S2, constructing three variant matrices of multiple distance deviations by calculating Euclidean distance, mean removal processing and difference operation according to the global position set, and representing the spatial structure of the UAV group from multiple angles; S3, defining a Pearson correlation coefficient standard, and calculating a correlation coefficient matrix and an autocorrelation value of each variant matrix; S4, mapping the autocorrelation value to structural regularity by an improved Chaddock scale function, calculating confidence by combining a consistency index, and outputting a core parameter set and a detailed data set of structural regularity determination; S5, constructing a heterogeneous ensemble learning framework based on a spiking neural network and a gradient boosting decision tree model, cooperatively optimizing the training framework, inputting the autocorrelation value into the spiking neural network, inputting the autocorrelation value and the structural regularity into the gradient boosting decision tree model, and executing a heterogeneous model fusion strategy to obtain an integrated evaluation result; S6, based on the integrated evaluation result, verifying and evaluating system performance to generate a core output parameter set of the integrated system.

2. The method of claim 1, wherein the method is performed by a computer system. S1 comprises the following steps: S11, setting a UAV cluster size and an observation time sequence, defining a single-machine state vector and a system global state vector, and integrating multi-source sensor data, wherein the multi-source sensor data comprises odometer observation, RTK global positioning observation and relative observation; S12, constructing target functions of the odometer observation, the RTK global positioning observation and the relative observation, and establishing a factor graph model by comprehensively considering the target functions; S13, linearizing an error function at a current estimate by using a sparse structure of the factor graph model, solving a normal equation by using an incremental optimization algorithm, and obtaining optimal pose estimations of all UAVs at all times; S14, extracting a global position set at a selected time according to the optimal pose estimations of all UAVs at all times.

3. The method of claim 2, wherein the method further comprises: In S11, the expression of the single-machine state vector and the system global state vector is defined as follows: wherein is the UAV i is the horizontal axis coordinate of the UAV at time t is the UAV i is the vertical axis coordinate of the UAV at time t is the UAV i is the vertical axis coordinate of the UAV at time t is the UAV i is the roll angle of the UAV at time t is the UAV i is the pitch angle of the UAV at time t is the UAV i is the yaw angle of the UAV at time t is the transpose symbol, is the UAV cluster size, is the observation time;​​​​​​ integrating multi-source sensor data, which specifically includes odometer observation , representing the unmanned aerial vehicle relative motion observation from time to time RTK global positioning observation , representing the unmanned aerial vehicle absolute global position coordinates of the unmanned aerial vehicle at time , representing the unmanned aerial vehicle relative position vector of the unmanned aerial vehicle at time ​ In S12, an expression of the factor graph model is specifically as follows: wherein, is the optimal pose estimate of all UAVs at all time instants, is the objective function of odometry observations, is the objective function of RTK global positioning observations, is the objective function of relative observations, is the error function of relative observations, is the error function of RTK global positioning observations, is the error function of odometry observations, , and denotes the Mahalanobis distance, is the covariance matrix of odometry observation noise, is the covariance matrix of RTK observation noise, is the covariance matrix of relative observation noise, is the set of UAVs for which there are RTK observations, is the set of UAV pairs for which there are relative observations at time instant t; wherein is the relative transformation matrix from the state predicted adjacent time instant; wherein is a transformation matrix converted from a transformation matrix converted from is a transformation matrix converted from a transformation matrix converted from wherein is a rotation matrix composed of , is a three-dimensional position coordinate of the UAV , , is a three-dimensional special Euclidean group, representing a group composed of all rigid body transformation matrices in three-dimensional space; In the formula, a drone three-dimensional position coordinates; In S13, an expression of the normal equation is specifically as follows: In the formula, is a Jacobian matrix, is a parameter increment, is a block diagonal noise covariance matrix, , is a block diagonal matrix constructor, which arranges the input matrix parameters along the main diagonal to form a block diagonal matrix, is a block diagonal matrix, is an error vector calculated according to the odometer, RTK and relative observation at is a current estimated value; wherein e is the total error vector, is the state error component, is the observation error component; A method for solving the normal equation by using the incremental optimization algorithm is specifically as follows: S131, building a Hessian matrix according to the sparse structure of the factor graph model ; In the formula, corresponding to the UAV and the state variable of the UAV , is the state variable of the UAV itself, is composed of the odometer observation and the RTK global positioning observation of the UAV itself, is only non-zero when there is a relative observation between the UAV and the UAV ; In S132, a first linear equation is established by using a Levenberg-Marquardt algorithm, and an expression of the first linear equation is specifically as follows: wherein is a diagonal matrix whose main diagonal elements are taken from the matrix, is a damping factor; S133, solve the first linear equation to obtain the increment , and perform iterative updating according to the increment by the following formula to obtain the optimal pose estimation of all unmanned aerial vehicles at all times ; wherein is an additive operation on the manifold, is the system state vector estimate at the th iteration, is the system state vector estimate at the th iteration; In S14, the global position set of the selected time point is extracted The expression is specifically: In the formula, is a UAV at time is a three-dimensional position coordinate of the UAV, , is a UAV i at time is a horizontal axis coordinate of the UAV, is a UAV i at time is a vertical axis coordinate of the UAV, is a UAV i at time is a vertical axis coordinate of the UAV.

4. The method of claim 3, wherein the method further comprises: S2 comprises the following steps: S21, calculating Euclidean distance and extracting ordered neighbor distance according to the global position set, and constructing a first variant matrix; S22, calculating a row mean of the first variant matrix, and constructing a second variant matrix for measuring distance deviation by mean removal processing; S23, performing difference processing on the ordered neighbor distance, and constructing a third variant matrix for enhancing sensitivity of a local dense structure.

5. The method of claim 4, wherein, S21 is specifically as follows: S211、According to the global position set, the Euclidean distance between any two unmanned aerial vehicles is calculated, and a symmetric distance matrix is constructed R ; In the formula, is a drone k with the drone l between the drone and the object; S212, based on the symmetric distance matrix, obtain a distance set of any unmanned aerial vehicle to all other unmanned aerial vehicles, perform ascending sequence arrangement on elements in the distance set through an ordered neighbor distance extraction operation, and select a value of the number of elements to generate a neighbor distance set, wherein the neighbor distance set of the unmanned aerial vehicle is obtained through the ordered neighbor distance extraction operation m The expression of the number of elements is specifically k The expression of the number of elements is specifically ​ wherein is a drone k a set of distances to all other drones, , is an ordered neighbor distance extraction operation, is a drone k a distance to its first nearest neighbor, is a drone k a distance to its second nearest neighbor, is a drone k a distance to its nearest neighbor, , is a min function; S213. Constructing a first variant matrix from the set of neighbor distances ; wherein is the relative position vector of the UAV k , is the relative position vector of the UAV N to its first nearest neighbor; In S22, the second variant matrix The expression of the second variant matrix is specifically: wherein is the first variant matrix is the first row mean of the first variant matrix is the second row mean of the first variant matrix is the third row mean of the first variant matrix is the fourth row mean of the first variant matrix is the fifth row mean of the first variant matrix is the sixth row mean of the first variant matrix is the seventh row mean of the first variant matrix is the eighth row mean of the first variant matrix is the ninth row mean of the first variant matrix wherein is the first variant matrix is the first row, the first column element of the first variant matrix , is the distance of the UAV to its first neighbor. In S23, the third variant matrix The expression of the third variant matrix is specifically: wherein is a first variant matrix is a first variant matrix is a first variant matrix wherein is the difference of the first variant matrix is the difference of the first variant matrix is the difference of the first variant matrix is the distance of the UAV to its first nearest neighbor, is the distance of the UAV to its first nearest neighbor, is the distance of the UAV to its first nearest neighbor, is the distance of the UAV to its first nearest neighbor, is the distance of the UAV to its first nearest neighbor, is the distance of the UAV to its first nearest neighbor.

6. The method of claim 5, wherein the method further comprises: S3 comprises the following steps: S31, calculating a correlation coefficient matrix corresponding to each variant matrix according to a Pearson correlation coefficient based on the three constructed variant matrices, and quantifying pairwise correlation within the cluster; S32, aggregating the correlation coefficient matrix, calculating an autocorrelation value of each variant matrix, and obtaining a macro regularity index; S33, analyzing autocorrelation value characteristics, and outputting an autocorrelation value set and a correlation coefficient matrix set; In S31, a correlation coefficient matrix corresponding to each variant matrix is calculated, and an element of the correlation coefficient matrix is The expression of the element of the correlation coefficient matrix is specifically: wherein is the number of variants is the number of variants i is the number of variants k is the number of variants is the number of variants is the number of variants i is the number of variants l is the number of variants is the number of variants is the number of variants k is the number of variants is the number of variants is the number of variants l is the number of variants m is the number of variants In S32, the autocorrelation values of the respective variant matrices are calculated The expression of the autocorrelation value is specifically: In S33, the value range of the autocorrelation value is The physical meaning is defined as follows: The autocorrelation value is about equal to 1, representing a highly regular spatial structure, i.e., the UAVs are uniformly and orderly distributed; The autocorrelation value is about equal to 0, representing random distribution, without obvious regularity; The autocorrelation value is about equal to -1, representing a highly irregular distribution, with obvious repulsion or aggregation phenomenon; An autocorrelation value set is obtained according to the autocorrelation values of the variant matrices, and a correlation coefficient matrix set is obtained according to the correlation coefficient matrices corresponding to the variant matrices.

7. The method of claim 6, wherein the method further comprises: S4 includes the following sub-steps: S41, calculating the structural regularity of each variant matrix based on the improved Chaddock scale function, and fusing to obtain a comprehensive structural regularity and consistency index; S42, determining the cluster structure level according to the comprehensive structural regularity, calculating the confidence according to the comprehensive structural regularity and consistency index, and outputting a core parameter set and a detailed data set of the structural regularity determination; In S41, the structural regularity of each variant matrix is calculated based on the improved Chaddock scale function The expression is specifically: wherein is the improved Chaddock scale function; Computing weighted integrated structural regularity The expression for the weighted integrated structural regularity is given by: wherein is a first variant matrix structure regularity, is a second variant matrix structure regularity, is a third variant matrix structure regularity, is a base distance structure regularity weight, is a correction distance structure regularity weight, is a neighbor distance difference structure regularity weight; Computing a consistency index of the overall structural regularity The expression of the consistency index of the overall structural regularity is specifically: In the formulae, is the mean value of the regularity of the variant structure, ; In S42, the cluster structure level is determined according to the comprehensive structure regularity The expression is specifically: The confidence is calculated according to the comprehensive structure regularity and consistency index The expression is specifically: Core parameter set for structural regularity determination and detailed data set The expression is specifically: In the formula, is a timestamp, is a first autocorrelation value, is a second autocorrelation value, is a third autocorrelation value.

8. The method of claim 7, wherein the method further comprises: In S5, a pulse neural network based on a small-world topology is constructed, and a neuron dynamics model of the pulse neural network is defined to simulate the discharge characteristics of biological neurons; The pulse neural network includes an input layer, a hidden layer and an output layer connected in sequence, the input layer is provided with 3 neurons for inputting the first autocorrelation value to the third autocorrelation value; the hidden layer includes a first hidden layer and a second hidden layer connected with each other, the first hidden layer is provided with 128 LIF neurons, the second hidden layer is provided with 64 LIF neurons, and the output layer is provided with 5 neurons for outputting a core parameter set of brain-like computing; The hidden layer adopts a small-world network connection mode, and the expression of connection probability is specifically wherein is the neuron and the functional distance between neurons , is the decay constant, is the small-world parameter, is the average connectivity degree; The expression of the neuron dynamics model is specifically: wherein, is the membrane potential, is the adaptation current, is the membrane capacitance, is the leak conductance, is the slope factor, is the leak reversal potential, is the threshold potential, is the synaptic current, is the time constant of the adaptation current, is the sensitivity coefficient of the adaptation current, is the increment of the adaptation current after a spike trigger, is the spike firing time, is the Dirac function, used to represent a spike at time. The working process of the pulse neural network is specifically: A1, input the autocorrelation value into the pulse neural network, and execute the spatiotemporal coding mechanism to convert the continuous autocorrelation value into a discrete pulse sequence pattern; A2, dynamically updating the synaptic weight by adopting a multi-factor pulse temporal dependence plasticity learning rule to extract deep features; A3, combine multi-time scale integration with attention mechanism for brain-like decision making, map the decision output vector to a category probability distribution through a Softmax function , and define the multi-scale feature vector as a brain-like feature , while calculating the pulse activity matrix and multi-index confidence A4, optimizing the network performance based on energy efficiency constraint and noise robustness, and outputting a core evaluation parameter set of brain-like computing; In A1, the method for executing the spatiotemporal coding mechanism is specifically: Three parallel encoding strategies are used to encode the autocorrelation values, including rate encoding, time encoding and phase encoding, and a distributed swarm encoding strategy is used to integrate the encoded features to output the pulse sequence pattern ; wherein is a nonlinear mapping function of the features to the pulse pattern, is the number of encoding neurons for each feature, is a background noise term, is the weight of the encoding neurons, is the autocorrelation value of the input; In A2, the method for dynamically updating the synaptic weight by adopting the multi-factor pulse temporal dependence plasticity learning rule is specifically: Calculate the synaptic weight increment, and update the synaptic weights using the synaptic weight increment. The specific expression is: wherein, is a global learning rate, is a fusion coefficient for the pulse timing dependent plasticity factor, is a fusion coefficient for the homeostatic plasticity factor, is a fusion coefficient for the reward modulation factor, is an STDP weight, is a homeostatic plasticity weight, is a reward modulation weight; wherein is the STDP potentiation amplitude, is the STDP depression amplitude, is the potentiation time constant, is the depression time constant, is the upper threshold of the synaptic weight, is the synaptic weight before update, is the time difference between the pre- and post-synaptic neuron spikes; wherein is a steady-state plasticity adjustment coefficient, is a firing rate of a neuron is a firing rate of a neuron is a parameter to maintain a target firing rate; In the formula, is a dynamic reward signal based on classification accuracy; A3 is specifically: (1) Multi-time scale integration of the pulse activities of the output layer neurons to obtain an output vector of the brain-like decision ; wherein and are short and long integration windows, respectively, is a short integration weight, is a long integration weight, is a pulse emission function, is an integration variable; (2) The output vector is normalized by using a Softmax function to output a class probability distribution ;​ wherein denotes the class probability distribution of the output of the pulse neural network, denotes the normalized exponential function, is the Softmax function, is the integral output of the output layer neuron number is the integral output of the output layer neuron number is the integral output of the output layer neuron number is the integral output of the output layer neuron number is the number of output layer neurons; The maximum value indexing method is used to index the output vector. Decoding is performed to obtain the structural hierarchy of the spiking neural network output. ; In the formula, to extract the index at which the function attains a maximum value j ; (3) Using multi-scale integration strategy, the pulse activity of hidden layer neurons is counted and integrated with time windows of different lengths respectively to obtain integration results of different time scales The expression of the integration result of the first time scale is as follows: k ​​​ wherein, is the first time scale, k is the second time scale, is the pulse firing state vector of the corresponding hidden layer neuron at time , is the pulse firing state vector of the first hidden layer neuron at time , is the pulse firing state vector of the second hidden layer neuron at time , is the pulse firing state vector of the Hth hidden layer neuron at time , is the integral window width corresponding to the th time scale; Vector splicing all integral results, get multi-scale feature vector ; wherein is the integral result of the 1st time scale, is the integral result of the 2nd time scale, is the integral result of the 1st time scale, is the integral result of the 1st time scale; (4) A pulse activity matrix is dynamically constructed through an attention neural modulation mechanism, used to realize selective enhancement of a specific feature channel, and the expression of the pulse activity matrix is specifically as follows: The expression is specifically as follows: In the formula, For the first One neuron in The modulation intensity of the pulse activity at time t, For the first One neuron in The modulation intensity of the pulse activity at time t, For the first One neuron in The modulation intensity of the pulse activity at time t, For the first One neuron in The modulation intensity of the pulse activity at time t, For diagonalization operation, The Sigmoid activation function maps the calculation result to the (0, 1) interval, representing the modulation intensity. Here is the learnable weight matrix in the attention subnetwork. For learnable biases in attention subnetworks, For the first The average impulse firing rate of each neuron within a recent time window; In the formula, , is the hidden layer first i neuron at time the pulse firing state vector; (5) calculating a multi-index confidence degree containing synchronism, entropy value and stability, the expression of the multi-index confidence degree is specifically: wherein is a synchronicity index, is an entropy index, is a stability index; wherein is a reference signal, is an output signal, is an output signal standard deviation, is a reference signal standard deviation; wherein is the Shannon entropy of the output distribution, is the class probability distribution vector; wherein is the output probability distribution at time t, is the output probability distribution at the previous time. In A4, the method for optimizing the network performance based on the energy efficiency constraint and the noise robustness is specifically: An optimization objective function is constructed based on energy efficiency constraints, which is used to improve network performance The expression of the optimization objective function is specifically as follows: wherein, is the total energy consumption, is the single pulse energy consumption, is the firing rate of a neuron is the firing rate of a neuron is the leakage current energy consumption rate, is the synaptic transmission energy consumption coefficient, is the synaptic weight, is the firing rate of a neuron is the firing rate of a neuron A noise robustness process is introduced to enhance the stability of the network in uncertain environments, the noise robustness process The expression of the noise robustness process is specifically: wherein is the original input, is the Gaussian noise, is the uniform distribution noise, is the uniform noise amplitude, is the Gaussian noise standard deviation; Core evaluation parameter set of output brain-like computing The expression is specifically: wherein is a structure level output by the SNN, is a multi-scale feature vector, is a spike activity matrix; In S5, the working process of the gradient boosting decision tree model is specifically: B1、the autocorrelation value set with the structural regularity determination parameter dimension alignment and splicing are performed, and an input feature vector of the gradient boosting decision tree model is constructed The expression is specifically: B2. By sequentially constructing decision trees, each new tree specially corrects the residual error of the previous tree, and the optimization objective function The expression of the objective function is specifically: wherein is the predicted loss of the UAV, is the regularization term of the th tree, is the predicted output of the ensemble model for the UAV, is the true label of the UAV, K is the total number of decision trees, is the input feature vector of the UAV, is the input feature vector of the UAV,​​​ The gradient boosting decision tree model adopts an ordered boosting strategy to avoid target leakage, and the gradient estimation formula is: wherein is the prediction loss of the UAV , is the prediction output of the integrated model for the UAV , is the input feature vector of the UAV , is the true label of the UAV , is the gradient operator, and is the random permutation order of the training samples; B3. Input feature vector Input to Decision Tree In the ensemble model, for each decision tree The mapping function is defined using its internal splitting criterion and topological structure. This maps the continuous high-dimensional feature space to a discrete leaf node index space. ,in For the first The total number of leaf nodes in the tree, and the input feature vector. Index of active leaf nodes in each tree It is uniquely determined by the following formula: B4. For each leaf node index output by the decision tree, generate a corresponding sparse one-hot encoding vector , where the value of the th element of the vector is defined as: if and only if , otherwise ;​​ B5, will One-hot encoded vectors generated by decision trees The final statistical features are synthesized by concatenating the vertical vectors according to the generation order of the decision tree. ; wherein is the decision tree for the one-hot encoded vector for the decision tree for the is the decision tree for the one-hot encoded vector for the decision tree for the is the decision tree for the one-hot encoded vector for the decision tree for the 9. The method of claim 8, wherein the method further comprises: In S5, a heterogeneous model fusion strategy is executed by a heterogeneous ensemble learning framework, for the brain-like features output by the pulse neural network and the statistical features output by the gradient boosting decision tree model, the brain-like features and the statistical features are spliced at the feature level to generate a fusion feature vector, a decision-level fusion mechanism is executed at the decision level to weight and fuse the output probabilities of the pulse neural network and the gradient boosting decision tree, to obtain a final probability distribution, and the fusion feature vector and the final probability distribution are taken as an integrated evaluation result; Wherein, the expression of the fusion feature vector is specifically: In the formula, is a brain-like feature, is a statistical feature; Final probability distribution The expression for the final probability distribution is given by wherein is a class probability distribution output by the pulse neural network, is a class probability distribution output by the gradient boosting decision tree model, is a dynamic fusion weight; wherein is the tuning parameter, is the classification accuracy of the pulse neural network on the validation set, is the classification accuracy of the gradient boosting decision tree model on the validation set; In S5, the method for the collaborative optimization training framework is specifically: (1) Two-stage training strategy is adopted: independent pre-training stage: the spiking neural network is pre-trained unsupervisedly through the STDP rule, and the gradient boosting decision tree model is trained supervisedly on traditional features; joint fine-tuning stage: the parameters of the spiking neural network are fixed, the gradient boosting decision tree model is trained using the fusion features, and the fusion parameters are adjusted based on the performance of the validation set; (2) defining an overall loss function for the co-optimization training ; wherein, is a prediction classification loss of the gradient boosting decision tree model, is a compound loss of the spiking neural network, is a consistency penalty term of the output probability distribution of the heterogeneous model, used to constrain the consistency of the decision logic of the SNN branch and the CatBoost branch, is a regulation weight of the SNN loss, is a regulation weight of the consistency loss.

10. The method of claim 8, wherein the method further comprises: S6 comprises the following sub-steps: S61、According to the confidence and the final probability distribution The final confidence is calculated, and the expression of the final confidence is specifically: wherein is a set weight coefficient, , is the maximum probability value taken in the vector; S62, calculating a final cluster structure level according to the final probability distribution, the final cluster structure level The expression is specifically: wherein is the final probability distribution corresponding to the probability component of the is the maximum element index function is the discrete level space denotes highly structured denotes well structured denotes medium structured denotes weak structured denotes unstructured S63、based on the final cluster structure level, the overall accuracy rate is calculated by comparing the true value label to verify the system accuracy, and the overall accuracy rate The expression of the overall accuracy rate is specifically: wherein is the total number of test samples, is the test sample is the final cluster structure level of the test sample is the contrast true value label of the test sample is the contrast true value label of the test sample is the Kronecker function; S64, in order to quantify the certainty of the evaluation result, the final probability distribution Shannon entropy calculation is performed to calculate the Shannon entropy value The expression is specifically: S65, introducing Gaussian noise with variance in the input data , observing the fluctuations of the final confidence, the expression of the robustness stability index is: wherein is the final confidence of the ensemble system with Gaussian noise introduced, is the final confidence of the ensemble system under the original input data; S66, when the overall accuracy , the shannon entropy value and the robust stability index When the preset threshold is exceeded, it is determined that the performance is verified, and then a core output parameter set of the integrated system is generated ; 。

Images