An anti-interference identification method for aircraft behavior based on track noise modeling

By splitting aircraft trajectory data into multi-scale data and constructing a recursive graph image, and combining dual-channel noise robustness assessment and adaptive fusion classifier, the problem of poor noise robustness in existing technologies is solved, and high-precision aircraft behavior recognition in complex environments is achieved.

CN123115663APending Publication Date: 2026-11-17BEIHANG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611291766.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-25
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing aircraft trajectory behavior recognition methods have poor noise robustness in complex environments, cannot effectively suppress multi-source physical noise, and lack feature information distillation mechanisms, resulting in a decline in model generalization ability.

Method used

Based on track noise modeling, the aircraft track data is split into microscopic space, microscopic attitude and macroscopic global multi-scale data, which are then transformed into recursive graph images and input into a decoupled expert network. Combined with a dual-channel noise resistance evaluation mechanism and an adaptive fusion classifier, the recognition model parameters are optimized.

Benefits of technology

It significantly improves the model's noise resistance and recognition accuracy, and can maintain high robustness and accuracy in extreme scenarios such as severe airflow interference or signal obstruction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115663A_ABST
    Figure CN123115663A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of aircraft behavior anti-interference recognition technology, and relates to an aircraft behavior anti-interference recognition method based on track noise modeling. The method includes: processing aircraft track data based on an aircraft behavior anti-interference recognition model to obtain microscopic spatial data, microscopic attitude data, and macroscopic global data, and converting them into recursive graph images; based on each recursive graph image, obtaining the information compression term and purity posterior probability corresponding to each dimension, constructing a comprehensive distortion evaluation index corresponding to each dimension, and obtaining multi-dimensional weighted fusion features; constructing a total objective loss function based on the predicted probability distribution and each information compression term, and jointly optimizing the parameters of the aircraft behavior anti-interference recognition model; processing real-time aircraft track data, and outputting the aircraft track behavior recognition result. This invention can effectively suppress high-frequency noise and redundant features, significantly improving the noise robustness and accuracy of aircraft track behavior recognition in complex scenarios.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of aircraft behavior anti-interference recognition technology, and in particular to an aircraft behavior anti-interference recognition method based on flight path noise modeling. Background Technology

[0002] With the rapid development of integrated air and space monitoring and air traffic management technologies, accurately identifying the intentions and behavioral patterns of aircraft using multi-dimensional trajectory data acquired by radar, ADS-B, and various photoelectric sensors has become a core component of situational awareness and air defense early warning systems. However, existing aircraft trajectory behavior recognition methods still face severe challenges in terms of noise robustness in complex real-world application environments. First, existing trajectory feature extraction methods ignore the inherent decoupling characteristics of different sensor physical dimensions, failing to distinguish the differences in physical intent expression between spatial trajectories and aircraft attitude. In actual flight, the external environmental interference mechanisms affecting spatial position and aircraft attitude are drastically different, making it difficult for traditional single networks to specifically decouple and suppress physical noise in different dimensions. Second, existing behavior recognition networks lack effective distillation mechanisms for input physical redundancy and random noise during feature extraction. When the input trajectory exhibits strong high-frequency noise due to sensor distortion or severe airflow interference, traditional feature extractors directly retain and pass the noise features to subsequent classification layers, resulting in severe degradation of the quality of fused features from different physical dimensions and a significant decrease in model generalization ability, failing to meet the high robustness requirements in complex noise adversarial scenarios. Summary of the Invention

[0003] Based on the above analysis, the embodiments of the present invention aim to provide an anti-interference identification method for aircraft behavior based on track noise modeling, in order to solve the problems of existing technologies that are difficult to suppress multi-source physical noise, lack feature information distillation mechanism, and have poor noise robustness in complex noise scenarios.

[0004] This invention provides a method for anti-interference identification of aircraft behavior based on track noise modeling, comprising: Based on the pre-set aircraft behavior anti-interference recognition model, the pre-collected aircraft trajectory data is processed to obtain micro-space data, micro-attitude data and macro-global data, and then converted into recursive graph images of the corresponding dimensions. Based on each of the recursive graph images, the image features corresponding to each dimension are obtained, and based on each of the image features, the information compression term and purity posterior probability corresponding to each dimension are obtained. Based on the information compression terms and the posterior probability of purity, a comprehensive distortion evaluation index corresponding to each dimension is constructed and a multi-dimensional weighted fusion feature is obtained. The predicted probability distribution is calculated based on the multidimensional weighted fusion features. A total objective loss function is constructed based on the predicted probability distribution and each of the information compression terms. The parameters of the aircraft behavior anti-interference recognition model are jointly optimized using the total objective loss function. Based on the jointly optimized aircraft behavior anti-interference recognition model, the real-time data of the aircraft trajectory collected in real time is processed, and the aircraft trajectory behavior recognition results are output.

[0005] Furthermore, the pre-collected aircraft trajectory data is processed to obtain microscopic spatial data, microscopic attitude data, and macroscopic global data, including: Longitude, latitude, and altitude information are extracted from the aircraft trajectory data to obtain the microscopic spatial data used to characterize the evolution of the aircraft's position in a three-dimensional spatial coordinate system. Pitch angle, roll angle and yaw angle information are extracted from the aircraft trajectory data to obtain the micro attitude data used to characterize the aircraft body attitude change characteristics; Global trajectory information is extracted from the aircraft trajectory data to obtain the macroscopic global data used to characterize the overall motion state of the trajectory.

[0006] Furthermore, the aircraft behavior anti-interference recognition model includes: an adaptive Gaussian kernel smoothing recursive image conversion module; the conversion into recursive graph images of corresponding dimensions includes: Based on the adaptive Gaussian kernel smoothing recursive image processing module, the microscopic spatial data, microscopic attitude data and macroscopic global data are reconstructed into phase space respectively, and the recursive matrix is ​​calculated based on the preset distance metric threshold and the reconstructed phase space respectively. Based on the recursion matrices corresponding to each dimension, the corresponding microscopic spatial recursion graph image, microscopic posture recursion graph image, and macroscopic global recursion graph image are generated respectively.

[0007] Further, generating the microscopic spatial recursive graph image, the microscopic pose recursive graph image, and the macroscopic global recursive graph image specifically includes: The recursive matrices corresponding to each dimension are numerically normalized. The normalized values ​​corresponding to each dimension are mapped to the grayscale range and converted into standard three-channel RGB images containing red, green and blue channels respectively. Each of the standard three-channel RGB images is used as the corresponding microscopic spatial recursive graph image, microscopic pose recursive graph image and macroscopic global recursive graph image respectively.

[0008] Furthermore, the aircraft behavior anti-interference recognition model also includes: a micro-space expert network, a micro-attitude expert network, a macro-global expert network, and a first dual-channel noise estimator, a second dual-channel noise estimator, and a third dual-channel noise estimator respectively connected to each expert network; Based on each of the recursive graph images, image features corresponding to each dimension are obtained, and based on each of the image features, information compression terms and purity posterior probabilities corresponding to each dimension are obtained, including: The micro-space recursive graph image is input into the micro-space expert network to extract micro-space image features. After being processed by the first dual-channel noise evaluator, the information compression term and purity posterior probability of the micro-space dimension are obtained. The micro-pose recursive graph image is input into the micro-pose expert network to extract micro-pose image features. After being processed by the second dual-channel noise evaluator, the information compression term and purity posterior probability of the micro-pose dimension are obtained. The macro-global recursive graph image is input into the macro-global expert network to extract macro-global image features. After processing by the third dual-channel noise estimator, the information compression term and purity posterior probability of the macro-global dimension are obtained.

[0009] Furthermore, the first dual-channel noise estimator, the second dual-channel noise estimator, and the third dual-channel noise estimator all include: a variational information bottleneck evaluation unit and a Gaussian mixture model evaluation unit, wherein: The variational information bottleneck evaluation unit is used to: map the image features corresponding to each dimension to the mean vector and standard deviation vector of the corresponding latent space Gaussian distribution through two parallel fully connected layers; calculate the KL divergence between each latent space Gaussian distribution and the preset standard Gaussian prior distribution based on the mean vector and standard deviation vector corresponding to each dimension; and use the KL divergence as the information compression term corresponding to each dimension after dimension normalization. The Gaussian mixture model evaluation unit is used to: calculate the single-sample prediction distortion scalar for each dimension based on the image features corresponding to each dimension; and calculate the purity posterior probability that the image features corresponding to each dimension belong to the pure signal component based on the single-sample prediction distortion scalar and using preset two-component Gaussian mixture model parameters.

[0010] Furthermore, for any dimension of information compression, the following calculation formula is satisfied: ; in, This represents the information compression term corresponding to the i-th expert network, used to quantitatively characterize the noise intensity of the image features extracted by the i-th expert network; Let x represent the mean vector of the Gaussian distribution in the latent space corresponding to the i-th expert network in the x-th dimension. Let X represent the standard deviation vector of the Gaussian distribution in the latent space corresponding to the i-th expert network in the x-th dimension; X represents the total dimension of the latent space. , respectively, correspond to the microscopic spatial expert network s, the microscopic attitude expert network p, and the macroscopic global expert network g; The posterior probability of purity for any dimension is calculated using the following formula: ; in, This represents the posterior probability of purity corresponding to the i-th expert network; Represents the Gaussian probability density function; Let represent the single-sample prediction distortion scalar corresponding to the i-th expert network; This represents the mean of the Gaussian distribution of the pure signal components; The variance represents the Gaussian distribution of the pure signal components; This represents the mixing coefficient of the pure signal components in a Gaussian mixture model; This represents the mean of the Gaussian distribution of the noise signal components; The variance represents the Gaussian distribution of the noise signal components; This represents the mixing coefficient of the noise signal components in the Gaussian mixture model.

[0011] Furthermore, the construction of the comprehensive distortion assessment index corresponding to each dimension and the acquisition of multi-dimensional weighted fusion features include: Based on the information compression terms and purity posterior probabilities, a comprehensive distortion evaluation index is constructed for each dimension; dynamic normalized weights are calculated for each dimension based on the comprehensive distortion evaluation index; and the image features corresponding to each expert network are adaptively weighted and fused using the dynamic normalized weights to obtain the multi-dimensional weighted fusion features.

[0012] Furthermore, the comprehensive distortion assessment index for any dimension satisfies the following formula: ; in, This represents the overall distortion assessment index corresponding to the i-th expert network; The pre-set smoothing balance coefficient between decision distortion and latent space compressibility divergence; For any dimension, the dynamically normalized weights satisfy the following formula: ; in, Indicates the overall distortion assessment index The corresponding dynamic normalized weights; This represents a preset constant to prevent division by zero; represents the comprehensive distortion assessment index variable; j represents the traversal index used for summing the denominator.

[0013] Furthermore, the aircraft behavior anti-interference recognition model also includes: an adaptive fusion classifier connected to the first dual-channel noise estimator, the second dual-channel noise estimator, and the third dual-channel noise estimator respectively; the predicted probability distribution is calculated and output through the adaptive fusion classifier; The construction of the overall objective loss function based on the predicted probability distribution and each of the information compression terms includes: The classification prediction loss term is calculated based on the cross-entropy of the predicted probability distribution and the preset real track behavior labels; the variational regularization loss term is calculated based on the sum of the information compression terms corresponding to each expert network; and the total objective loss function is constructed based on the classification prediction loss term and the variational regularization loss term.

[0014] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: First, unlike related technologies where splicing multi-dimensional temporal features leads to difficulty in suppressing noise, this invention decomposes the trajectory into micro-space, micro-attitude, and macro-global multi-scale data according to physical semantics, and transforms them into expert networks with decoupled recursive graph image inputs, thereby achieving decoupled expression of multi-dimensional physical features and significantly improving the model's noise resistance.

[0015] Second, unlike related technologies which suffer from problems such as noise-contaminated fusion expression and poor weight allocation capabilities, this invention employs a dual-channel noise resistance evaluation mechanism. It combines a variational information bottleneck evaluator to explicitly calculate the information compression term of latent features and introduces an offline-fixed dual-component Gaussian mixture model evaluation unit to estimate the posterior probability of purity in predictive distortion. By calculating the comprehensive distortion evaluation index through collaborative decision distortion and latent space compression divergence, a dynamic adaptive weighted fusion mechanism is achieved. This reduces the weight of heavily noise-affected expert features and makes pure features dominant, significantly improving the algorithm's noise resistance robustness in extreme scenarios such as severe airflow interference or signal obstruction.

[0016] Third, unlike related technologies where classification models are prone to fitting irrelevant noise and have poor noise robustness, this invention jointly optimizes the aircraft behavior anti-interference recognition model by combining prediction classification loss with variational regularization loss based on information compression terms. At the information theory level, it maximizes the mutual information of the target for track behavior recognition while strictly constraining the complexity of the latent feature expressions of each expert. This fundamentally curbs the overfitting of the model to noise and redundant features, and significantly improves the recognition accuracy and generalization ability of aircraft track behavior in complex scenarios.

[0017] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0018] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 This is a flowchart of an anti-interference identification method for aircraft behavior based on track noise modeling, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the overall framework of the anti-interference recognition method for aircraft behavior based on track noise modeling according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the framework of a dual-channel noise evaluator according to an embodiment of the present invention. Detailed Implementation

[0019] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0020] A specific embodiment of the present invention discloses a method for anti-interference identification of aircraft behavior based on track noise modeling, such as... Figure 1 As shown, the steps S1 to S5 are as follows: Step S1: Based on the preset aircraft behavior anti-interference recognition model, the pre-collected aircraft trajectory data is processed to obtain microscopic spatial data, microscopic attitude data and macroscopic global data, and then converted into recursive graph images of the corresponding dimensions.

[0021] Specifically, such as Figure 2 As shown, the aircraft behavior anti-interference recognition model includes at least: an adaptive Gaussian kernel smoothing recursive image processing module, a multi-expert network structure (including a micro-space expert network, a micro-attitude expert network, and a macro-global expert network), a dual-channel noise evaluation module (including a first dual-channel noise estimator, a second dual-channel noise estimator, and a third dual-channel noise estimator) connected to each expert network respectively, and an adaptive fusion classifier connected to the dual-channel noise evaluation module.

[0022] During implementation, the pre-collected aircraft trajectory data is processed, specifically including: extracting longitude, latitude, and altitude information from the aircraft trajectory data to obtain microscopic spatial data that characterizes the aircraft's position evolution in a three-dimensional spatial coordinate system; extracting pitch, roll, and yaw angle information from the aircraft trajectory data to obtain microscopic attitude data that characterizes the aircraft's attitude change characteristics; and extracting global trajectory information from the aircraft trajectory data to obtain macroscopic global data that characterizes the overall motion state of the trajectory.

[0023] Furthermore, based on the adaptive Gaussian kernel smoothing recursive image generation module, the microscopic spatial data, microscopic attitude data, and macroscopic global data are reconstructed into phase spaces, and recursive matrices are calculated based on preset distance metric thresholds and the reconstructed phase spaces. Based on the recursive matrices corresponding to each dimension, corresponding microscopic spatial recursive graph images, microscopic attitude recursive graph images, and macroscopic global recursive graph images are generated.

[0024] In some implementations, during the data preprocessing stage, the aircraft trajectory data (one-dimensional multivariable temporal trajectory) is up-dimensioned into a two-dimensional image through physical semantic decoupling and recursive graph mapping to express topological texture features. This design addresses the physical characteristics of the different disturbance mechanisms between spatial coordinates and aircraft attitude in the aircraft trajectory data, and provides an input basis for subsequent noise suppression and information bottleneck noise assessment through multi-dimensional decoupling.

[0025] For example, the microscopic spatial data, microscopic attitude data, and macroscopic global data can be denoted as time series. , and Suppose the extracted time-series vector sequence Where T is the length of the time series, the kernel bandwidth for local density estimation is adaptively calculated based on the average distance of K nearest neighbors at each time point, and the Euclidean distance between each time point is mapped to continuous probability similarity using a Gaussian kernel function to construct the corresponding adaptive Gaussian kernel recursive matrix. The adaptive Gaussian kernel recursive matrix is ​​then converted into a continuous probability similarity image, where, for the th recursive matrix... The formula for calculating each element is: ; in, Let J and K represent the K-nearest neighbor average distance kernel bandwidths at time j and time k, respectively. express Euclidean distance norm; Represents the state point at time j State point at time k The difference vector between them.

[0026] Preferably, the recursive matrices corresponding to each dimension are numerically normalized; the normalized values ​​corresponding to each dimension are mapped to the grayscale range and converted into standard three-channel RGB images containing red, green, and blue channels respectively; and each of the standard three-channel RGB images is used as the corresponding microscopic spatial recursive graph image, microscopic pose recursive graph image, and macroscopic global recursive graph image respectively.

[0027] For example, the recursion matrix is ​​numerically mapped to the grayscale range and converted into a continuous probability similarity image containing three channels: red (R), green (G), and blue (B), thereby generating corresponding microscopic spatial recursion map images. Microscopic posture recursion image and macro-global recursive graph image in Through the above decoupling and recursive graph mapping, the local jitter, trend evolution and abrupt changes in a one-dimensional track sequence can be mapped to the diagonal and block texture of a two-dimensional image.

[0028] Therefore, compared with traditional recursive graphs, this embodiment uses an adaptive Gaussian mapping based on K-nearest neighbor local density to transform hard distances into continuous probability similarity, achieving smooth filtering of high-frequency jitter and positioning drift of the sensor; at the same time, it maintains the coherence of the diagonal and block textures representing maneuvering behavior, and adaptively eliminates the dimensional differences between spatial and attitude dimensions, providing a high signal-to-noise ratio high-dimensional topological feature foundation for subsequent expert networks and dual-channel noise estimators.

[0029] Step S2: Based on each of the recursive graph images, obtain the image features corresponding to each dimension, and based on each of the image features, obtain the information compression term and purity posterior probability corresponding to each dimension.

[0030] In practice, the specific steps include: inputting the micro-space recursive graph image into the micro-space expert network, extracting micro-space image features, and then processing them through the first dual-channel noise evaluator to obtain the information compression term and purity posterior probability of the micro-space dimension. The micro-pose recursive graph image is input into the micro-pose expert network to extract micro-pose image features. After being processed by the second dual-channel noise evaluator, the information compression term and purity posterior probability of the micro-pose dimension are obtained. The macro-global recursive graph image is input into the macro-global expert network to extract macro-global image features. After processing by the third dual-channel noise estimator, the information compression term and purity posterior probability of the macro-global dimension are obtained.

[0031] It can be understood that the dual-channel noise evaluation in this embodiment refers to two evaluation channels: variational information bottleneck evaluation and Gaussian mixture model evaluation. That is, the dual-channel noise evaluator can evaluate the feature quality from two complementary dimensions: noise redundancy concentration and signal purity. In addition, variational information bottleneck refers to a regularization framework that constrains the input features to follow a standard Gaussian prior distribution after latent space probability encoding, in order to compress redundant information and suppress noise.

[0032] Preferably, the first dual-channel noise estimator, the second dual-channel noise estimator, and the third dual-channel noise estimator each include: a variational information bottleneck evaluation unit and a Gaussian mixture model evaluation unit; wherein: The variational information bottleneck evaluation unit is used to: map the image features corresponding to each dimension to the mean vector and standard deviation vector of the corresponding latent space Gaussian distribution through two parallel fully connected layers; calculate the KL divergence between each latent space Gaussian distribution and the preset standard Gaussian prior distribution based on the mean vector and standard deviation vector corresponding to each dimension; and use the KL divergence as the information compression term corresponding to each dimension after dimension normalization. The Gaussian mixture model evaluation unit is used to: calculate the single-sample prediction distortion scalar for each dimension based on the image features corresponding to each dimension; and calculate the purity posterior probability that the image features corresponding to each dimension belong to the pure signal component based on the single-sample prediction distortion scalar and using preset two-component Gaussian mixture model parameters.

[0033] In some implementations, the microscopic spatial recursion graph image generated by the aforementioned steps... Microscopic posture recursion image and macro-global recursive graph image The data is then input into the corresponding expert networks. The microscopic spatial expert network, microscopic pose expert network, and macroscopic global expert network can all use models such as Vision Transformer as encoder networks to extract the corresponding high-dimensional image feature vectors. ( ).

[0034] Furthermore, to explicitly quantify and suppress noise mixed in each physical dimension feature, each expert network terminal integrates a corresponding dual-channel noise estimator. The image feature vector... The variational information bottleneck evaluation unit corresponding to the noise estimator is input. This variational information bottleneck evaluation unit estimates the mean vector of the Gaussian distribution in the latent space through two parallel fully connected layers. and standard deviation vector The calculation formula includes: ; ; in, and These are the weights and bias terms of the mean linear layer, respectively. and These are the weights and biases of the variance linear layer; It is a non-negative activation function.

[0035] Furthermore, each of the aforementioned dual-channel noise estimators calculates the latent space Gaussian distribution and the standard Gaussian prior distribution. The KL divergence between them generates the information compression term of the corresponding expert network. Preferably, for any dimension of information compression, the following calculation formula is satisfied: ; in, This represents the information compression term corresponding to the i-th expert network, used to quantitatively characterize the noise intensity of the image features extracted by the i-th expert network; Let x represent the mean vector of the Gaussian distribution in the latent space corresponding to the i-th expert network in the x-th dimension. Let X represent the standard deviation vector of the Gaussian distribution in the latent space corresponding to the i-th expert network in the x-th dimension; X represents the total dimension of the latent space. , respectively, correspond to the micro-space expert network s, the micro-attitude expert network p, and the macro-global expert network g.

[0036] Furthermore, the parameter calibration and single-sample inference calculation process of the Gaussian mixture model evaluation unit in each of the dual-channel noise estimators includes: During the training phase, collect the scalar of prediction distortion generated by each expert network from all training samples. This refers to cross-entropy loss, which uses a two-component Gaussian mixture model (GMM) for unsupervised fitting via the expectation-maximization algorithm. The mean and variance parameters of the low-distortion pure Gaussian components are then calculated and fixed. Mean and variance parameters of the Gaussian components of high-distortion noise ( ), and the corresponding mixed prior coefficients ( This refers to the ratio of clean samples to noisy samples; among them, the Gaussian component with the smaller mean... The Gaussian component with a relatively large mean corresponds to a pure data distribution. The corresponding noise data distribution; in the subsequent training phase, for a single sample without a real label input, the class probability distribution output by the expert network i can be used. ] Calculate its prediction entropy, which serves as the scalar of prediction distortion for the current single sample. : ; Where K is the total number of track behavior categories; In other words, during the training phase, the prediction distortion scalar The cross-entropy loss is calculated between the predicted probability distribution of the categories and the true labels; during the inference phase, the prediction distortion scalar is used. The prediction entropy for predicting the probability distribution of the category.

[0037] Furthermore, during training, the single-sample prediction distortion scalar is... Substitute into the parameters of the two-component Gaussian mixture model ( , , The posterior probability that the current expert feature belongs to the pure signal is calculated using Bayes' theorem. Preferably, the posterior probability of purity in any dimension satisfies the following calculation formula: ; in, This represents the posterior probability of purity corresponding to the i-th expert network; Represents the Gaussian probability density function; Let represent the single-sample prediction distortion scalar corresponding to the i-th expert network; This represents the mean of the Gaussian distribution of the pure signal components; The variance represents the Gaussian distribution of the pure signal components; This represents the mixing coefficient of the pure signal components in a Gaussian mixture model; This represents the mean of the Gaussian distribution of the noise signal components; The variance represents the Gaussian distribution of the noise signal components; This represents the mixing coefficient of the noise signal components in the Gaussian mixture model.

[0038] Therefore, this embodiment of the invention combines a variational information bottleneck evaluator to explicitly calculate the information compression term of latent features, and introduces an offline-fixed two-component Gaussian mixture model evaluation unit to estimate the posterior probability of purity of prediction distortion; it realizes the quantitative evaluation of feature quality by two complementary dimensions of noise redundancy concentration and signal purity, providing an accurate basis for subsequent adaptive weighted fusion.

[0039] Step S3: Based on each information compression term and the posterior probability of purity, construct the comprehensive distortion evaluation index corresponding to each dimension and obtain the multi-dimensional weighted fusion feature.

[0040] In practice, the specific steps include: each expert network constructs a comprehensive distortion evaluation index corresponding to each dimension based on each of the aforementioned information compression terms and the posterior probability of purity; calculates the dynamic normalization weights corresponding to each dimension based on the comprehensive distortion evaluation indices; and uses the dynamic normalization weights to adaptively weight and fuse the image features corresponding to each expert network to obtain the multi-dimensional weighted fusion features.

[0041] For example, the purity posterior probability output by the Gaussian mixture model evaluation unit is collaboratively fused. Information compression terms output by the variational information bottleneck evaluation unit Construct the comprehensive distortion assessment index for the i-th expert network. Preferably, the comprehensive distortion assessment index for any dimension satisfies the following formula: ; in, This represents the overall distortion assessment index corresponding to the i-th expert network; The pre-defined smoothing balance coefficient between decision distortion and latent space compressibility divergence is used; based on this, the comprehensive distortion assessment index is calculated. The input is fed into the adaptive fusion classifier.

[0042] In some implementations, reference Figure 3 As shown, the comprehensive distortion assessment index In terms of physical semantics, it directly reflects the intensity of noise interference and the redundancy of invalid information in the current dimension's features. When a physical dimension (such as the micro-attitude branch) is affected by high-frequency interference such as airflow, leading to a decrease in feature quality, the comprehensive distortion evaluation index output by its dual-channel noise estimator is... The corresponding increase will occur; conversely, the comprehensive distortion assessment index of dimensions with higher feature purity will increase accordingly. The smaller the value, the better.

[0043] Furthermore, the comprehensive distortion assessment index is calculated based on the micro-space expert network, the micro-attitude expert network, and the macro-global expert network. ( ), to obtain the dynamic normalized weights corresponding to each expert network. Preferably, the dynamically normalized weights for any dimension satisfy the following formula: ; in, Indicates the overall distortion assessment index The corresponding dynamic normalized weights; This represents a preset minimum constant to prevent division by zero; Let represent the comprehensive distortion assessment index variable; j represents the traversal index used for summing the denominator; simultaneously, ensure that the sum of all weights strictly satisfies the normalization constraint. = 1.

[0044] Furthermore, the image features of the corresponding expert network are weighted and summed using the dynamically normalized weights to obtain multidimensional weighted fusion features. The calculation formula includes: ; Thus, through the aforementioned dynamic weighting mechanism, adaptive pure feature enhancement and noise feature suppression are achieved in the feature fusion stage; micro-spatial branches are automatically suppressed in severe positioning drift scenarios, and micro-attitude branches are automatically suppressed in strong airflow interference scenarios, ensuring that subsequent behavior recognition is always dominated by rich image features with high signal-to-noise ratio, thereby significantly improving the noise resistance performance in complex adversarial environments.

[0045] Step S4: Calculate the predicted probability distribution based on the multi-dimensional weighted fusion features, construct the total objective loss function based on the predicted probability distribution and each of the information compression terms, and use the total objective loss function to jointly optimize the parameters of the aircraft behavior anti-interference recognition model.

[0046] Specifically, the predicted probability distribution is calculated and output by the adaptive fusion classifier (flight behavior classifier); the construction of the overall objective loss function includes: calculating a classification prediction loss term based on the cross-entropy of the predicted probability distribution and the preset real flight behavior labels; calculating a variational regularization loss term based on the sum of the information compression terms corresponding to each expert network; and constructing the overall objective loss function based on the classification prediction loss term and the variational regularization loss term.

[0047] In some implementations, the multidimensional weighted fusion features calculated in the aforementioned steps are... The input adaptive fusion classifier is processed by a fully connected layer and a Softmax activation function to calculate the predicted probability distribution of the current trajectory belonging to each predefined behavior category (e.g., surveying, transportation).

[0048] Preferably, a classification prediction loss term is constructed by combining the predicted probability distribution with the actual track behavior labels. And the information compression items calculated by each expert network Accumulate to construct variational information bottleneck regularization loss term The classification prediction loss term With the variational information bottleneck regularization loss term By combining weighted methods, the overall objective loss function for end-to-end training is constructed. Based on the overall objective loss function The network parameters of the aircraft behavior anti-interference recognition model are updated using the backpropagation algorithm to obtain the jointly optimized aircraft behavior anti-interference recognition model.

[0049] For example, during the training phase, the process of constructing the overall target loss function includes: First, a cross-entropy classification prediction loss term is constructed based on the predicted probability distribution and the actual trajectory behavior labels. The formula is used to constrain the model to accurately identify the target behavior category. ; Where K is the total number of track behavior categories. The unique hot code value is taken from the actual track behavior label. This represents the predicted probability of the k-th behavior category output by the behavior classifier. Secondly, the information compression items calculated by the dual-channel noise evaluator corresponding to each expert network are... Accumulate the values ​​to construct the variational information bottleneck regularization loss term. The formula includes: ; Finally, the classification prediction loss term Variational information bottleneck regularization loss term We perform weighted combination to construct the overall objective loss function. The formula includes:

[0050] in, These are hyperparameters with preset weights greater than zero used to balance classification accuracy and feature compression strength; thus, the overall objective loss function is used. In conjunction with the backpropagation algorithm, the model parameters of the aircraft behavior anti-interference recognition model are jointly optimized end-to-end.

[0051] For example, in the model optimization process, gradient descent and backpropagation algorithms are used, utilizing the overall objective loss function. The parameters of three expert networks, three dual-channel noise estimators, and an adaptive fusion classifier are updated end-to-end.

[0052] Therefore, under the constraint of the loss function, the model can not only learn to extract highly discriminative classification features (minimize) It can also compress high-frequency jitter and sensor noise in each physical dimension that are irrelevant to the recognition task (minimize) While ensuring the accuracy of behavior classification, we minimize useless and redundant information in the feature extraction process of each expert network. Step S5: Based on the jointly optimized aircraft behavior anti-interference recognition model, process the real-time data of the aircraft trajectory collected in real time, and output the aircraft trajectory behavior recognition result.

[0053] Specifically, after completing the joint optimization of the network parameters, a trained trajectory behavior recognition model is obtained, which can be deployed in actual aircraft trajectory recognition tasks. The real-time data of the aircraft trajectory to be identified is input into the jointly optimized aircraft behavior anti-interference recognition model. After executing steps S1 to S4 sequentially, the real-time distribution of predicted probabilities output by the adaptive fusion classifier is obtained, and the category corresponding to the maximum probability is output as the final aircraft trajectory behavior recognition result.

[0054] For example, in the actual inference phase, real-time data of the aircraft trajectory is acquired. First, the real-time data of the aircraft trajectory is decoupled and decomposed into real-time data of micro-space, micro-attitude, and macro-global aspects according to physical semantics, and each dimension of real-time data is transformed into corresponding recursive graph real-time images. Second, each recursive graph real-time image is input into a pre-trained micro-space expert network, micro-attitude expert network, and macro-global expert network, respectively. Each expert network extracts corresponding real-time image features, and the real-time information compression term and real-time purity posterior probability of each dimension are calculated using the corresponding dual-channel noise estimator of each expert network. Based on this, a corresponding comprehensive distortion real-time evaluation index is constructed according to the real-time information compression term and real-time purity posterior probability of each dimension, and the dynamic normalized weights of each expert network are calculated. The real-time image features extracted by each expert network are adaptively weighted and fused using the dynamic normalized weights to obtain multi-dimensional weighted fused real-time features. Finally, the multidimensional weighted fusion real-time features are input into the adaptive fusion classifier, which outputs the probability distribution of the aircraft's current trajectory behavior belonging to each trajectory behavior category. The category corresponding to the highest probability value is taken as the final trajectory behavior recognition result and output.

[0055] Therefore, through the above methods, in actual noise interference environments, the model can dynamically adjust the fusion weights based on the real-time quality of features in each dimension, effectively suppressing the interference of noise features on the recognition results and outputting accurate aircraft trajectory behavior recognition results; the aircraft trajectory behavior categories can be predefined according to actual task requirements, including but not limited to destruction, reconnaissance, transportation, patrol, etc.

[0056] It is understood that the above embodiments are only for ease of understanding and simplification of description, and should not be construed as limitations on the present invention. The present invention does not specifically limit the type of aircraft trajectory data, the network structure of the aircraft behavior anti-interference identification model, the calculation method, etc.

[0057] Therefore, it can be seen that the embodiments of the present invention can achieve at least one of the following beneficial effects: First, this invention decomposes the trajectory into microscopic space, microscopic attitude and macroscopic global multi-scale data according to physical semantics, and transforms them into expert networks with decoupled recursive graph image inputs, thereby achieving decoupled expression of multi-dimensional physical features and significantly improving the model's noise resistance.

[0058] Second, the present invention uses a dynamic adaptive weighted fusion mechanism to reduce the weight of expert features that are heavily affected by noise and make pure features dominant, which significantly improves the noise robustness of the algorithm in extreme scenarios such as severe airflow interference or signal blockage.

[0059] Third, while maximizing the mutual information of the target in track behavior recognition at the information theory level, this invention strictly constrains the complexity of the latent feature expressions of each expert, fundamentally curbing the overfitting of the model to noise and redundant features, and significantly improving the recognition accuracy and generalization ability of aircraft track behavior in complex scenarios.

[0060] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0061] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for anti-interference identification of aircraft behavior based on track noise modeling, characterized in that, include: Based on the pre-set aircraft behavior anti-interference recognition model, the pre-collected aircraft trajectory data is processed to obtain micro-space data, micro-attitude data and macro-global data, and then converted into recursive graph images of the corresponding dimensions. Based on each of the recursive graph images, the image features corresponding to each dimension are obtained, and based on each of the image features, the information compression term and purity posterior probability corresponding to each dimension are obtained. Based on the information compression terms and the posterior probability of purity, a comprehensive distortion evaluation index corresponding to each dimension is constructed and a multi-dimensional weighted fusion feature is obtained. The predicted probability distribution is calculated based on the multidimensional weighted fusion features. A total objective loss function is constructed based on the predicted probability distribution and each of the information compression terms. The parameters of the aircraft behavior anti-interference recognition model are jointly optimized using the total objective loss function. Based on the jointly optimized aircraft behavior anti-interference recognition model, the real-time data of the aircraft trajectory collected in real time is processed, and the aircraft trajectory behavior recognition results are output.

2. The method according to claim 1, characterized in that, The aforementioned process involves collecting pre-collected aircraft trajectory data to obtain microscopic spatial data, microscopic attitude data, and macroscopic global data, including: Longitude, latitude, and altitude information are extracted from the aircraft trajectory data to obtain the microscopic spatial data used to characterize the evolution of the aircraft's position in a three-dimensional spatial coordinate system. Pitch angle, roll angle and yaw angle information are extracted from the aircraft trajectory data to obtain the micro attitude data used to characterize the aircraft body attitude change characteristics; Global trajectory information is extracted from the aircraft trajectory data to obtain the macroscopic global data used to characterize the overall motion state of the trajectory.

3. The method according to claim 2, characterized in that, The aircraft behavior anti-interference recognition model includes: an adaptive Gaussian kernel smoothing recursive image conversion module; the conversion into recursive graph images of corresponding dimensions includes: Based on the adaptive Gaussian kernel smoothing recursive image processing module, the microscopic spatial data, microscopic attitude data and macroscopic global data are reconstructed into phase space respectively, and the recursive matrix is ​​calculated based on the preset distance metric threshold and the reconstructed phase space respectively. Based on the recursion matrices corresponding to each dimension, the corresponding microscopic spatial recursion graph image, microscopic posture recursion graph image, and macroscopic global recursion graph image are generated respectively.

4. The method according to claim 3, characterized in that, Generating the microscopic spatial recursive graph image, the microscopic pose recursive graph image, and the macroscopic global recursive graph image specifically includes: The recursive matrices corresponding to each dimension are numerically normalized. The normalized values ​​corresponding to each dimension are mapped to the grayscale range and converted into standard three-channel RGB images containing red, green and blue channels respectively. Each of the standard three-channel RGB images is used as the corresponding microscopic spatial recursive graph image, microscopic pose recursive graph image and macroscopic global recursive graph image respectively.

5. The method according to claim 4, characterized in that, The aircraft behavior anti-interference recognition model also includes: a micro-space expert network, a micro-attitude expert network, a macro-global expert network, and a first dual-channel noise estimator, a second dual-channel noise estimator, and a third dual-channel noise estimator, which are respectively connected to each expert network. Based on each of the recursive graph images, image features corresponding to each dimension are obtained, and based on each of the image features, information compression terms and purity posterior probabilities corresponding to each dimension are obtained, including: The micro-space recursive graph image is input into the micro-space expert network to extract micro-space image features. After being processed by the first dual-channel noise evaluator, the information compression term and purity posterior probability of the micro-space dimension are obtained. The micro-pose recursive graph image is input into the micro-pose expert network to extract micro-pose image features. After being processed by the second dual-channel noise evaluator, the information compression term and purity posterior probability of the micro-pose dimension are obtained. The macro-global recursive graph image is input into the macro-global expert network to extract macro-global image features. After processing by the third dual-channel noise estimator, the information compression term and purity posterior probability of the macro-global dimension are obtained.

6. The method according to claim 5, characterized in that, The first dual-channel noise estimator, the second dual-channel noise estimator, and the third dual-channel noise estimator all include: a variational information bottleneck evaluation unit and a Gaussian mixture model evaluation unit, wherein: The variational information bottleneck evaluation unit is used to: map the image features corresponding to each dimension to the mean vector and standard deviation vector of the corresponding latent space Gaussian distribution through two parallel fully connected layers; calculate the KL divergence between each latent space Gaussian distribution and the preset standard Gaussian prior distribution based on the mean vector and standard deviation vector corresponding to each dimension; and use the KL divergence as the information compression term corresponding to each dimension after dimension normalization. The Gaussian mixture model evaluation unit is used to: calculate the single-sample prediction distortion scalar for each dimension based on the image features corresponding to each dimension; and calculate the purity posterior probability that the image features corresponding to each dimension belong to the pure signal component based on the single-sample prediction distortion scalar and using preset two-component Gaussian mixture model parameters.

7. The method according to claim 6, characterized in that, For any dimension of information compression, the following calculation formula is satisfied: ; in, This represents the information compression term corresponding to the i-th expert network, used to quantitatively characterize the noise intensity of the image features extracted by the i-th expert network; Let x represent the mean vector of the Gaussian distribution in the latent space corresponding to the i-th expert network in the x-th dimension. Let X represent the standard deviation vector of the Gaussian distribution in the latent space corresponding to the i-th expert network in the x-th dimension; X represents the total dimension of the latent space. , respectively, correspond to the microscopic spatial expert network s, the microscopic attitude expert network p, and the macroscopic global expert network g; The posterior probability of purity for any dimension is calculated using the following formula: ; in, This represents the posterior probability of purity corresponding to the i-th expert network; Represents the Gaussian probability density function; Let represent the single-sample prediction distortion scalar corresponding to the i-th expert network; This represents the mean of the Gaussian distribution of the pure signal components; The variance represents the Gaussian distribution of the pure signal components; This represents the mixing coefficient of the pure signal components in a Gaussian mixture model; This represents the mean of the Gaussian distribution of the noise signal components; The variance represents the Gaussian distribution of the noise signal components; This represents the mixing coefficient of the noise signal components in the Gaussian mixture model.

8. The method according to any one of claims 1 to 7, characterized in that, The construction of the comprehensive distortion assessment index corresponding to each dimension and the acquisition of multi-dimensional weighted fusion features include: Based on the information compression terms and purity posterior probabilities, a comprehensive distortion evaluation index is constructed for each dimension; dynamic normalized weights are calculated for each dimension based on the comprehensive distortion evaluation index; and the image features corresponding to each expert network are adaptively weighted and fused using the dynamic normalized weights to obtain the multi-dimensional weighted fusion features.

9. The method according to claim 8, characterized in that, For any dimension of the overall distortion assessment index, the following formula is satisfied: ; in, This represents the overall distortion assessment index corresponding to the i-th expert network; The pre-set smoothing balance coefficient between decision distortion and latent space compressibility divergence; For any dimension, the dynamically normalized weights satisfy the following formula: ; in, Indicates the overall distortion assessment index The corresponding dynamic normalized weights; This represents a preset constant to prevent division by zero; represents the comprehensive distortion assessment index variable; j represents the traversal index used for summing the denominator.

10. The method according to claim 9, characterized in that, The aircraft behavior anti-interference recognition model further includes: an adaptive fusion classifier connected to the first dual-channel noise estimator, the second dual-channel noise estimator, and the third dual-channel noise estimator respectively; the predicted probability distribution is calculated and output through the adaptive fusion classifier; The construction of the overall objective loss function based on the predicted probability distribution and each of the information compression terms includes: The classification prediction loss term is calculated based on the cross-entropy of the predicted probability distribution and the preset real track behavior labels; the variational regularization loss term is calculated based on the sum of the information compression terms corresponding to each expert network; and the total objective loss function is constructed based on the classification prediction loss term and the variational regularization loss term.