Unmanned aerial vehicle ads-b regulation command execution difference intelligent evaluation system fusing deep learning

Abstract

The application relates to the technical field of unmanned aerial vehicle control, and discloses a difference intelligent evaluation system for unmanned aerial vehicle ADS-B control instruction execution, which is fused with deep learning and comprises the following modules: a data acquisition module, which is used for acquiring ADS-B control instruction runtime sequence data of an unmanned aerial vehicle and a manned aircraft; a data preprocessing module, which is used for cleaning, standardizing and time sequence aligning the sequence data; a time sequence feature reconstruction module, which is used for constructing the preprocessed time sequence data into a multidimensional time sequence feature matrix; and a time sequence VAE evaluation module, which is used for constructing a time sequence VAE evaluation model and training and optimizing the model, wherein the time sequence VAE evaluation model comprises an encoder module, a decoder module and a difference quantification module, and the model is used for outputting a quantitative difference evaluation result of the unmanned aerial vehicle and the manned aircraft in an ADS-B control instruction execution process. The application can realize all-around, high-precision and interpretable quantitative evaluation of the difference between a medium / large unmanned aerial vehicle and a manned aircraft in an ADS-B control instruction execution process.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) control technology, and in particular to an intelligent evaluation system for UAV ADS-B control instruction execution differences that integrates deep learning. Background Technology

[0002] With the rapid development of unmanned aerial vehicle (UAV) technology, medium and large UAVs are increasingly being used in logistics transportation, emergency rescue, power line inspection, and other fields. Their integration into the national airspace system and their combined operation with manned aircraft has become a key direction for promoting the efficient development of the aviation industry. ADS-B (Automatic Dependent Surveillance-Broadcast) technology, with its advantages of strong real-time performance and high surveillance accuracy, has become the core surveillance means for manned aircraft airspace management. Medium and large UAVs already possess the basic conditions for installing ADS-B equipment and establishing two-way communication with control units in terms of onboard equipment payload and communication link construction. Therefore, exploring the sharing of ADS-B control rules between medium and large UAVs and manned aircraft has significant practical implications.

[0003] However, significant differences exist between medium and large-sized UAVs and manned aircraft in the execution of ADS-B control commands, making it difficult for existing ADS-B control rules designed for manned aircraft to be directly adapted to mixed airspace operation scenarios. Specifically, medium and large-sized UAVs suffer from heterogeneous pilot loops, differences in command and control link stability, and insufficient onboard perception and collision avoidance capabilities, causing their safety level of operation (ELOS) to not yet reach the equivalent standard for manned aircraft. Furthermore, communication delays between UAV ground stations and aircraft exhibit temporal fluctuations, ADS-B surveillance signals (position, altitude accuracy, and stability) are dynamically changing due to terrain obstruction, and UAVs show temporal correlation differences in their response to control commands such as heading, altitude, and speed. These dynamic temporal differences directly lead to risks such as conflicts in mixed airspace operations and low airspace utilization.

[0004] In existing technologies, traditional evaluation methods mostly rely on static statistical models or single-dimensional feature analysis, which cannot effectively capture the dynamic temporal correlation features during the execution of control commands, and are difficult to quantify the execution differences between UAVs and manned aircraft in the temporal dimension. While conventional deep learning models (such as CNNs and basic VAEs) can handle high-dimensional data, they lack the ability to model dependencies in long-term data, and cannot accurately characterize dynamic differences such as communication delay fluctuations and command response temporal evolution. In addition, existing models have many technical shortcomings: First, they lack a modeling module that can efficiently linearize nonlinear temporal features, making it difficult to balance feature representation ability and computational efficiency; second, the encoder structure is simple and cannot simultaneously achieve dynamic error correction, long-term dependency capture, and interpretability labeling functions; third, the decoder fails to fully link the key information of the encoder, resulting in poor reconstruction accuracy and retention of differential features; fourth, the loss function fails to take into account reconstruction accuracy, distribution rationality, and interpretability, and the model training is prone to bias; fifth, the optimization algorithm is prone to getting trapped in local optima when dealing with high-dimensional key parameters, making it difficult to balance the relationship between evaluation accuracy, computational efficiency, and interpretability. These issues prevent existing technologies from achieving accurate, efficient, and interpretable assessments of discrepancies in the execution of ADS-B control instructions for drones.

[0005] Therefore, there is an urgent need to develop a deep learning system that integrates innovative modeling modules, optimized network structures, and efficient optimization algorithms to provide technical support for the optimization and safety management of hybrid airspace control rules. Summary of the Invention

[0006] This invention aims to provide an intelligent evaluation system for the differences in ADS-B control instruction execution between UAVs and manned aircraft by integrating deep learning. It addresses the shortcomings of existing technologies in dynamic temporal modeling, network structure design, and parameter optimization. Through a series of innovative designs, it achieves a comprehensive, high-precision, and interpretable quantitative evaluation of the differences in ADS-B control instruction execution between medium and large UAVs and manned aircraft.

[0007] To achieve the above objectives, the following technical solution is provided: a deep learning-integrated intelligent evaluation system for ADS-B control instruction execution differences in unmanned aerial vehicles (UAVs), comprising: a data acquisition module for acquiring time-series data of ADS-B control instruction execution between UAVs and manned aircraft; a data preprocessing module for cleaning, standardizing, and aligning the time-series data; a time-series feature reconstruction module for constructing a multi-dimensional time-series feature matrix from the preprocessed time-series data; and a time-series VAE evaluation module for constructing and training a time-series VAE evaluation model, wherein the time-series VAE evaluation model includes an encoder module, a decoder module, and a difference quantization module. The module comprises: an encoder module for converting a multi-dimensional temporal feature matrix into a low-dimensional latent vector and outputting the latent vector with interpretable labels and its probability distribution parameters; a decoder module for reconstructing the latent vector with interpretable labels back to the original temporal space and outputting the reconstructed temporal feature matrix; and a difference quantification module for comprehensively analyzing, based on the probability distribution parameters of the latent vector with interpretable labels and the reconstructed temporal feature matrix, the differences in distribution, temporal dynamics, and key feature contributions, and outputting the quantitative difference assessment results between UAVs and manned aircraft during the execution of ADS-B control instructions.

[0008] Compared with the prior art, the present invention achieves the following beneficial effects:

[0009] 1. The Koopman_Net module of this invention constructs a linearization mapping mechanism based on Koopman theory, which transforms the nonlinear temporal characteristics (such as communication delay fluctuations and nonlinear deviations in command response) in the execution data of control instructions into a linear system in the measurement space, solving the problem that traditional models are difficult to handle nonlinear dynamic characteristics, and laying the foundation for subsequent dynamic optimization and differential quantization.

[0010] 2. The Kalman_Net encoder of the present invention introduces a data-driven Kalman filtering mechanism for the "biased linear system" generated by Koopman_Net. By integrating nonlinear residuals, it realizes state estimation correction and process uncertainty modeling, solves the error accumulation problem that still exists after linearization, and outputs a refined latent state containing covariance information.

[0011] 3. The SHAPAttenEncoder of this invention integrates a multi-attention mechanism and Shapley value calculation. On the one hand, it captures long-term temporal dependencies in the execution data of control commands (such as the chain relationship of "communication delay → command response delay → flight accuracy fluctuation"). On the other hand, it quantifies the contribution of each input feature to the execution difference, thus solving the problems of insufficient interpretability and weak ability to capture long-term temporal dependencies in existing models.

[0012] 4. The decoder of the present invention, through Koopman inverse mapping, temporal attention decoding and interpretability preservation mechanism, links key information such as attention weights and feature contribution output by the encoder to achieve accurate reconstruction from latent space to the original temporal space, while preserving difference features and interpretability labels, thus solving the problems of low reconstruction accuracy and insufficient utilization of key information in traditional decoders.

[0013] 5. This invention designs a customized ELBO loss function that integrates reconstruction loss, KL divergence regularization term and interpretability regularization term, balancing the reconstruction accuracy, latent distribution rationality and feature contribution distribution rationality of the model, avoiding overfitting or bias during model training, and improving the model's generalization ability and evaluation reliability.

[0014] 6. This invention is based on a parameter optimization algorithm of a non-dominated sorting moth flame optimizer. It achieves multi-objective optimization for key model parameters (such as the initialization range of the Koopman operator, the number of attention heads, and the regularization coefficient), while minimizing the evaluation error, minimizing the computation time, and maximizing the interpretability score. This solves the problems of traditional optimization algorithms being prone to getting trapped in local optima and struggling to balance multiple optimization objectives.

[0015] Through the above innovative design, this invention can provide a scientific basis for adapting ADS-B control rules for mixed airspace, formulating customized interval standards, and optimizing control measures, ultimately improving the safety and airspace utilization efficiency of the integrated operation of medium and large UAVs and manned aircraft.

[0016] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0017] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0018] Figure 1 This is a schematic diagram of the module of the intelligent evaluation system for the difference in execution of ADS-B control instructions for unmanned aerial vehicles that integrates deep learning, provided in an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of the workflow of the intelligent assessment system for differences in the execution of ADS-B control instructions for unmanned aerial vehicles that integrates deep learning, provided in an embodiment of the present invention.

[0020] Figure 3This is a schematic diagram of the internal structure of the time-series VAE evaluation model according to an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0023] Figure 1 This is a schematic diagram of the module of the intelligent evaluation system for the difference in execution of ADS-B control instructions for unmanned aerial vehicles that integrates deep learning, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the workflow of the intelligent assessment system for differences in ADS-B control command execution of unmanned aerial vehicles (UAVs) provided in an embodiment of the present invention. Figure 1 and Figure 2 As shown, the present invention provides an intelligent evaluation system 100 for the difference in execution of ADS-B control instructions for unmanned aerial vehicles (UAVs) that integrates deep learning. The system includes a data acquisition module 110, a data preprocessing module 120, a temporal feature reconstruction module 130, and a temporal VAE evaluation module 140. These modules are connected sequentially and work collaboratively. Their specific structure and functions are as follows:

[0024] Data acquisition module 110 is used to acquire the runtime sequence data of ADS-B control commands for UAVs and manned aircraft;

[0025] Step 1: Data Collection

[0026] The data acquisition module 110 is used to collect operational data of medium and large-sized UAVs and manned aircraft under ADS-B control rules, focusing on capturing dynamic characteristics in the time series dimension. The data collection targets are mature medium and large-sized UAVs (such as models under Tengdun Company) actually operating at Zigong Airport and manned aircraft operating at the same airport. The collected data includes:

[0027] (1) Time-series control command data: a sequence of heading change commands, altitude adjustment commands, speed control commands and autonomous navigation recovery commands recorded by timestamp (sampling frequency 1 time / second, single collection duration not less than 30 minutes).

[0028] (2) Timing execution feedback data: flight accuracy timing sequence synchronized with control instructions, dynamic recording of minimum heading change value, real-time turning radius at different altitudes / speeds, timing data of altitude / speed adjustment response delay, and curve of autonomous navigation recovery time variation;

[0029] (3) Timing communication data: communication delay time sequence between ground station and aircraft (including instantaneous peak delay), VHF communication link stability parameters (signal strength, interruption frequency and duration).

[0030] (4) Time-series ADS-B monitoring data: time-series changes in position signal accuracy, dynamic fluctuations in altitude signal accuracy, monitoring signal stability index (signal loss rate), real-time data of transmission distance, and time-series sequence of terrain occlusion influence coefficient (recorded by airspace partition).

[0031] The data preprocessing module 120 is used to clean, standardize, and align the time-series data.

[0032] Step 2: Data Preprocessing

[0033] The collected time-series data undergoes cleaning, standardization, and time-series alignment processes. Specific steps include:

[0034] (1) Outlier removal: The 3σ criterion is used to remove extreme outliers in communication delay and signal accuracy, and the instantaneous fluctuation data is corrected by the sliding window smoothing method;

[0035] (2) Missing value supplementation: For short-term missing time series data (missing duration ≤ 5 seconds), linear interpolation is used to supplement the missing data. For long-term missing data (missing duration > 5 seconds), the missing time series segments are marked as invalid time periods and the corresponding time series segments are removed.

[0036] (3) Standardization processing: Min-Max normalization is used to convert time series data with different dimensions (such as communication delay in milliseconds and flight accuracy in meters) into standardized data in the [0,1] interval to eliminate the influence of dimensions.

[0037] (4) Time alignment: Based on the time stamp of the control command issuance, the time data of UAVs and manned aircraft are aligned on the time axis to ensure that the features at the same time point can be directly compared.

[0038] The time series feature reconstruction module 130 is used to construct a multi-dimensional time series feature matrix from the preprocessed time series data;

[0039] Step 3: Temporal Feature Reconstruction

[0040] The preprocessed time series data is reconstructed into a format suitable for input to a time series VAE, specifically including:

[0041] Step 3.1: Timing Window Division

[0042] The sliding window method (window size 60 seconds, step size 10 seconds) is used to divide long time series data into multiple time series subsequences, each subsequence containing feature data of 60 time steps;

[0043] Step 3.2: Feature Dimension Integration

[0044] Control commands, execution feedback, communication, and monitoring data within the same time window are integrated into a multi-dimensional time-series feature matrix (e.g., each time window corresponds to a 60×20 feature matrix, where 20 is the feature dimension of a single time step), which serves as the input sample for the time-series VAE.

[0045] The temporal VAE evaluation module 140 constructs and trains and optimizes the temporal VAE evaluation model, which includes an encoder module 141, a decoder module 142, and a differential quantization module 143.

[0046] Step 4: Construct a time-series VAE evaluation model

[0047] The temporal VAE evaluation module 140 is the core of the system. By integrating Koopman linearization modeling, Kalman dynamic optimization, SHAP interpretability mechanism and temporal attention mechanism, it constructs a variational autoencoder with long temporal modeling capability, dynamic error correction and inter-layer interpretability, so as to achieve accurate quantification of the differences in the execution of ADS-B control commands between UAVs and manned aircraft. Figure 3 This is a schematic diagram of the internal structure of the time-series VAE evaluation model according to an embodiment of the present invention. Figure 3 As shown, the temporal VAE evaluation model consists of three parts: an encoder module 141, a decoder module 142, and a difference quantification module 143. The encoder module 141 converts the multi-dimensional temporal feature matrix into a low-dimensional latent vector and outputs the latent vector with interpretable labels and its probability distribution parameters. The decoder module 142 reconstructs the refined latent state sequence back to the original temporal space and outputs the reconstructed temporal feature matrix. The difference quantification module 143, based on the probability distribution parameters of the latent vector with interpretable labels and the reconstructed temporal feature matrix, performs a comprehensive analysis from three dimensions: distribution difference, temporal dynamic difference, and key feature contribution difference, outputting the quantitative difference evaluation results between UAVs and manned aircraft during ADS-B control instruction execution. The quantitative difference evaluation results include specific difference indicators and a list of key features, providing quantitative decision-making basis for airspace control rule adaptation and the formulation of customized UAV interval standards. The specific design is as follows:

[0048] Step 4.1: Construct the encoder module 141

[0049] The core function of the encoder is to convert high-dimensional, non-linear control command execution time series data into low-dimensional, linear latent vectors containing uncertain information, while preserving the temporal correlation features and interpretability labels. It is divided into three sub-modules that work together: the encoder module 141 includes the Koopman_Net sub-module 1411, the Kalman_Net sub-module 1412, and the SHAPAttenEncoder sub-module 1413.

[0050] Step 4.1.1: Construct submodule 1 – Koopman_Net

[0051] To address the nonlinear characteristics of control command execution data (such as communication delay fluctuations and command response timing), a linearization mapping is constructed based on Koopman theory to transform the nonlinear dynamics of the original time-series space into a linear system in the measurement space.

[0052] The Koopman_Net submodule 1411 constructs a linearization mapping mechanism based on Koopman theory, transforming the nonlinear temporal features in a multi-dimensional temporal feature matrix into a linear system in the measurement space. Specifically, this includes: vectorizing the multi-dimensional temporal feature matrix; mapping the vectorized features to the measurement space using a learnable measurement function to obtain measurement vectors; fitting local Koopman operators through extended dynamic mode decomposition and combining them with globally learnable operators to construct a complete Koopman operator; and iteratively predicting the measurement vectors based on the complete Koopman operator to generate linearized temporal features. The specific implementation is as follows:

[0053] Step 4.1.1.1: Input Processing

[0054] Koopman_Net submodule 1411 receives the multi-dimensional temporal feature matrix output by temporal feature reconstruction module 130. The temporal features of each window are converted into vector form through the flatten operation. .

[0055] : A multi-dimensional temporal feature matrix, output by the temporal feature reconstruction module 130, containing temporal features related to the execution of UAV / manned aircraft control commands (such as communication delay, command response delay, flight accuracy, etc.), with dimensions of . Feature dimension, which is the number of features contained in a single time step (such as control instruction type, communication delay, signal accuracy, etc., total dimension). The number of time steps within a window, i.e., the number of time steps contained in each time series window (window size is 60 seconds, sampling frequency is 1 time / second, therefore...). ). The vector form features of a single time-series window, by Obtained through flattening operation, with dimensions of , used to perform linearization mapping on the input measurement function.

[0056] Step 4.1.1.2: Measurement Function Learning

[0057] A two-layer MLP is used as the learnable measurement function. ,Will Mapping to the measurement space yields the measurement vector. . It can learn measurement functions and is implemented using a 2-layer MLP (512 hidden layer neurons, sigmoid activation function). Its core function is to map nonlinear temporal features to a linear measurement space. : Measurement vector in measurement space, by Measuring function The mapping yields a dimension of This enables the linearization of nonlinear temporal characteristics. The measurement space dimension is the linearized feature dimension (set to 256 to balance computational efficiency and feature representation capability). Preliminary experiments have verified that this dimension setting can achieve a balance between ensuring the linearization representation capability of nonlinear features and controlling the computational complexity of the model. When the dimension is less than 200, the representation capability is insufficient, and when it is more than 300, the model training efficiency decreases significantly.

[0058] Step 4.1.1.3: Koopman operator fitting

[0059] Fitting the local Koopman operator using extended dynamic mode decomposition (eDMD) Then, superimpose globally learnable operators. (Initialization uses Xavier uniform distribution) to form a complete Koopman operator. This is used to convert nonlinear time-series characteristics into linear dynamic systems. The calculation method is as follows:

[0060]

[0061] The local Koopman operator, obtained through extended dynamic mode decomposition (eDMD) fitting, captures the local linear dynamic laws of temporal features. Globally learnable operators, initialized using a uniform Xavier distribution, are used to supplement the global dynamic representation capabilities of local operators. The superposition forms a complete operator. The backward subsequence of the measurement vector, containing Middle 1 to 1 The feature of each time step is expressed as follows: , used to fit the local Koopman operator. : The forward subsequence of the measurement vector, containing The second to the third The feature of each time step is expressed as follows: ,and Combined with the calculation of local Koopman operators. Moore-Penrose pseudo-inverse operation is used to solve for the inverse of a non-full-rank eigenma matrix, ensuring... The stability of the calculation (avoiding solution failure due to linear dependence of the characteristic matrix).

[0062] Step 4.1.1.4: Generation of Linear Systems

[0063] Based on Koopman operator Generate a linear dynamic system for the measurement vector Perform iterative prediction to obtain linearized time-series features. .

[0064] The linearized time series characteristics are derived from the Koopman operator. right Iterative prediction yields the following expression: This is used for subsequent nonlinear residual calculations. : The zero power of the Koopman operator, i.e. the identity matrix, corresponds to the linear mapping of the initial time step. Koopman operators The power of 1 is used for iterative prediction of the 1st power. The linearization of time steps enables a linear representation of temporal dynamics.

[0065] Step 4.1.2: Construct submodule 2 – Kalman_Net

[0066] Kalman_Net submodule 1412 introduces a data-driven Kalman filtering mechanism to perform state estimation correction and process uncertainty modeling on the linear system output by the Koopman_Net submodule, and outputs a refined latent state sequence containing a sequence of covariance matrices.

[0067] For linear systems generated by Koopman_Net, or "biased linear systems" (errors caused by incomplete linearization of the measurement function), a data-driven Kalman filtering mechanism is introduced. By integrating nonlinear residuals to correct state estimation, and modeling process uncertainties, a refined latent state containing covariance information is output.

[0068] Furthermore, the Kalman_Net submodule 1412 is configured to perform the following operations: calculate the residual between the linearized time-series features and the original measurement vector; integrate the residual into the control input using a Transformer-based residual integrator; construct a state-space model containing process noise and observation noise; and, using the control input, correct the estimation of the latent state and update its covariance matrix through iterative prediction and update steps, thereby outputting a refined latent state sequence containing the covariance matrix sequence. The specific implementation is as follows:

[0069] Step 4.1.2.1 Residual Calculation and Integration

[0070] Calculate the linearized temporal features of the output of Koopman_Net submodule 1411 Compared with the original measurement vector residual The residuals are integrated into the control input through a "residual integrator" consisting of an Encoder-Only Vanilla Transformer (8 attention heads, 3 layers). ( The prediction time steps are consistent with the window size.

[0071] The linearized residual matrix of Koopman_Net reflects the nonlinear errors that still exist after linearization by the Koopman operator (such as communication delay fluctuations, nonlinear deviations in command response, etc.). Its dimensions are... Consistent. The original measurement vector in the measurement space, i.e., the unlinearized feature matrix input to the Koopman_Net submodule 1411 (already passed through the measurement function). Mapping). Linearized temporal features are used to compare with the original measurement vector to calculate the residuals. Encoder-OnlyVanillaTransformer: The core network structure of the residual integrator, which uses 8 attention heads and 3 layers of network depth to capture long temporal dependencies in the residuals (such as the correlation between "residual accumulation and instruction execution bias amplification"). The integrated control input matrix is ​​processed by the residual integrator. The processed data contains nonlinear residual information that guides the Kalman_Net correction state, expressed as follows: , dimension . : No. The control input vector for each time step corresponds to the single-time-step nonlinear residual information output by the residual integrator, which is used to correct the state transition process of Kalman_Net. The prediction time steps are consistent with the time series window size. This ensures that the control input matches the prediction time series length. :Measure the spatial dimension (consistent with the previous text, set to 256) to ensure that the control input matrix is ​​compatible with the state dimension of Kalman_Net.

[0072] Step 4.1.2.2: State transition and observation model construction

[0073] The state transition model is as follows:

[0074]

[0075] : No. The latent state vector of each step represents the refined low-dimensional state information of Kalman_Net (including the core dynamic features and uncertainties of command execution), with a dimension of [missing information]. . : No. The latent state vector of step 1, as the first step The input for each state transition ensures the continuity of the timing dynamics. : State transition matrix, dimension , which are learnable parameters, characterize the transition pattern of the latent state in adjacent time steps (such as the temporal evolution of the instruction execution state). : Control input matrix, dimension is , is a learnable parameter used to adjust the control input. The weights of the impact of state transitions ensure the control input. (dimension) ×1) can be added to the output of the state transition matrix A. : Process noise vector, with a mean of 0 and a covariance matrix of normal distribution ( This simulates random interference during the execution of control commands (such as instantaneous communication jitter, sensor noise, etc.). The process noise covariance matrix is ​​expressed as follows: This ensures the positive definiteness of the matrix (which conforms to the mathematical properties of the covariance matrix). It can learn the lower triangular matrix and optimize it to adapt to the process noise distribution in actual operation (such as the statistical characteristics of UAV communication noise) through training.

[0076] The observation model is:

[0077]

[0078] : No. The observation vector for each step is taken from the linearized temporal features output by Koopman_Net. Used for comparison with the state predictions of Kalman_Net, with dimensions of . The linearized temporal features (prediction part) output by Koopman_Net serve as the observation input for Kalman_Net, providing the basic temporal information after linearization, i.e., the actual observation results of the linearized temporal features, which are used to correct the prediction state bias. Observation matrix, dimension 1 , which are learnable parameters, to realize the mapping from latent state vector to observation vector (such as restoring low-dimensional state to linearized temporal features). The observed noise vector follows a pattern with a mean of 0 and a covariance matrix of... normal distribution The simulation includes errors in the observation process (such as ADS-B signal measurement noise, data transmission delay errors, etc.). The observation noise covariance matrix is ​​expressed as follows: This ensures the positive definiteness of the matrix. It can learn the lower triangular matrix and optimize it through training to adapt to the statistical distribution of actual observed noise (such as the accuracy fluctuation characteristics of ADS-B signals).

[0079] Step 4.1.2.3: Iterative Prediction - Update

[0080] Prediction steps: Predict the current state and covariance based on the previous state estimate:

[0081]

[0082] : No. The predicted state vector of the step is derived from the state transition matrix. Previous state and control input Calculated (expression) This reflects the latent state prediction results based on historical state and residual information, with the following dimensions: . State transition matrix (consistent with the previous text, dimension) This characterizes the temporal transition patterns of latent states and provides a core mapping relationship for predicting states. : No. The latent state vector of the first step (consistent with the previous step) serves as the historical state input for the prediction step, ensuring temporal continuity. B: Control input matrix (consistent with the previous step, dimension...) Adjust control input The weight of the influence on the predicted state. : No. The control input vector for each step (consistent with the previous text) provides nonlinear residual information to help correct prediction state bias. : No. The prediction covariance matrix of the step is expressed as follows: Characterizing the uncertainty distribution of the predicted state, with dimensions of . : No. The corrected covariance matrix of the previous step serves as the historical input for predicting the covariance, reflecting the level of uncertainty of the previous state. State transition matrix The transpose of the matrix is ​​used for time-series propagation calculation of the covariance matrix, ensuring dimensional compatibility of matrix operations. The process noise covariance matrix (consistent with the previous text) characterizes the uncertainty of random disturbances during the prediction process and participates in the prediction covariance update.

[0083] Update steps: Fuse observed and predicted values ​​using Kalman gain to correct state and covariance:

[0084]

[0085] : No. The Kalman gain matrix of the step is expressed as follows: Used to balance the weights of observed and predicted values, with dimensions of . Observation matrix The transpose of the matrix is ​​used for matrix operations on Kalman gain to ensure dimension matching. The product of the observation matrix, the prediction covariance matrix, and the transpose of the observation matrix describes the uncertainty after the predicted state is mapped to the observation space. The observation noise covariance matrix (consistent with the previous text) characterizes the error uncertainty in the observation process and participates in the Kalman gain calculation to balance the observation reliability. The inverse of the matrix within the brackets provides core inverse operation support for Kalman gain calculation, ensuring that the gain value is reasonable. : No. The corrected state vector after one step is expressed as follows: By fusing the deviation between observed and predicted values ​​using Kalman gain, a more accurate latent state is obtained, with dimensions of [missing information]. . The vector that maps the predicted state to the observation space, defined by the observation matrix. With predicted state The product is used to compare with the actual observed values. Compare the calculated deviations. The deviation vector between the observed value and the predicted mapping value reflects the difference between the predicted state and the actual observation, providing a basis for state correction. : No. The corrected covariance matrix after the step is expressed as follows: Characterizes the uncertainty distribution of the corrected state, with dimension . . The symmetry treatment term for the covariance matrix ensures that the matrix satisfies the symmetry property (a necessary mathematical condition for the covariance matrix) and avoids asymmetry problems caused by numerical calculations. The product of Kalman gain, observation noise covariance matrix, and gain transpose is used to correct the covariance matrix, ensuring its positive definiteness and improving numerical stability.

[0086] Step 4.1.2.4: Output

[0087] We obtain the refined latent state sequence, i.e., the refined latent state sequence. and the corresponding covariance matrix sequence ,in, Used for subsequent decoding Used to quantify uncertainty.

[0088] The refined latent state sequence is expressed as follows: It contains the corrected latent states for all time steps, with dimensions of This is used for timing reconstruction in subsequent decoders. The corrected state vector at each time step (as described above) (Consistent) to form a complete temporal state sequence, preserving the core dynamic characteristics of control command execution. The corresponding sequence of covariance matrices is expressed as follows: The corrected covariance matrix containing all time steps, with dimensions of . It is used to quantify the uncertainty of the latent state at each time step (such as the confidence interval of instruction execution deviation). The corrected covariance matrix at each time step constitutes a complete uncertainty sequence, providing a confidence basis for differential quantification.

[0089] Step 4.1.3 Construct submodule 3 – SHAPAttenEncoder

[0090] To capture long-term time dependencies in control command execution data (such as "communication delay - command response delay") The chain correlation of "flight accuracy fluctuations" is established, and the contribution of each feature to the performance difference is quantified. A multi-attention mechanism and Shapley value calculation are introduced.

[0091] The SHAPAttenEncoder submodule 1413 integrates a multi-head self-attention mechanism and Shapley value computation to capture long-term temporal dependencies and quantify the contribution of each input feature to the performance differences. It outputs a latent vector with interpretable labels, as well as a mean vector and a log-variance vector representing its probability distribution parameters. Furthermore, the SHAPAttenEncoder submodule 1413 is configured to perform the following operations: refine the latent state sequence... Multi-head self-attention computation is performed to generate attention-weighted states that highlight key temporal segments. Based on the Shapley value calculation mechanism, the contribution of each input feature to the generation of the refined latent state sequence is calculated. The attention-weighted states and the feature contributions are fused to generate a latent vector with interpretable labels. The mean vector and log-variance vector of this latent vector with interpretable labels are then output through a multilayer perceptron. The specific implementation is as follows:

[0092] Step 4.1.3.1: Construct the temporal attention layer

[0093] The refined latent state sequence output by Kalman_Net submodule 142 The input to the temporal attention layer includes the core dynamic features and temporal correlation information of the execution of control instructions. Multi-head self-attention (8 heads, 32 key / query dimensions) is used to calculate the temporal attention weights. The formula is:

[0094]

[0095] Temporal attention weight matrix, dimension 1 The weight values ​​reflect the degree of influence of each time step on the target time step, and are used to highlight key time segments (such as the response phase 10-20 seconds after the instruction is issued). : These are the query, key, and value matrices, respectively, composed of The linear transformation is used to highlight key time segments (such as the response phase 10-20 seconds after the instruction is issued) by attention weighting. QueryMatrix, with dimensions of ( (For query dimensions), used to calculate attention scores by matching with the key matrix; Key matrix, with dimensions of ( (as a key dimension), used in conjunction with the query matrix to capture the relationships between time steps; Value matrix, with dimensions of ( (Value dimension) stores feature information for each time step, which is used to generate the attention output in a weighted manner. : The product of the query matrix and the transpose of the key matrix, with dimension 1. We obtain the raw attention scores between each time step, which characterizes the correlation strength between time steps. : key / square root of query dimension ( Therefore This is used to scale the original attention score, preventing the gradient from vanishing due to excessive dimensionality in the Softmax function. Softmax The Softmax activation function is used to normalize the original attention score to... The weight values ​​of the intervals are set to ensure that the sum of the weights is 1, representing the relative importance of each time step.

[0096] Step 4.1.3.2 Calculation of Shapley value

[0097] Based on the Shapley Net mechanism, the attention-weighted latent state Associated with the original input features, calculate the contribution of each input feature (such as communication delay, signal precision) to the latent state. The formula is:

[0098]

[0099] Attention-weighted latent state sequence, with dimension 1. It integrates feature enhancement information from key temporal segments to highlight long-term temporal dependencies. The Shapley_Net mechanism is an interpretable modeling framework based on Shapley values, used to quantify the contribution of each input feature to the latent state, thus achieving interpretability of the model's decision-making process. : No. The Shapley contribution of each input feature is used to quantify the influence of that feature (such as communication latency, ADS-B signal precision, instruction response latency, etc.) on the generation of latent states. The larger the value, the more significant the contribution of that feature to the execution difference. Input feature set, which includes all time-series features involved in the evaluation (such as control instruction type, communication delay, signal accuracy, flight accuracy, etc.). : Feature subset, for Excluding the first Any subset after a feature This is used to calculate the difference in the impact of the presence or absence of this feature on the model output. Feature subset The cardinality (i.e., the number of features contained in the subset) is used for the average weight normalization in Shapley value calculation. Feature set Total cardinality (i.e., total dimension of input features) ). Combinations, representing the number of combinations from... Select from features The number of ways to represent each feature is used to adjust the weighting of the influence of different sized subsets on the contribution calculation. The reciprocal of the number of combinations is used as a weighting coefficient for the contribution of subsets, ensuring that the influence of subsets of different sizes is evenly distributed. The latent state generation function takes a subset of features as input and outputs the corresponding latent state fragments. It is used to calculate the output difference when a feature is present or absent. . : Including the A subset of features The corresponding latent state generation result. : Not including the first A subset of features The corresponding latent state generation result. : No. The difference in the impact of the presence or absence of a feature on the generation of latent states is the core factor in calculating the Shapley contribution.

[0100] Optionally, in specific implementations of this invention, to avoid the high complexity of precise calculations, the integral gradient method is used to efficiently approximate the Shapley value. The specific steps are as follows: Use the all-zero vector as the baseline input. The current input is a multi-dimensional temporal feature matrix. For the encoder network f(·), the approximate contribution of the i-th dimension feature. The calculation formula is as follows:

[0101]

[0102] in, and Let be the i-th feature vectors of the current input and the baseline input, respectively, and α be the interpolation parameter along a straight path from the baseline to the current input. This integral is numerically approximated by sampling a finite number of α values ​​(e.g., 20). This method efficiently computes the latent vector of all input features for the encoder's final output. Approximate Shapley contribution.

[0103] Step 4.1.3.3: Final Output

[0104] Generate latent vectors with interpretable labels. The mean is output through two layers of MLP. With log variance This is used for variational inference. Variational inference utilizes... and Modeling the probability distribution of the latent space enables a probabilistic assessment of differences in the execution of control orders, balancing model complexity with assessment accuracy.

[0105] Latent vectors containing interpretable tags, through fusion (Attention-weighted temporal features) and (Feature contribution) is obtained, with dimensions of It retains both temporal dynamic features and interpretable information. A 2-layer MLP: a network structure used to generate variational inference parameters, with input being... The output is the mean. With log variance This enables the mapping of latent vectors to distribution parameters. The mean vector of the latent distribution, with dimension 1. It is used to represent the distribution center in variational inference, reflecting the standard state characteristics of the execution of control instructions. The log-variance vector of the latent distribution, with dimension . It is used to characterize the uncertainty of the latent state (avoiding the numerical problem of negative variance) and to support subsequent reparameterization sampling.

[0106] Step 4.2: Construct decoder module 142

[0107] The core function of the decoder is to map the latent vector output by the encoder back to the original time space, reconstruct the timing data of the execution of control instructions, and at the same time retain the difference features and interpretability information.

[0108] The decoder module 142 is configured to perform the following operations: receive the latent vector and its distribution parameters output by the encoder module 141, and generate a sampled latent vector using a reparameterization technique; map the sampled latent vector from the measurement space back to the original temporal feature space using a Koopman inverse mapping to obtain a preliminary reconstructed vector; perform temporal attention decoding on the preliminary reconstructed vector based on the attention weights provided by the encoder module 141 to restore the features of key temporal segments; and use the feature contribution calculated by the encoder module 141 to weight the decoded reconstructed vector by feature dimension, highlighting the reconstruction accuracy of high-contribution features, and outputting a reconstructed temporal feature matrix. The specific design is as follows:

[0109] Step 4.2.1: Input Processing

[0110] Receive encoder output The sampled latent vector is generated using reparameterization techniques:

[0111]

[0112] The log-variance vector of the latent distribution output by the encoder. A reparameterization technique is used to circumvent the gradient non-differentiability problem during latent vector sampling by introducing random noise for indirect sampling, ensuring end-to-end model training. The sampled latent vector has dimensions of... It retains the statistical characteristics of the latent distribution while supporting gradient backpropagation. Random noise vector, following a standard normal distribution ( (The identity matrix) is used to introduce sampling randomness to simulate random fluctuations during the execution of control commands. The variance vector of the latent distribution, given by Obtained through exponential transformation This ensures that the variance is positive, conforming to the properties of a probability distribution.

[0113] Step 4.2.2: Koopman Inverse Mapping

[0114] Using Koopman_Net measurement functions Symmetric inverse mapping function Multilayer MLP, 512 hidden layer neurons, sigmoid activation function. Mapping back from the measurement space to the original temporal feature space yields a preliminary reconstructed vector. . : Koopman inverse mapping function, and Koopman_Net measurement function The structure is symmetric (2-layer MLP, 512 hidden layer neurons, sigmoid activation function). Its core function is to map the latent vector in the measurement space back to the original temporal feature space, thus achieving the inverse transformation of linearized features.

[0115] Step 4.2.3: Temporal Attention Decoding

[0116] Introduce a temporal attention layer that matches the encoder, based on transpose matrix Correction The focus is on reconstructing the features of key time segments to obtain... .

[0117] : Initially reconstruct the vector, from through The mapping yields dimensions consistent with the original temporal features. The system initially reconstructs the temporal characteristics of the execution of control instructions. The temporal attention decoding layer matches the structure of the encoder's temporal attention layer (8 attention heads, key / query dimension 32) and is used to reconstruct the feature details of key temporal segments. Attention weight matrix output by the encoder transpose matrix (dimension) This is used in the decoder to highlight key timing segments (such as the instruction response phase) in reverse, ensuring that core dynamic features are preserved during reconstruction. The reconstructed vector after attention correction is... right The weighted result is obtained, with the dimension being The focus is on restoring the feature accuracy of high-profile time series segments.

[0118] Step 4.2.4: Interpretability Reservation

[0119] The contribution calculated by the encoder As a weight, for The various feature dimensions are weighted to highlight the reconstruction accuracy of high-contribution features (such as instruction response latency), and the final output is a reconstructed time-series feature matrix. . The encoder calculates the first... The Shapley contribution of each input feature is used as a weighting coefficient for the feature dimension to highlight the reconstruction accuracy of high-contribution features (such as communication delay and instruction response delay). The final output is the reconstructed temporal feature matrix, with dimensions of . Through the By weighting the feature dimensions, the data not only restores the temporal dynamics of the execution of control instructions, but also retains the difference features and interpretability markers, which can be used to compare with the original time series data to calculate the differences.

[0120] Step 4.2.5: Distribution Modeling

[0121] Decoder module 142 is also configured to perform the following operations: perform distribution modeling on the reconstructed temporal feature matrix, and output the mean vector and standard deviation vector of the reconstructed data to characterize the predicted distribution of the reconstructed temporal data; wherein, the Koopman inverse mapping is implemented through an inverse mapping function symmetric to the measurement function structure of the Koopman_Net submodule 1411 in the encoder; temporal attention decoding uses the transpose matrix of the temporal attention weights computed by the SHAPAttenEncoder submodule in the encoder. The specific implementation is as follows:

[0122] The mean of the reconstructed data is output separately using two independent MLP layers. with standard deviation Modeling the final predicted distribution This is used for subsequent difference quantification. A separate 2-layer MLP is used to model the probability distribution of the reconstructed data, with the input being... The output is the mean of the reconstructed distribution. Standard deviation The network structure consists of two fully connected layers (256 hidden neurons, ReLU activation function).

[0123] in, : Reconstruct the mean vector of the data, with dimension . It represents the central trend of the reconstructed time series data and reflects the core of the model's prediction of the timing of the execution of control orders. : Reconstruct the standard deviation vector of the data, with dimension . It characterizes the uncertainty of reconstructed time series data and reflects the model's prediction confidence for each time series feature. The final predicted distribution follows a mean of . variance is normal distribution It is used to quantify the distribution differences between the reconstructed data and the original data, providing a probabilistic basis for performing difference assessment. The decoder's output variables (i.e., reconstructed time-series data) ), representing the timing prediction results of the execution of control instructions. : The input variable of the decoder, representing the sampling state of the latent space.

[0124] Step 4.3: Construct the differential quantification module 143

[0125] Based on the latent distribution of the Encoder and the reconstruction results of the Decoder, the difference quantification module 143 quantifies the execution differences between the UAV and the manned aircraft from three dimensions: distribution difference, temporal dynamic difference, and key feature contribution difference. This includes: calculating the KL divergence between the potential distributions of the UAV and the manned aircraft based on the probability distribution parameters of the refined latent state sequence to obtain a distribution difference index; calculating the dynamic time warping distance and temporal similarity between the UAV reconstruction temporal feature matrix output by the decoder module and the original temporal data of the manned aircraft obtained by the data acquisition module to obtain a temporal dynamic difference index; comparing the Shapley contribution of the UAV and the manned aircraft on each input feature, and filtering key features according to a preset threshold to obtain a key feature contribution difference; the distribution difference index, the temporal dynamic difference index, and the key feature contribution difference together constitute the quantified difference evaluation result. The specific implementation is as follows:

[0126] Step 4.3.1: Distribution Differences

[0127] Calculate the latent distribution of drones Latent distribution of manned aircraft The KL divergence is given by the formula:

[0128]

[0129] The latent distribution of the UAV is determined by the encoder based on the UAV's temporal features. The modeling yields a state distribution pattern reflecting the execution of drone control commands, with parameters representing the mean. With covariance matrix . The latent distribution (reference distribution) of manned aircraft is determined by the encoder based on the temporal features of manned aircraft. The training yields a standard state distribution reflecting the execution of control commands, with the parameter being the mean. With covariance matrix . Kullback-Leibler divergence is used to quantify the degree of difference between the latent distributions of UAVs and manned aircraft. The larger the value, the more significant the deviation in the distribution of the execution states of the two, and it is the core evaluation indicator of the "distribution difference" dimension. The mean vector of the latent distribution of drones, with dimension 1. This represents the central trend in the execution status of drone control directives. : The mean vector of the latent distribution of human and machine representations (reference mean), with dimensions of , representing the standard state center for the execution of control instructions. The covariance matrix of the latent distribution of drones, with dimensions of This characterizes the uncertainty distribution of the drone's execution state. : The covariance matrix of the latent distribution of human and machine systems (refer to covariance), with dimensions of This characterizes the uncertainty distribution of the standard's execution status. The determinant of the human-machine covariance matrix is ​​used to calculate the difference in the distribution "volume" in KL divergence, reflecting the overall scale of standard state uncertainty. The determinant of the covariance matrix of unmanned aircraft, compared with that of manned aircraft, reflects the difference in the scale of uncertainty between the two. The logarithmic term of the determinant ratio is used to quantify the difference in the distribution "volume" in KL divergence, reflecting the logarithmic difference in the overall scale of uncertainty between the two. The trace of a matrix is ​​the sum of the elements on the main diagonal of the matrix, used to calculate the difference in the "shape" of the covariance matrix in KL divergence. The inverse of the human-machine covariance matrix is ​​used for vector and matrix operations in KL divergence to ensure dimensionality compatibility and computational rationality. The trace of the product of the covariance matrix of the unmanned aircraft and the inverse covariance matrix of the manned aircraft This characterizes the degree of difference in the "shape" of the covariance matrices of the two entities. : The vector of differences between the latent distribution means of manned and unmanned aircraft, with dimension . This characterizes the degree of deviation between the execution state centers of the two. : The transpose of the mean difference vector, used for matrix multiplication with the inverse covariance matrix to quantify the weighted difference of mean deviation. The dimension of the latent distribution (set to 256) is used for normalization of the KL divergence formula to ensure that the index scale is reasonable.

[0130] Step 4.3.2: Temporal Dynamic Differences

[0131] Computational reconstruction timing Original timing sequence of manned aircraft The Dynamic Time Warped (DTW) distance is calculated using the following formula:

[0132] Simultaneously calculate the temporal similarity between the two (1 / DTW distance).

[0133] The reconstructed temporal feature matrix of the UAV is obtained by the decoder based on the UAV latent vector, and has a dimension of [missing information]. This reflects the execution sequence of drone control commands as reconstructed by the model. : The original temporal feature matrix of the manned machine (reference temporal series), with dimensions of It reflects the standard time-series dynamics of the execution of control orders and serves as a reference benchmark for the "time-series dynamics difference" dimension. ): Dynamic time warping distance, used to quantify the dynamic morphological differences between two time series, allowing for flexible alignment of time series lengths (such as solving the time series misalignment problem caused by the different command response speeds of UAVs and manned aircraft). The smaller the value, the more similar the dynamic execution processes of the two. Time-warped path matrix, dimension 1 Matrix elements Indicates manned aircraft Time step and drones Do the time steps match and satisfy monotonicity (if) but correspond ) and boundary constraints and To ensure that the regular path is reasonable. L2 norm (Euclidean distance) is used to calculate the distance between manned and aircraft. Time step and drones Instantaneous differences in the eigenvectors at each time step. The sum of temporal differences based on regular paths. By filtering out matching time step pairs, the summation yields the overall dynamic difference between the two time series. The minimum value is taken for all compliant adjustment paths to ensure that the DTW distance is the minimum dynamic difference between two time series, thus avoiding evaluation bias caused by path selection.

[0134] Furthermore, by The temporal similarity score was calculated; a higher value indicates greater similarity in temporal dynamics between the UAV and the manned aircraft. The range is [range missing]. It is an auxiliary evaluation indicator for the "time-series dynamic differences" dimension.

[0135] Step 4.3.3: Differences in Key Feature Contributions

[0136] Comparing the Shapley contributions of drones and manned aircraft on the same feature dimensions. ,filter ( Features with an empirical threshold (set to 0.1) are used as "key features of difference" (such as communication delay, height adjustment response delay), and a heatmap of contribution difference is output. : Drone The Shapley contribution of each input feature is calculated by the encoder, quantifying the degree of influence of the feature on the drone's execution state. Manned aircraft The Shapley contribution (reference contribution) of each input feature is calculated by the encoder based on human-machine data, quantifying the degree of influence of the feature on the standard execution state. : Difference in feature contribution, expressed as It is used to quantify the difference in contribution between UAVs and manned aircraft on the same feature, and is the core calculation item of the "key feature contribution difference" dimension. : An empirical threshold, set to 0.1, is used to filter "key features of difference". This indicates that this feature significantly contributes to the performance differences between UAVs and manned aircraft (e.g., communication delay, altitude adjustment response delay, etc.).

[0137] Key features of the difference: satisfying The input characteristics are the core factors leading to the differences in the performance of drones and manned aircraft, providing a targeted basis for optimizing regulatory rules.

[0138] Contribution Difference Heatmap: A visual representation of all features The distribution uses color intensity to represent the magnitude of differences, visually presenting the contribution of each feature to the performance differences and improving the interpretability of the evaluation results.

[0139] Step 4.4: Training and Optimization of the Time-Series VAE Evaluation Model

[0140] Step 4.4 ensures the model converges to its optimal state through phased training, adversarial enhancement, and heuristic optimization, while balancing evaluation accuracy, computational efficiency, and interpretability. Specifically, it consists of three parts: the training phase, the adversarial enhancement phase, and the parameter optimization phase.

[0141] Step 4.4.1: Temporal VAE Evaluation Model Training

[0142] Using manned aircraft control command execution time-series data as training samples, the model learns the time-series distribution of the "standard execution pattern." The first stage involves using manned aircraft time-series data as samples and employing a customized loss function that combines reconstruction loss, latent distribution KL divergence regularization, and feature contribution distribution regularization for basic training. The specific steps are as follows:

[0143] Step 4.4.1.1: Dataset Partitioning

[0144] The preprocessed data from both human and machine teams is divided into training, validation, and test sets in a 7:2:1 ratio. The training set is used for parameter updates, the validation set is used for early stopping and hyperparameter tuning, and the test set is used for initial performance evaluation.

[0145] Step 4.4.1.2 Optimizer and Learning Rate Scheduling

[0146] The Adam optimizer is used, with an initial learning rate of 1e-4, and cosine annealing scheduling is employed to avoid oscillations during later training stages.

[0147] Adam optimizer: An optimization algorithm for model training, with parameters set to... (First-order moment estimation of exponential decay rate) (Second-order moment estimation of exponential decay rate), weight decay (Regularization term to prevent overfitting), used for efficient updating of model parameters.

[0148] Initial learning rate: The initial step size of the optimizer, set to... This controls the magnitude of parameter updates to ensure stable convergence in the early stages of training.

[0149] Cosine annealing scheduling: learning rate adjustment strategy, parameters (Cosine period) (Minimum learning rate) makes the learning rate decay according to a cosine curve with each training round, avoiding oscillations in later training stages and improving convergence accuracy.

[0150] Step 4.4.1.3: Loss Function Optimization

[0151] Using customized ELBO as the objective function, the formula is:

[0152]

[0153] A customized evidence lower bound loss function serves as the core objective function for model training, balancing reconstruction accuracy, the rationality of the latent distribution, and interpretability. The expected term of the reconstruction loss, based on the latent vector. For input The expected value of the logarithm of the reconstruction probability measures the model's ability to reconstruct the original temporal features. The logarithm of the reconstruction probability, calculated using MSE loss, is expressed as follows: The larger the value, the higher the reconstruction accuracy (the negative sign minimizes the loss and corresponds to optimal reconstruction). : Original input time-series feature matrix (control command execution data of manned or unmanned aircraft), dimension is , as the benchmark for reconstruction. The reconstructed temporal feature matrix output by the decoder, with dimensions of... Used with Compare and calculate the reconstruction error. L2 norm squared: quantifies the sum of element-wise differences between the original input and the reconstructed result, reflecting the accuracy of the reconstruction. KL divergence regularization coefficient, initially set to 1.0, is linearly reduced by 0.001 per epoch during the first 500 epochs of training, until it drops to 0.5, and then remains unchanged. It is used to balance the weights of reconstruction loss and KL divergence, and to control the constraint strength of latent distribution. KL divergence regularization term, constraining the latent distribution of the model. Approximately standard normal distribution This avoids redundancy in the latent space and improves the generalization of feature representation. The standard normal prior distribution is expressed as follows: (The mean is 0, and the covariance matrix is ​​the identity matrix) ), serving as a reference benchmark for the latent distribution. : Interpretability regularization coefficient, set to 0.1, used to balance the rationality of reconstruction performance and feature contribution distribution. : Interpretability regularization term, expression is By constraining the dispersion of Shapley values, we can ensure that the distribution of feature contributions is reasonable (avoiding the dominance of individual features or the homogenization of contributions). : No. Shapley contribution of each feature. : The mean Shapley contribution of all features, expressed as: , serving as the central benchmark for the contribution distribution.

[0154] Step 4.4.1.4: Training Iteration and Early Stopping

[0155] The number of iterations is set to 1000 rounds. After each round, the validation set ELBO is calculated. If the validation set ELBO does not decrease for 20 consecutive rounds, early stopping is triggered, and the current optimal model parameters are saved.

[0156] Step 4.4.2: Temporal VAE Evaluation Model Adversarial Enhancement

[0157] Because the model performs well on clean data after basic training, but is sensitive to signal fluctuations in real-world environments (such as those caused by terrain occlusion), a second adversarial training phase is introduced to enhance model robustness, addressing noise in the UAV operating environment (such as ADS-B signal fluctuations caused by terrain occlusion and momentary communication link interruptions). The second phase involves generating adversarial examples by injecting noise into the training data and minimizing the difference between the model's reconstruction of the original and adversarial examples to enhance model robustness. The specific steps are as follows:

[0158] Step 4.4.2.1: Noise Injection

[0159] Add Gaussian noise to the training set samples Generate adversarial examples .

[0160] Gaussian noise vector, with mean 0 and variance . normal distribution , Set as The initial value in the first round of the training was 0.01, and it increased by (0.05-0.01) / 300 with each round, reaching 0.05 in the 300th round, gradually increasing the intensity of the training. Adversarial examples, expressed as: From the original sample Noise is added to train the model's robustness to noise interference. The injected Gaussian noise has a variance ranging from 0.01 to 0.05, used to simulate typical signal fluctuation intensities. Specifically, by analyzing historical ADS-B surveillance data from a certain airport, it was statistically determined that the fluctuation variance of the instantaneous signal error mainly falls within the range of 0.01 to 0.05, providing a basis for setting noise parameters in the adversarial training of this invention.

[0161] Step 4.4.2.2: Design against loss

[0162] Add an adversarial loss term The model's sensitivity to noise is constrained, and the updated total loss is: .

[0163] : Adversarial loss term, expressed as The value quantifies the difference between the original sample and the adversarial sample reconstruction by the model. The smaller the value, the less sensitive the model is to noise (the stronger its robustness). Original sample The reconstruction result. Adversarial examples The reconstruction result. The total loss function during the adversarial training phase is expressed as follows: To balance reconstruction accuracy and model robustness. The adversarial loss coefficient is set to 0.2 to control the weight of the adversarial loss in the total loss, avoiding an overemphasis on robustness that could lead to a decrease in reconstruction accuracy.

[0164] Step 4.4.2.3: Adversarial Training Iteration

[0165] The adversarial training rounds are set to 300, using mini-batch gradient descent. The model's performance on a noisy test set is evaluated every 50 rounds (in terms of reconstruction error). (For the qualified standard).

[0166] Step 4.4.3: Optimization of Time-Series VAE Evaluation Model Parameters

[0167] The third stage is heuristic parameter optimization, which specifically includes: after the basic training in the first stage and the adversarial training in the second stage, a multi-objective optimization algorithm is used to jointly optimize the key parameters of the model; wherein, the multi-objective optimization algorithm aims to minimize the evaluation error, minimize the computation time, and maximize the interpretability score; from the optimized Pareto front, the parameter combination that simultaneously satisfies the preset thresholds of evaluation error, computation time, and interpretability score is selected as the final parameters of the time-series VAE evaluation model.

[0168] Furthermore, for the key parameters of the model (which affect the evaluation accuracy and efficiency), a non-dominated ranking moth flame optimizer is used for multi-objective optimization. The specific steps are as follows:

[0169] Step 4.4.3.1: Optimize parameter selection

[0170] Determine the set of parameters to be optimized Initialization range, number of attention heads, Number of hidden layer neurons Each parameter should be set with a reasonable search range (e.g., the number of attention heads). ). The set of parameters to be optimized includes key parameters that affect the accuracy and efficiency of model evaluation, specifically: Initialization range, number of attention heads, , Number of neurons in the hidden layer of an MLP. Initialization range: The search range for initial values ​​of the global Koopman operator, used to optimize the global dynamic representation capability of linearized mappings. : KL divergence regularization coefficient, search range is Balance the reconstruction loss with the latent distribution constraint strength. : Interpretability regularization coefficient, the search range is adapted to the basic training stage (fine-tuned around 0.1) to ensure a reasonable distribution of feature contributions. : Adversarial loss coefficient, search range adapted to the adversarial training phase (fine-tuned around 0.2), balancing reconstruction accuracy and model robustness.

[0171] Number of hidden layer neurons in MLP: The number of hidden layer neurons in the MLP network of the model. The search range is adapted according to the network structure (e.g., ...). This affects the feature mapping capability.

[0172] Step 4.4.3.2: Define the objective function

[0173] The three optimization objectives are: "minimizing evaluation error", "minimizing computation time", and "maximizing interpretability score".

[0174] 1. Evaluation error:

[0175]

[0176] Err: Evaluation error, one of the core objectives of multi-objective optimization (minimization), expressed as: The accuracy of the differential assessment of the comprehensive quantitative model. The number of test samples is the total number of independent samples used in the error calculation to ensure the reliability of error statistics. : No. Latent distribution of each test sample (corresponding to drone samples) Human-machine sample correspondence ). Manned machine reference distribution, i.e. This serves as the benchmark distribution for evaluating error calculation. : No. The KL divergence between each test sample and the reference distribution, quantifying the contribution of distribution differences. : No. The original temporal feature matrix of each test sample. : No. The reconstructed temporal feature matrix of each test sample. : No. The dynamic time warping distance between the original and reconstructed time series of each test sample is used to quantify the contribution of dynamic time series differences.

[0177] Calculation time: The average inference time for a single evaluation of the statistical model (including data preprocessing and difference calculation).

[0178] 3. Interpretability score:

[0179] Calculate the Shapley contribution of all features The absolute values ​​of the features are sorted in descending order. The interpretability score is defined as the ratio of the cumulative contribution of the top K features (e.g., K=3) to the sum of the absolute values ​​of the total contributions:

[0180]

[0181] Interpret: Interpretability score, one of the objectives of multi-objective optimization (maximization). The higher the score, the more concentrated the performance differences are driven by a few key features, and the clearer and more interpretable the model's decision-making process is. : Represents the i-th largest contribution The absolute value of.

[0182] Step 4.4.3.3 NSMFO Optimization Process

[0183] Step 4.4.3.3.1: Population Initialization and Parameter Encoding

[0184] Each continuous parameter in the set of parameters to be optimized Θ (such as...) Within its search range, the values ​​are normalized to the [0,1] interval; discrete parameters (such as the number of attention heads) are mapped sequentially to integer indices 0, 1, 2... before being normalized. The position of an individual moth is a vector of length |Θ|, where each element represents a normalized value of a corresponding parameter. For example, 100 individual moths are generated (each corresponding to a set of parameter combinations), and the flame population (of the same size as the moths) is initialized.

[0185] Step 4.4.3.3.2: Moth location update

[0186] The moth's position is updated based on a spiral function, using the following formula:

[0187]

[0188]

[0189] : No. The position of each individual moth (corresponding to a set of parameter combinations) is updated through a spiral function, gradually approaching the optimal parameter combination. : No. A moth and the target flame The distance is used to quantify the degree of deviation of the moth from the optimal direction. : No. The position of each individual flame (corresponding to a set of optimal parameter combinations) serves as a guiding benchmark for updating the moth's position. : Spiral shape constant, set to 1, defines the spiral trajectory shape of the moth flying around the flame. : Helical parameter, value range is The spiral trajectory is linearly decreased from -1 to -2, controlling the contraction and search range of the spiral trajectory. : Exponential term, following The speed at which the moth approaches the flame is adjusted by changing the angle of motion. The cosine term causes the moth to spiral around the flame, expanding the search range and avoiding local optima.

[0190] Step 4.4.3.3.3: Non-dominated sorting and crowding distance calculation

[0191] Rapid nondominated ordination (FNSM) was performed on the moth population to delineate the Pareto front; crowding distance for each individual moth was calculated. The formula is:

[0192]

[0193] Fast Non-Dominated Ranking (FNSM): A stratified ranking method for moth populations, defining the Pareto front based on individual dominance across three optimization objectives (Err, Time, Interpret). If individual A is no worse than individual B in all objectives and better than individual B in at least one objective... ,but Dominate Non-dominated individuals form the first frontier, individuals dominated by the first frontier form the second frontier, and so on, used to select better parameter combinations. Pareto front: The solution set of each layer obtained after non-dominated sorting. The first frontier contains the best non-dominated individuals in the current population (no other individuals can optimize all objectives at the same time) and is the core candidate set for parameter selection. : No. The crowding distance of an individual is used to quantify the sparsity of the distribution of individuals on the Pareto front. The larger the value, the sparser the solutions around that individual are. Keeping that individual can maintain the diversity of the solution set. Optimize the target index. These correspond to the evaluation error (Err), computation time (Time), and interpretability score (Interpret), respectively. : No. The moth individual was in the first The function value on each objective (e.g.) At that time ). : No. The moth individual was in the first The function values ​​of adjacent individuals on each target (arranged in ascending order of target value). : No. The maximum value of an objective in the current population. : No. The minimum value of each objective in the current population. : No. The standardized distance between adjacent individuals on each target is used to eliminate the influence of differences in the dimensions of different targets.

[0194] Step 4.4.3.3.4: Fire Population Update

[0195] The number of flames decreases with each iteration, as shown by the formula:

[0196] round

[0197] The number of flames in the current iteration decreases with each iteration, balancing global exploration with local development capabilities. The rounding function ensures that the number of flames is an integer. Initial number of flames (consistent with the size of the moth population, set to 100). The current number of optimization iterations, ranging from [number] to [number]. . : Maximum number of optimization iterations, set to 50, controls the total time of parameter optimization. The rate of decrease in the number of flames ensures that the number of flames is 1 at the end of the iteration (focusing on the optimal solution).

[0198] Step 4.4.3.3.5: Optimal Parameter Selection

[0199] After the iteration is complete, select "Err" from the optimal Pareto front. Time Interpret The parameter combination is used as the final model parameters. Optimal Pareto front: after... The first Pareto front obtained after the iteration contains the set of parameter combinations that balance the three objectives. The evaluation error screening threshold requires that the average evaluation error of the final parameter combination does not exceed [a certain threshold]. To ensure the accuracy of difference assessment. Time The threshold for computational time consumption requires that the average inference time of a single model evaluation does not exceed 0.5 seconds to meet real-time requirements. The interpretability score is set as a screening threshold, requiring the interpretability score of the feature contribution distribution to be no less than 0.8 to ensure that the model's decision-making process is interpretable.

[0200] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0201] It should also be noted that, in the embodiments of this application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0202] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in the embodiments of this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown in this application, but is to be accorded the widest scope consistent with the principles and novel features disclosed in the embodiments of this application.

Claims

1. A deep learning-integrated intelligent evaluation system for evaluating differences in the execution of ADS-B control instructions by unmanned aerial vehicles (UAVs), characterized in that, include: The data acquisition module is used to acquire the runtime sequence data of ADS-B control commands for both UAVs and manned aircraft. The data preprocessing module is used to clean, standardize, and align the time-series data. The time series feature reconstruction module is used to construct a multi-dimensional time series feature matrix from the preprocessed time series data; The temporal VAE evaluation module constructs, trains, and optimizes a temporal VAE evaluation model, which includes an encoder module, a decoder module, and a differential quantization module. The encoder module is used to convert a multi-dimensional temporal feature matrix into a low-dimensional latent vector and output the latent vector with interpretable labels and its probability distribution parameters; the decoder module is used to reconstruct the latent vector with interpretable labels back to the original temporal space and output the reconstructed temporal feature matrix. The difference quantification module is used to perform a comprehensive analysis from three dimensions—distribution difference, temporal dynamic difference, and key feature contribution difference—based on the probability distribution parameters of the latent vector with interpretable labels and the reconstructed temporal feature matrix, and output the quantitative difference assessment results between UAVs and manned aircraft in the ADS-B control command execution process.

2. The intelligent evaluation system according to claim 1, characterized in that, The encoder module includes a Koopman_Net submodule, a Kalman_Net submodule, and a SHAPAttenEncoder submodule; The Koopman_Net submodule is based on Koopman theory to construct a linearization mapping mechanism, which transforms the nonlinear time series features in the multi-dimensional time series feature matrix into a linear system in the measurement space. The Kalman_Net submodule introduces a data-driven Kalman filtering mechanism to perform state estimation correction and process uncertainty modeling on the linear system output by the Koopman_Net submodule, and outputs a refined latent state sequence containing a covariance matrix sequence. The SHAPAttenEncoder submodule integrates a multi-head self-attention mechanism and Shapley contribution value calculation to capture long-term temporal dependencies and quantify the contribution of each input feature to the execution difference. It outputs a latent vector with interpretable labels and a mean vector and a log-variance vector representing its probability distribution parameters.

3. The intelligent evaluation system according to claim 2, characterized in that, The Koopman_Net submodule is configured to perform the following operations: Receive the multi-dimensional temporal feature matrix and vectorize it; The vectorized features are mapped to the measurement space using a learnable measurement function to obtain the measurement vector; By fitting local Koopman operators through extended dynamic mode decomposition and combining them with globally learnable operators, a complete Koopman operator is constructed. Based on the complete Koopman operator, the measurement vector is iteratively predicted to generate linearized time-series features, which are then output to the Kalman_Net submodule.

4. The intelligent evaluation system according to claim 3, characterized in that, The Kalman_Net submodule is configured to perform the following operations: Calculate the residual between the linearized time-series features and the original measurement vector; The residuals are integrated into a control input using a Transformer-based residual integrator; Construct a state-space model that includes process noise and observation noise; Using the control input, through iterative prediction and update steps, the estimation of the latent state is corrected and its covariance matrix is ​​updated, thereby outputting the refined latent state sequence containing the covariance matrix sequence.

5. The intelligent evaluation system according to claim 2, characterized in that, The SHAPAttenEncoder submodule is configured to perform the following operations: Multi-head self-attention computation is performed on the refined latent state sequence to generate attention-weighted states that highlight key temporal segments; The Shapley contribution of each input feature to the generation of the latent vector is calculated using an approximation method based on gradient integral or expected gradient. By fusing the attention-weighted state with the Shapley contribution, a latent vector with interpretable labels is generated, and the mean vector and log-variance vector of the latent vector with interpretable labels are output by a multilayer perceptron.

6. The intelligent evaluation system according to claim 1, characterized in that, The decoder module is configured to perform the following operations: The encoder module outputs a latent vector with interpretable labels and its distribution parameters, and generates a sampled latent vector using a reparameterization technique. The sampled latent vectors are mapped back from the measurement space to the original temporal feature space using the Koopman inverse mapping to obtain a preliminary reconstructed vector; The initial reconstructed vector is temporally attention-decoded based on the attention weights provided by the encoder module to restore the features of key temporal segments; The feature contribution calculated by the encoder module is used to weight the feature dimension of the decoded reconstructed vector, highlighting the reconstruction accuracy of high contribution features, and outputting the reconstructed temporal feature matrix.

7. The intelligent evaluation system according to claim 6, characterized in that, The decoder module is also configured to perform the following operations: The reconstructed time series feature matrix is ​​modeled for distribution, and the mean vector and standard deviation vector of the reconstructed data are output to characterize the predicted distribution of the reconstructed time series data; The Koopman inverse mapping is implemented through an inverse mapping function that is symmetrical to the measurement function structure of the Koopman_Net submodule in the encoder. The temporal attention decoding uses the transpose matrix of the temporal attention weights computed by the SHAPAttenEncoder submodule in the encoder.

8. The intelligent evaluation system according to claim 1, characterized in that, The difference quantification module performs the following comprehensive analysis to output the quantitative difference assessment results between UAVs and manned aircraft during the execution of ADS-B control instructions: Based on the probability distribution parameters of the latent vectors with interpretable labels, the KL divergence between the latent distributions of UAVs and manned aircraft is calculated to obtain the distribution difference index. Based on the UAV reconstructed temporal feature matrix output by the decoder module and the original manned UAV temporal data obtained by the data acquisition module, the dynamic time warping distance and temporal similarity between the two are calculated to obtain the temporal dynamic difference index. The Shapley contribution of UAVs and manned aircraft on each input feature was compared, and key features were selected according to a preset threshold to obtain the difference in the contribution of key features. The distribution difference index, the time-series dynamic difference index, and the contribution difference of key features together constitute the quantitative difference assessment result.

9. The intelligent evaluation system according to claim 1, characterized in that, The training and optimization of the time-series VAE evaluation model includes: Phase 1: Using manned and machine time series data as samples, basic training is performed using a customized loss function that combines reconstruction loss, latent distribution KL divergence regularization term and feature contribution distribution regularization term; The second stage involves generating adversarial examples by injecting noise into the training data and minimizing the difference between the model's reconstruction of the original and adversarial examples to enhance the model's robustness.

10. The intelligent evaluation system according to claim 9, characterized in that, The training and optimization of the time-series VAE evaluation model also includes: a third stage: heuristic parameter optimization, specifically including: After the basic training and the adversarial training, a multi-objective optimization algorithm is used to jointly optimize the key parameters of the model; wherein, the multi-objective optimization algorithm aims to minimize the evaluation error, minimize the computation time, and maximize the interpretability score. The parameter combinations that simultaneously satisfy the preset thresholds for evaluation error, computation time, and interpretability score are selected from the optimized Pareto fronts and used as the final parameters of the time-series VAE evaluation model.

Images