Unmanned aerial vehicle electronic feature recognition method and system based on deep convolutional network

Abstract

The invention provides an unmanned aerial vehicle electronic feature recognition method and system based on a deep convolutional network, and relates to the technical field of unmanned aerial vehicle recognition. Inputting the communication signal data into a convolutional layer to extract a feature vector; inputting the flight state data into a bidirectional gating recursion unit to generate a state feature vector; projecting the feature vectors to unrelated semantic subspaces to realize feature decoupling; and reconstructing features by using a sparse coding dictionary and outputting an identification result through a full-connection classifier. The method achieves the efficient recognition of the electronic features of the unmanned aerial vehicle, and improves the recognition accuracy.

Description

Technical Field

[0001] The present invention relates to the field of drone identification technology, and in particular to a drone electronic feature recognition method and system based on a deep convolutional network. Background Art

[0002] Drone electronic signature recognition aims to automatically identify and classify drone models, manufacturers, and operation types by analyzing the communication signals emitted by drones and their flight status data. With the rapid increase in the number of drones and the diversification of their application scenarios, accurately and efficiently identifying different types of drones has become a key research topic.

[0003] Existing drone electronic feature recognition technologies are mainly based on traditional feature engineering methods or simple deep learning models, but they have obvious shortcomings in practical applications. Existing methods usually use fixed-structure convolutional networks to process drone communication signals, and are unable to adaptively adjust the convolution kernel parameters according to different frequency characteristics, resulting in insufficient signal feature extraction in complex electromagnetic environments. The time series modeling capabilities of drone flight status data are limited, and a one-way processing mode is mostly used, which makes it difficult to capture complete time series dependencies, affecting the accuracy of state feature extraction. Existing technologies lack an effective feature decoupling mechanism, and there is semantic aliasing between communication signal features and flight status features, which reduces the model's robustness in identifying drones in different working modes. In particular, the recognition accuracy drops significantly in interference environments. Summary of the Invention

[0004] The embodiments of the present invention provide a method and system for drone electronic feature recognition based on a deep convolutional network, which can solve the problems in the prior art.

[0005] A first aspect of an embodiment of the present invention provides a method for identifying electronic features of a drone based on a deep convolutional network, comprising: Acquiring electronic characteristic data of the drone, the electronic characteristic data including communication signal data and flight status data; The communication signal data is input into a signal processing channel composed of multiple convolutional layers in series. The size of each convolution kernel is adaptively adjusted according to the frequency characteristics of the current layer signal to extract the signal feature vector; The flight status data is input into a bidirectional gated recursive unit, which processes the input data from both the forward and backward directions, controls the information flow using a gating mechanism, and generates a state feature vector that incorporates bidirectional temporal information. Projecting the signal feature vector and the state feature vector into mutually unrelated semantic subspaces, respectively, calculating the correlation score of the semantic subspace features, and completing feature decoupling when the correlation score is less than a decoupling completion threshold to obtain decoupled signal features and state features; The decoupled signal features and state features are reconstructed using a sparse coding dictionary, and the reconstructed features are input into a fully connected classifier. The drone attribute information corresponding to the feature with the highest confidence is used as the final recognition result, where the final recognition result includes at least one of the drone model, manufacturer, and operation type.

[0006] The communication signal data is input into a signal processing channel composed of multiple convolutional layers in series. The size of each convolution kernel is adaptively adjusted according to the frequency characteristics of the current layer signal to extract the signal feature vector, including: Performing Fourier transform on the communication signal data to obtain a time-frequency spectrum, calculating the energy density of the communication signal at each frequency point, and taking the maximum value of the energy density as the maximum value of the spectrum energy; Based on the energy density, a dynamic convolution kernel is constructed; a base convolution kernel size is multiplied by the signal frequency value and the energy value respectively, and a weighted sum is performed to obtain a dynamic convolution kernel size; Using the dynamic convolution kernel to extract features from the communication signal, performing dense convolution on feature areas with energy density higher than an energy threshold, and using sparse convolution on feature areas with energy density lower than the energy threshold, to generate a multi-scale feature map sequence; Performing feature enhancement on each feature map in the multi-scale feature map sequence, adaptively fusing the current feature map with two adjacent feature maps, and calculating the fusion weight through a feature attention mechanism to obtain a fused feature map; The fused feature map is used as the input of the next convolutional layer, and the convolution operation and feature fusion process are repeated until the feature extraction of all convolutional layers is completed; global feature extraction is performed on the last layer of fused feature map to obtain a signal feature vector containing signal frequency, energy and time domain characteristics.

[0007] The flight status data is input into a bidirectional gated recursive unit, which processes the input data from both the forward and backward directions, controls the information flow using a gating mechanism, and generates a state feature vector that integrates bidirectional temporal information, including: Reconstruct the flight status data into a state feature sequence in chronological order; Calculating the cosine similarity between the state features of the current moment and the state features of the two adjacent time steps to obtain a local correlation matrix; normalizing the local correlation matrix to obtain local feature weights; multiplying the local feature weights by the state feature sequence to generate a short-term feature sequence; Mapping the state feature sequence to a semantic space through a linear transformation; calculating the dot product similarity of feature vectors in the semantic space to generate a semantic similarity matrix; normalizing the semantic similarity matrix to obtain semantic feature weights; and multiplying the semantic feature weights by the state feature sequence to generate a semantic feature sequence; Inputting the state feature sequence into a multilayer perceptron to obtain a forward feature sequence and a backward feature sequence; calculating the Wasserstein distance between the forward feature sequence and the backward feature sequence as a difference; performing a gradient update on the feature extraction parameters in the multilayer perceptron based on the difference to obtain updated forward feature sequence and backward feature sequence, and generating an adversarial feature sequence; The short-term feature sequence, the semantic feature sequence and the adversarial feature sequence are weightedly combined to obtain a state feature vector.

[0008] The signal feature vector and the state feature vector are respectively projected into mutually unrelated semantic subspaces, and a correlation score of the semantic subspace features is calculated. When the correlation score is less than a decoupling completion threshold, feature decoupling is completed, and decoupled signal features and state features are obtained, including: The signal feature vector and the state feature vector are respectively transformed into feature sequences of multiple scale levels by wavelet transform to obtain a multi-scale signal feature sequence and a multi-scale state feature sequence; The ratio of the variance of each scale feature in the multi-scale signal feature sequence to the mean of the variance of the multi-scale signal feature sequence is used as a regularization coefficient, the product of the regularization coefficient and the transposed matrix of the scale feature is used as a correction term, the correction term is summed with the autocorrelation matrix of the scale feature and then inverted to obtain an adaptive projection matrix of the scale; the adaptive projection matrix is ​​applied to the corresponding scale feature to obtain a signal subspace feature of the scale; Using the same calculation method, the corresponding adaptive projection matrix is ​​calculated for the features of each scale in the multi-scale state feature sequence, and the corresponding state subspace features are obtained; performing weighted fusion on the signal subspace features and the state subspace features of each scale level respectively to obtain fused signal features and fused state features, inputting the fused signal features and the fused state features into a generation network to generate decoupled signal features and decoupled state features, and calculating correlation scores of the decoupled signal features and the decoupled state features; Iteratively optimize the parameters of the generation network until the feature correlation score is less than the decoupling completion threshold, and output the decoupled signal features and the decoupled state features.

[0009] The decoupled signal features and state features are reconstructed using a sparse coding dictionary. The reconstructed features are input into a fully connected classifier, and the drone attribute information corresponding to the feature with the highest confidence is used as the final recognition result, including: The decoupled signal features and the decoupled state features are concatenated to obtain decoupled features; the intra-class feature variance of the decoupled features within different drone categories and the inter-class feature variance between different drone categories are calculated; and the ratio of the intra-class feature variance to the inter-class feature variance is used as a discriminant factor; Based on the discriminant factor, the decoupled features are divided into multiple groups; for each feature group, a feature similarity matrix within the group is calculated, and based on the similarity matrix, the number of atoms in the sub-dictionary corresponding to the group is determined by a clustering algorithm; Cascading the sub-dictionaries corresponding to all feature groups in order of the size of the discriminant factors to construct a sparse coding dictionary; calculating the sparse representation coefficient of the decoupled feature under the sparse coding dictionary, and taking the product of the sparse representation coefficient and the sparse coding dictionary as the reconstructed decoupled feature; A multi-branch fully connected classification network is constructed, where each branch corresponds to the reconstructed decoupled features of a feature group; the reconstructed decoupled features of each branch are input into the corresponding branch network to obtain the branch recognition results, the confidence scores of each branch recognition results are calculated, and the branch recognition result with the highest confidence score is selected as the final recognition result.

[0010] Based on the discriminant factor, the decoupled features are divided into multiple groups; for each feature group, a feature similarity matrix within the group is calculated, and based on the similarity matrix, the number of atoms in the sub-dictionary corresponding to the group is determined by a clustering algorithm, including: Arranging the discriminant factors in descending order of numerical value to obtain a discriminant factor sequence, calculating the first-order difference of the discriminant factor sequence to obtain a difference sequence, and calculating the global mean of the difference sequence; dividing the difference sequence into a plurality of subintervals and calculating the interval mean, determining the optimal segmentation point of the difference sequence based on the difference between the interval mean and the global mean, and dividing the feature vector into a plurality of feature groups based on the optimal segmentation point; Calculating an initial similarity matrix corresponding to the eigenvectors in each of the feature groups; calculating the cosine similarity between the eigenvectors in each of the feature groups, multiplying the difference between the cosine similarity and the similarity threshold by the product of the initial similarity matrix to obtain an optimized similarity matrix, and calculating the eigenvalue sequence of the optimized similarity matrix; The adjacent eigenvalue differences of the eigenvalue sequence are calculated to obtain a spectral interval sequence, and the spectral interval sequence is divided by the corresponding eigenvalue to obtain a significance score sequence; the product of the maximum value of the significance score sequence and the number of samples of the feature group is determined as the number of atoms in the sub-dictionary corresponding to the feature group.

[0011] A second aspect of an embodiment of the present invention provides a drone electronic feature recognition system based on a deep convolutional network, comprising: The first unit is used to obtain electronic characteristic data of the UAV, wherein the electronic characteristic data includes communication signal data and flight status data; The second unit is used to input the communication signal data into the signal processing channel composed of multiple convolution layers connected in series. The size of each convolution kernel is adaptively adjusted according to the frequency characteristics of the signal in the current layer to extract the signal feature vector; The third unit is used to input the flight status data into the bidirectional gated recursive unit, which processes the input data from both the forward and backward directions, controls the information flow using a gating mechanism, and generates a state feature vector that integrates the bidirectional time series information; A fourth unit is configured to project the signal feature vector and the state feature vector into mutually unrelated semantic subspaces, calculate a correlation score of the semantic subspace features, and complete feature decoupling when the correlation score is less than a decoupling completion threshold to obtain decoupled signal features and state features; The fifth unit is used to reconstruct the decoupled signal features and state features using a sparse coding dictionary, input the reconstructed features into a fully connected classifier, and use the drone attribute information corresponding to the feature with the highest confidence as the final recognition result, wherein the final recognition result includes at least one of the drone model, manufacturer, and operation type.

[0012] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0013] According to a fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0014] The beneficial effects of this application are as follows: The present invention achieves high-precision recognition of UAV electronic features through feature decoupling technology and adaptive convolution kernel design, effectively reduces the impact of signal interference and environmental noise on recognition results, and improves the robustness of the system in complex electromagnetic environments.

[0015] The present invention uses a bidirectional gated recursive unit to process flight status data, which can simultaneously capture forward and backward temporal dependencies, making the system more accurate in identifying the flight trajectory and state changes of drones, helping to distinguish the behavioral characteristics of different types of drones and enhancing the reliability of identification.

[0016] The present invention combines sparse coding dictionary technology to reconstruct decoupled features, which can not only effectively compress feature representation and reduce computational complexity, but also filter redundant data while retaining key identification information, making the system have higher processing efficiency and better real-time performance, and is suitable for practical application scenarios with limited resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of a method for identifying electronic features of drones based on a deep convolutional network according to an embodiment of the present invention; Figure 2 Schematic diagram of the system framework for processing flight status data using a bidirectional gated recursive unit. DETAILED DESCRIPTION

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0019] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0020] Figure 1 Schematic diagram of the process of the embodiment of the present invention, as shown in FIG. Figure 1 As shown, the UAV electronic feature recognition method based on deep convolutional network includes: Acquiring electronic characteristic data of the drone, the electronic characteristic data including communication signal data and flight status data; The communication signal data is input into a signal processing channel composed of multiple convolutional layers in series. The size of each convolution kernel is adaptively adjusted according to the frequency characteristics of the current layer signal to extract the signal feature vector; The flight status data is input into a bidirectional gated recursive unit, which processes the input data from both the forward and backward directions, controls the information flow using a gating mechanism, and generates a state feature vector that incorporates bidirectional temporal information. Projecting the signal feature vector and the state feature vector into mutually unrelated semantic subspaces, respectively, calculating the correlation score of the semantic subspace features, and completing feature decoupling when the correlation score is less than a decoupling completion threshold to obtain decoupled signal features and state features; The decoupled signal features and state features are reconstructed using a sparse coding dictionary, and the reconstructed features are input into a fully connected classifier. The drone attribute information corresponding to the feature with the highest confidence is used as the final recognition result, where the final recognition result includes at least one of the drone model, manufacturer, and operation type.

[0021] In an optional embodiment, the communication signal data is input into a signal processing channel composed of multiple convolutional layers in series, the size of each convolution kernel is adaptively adjusted according to the frequency characteristics of the signal in the current layer, and the signal feature vector is extracted, including: Performing Fourier transform on the communication signal data to obtain a time-frequency spectrum, calculating the energy density of the communication signal at each frequency point, and taking the maximum value of the energy density as the maximum value of the spectrum energy; Based on the energy density, a dynamic convolution kernel is constructed; a base convolution kernel size is multiplied by the signal frequency value and the energy value respectively, and a weighted sum is performed to obtain a dynamic convolution kernel size; Using the dynamic convolution kernel to extract features from the communication signal, performing dense convolution on feature areas with energy density higher than an energy threshold, and using sparse convolution on feature areas with energy density lower than the energy threshold, to generate a multi-scale feature map sequence; Performing feature enhancement on each feature map in the multi-scale feature map sequence, adaptively fusing the current feature map with two adjacent feature maps, and calculating the fusion weight through a feature attention mechanism to obtain a fused feature map; The fused feature map is used as the input of the next convolutional layer, and the convolution operation and feature fusion process are repeated until the feature extraction of all convolutional layers is completed; global feature extraction is performed on the last layer of fused feature map to obtain a signal feature vector containing signal frequency, energy and time domain characteristics.

[0022] The present invention provides a communication signal feature extraction method, which realizes efficient signal feature representation through adaptive convolution kernel processing. In a specific embodiment, the system receives communication signal data, which can be I / Q data in complex form or waveform data in real form. Fourier transform is performed on the input communication signal data, and the time domain signal is converted into frequency domain representation to generate a time-frequency spectrum. Taking the LTE signal with a sampling rate of 20MHz as an example, a 512-point FFT transform is performed on the 1024-point sampling data to obtain a time-frequency spectrum with a frequency resolution of approximately 39.06kHz. The energy density E(f) is calculated for each frequency point f in the time-frequency spectrum, specifically by calculating the square of the modulus of the complex value of the frequency point. In actual applications, for a 10ms LTE signal, the system scans all frequency points and finds the maximum energy density E max For example, when the energy density at the center frequency is detected to be 86.5 dB, it is recorded as the maximum spectrum energy.

[0023] Based on the above energy density distribution, the system constructs a dynamic convolution kernel. The base convolution kernel size K0 is set to 3×3. For each frequency point, the signal frequency value S is calculated. f and energy value S e Two factors. Signal frequency value S f Defined as the current frequency f and the signal bandwidth center frequency f c The ratio is taken as a logarithm and then multiplied by the frequency adjustment coefficient α. In the embodiment, for a 20 MHz bandwidth LTE signal, the center frequency f c When the signal frequency is 2.4GHz, the frequency adjustment coefficient α is set to 0.8. When processing the signal area with a frequency of 2.395GHz, the calculated S f The value is about -0.002. Energy value S e It is defined as the ratio of the energy density E(f) at the current frequency point to the maximum spectrum energy Emax multiplied by the energy adjustment coefficient β. The energy adjustment coefficient β is set to 0.5. If the energy density at the current frequency point is 80.2dB and the maximum value is 86.5dB, then S e About 0.46.

[0024] The dynamic convolution kernel size K is calculated by the following method: the base convolution kernel size K0 is respectively f and energy value S e Multiply and perform a weighted sum. The weight coefficients are w1 = 0.6 and w2 = 0.4, respectively. Taking the above parameters as an example, for a signal region with a frequency of 2.395 GHz and an energy density of 80.2 dB, the calculated dynamic convolution kernel size is approximately 3.18 × 3.18, which the system rounds to 3 × 3. For a region with a frequency of 2.41 GHz and an energy density of 85.3 dB, the calculated convolution kernel size is approximately 4.2 × 4.2, which the system rounds to 4 × 4.

[0025] The calculated dynamic convolution kernel is used to extract features of the communication signal. The system sets the energy threshold to E max 75% of the total energy density, corresponding to 64.9dB in this example. For feature regions with energy density above 64.9dB, the system performs dense convolution with a stride of 1 to finely extract features. For feature regions with energy density below 64.9dB, sparse convolution with a stride of 2 is used to reduce computational complexity. This adaptive convolution strategy generates a multi-scale feature map sequence containing feature representations of different receptive field sizes.

[0026] Perform feature enhancement operations on each feature map in the generated multi-scale feature map sequence. Specifically, the current feature map F i With two adjacent feature maps F i-1 and F i+1Perform adaptive fusion. The fusion weight is calculated through the feature attention mechanism. The system first calculates the global average pooling value of each feature map, and then generates the attention weight through a two-layer fully connected network. Assume that for the feature map F i-1 、F i and F i+1 The calculated attention weights are 0.25, 0.5, and 0.25 respectively. The system weights the three feature maps and sums them according to these weights to obtain the fused feature map F' i .

[0027] The fused feature map F' i The convolution operation and feature fusion process is repeated as the input to the next convolutional layer. In one embodiment, the system includes four convolutional layers, with the number of channels in each layer being 64, 128, 256, and 512, respectively. The first layer input is the feature map of the original signal after initial processing, with a size of 64×64×1. After four layers of adaptive convolution processing, the final feature map is 8×8×512.

[0028] The system performs global feature extraction on the final fusion feature map using global average pooling and global maximum pooling, respectively extracting the global average feature and the global maximum feature. These two features are then concatenated to form a 1024-dimensional feature vector. Finally, a fully connected layer with a dimension of 512 compresses the features into a 512-dimensional signal feature vector, which contains the signal's frequency characteristics, energy distribution, and time domain characteristics.

[0029] Through the above implementation, the present invention can adaptively adjust the convolution kernel size according to the frequency characteristics of the signal, perform fine feature extraction in high-energy areas, and use sparse convolution for low-energy areas, thereby improving the efficiency and accuracy of feature extraction, and is suitable for feature analysis of various communication signals such as 5G, Wi-Fi, and Bluetooth.

[0030] Figure 2 This is a schematic diagram of the framework for a bidirectional gated recursive unit processing flight status data. In an optional embodiment, flight status data is input into the bidirectional gated recursive unit, which processes the input data in both the forward and backward directions, using a gating mechanism to control the flow of information and generate a state feature vector that incorporates bidirectional time series information, including: Reconstruct the flight status data into a state feature sequence in chronological order; Calculating the cosine similarity between the state features of the current moment and the state features of the two adjacent time steps to obtain a local correlation matrix; normalizing the local correlation matrix to obtain local feature weights; multiplying the local feature weights by the state feature sequence to generate a short-term feature sequence; Mapping the state feature sequence to a semantic space through a linear transformation; calculating the dot product similarity of feature vectors in the semantic space to generate a semantic similarity matrix; normalizing the semantic similarity matrix to obtain semantic feature weights; and multiplying the semantic feature weights by the state feature sequence to generate a semantic feature sequence; Inputting the state feature sequence into a multilayer perceptron to obtain a forward feature sequence and a backward feature sequence; calculating the Wasserstein distance between the forward feature sequence and the backward feature sequence as a difference; performing a gradient update on the feature extraction parameters in the multilayer perceptron based on the difference to obtain updated forward feature sequence and backward feature sequence, and generating an adversarial feature sequence; The short-term feature sequence, the semantic feature sequence and the adversarial feature sequence are weightedly combined to obtain a state feature vector.

[0031] This embodiment provides a method for inputting flight status data into a bidirectional gated recursive unit for processing. The method can process input data from both forward and backward directions, use a gating mechanism to control the flow of information, and generate a state feature vector that integrates bidirectional time series information.

[0032] In this method, the flight status data generated by the aircraft during flight are obtained, including but not limited to parameters such as altitude, speed, attitude angle, acceleration, angular velocity, etc. These raw data are reconstructed in chronological order to form a state feature sequence X=[x1, x2, ..., x n ], where n represents the time step, x i Represents the state feature vector at the i-th time step.

[0033] For the reconstructed state feature sequence, calculate the cosine similarity between the state feature at the current moment and the state feature of the two adjacent time steps to obtain the local correlation matrix L. Specifically, for time step i, calculate the feature vector x i with x i-1 and x i+1 The cosine similarity of the local correlation matrix L is used to form the elements in the local correlation matrix L. For example, when the speed of the aircraft at a certain moment is [12.5, 0, 2.3] m / s, the speed at the previous moment is [12.2, 0, 2.5] m / s, and the speed at the next moment is [12.8, 0, 2.1] m / s, the calculated cosine similarities are 0.998 and 0.997 respectively. The local correlation matrix L is normalized to obtain the local feature weight W local The normalization process can use the softmax function to make the sum of the weights of each time step equal to 1. local Multiply it with the state feature sequence X to generate a short-time feature sequence S=[s1, s2, ..., sn ].

[0034] Map the state feature sequence X to the semantic space through linear transformation, and obtain the semantic space feature Q=[q1, q2,..., q n ] and K=[k1, k2, ..., k n ]. The linear transformation can be achieved by two different weight matrices W q and W k Implementation, i.e. q i =x i W q , k i =x i W k . Calculate the dot product similarity of the feature vectors in the semantic space and generate the semantic similarity matrix M, where M ij Indicates q i and k j The semantic similarity matrix M is normalized to obtain the semantic feature weight W semantic Normalization is also processed by softmax function. The semantic feature weight W semantic Multiply it with the state feature sequence X to generate the semantic feature sequence V=[v1, v2, ..., v n ].

[0035] The state feature sequence X is input into the multi-layer perceptron MLP, which consists of three layers: input layer, hidden layer, and output layer. The number of input layer nodes is the same as the state feature dimension, the number of hidden layer nodes is 128, and the number of output layer nodes is 64. The forward feature sequence F is obtained by forward propagation. forward =[f1 f , f2 f , ..., f n f ]. At the same time, the reverse state feature sequence is input into the MLP with the same structure to obtain the backward feature sequence F_backward=[f1 b , f2 b , ..., f n b ].

[0036] Calculate the forward feature sequence F forward and the backward feature sequence F backward The Wasserstein distance is used as the difference degree D. The Wasserstein distance measures the difference by calculating the minimum transmission cost between two distributions. Taking the flight data of 32 time steps as an example, the calculated Wasserstein distance is 0.062. Based on the difference degree D, the feature extraction parameters in the multilayer perceptron are gradient updated. The update formula is , where η is the learning rate, set to 0.001. After the parameters are updated, the updated forward feature sequence F is recalculated forward ' and the backward feature sequence F backward ', concatenate them and map them through a linear layer to generate the adversarial feature sequence A=[a1, a2, ..., a n ].

[0037] The short-term feature sequence S, the semantic feature sequence V, and the adversarial feature sequence A are weighted together to produce the final state feature vector Z. The weighting coefficients are α, β, and γ, respectively. Experimental verification has verified that α = 0.3, β = 0.4, and γ = 0.3. The specific combination is Z = α·S + β·V + γ·A. In practice, for a flight state sequence containing 128 features and 64 time steps, the resulting state feature vector Z has a dimension of 128, encompassing both temporal and semantic information about the flight state.

[0038] The state feature vector Z, processed by the bidirectional gated recursive unit, can be used for tasks such as flight anomaly detection, flight mode recognition, and flight trajectory prediction. In the flight anomaly detection task, inputting the feature vector Z into the classifier achieves an anomaly detection accuracy of 97.8%, an 8.5 percentage point improvement over traditional methods. In the flight mode recognition task, the recognition accuracy reaches 94.6%, effectively distinguishing between flight states such as hovering, cruise control, and accelerated climb.

[0039] In an optional embodiment, the signal feature vector and the state feature vector are respectively projected into mutually unrelated semantic subspaces, and the correlation score of the semantic subspace features is calculated. When the correlation score is less than the decoupling completion threshold, feature decoupling is completed to obtain the decoupled signal features and state features, including: The signal feature vector and the state feature vector are respectively transformed into feature sequences of multiple scale levels by wavelet transform to obtain a multi-scale signal feature sequence and a multi-scale state feature sequence; The ratio of the variance of each scale feature in the multi-scale signal feature sequence to the mean of the variance of the multi-scale signal feature sequence is used as a regularization coefficient, the product of the regularization coefficient and the transposed matrix of the scale feature is used as a correction term, the correction term is summed with the autocorrelation matrix of the scale feature and then inverted to obtain an adaptive projection matrix of the scale; the adaptive projection matrix is ​​applied to the corresponding scale feature to obtain a signal subspace feature of the scale; Using the same calculation method, the corresponding adaptive projection matrix is ​​calculated for the features of each scale in the multi-scale state feature sequence, and the corresponding state subspace features are obtained; performing weighted fusion on the signal subspace features and the state subspace features of each scale level respectively to obtain fused signal features and fused state features, inputting the fused signal features and the fused state features into a generation network to generate decoupled signal features and decoupled state features, and calculating correlation scores of the decoupled signal features and the decoupled state features; Iteratively optimize the parameters of the generation network until the feature correlation score is less than the decoupling completion threshold, and output the decoupled signal features and the decoupled state features.

[0040] This embodiment provides a method for decoupling signal features and state features. First, the signal feature vector and state feature vector to be processed are obtained. Next, these two feature vectors are projected into mutually unrelated semantic subspaces. The orthogonality loss of each semantic subspace is calculated. When the orthogonality loss value is less than a preset orthogonality threshold, feature decoupling is completed, and the decoupled signal features and state features are obtained.

[0041] Specifically, the wavelet transform is used to process the signal feature vector and decompose it into feature sequences at multiple scale levels. For example, a signal feature vector with a dimension of 128 can be decomposed into four scale levels using the wavelet transform, corresponding to feature sequences from high to low frequency, with each scale level containing 32 feature elements. Similarly, the state feature vector is also decomposed into corresponding multi-scale state feature sequences using the wavelet transform. In practical applications, various wavelet basis functions such as Haar wavelet, Daubechies wavelet, or Symlet wavelet can be selected for transformation.

[0042] For each scale feature in the multi-scale signal feature sequence, an adaptive projection matrix is ​​calculated. The specific calculation process is as follows: calculate the variance of the scale feature. For example, for the first scale feature, assume its variance is 0.85; calculate the mean variance of all scales in the multi-scale signal feature sequence, assuming it is 0.75; use the ratio of the variance of the scale feature to the mean variance as the regularization coefficient, in this example 0.85 / 0.75 = 1.13; multiply the regularization coefficient by the transposed matrix of the scale feature to obtain the correction term; sum the correction term with the autocorrelation matrix of the scale feature and then invert it to obtain the adaptive projection matrix for that scale.

[0043] Taking the first scale feature as an example, assuming it is a 32-dimensional vector, the above calculation yields a 32×32 adaptive projection matrix. Applying this matrix to the corresponding scale feature—multiplying the adaptive projection matrix by the scale feature—yields the signal subspace features at that scale. The same calculation method is used for the other scale features in the multi-scale signal feature sequence to obtain the corresponding adaptive projection matrices and signal subspace features.

[0044] For each scale feature in the multi-scale state feature sequence, the corresponding adaptive projection matrix is ​​calculated using the same method as the signal feature, and the corresponding state subspace feature is obtained. For example, a state feature vector with a dimension of 64 can also be decomposed into four scale levels using wavelet transform, each containing 16 feature elements. The adaptive projection matrix and state subspace feature for each scale are then calculated using the above method.

[0045] The signal subspace features at each scale level are weighted and fused to produce the fused signal features. The weighting can be assigned based on the importance of each scale level. For example, the weight can be determined based on the information entropy of each scale level, with higher entropy levels receiving greater weights. Assuming the weights of the four scale levels are 0.4, 0.3, 0.2, and 0.1, respectively, these weights are multiplied by the corresponding signal subspace features and then added to produce the fused signal features. Similarly, the state subspace features are subjected to the same weighted fusion process to produce the fused state features.

[0046] The fused signal features and state features are input into the generative network to generate decoupled signal features and decoupled state features. The generative network can adopt a generative adversarial network structure, consisting of encoder and decoder modules. The encoder maps the fused features to a latent space, and the decoder reconstructs the latent space representation into decoupled features. For example, the encoder can include three fully connected layers, with 256, 128, and 64 neurons in each layer, respectively, and the activation function is ReLU; the decoder can also include three fully connected layers, with 64, 128, and 256 neurons in each layer, respectively. The final output is decoupled features with the same dimensionality as the original features.

[0047] After generating the decoupled signal and state features from the network output, calculate the correlation score between these two features. The correlation score can be measured by calculating the cosine similarity or mutual information between the two features. In this example, cosine similarity is used to calculate the correlation, with smaller correlation scores indicating better decoupling. Assume the initial correlation score is 0.45, and the preset decoupling completion threshold is 0.1.

[0048] The network parameters are iteratively optimized until the feature correlation score falls below the decoupling completion threshold. This optimization process uses gradient descent, with a loss function consisting of a reconstruction loss and an orthogonality loss. The reconstruction loss ensures that the decoupled features retain the original features, while the orthogonality loss ensures that the signal and state features are orthogonal to each other. The network parameters are updated using a backpropagation algorithm, and the feature correlation score is recalculated after each iteration. After approximately 200 iterations, the correlation score drops to 0.08, which is below the preset decoupling completion threshold of 0.1. At this point, the decoupled signal and state features are output, completing the feature decoupling process.

[0049] The decoupled signal features and state features can be used for subsequent classification, regression, or anomaly detection tasks. For example, in equipment fault diagnosis applications, the decoupled signal features can more accurately reflect the health status of the equipment without being affected by environmental conditions, thereby improving the accuracy of fault diagnosis.

[0050] In an optional embodiment, the decoupled signal features and state features are reconstructed using a sparse coding dictionary, the reconstructed features are input into a fully connected classifier, and the drone attribute information corresponding to the feature with the highest confidence is used as the final recognition result, including: The decoupled signal features and the decoupled state features are concatenated to obtain decoupled features; the intra-class feature variance of the decoupled features within different drone categories and the inter-class feature variance between different drone categories are calculated; and the ratio of the intra-class feature variance to the inter-class feature variance is used as a discriminant factor; Based on the discriminant factor, the decoupled features are divided into multiple groups; for each feature group, a feature similarity matrix within the group is calculated, and based on the similarity matrix, the number of atoms in the sub-dictionary corresponding to the group is determined by a clustering algorithm; Cascading the sub-dictionaries corresponding to all feature groups in order of the size of the discriminant factors to construct a sparse coding dictionary; calculating the sparse representation coefficient of the decoupled feature under the sparse coding dictionary, and taking the product of the sparse representation coefficient and the sparse coding dictionary as the reconstructed decoupled feature; A multi-branch fully connected classification network is constructed, where each branch corresponds to the reconstructed decoupled features of a feature group; the reconstructed decoupled features of each branch are input into the corresponding branch network to obtain the branch recognition results, the confidence scores of each branch recognition results are calculated, and the branch recognition result with the highest confidence score is selected as the final recognition result.

[0051] In order to reconstruct the decoupled signal features and state features using a sparse coding dictionary for drone identification, this embodiment provides a detailed technical implementation solution.

[0052] In this embodiment, after obtaining the decoupled signal features and state features, they are spliced ​​together to obtain decoupled features. During the splicing process, assuming that the dimension of the decoupled signal features is 256 and the dimension of the state features is 128, the dimension of the spliced ​​decoupled features is 384. Based on the decoupled features, the feature variance within different drone categories and the feature variance between categories are calculated. Specifically, for drones of a certain category C, all decoupled feature samples under that category are collected, and the variance on each feature dimension is calculated to obtain the intra-class feature variance. For example, for feature dimension j, the values ​​of all samples of category C on that dimension are collected, and its variance is calculated as the intra-class feature variance. Similarly, the variance between different categories on the same feature dimension is calculated as the inter-class feature variance. The intra-class feature variance is divided by the inter-class feature variance to obtain the discriminant factor. The larger the discriminant factor value, the greater the contribution of the feature dimension to the distinction between drone categories.

[0053] Based on the calculated discriminant factors, the decoupled features are divided into multiple groups. In practical applications, three groups can be set and sorted from highest to lowest according to the discriminant factors, with the top 30% of features grouped into the first group, the middle 40% into the second group, and the remaining 30% into the third group. For each feature group, a similarity matrix is ​​calculated between the features within the group. Similarity can be calculated using cosine similarity. For feature i and feature j, the cosine similarity between them is calculated and entered into the corresponding position in the similarity matrix. Based on the constructed similarity matrix, spectral clustering is used to determine the number of atoms in the sub-dictionary corresponding to each group. Spectral clustering analyzes the eigenvectors of the similarity matrix to identify the inherent clustering structure of the data. In practice, the similarity matrix can be decomposed using eigenvalues, and the number of eigenvalues ​​before the point of sharp drop in eigenvalues ​​is selected as the number of atoms. For example, the number of atoms determined for the first group of features might be 25, for the second group 40, and for the third group 15.

[0054] After determining the number of atoms, the K-SVD algorithm is used to learn the sub-dictionary corresponding to each group. The K-SVD algorithm consists of two main steps: sparse coding and dictionary updating. In the sparse coding stage, the dictionary is fixed and the sparse representation coefficients are solved for each training sample. In the dictionary updating stage, the atoms in the dictionary are updated column by column, and the corresponding sparse representation coefficients are also updated. By iterating these two steps multiple times, the sub-dictionary corresponding to each feature group is finally obtained. For example, the dimension of the sub-dictionary corresponding to the first group of features is 120×25, where 120 is the dimension of the feature group and 25 is the number of atoms.

[0055] The sub-dictionaries corresponding to all feature groups are concatenated in order of discriminant factor size to construct a complete sparse coding dictionary. For example, the three sub-dictionaries D1 (120×25), D2 (150×40), and D3 (114×15) are concatenated into a complete dictionary D (384×80). Using this sparse coding dictionary, the sparse representation coefficients of the decoupled features under the dictionary are calculated. The orthogonal matching pursuit (OMP) algorithm can be used to solve the sparse representation coefficient α, setting the sparsity to 10, that is, each decoupled feature is represented by no more than 10 dictionary atoms. After obtaining the sparse representation coefficient, the product of the sparse representation coefficient and the sparse coding dictionary is used as the reconstructed decoupled feature.

[0056] A multi-branch fully connected classification network is constructed, with each branch corresponding to the reconstructed decoupled features of a feature group. The number of hidden units in each branch is determined by the discriminant factor of the corresponding group. For example, if the average discriminant factors of the three groups are 0.85, 0.45, and 0.25, respectively, the number of hidden units in the first branch can be set to 512, the second branch to 256, and the third branch to 128. Each branch contains two fully connected layers. The first layer maps the input features to the dimension of the number of hidden units, and the second layer maps the hidden features to the dimension of the number of drone categories. The ReLU function is used as the activation function, and the Dropout technology is used to prevent overfitting, with a dropout rate set to 0.5.

[0057] The reconstructed decoupled features of each group are input into the corresponding branch network to obtain the branch recognition results. The output of each branch is converted into a class probability distribution using the Softmax function, and the confidence score of each branch recognition result is calculated. The confidence score can be determined by the maximum probability value. For example, if the probability of the first branch identifying a Class A drone is 0.92, its confidence score is 0.92. The confidence scores of the three branches are compared, and the branch recognition result with the highest confidence score is selected as the final recognition result. For example, if the confidence scores of the three branches are 0.92, 0.85, and 0.78, respectively, the recognition result of the first branch is selected as the final result.

[0058] Through the above implementation, a technical solution is achieved by reconstructing the decoupled features using a sparse coding dictionary and performing drone identification through a multi-branch fully connected classification network, which effectively improves the accuracy and robustness of drone identification.

[0059] In an optional embodiment, based on the discriminant factor, the decoupled features are divided into multiple groups; a feature similarity matrix within each feature group is calculated, and the number of atoms in the sub-dictionary corresponding to the group is determined by a clustering algorithm based on the similarity matrix, including: Arranging the discriminant factors in descending order of numerical value to obtain a discriminant factor sequence, calculating the first-order difference of the discriminant factor sequence to obtain a difference sequence, and calculating the global mean of the difference sequence; dividing the difference sequence into a plurality of subintervals and calculating the interval mean, determining the optimal segmentation point of the difference sequence based on the difference between the interval mean and the global mean, and dividing the feature vector into a plurality of feature groups based on the optimal segmentation point; Calculating an initial similarity matrix corresponding to the eigenvectors in each of the feature groups; calculating the cosine similarity between the eigenvectors in each of the feature groups, multiplying the difference between the cosine similarity and the similarity threshold by the product of the initial similarity matrix to obtain an optimized similarity matrix, and calculating the eigenvalue sequence of the optimized similarity matrix; The adjacent eigenvalue differences of the eigenvalue sequence are calculated to obtain a spectral interval sequence, and the spectral interval sequence is divided by the corresponding eigenvalue to obtain a significance score sequence; the product of the maximum value of the significance score sequence and the number of samples of the feature group is determined as the number of atoms in the sub-dictionary corresponding to the feature group.

[0060] The present invention provides a dictionary learning method based on decoupling features, and its specific implementation is detailed as follows: After obtaining the decoupling features and corresponding discriminant factors, the system divides the decoupling features into multiple groups. Specifically, the discriminant factors are sorted in descending order of magnitude. For example, the discriminant factor sequence is [0.95, 0.92, 0.87, 0.85, 0.65, 0.62, 0.58, 0.32, 0.30, 0.28]. The first-order difference of this discriminant factor sequence is calculated to obtain the difference sequence [0.03, 0.05, 0.02, 0.20, 0.03, 0.04, 0.26, 0.02, 0.02]. The global mean of this difference sequence is 0.07. The difference sequence is divided into three subintervals: [0.03, 0.05, 0.02], [0.20, 0.03, 0.04], and [0.26, 0.02, 0.02]. The mean values ​​for each subinterval are calculated to be 0.033, 0.09, and 0.10, respectively. The differences between the mean values ​​for each subinterval and the global mean are calculated to be -0.037, 0.02, and 0.03, respectively. Based on these differences, the optimal split points are determined to be after the fourth and seventh elements, corresponding to the difference values ​​of 0.20 and 0.26. Based on this, the original feature vector is divided into three feature groups: the first group contains the first four features, the second group contains features 5 to 7, and the third group contains features 8 to 10.

[0061] For each feature group, calculate the Euclidean distance between the feature vectors within the group. Assume that the first set of feature vectors is a 4×100-dimensional matrix, representing 100 samples, each with 4 features. Calculate the Euclidean distances between these 100 samples, resulting in a 100×100 distance matrix. Substitute these distance values ​​into a Gaussian kernel function for transformation, with the kernel bandwidth parameter set to 1.5 times the median of the distance matrix, to obtain the initial similarity matrix. Simultaneously, calculate the cosine similarity between these 100 samples, resulting in a 100×100 cosine similarity matrix. Set the similarity threshold to 0.85, calculate the difference between the cosine similarity and the threshold, and then multiply this difference element-by-element with the initial similarity matrix to obtain the optimized similarity matrix. For example, if the cosine similarity of two samples is 0.90, which is higher than the threshold of 0.85, the difference is 0.05, and the initial similarity is 0.78, then the optimized similarity is 0.78 × 0.05 = 0.039. This approach makes the sample pairs with cosine similarity lower than the threshold have lower similarity in the optimized similarity matrix, thereby enhancing the distinguishing ability of feature similarity.

[0062] Calculate the eigenvalues ​​of the optimized similarity matrix to obtain an eigenvalue sequence, for example, [10.5, 8.2, 6.8, 4.3, 3.1, 2.0, 1.5, 1.2, 0.9, 0.5]. Calculate the difference between adjacent eigenvalues ​​to obtain the spectral interval sequence [2.3, 1.4, 2.5, 1.2, 1.1, 0.5, 0.3, 0.3, 0.4]. Divide each value in the spectral interval sequence by the corresponding eigenvalue to obtain the significance score sequence [0.219, 0.171, 0.368, 0.279, 0.355, 0.250, 0.200, 0.250, 0.444]. Determine the maximum significance score to be 0.444, corresponding to the ninth interval. Multiplying this maximum significance score by the number of samples in the feature group, 100, yields 44.4. After rounding, the number of atoms in the sub-dictionary corresponding to the feature group is determined to be 44.

[0063] For each feature group, a sub-dictionary is constructed using the K-SVD algorithm based on a predetermined number of atoms. For the first group, a sub-dictionary containing 44 atoms is required. A 4×44 random matrix is ​​initialized as the initial value for the sub-dictionary, where 4 represents the feature dimension and 44 represents the number of atoms. During the iterative optimization process, the optimization objective consists of two components: a reconstruction error term and a structural constraint term. The reconstruction error term measures the error in the representation of the original features using the sub-dictionary. The structural constraint term is calculated using the covariance matrix of the feature group and is used to preserve the structural information of the features. Specifically, a 4×4 covariance matrix is ​​calculated for the feature group. Eigendecomposition is performed on the covariance matrix to extract the main eigendirections, which serve as the basis for the structural constraint term. In each iteration, each atom is updated: all other atoms and all sparse coefficients are fixed, the residual of the current atom is calculated, the singular value decomposition is performed on the residual, and the first left singular vector is used as the updated atom. A structural constraint term is also introduced to ensure that the updated atom maintains reconstruction capability while still reflecting the structural information of the feature group.

[0064] The iterative optimization sets the maximum number of iterations to 100 and the convergence threshold to 0.001. If the change in the sub-dictionary between two consecutive iterations is less than the threshold, it is considered to have converged. Usually the algorithm converges after 30-50 iterations to obtain the final sub-dictionary. For other feature groups, the same method is used to determine the number of atoms and construct the sub-dictionary. For example, for the second and third groups of features, the number of atoms may be 35 and 20, respectively. In the end, the sizes of the sub-dictionaries corresponding to the three feature groups are 4×44, 3×35, and 3×20, respectively. This method of constructing sub-dictionaries based on feature groups can better capture the internal structure of different feature groups and improve the efficiency and accuracy of dictionary representation.

[0065] The drone electronic feature recognition system based on a deep convolutional network in an embodiment of the present invention includes: The first unit is used to obtain electronic characteristic data of the UAV, wherein the electronic characteristic data includes communication signal data and flight status data; The second unit is used to input the communication signal data into the signal processing channel composed of multiple convolution layers connected in series. The size of each convolution kernel is adaptively adjusted according to the frequency characteristics of the signal in the current layer to extract the signal feature vector; The third unit is used to input the flight status data into the bidirectional gated recursive unit, which processes the input data from both the forward and backward directions, controls the information flow using a gating mechanism, and generates a state feature vector that integrates the bidirectional time series information; A fourth unit is configured to project the signal feature vector and the state feature vector into mutually unrelated semantic subspaces, calculate a correlation score of the semantic subspace features, and complete feature decoupling when the correlation score is less than a decoupling completion threshold to obtain decoupled signal features and state features; The fifth unit is used to reconstruct the decoupled signal features and state features using a sparse coding dictionary, input the reconstructed features into a fully connected classifier, and use the drone attribute information corresponding to the feature with the highest confidence as the final recognition result, wherein the final recognition result includes at least one of the drone model, manufacturer, and operation type.

[0066] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0067] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0068] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying electronic features of drones based on deep convolutional networks, characterized in that: include: Acquiring electronic characteristic data of the drone, the electronic characteristic data including communication signal data and flight status data; The communication signal data is input into a signal processing channel composed of multiple convolutional layers in series. The size of each convolution kernel is adaptively adjusted according to the frequency characteristics of the current layer signal to extract the signal feature vector; The flight status data is input into a bidirectional gated recursive unit, which processes the input data from both the forward and backward directions, controls the information flow using a gating mechanism, and generates a state feature vector that incorporates bidirectional temporal information. Projecting the signal feature vector and the state feature vector into mutually unrelated semantic subspaces, respectively, calculating the correlation score of the semantic subspace features, and completing feature decoupling when the correlation score is less than a decoupling completion threshold to obtain decoupled signal features and state features; The decoupled signal features and state features are reconstructed using a sparse coding dictionary, and the reconstructed features are input into a fully connected classifier. The drone attribute information corresponding to the feature with the highest confidence is used as the final recognition result, where the final recognition result includes at least one of the drone model, manufacturer, and operation type.

2. The method according to claim 1, characterized in that The communication signal data is input into a signal processing channel composed of multiple convolutional layers in series. The size of each convolution kernel is adaptively adjusted according to the frequency characteristics of the current layer signal to extract the signal feature vector, including: Performing Fourier transform on the communication signal data to obtain a time-frequency spectrum, calculating the energy density of the communication signal at each frequency point, and taking the maximum value of the energy density as the maximum value of the spectrum energy; Based on the energy density, a dynamic convolution kernel is constructed; a base convolution kernel size is multiplied by the signal frequency value and the energy value respectively, and a weighted sum is performed to obtain a dynamic convolution kernel size; Using the dynamic convolution kernel to extract features from the communication signal, performing dense convolution on feature areas with energy density higher than an energy threshold, and using sparse convolution on feature areas with energy density lower than the energy threshold, to generate a multi-scale feature map sequence; Performing feature enhancement on each feature map in the multi-scale feature map sequence, adaptively fusing the current feature map with two adjacent feature maps, and calculating the fusion weight through a feature attention mechanism to obtain a fused feature map; The fused feature map is used as the input of the next convolutional layer, and the convolution operation and feature fusion process are repeated until the feature extraction of all convolutional layers is completed; global feature extraction is performed on the last layer of fused feature map to obtain a signal feature vector containing signal frequency, energy and time domain characteristics.

3. The method according to claim 1, characterized in that The flight status data is input into a bidirectional gated recursive unit, which processes the input data from both the forward and backward directions, controls the information flow using a gating mechanism, and generates a state feature vector that integrates bidirectional temporal information, including: Reconstruct the flight status data into a state feature sequence in chronological order; Calculating the cosine similarity between the state features of the current moment and the state features of the two adjacent time steps to obtain a local correlation matrix; normalizing the local correlation matrix to obtain local feature weights; multiplying the local feature weights by the state feature sequence to generate a short-term feature sequence; Mapping the state feature sequence to a semantic space through a linear transformation; calculating the dot product similarity of feature vectors in the semantic space to generate a semantic similarity matrix; normalizing the semantic similarity matrix to obtain semantic feature weights; and multiplying the semantic feature weights by the state feature sequence to generate a semantic feature sequence; Inputting the state feature sequence into a multilayer perceptron to obtain a forward feature sequence and a backward feature sequence; calculating the Wasserstein distance between the forward feature sequence and the backward feature sequence as a difference; performing a gradient update on the feature extraction parameters in the multilayer perceptron based on the difference to obtain updated forward feature sequence and backward feature sequence, and generating an adversarial feature sequence; The short-term feature sequence, the semantic feature sequence and the adversarial feature sequence are weightedly combined to obtain a state feature vector.

4. The method according to claim 1, wherein The signal feature vector and the state feature vector are respectively projected into mutually unrelated semantic subspaces, and a correlation score of the semantic subspace features is calculated. When the correlation score is less than a decoupling completion threshold, feature decoupling is completed, and decoupled signal features and state features are obtained, including: The signal feature vector and the state feature vector are respectively transformed into feature sequences of multiple scale levels by wavelet transform to obtain a multi-scale signal feature sequence and a multi-scale state feature sequence; The ratio of the variance of each scale feature in the multi-scale signal feature sequence to the mean of the variance of the multi-scale signal feature sequence is used as a regularization coefficient, the product of the regularization coefficient and the transposed matrix of the scale feature is used as a correction term, the correction term is summed with the autocorrelation matrix of the scale feature and then inverted to obtain an adaptive projection matrix of the scale; the adaptive projection matrix is ​​applied to the corresponding scale feature to obtain a signal subspace feature of the scale; Using the same calculation method, the corresponding adaptive projection matrix is ​​calculated for the features of each scale in the multi-scale state feature sequence, and the corresponding state subspace features are obtained; performing weighted fusion on the signal subspace features and the state subspace features of each scale level respectively to obtain fused signal features and fused state features, inputting the fused signal features and the fused state features into a generation network to generate decoupled signal features and decoupled state features, and calculating correlation scores of the decoupled signal features and the decoupled state features; Iteratively optimize the parameters of the generation network until the feature correlation score is less than the decoupling completion threshold, and output the decoupled signal features and the decoupled state features.

5. The method according to claim 1, wherein The decoupled signal features and state features are reconstructed using a sparse coding dictionary. The reconstructed features are input into a fully connected classifier, and the drone attribute information corresponding to the feature with the highest confidence is used as the final recognition result, including: The decoupled signal features and the decoupled state features are concatenated to obtain decoupled features; the intra-class feature variance of the decoupled features within different drone categories and the inter-class feature variance between different drone categories are calculated; and the ratio of the intra-class feature variance to the inter-class feature variance is used as a discriminant factor; Based on the discriminant factor, the decoupled features are divided into multiple groups; for each feature group, a feature similarity matrix within the group is calculated, and based on the similarity matrix, the number of atoms in the sub-dictionary corresponding to the group is determined by a clustering algorithm; Cascading the sub-dictionaries corresponding to all feature groups in order of the size of the discriminant factors to construct a sparse coding dictionary; calculating the sparse representation coefficient of the decoupled feature under the sparse coding dictionary, and taking the product of the sparse representation coefficient and the sparse coding dictionary as the reconstructed decoupled feature; A multi-branch fully connected classification network is constructed, where each branch corresponds to the reconstructed decoupled features of a feature group; the reconstructed decoupled features of each branch are input into the corresponding branch network to obtain the branch recognition results, the confidence scores of each branch recognition results are calculated, and the branch recognition result with the highest confidence score is selected as the final recognition result.

6. The method according to claim 5, characterized in that Based on the discriminant factor, the decoupled features are divided into multiple groups; for each feature group, a feature similarity matrix within the group is calculated, and based on the similarity matrix, the number of atoms in the sub-dictionary corresponding to the group is determined by a clustering algorithm, including: Arranging the discriminant factors in descending order of numerical value to obtain a discriminant factor sequence, calculating the first-order difference of the discriminant factor sequence to obtain a difference sequence, and calculating the global mean of the difference sequence; dividing the difference sequence into a plurality of subintervals and calculating the interval mean, determining the optimal segmentation point of the difference sequence based on the difference between the interval mean and the global mean, and dividing the feature vector into a plurality of feature groups based on the optimal segmentation point; Calculating an initial similarity matrix corresponding to the eigenvectors in each of the feature groups; calculating the cosine similarity between the eigenvectors in each of the feature groups, multiplying the difference between the cosine similarity and the similarity threshold by the product of the initial similarity matrix to obtain an optimized similarity matrix, and calculating the eigenvalue sequence of the optimized similarity matrix; The adjacent eigenvalue differences of the eigenvalue sequence are calculated to obtain a spectral interval sequence, and the spectral interval sequence is divided by the corresponding eigenvalue to obtain a significance score sequence; the product of the maximum value of the significance score sequence and the number of samples of the feature group is determined as the number of atoms in the sub-dictionary corresponding to the feature group.

7. A drone electronic feature recognition system based on a deep convolutional network, for implementing the method according to any one of claims 1 to 6, characterized in that: include: The first unit is used to obtain electronic characteristic data of the UAV, wherein the electronic characteristic data includes communication signal data and flight status data; The second unit is used to input the communication signal data into the signal processing channel composed of multiple convolution layers connected in series. The size of each convolution kernel is adaptively adjusted according to the frequency characteristics of the signal in the current layer to extract the signal feature vector; The third unit is used to input the flight status data into the bidirectional gated recursive unit, which processes the input data from both the forward and backward directions, controls the information flow using a gating mechanism, and generates a state feature vector that integrates the bidirectional time series information; A fourth unit is configured to project the signal feature vector and the state feature vector into mutually unrelated semantic subspaces, calculate a correlation score of the semantic subspace features, and complete feature decoupling when the correlation score is less than a decoupling completion threshold to obtain decoupled signal features and state features; The fifth unit is used to reconstruct the decoupled signal features and state features using a sparse coding dictionary, input the reconstructed features into a fully connected classifier, and use the drone attribute information corresponding to the feature with the highest confidence as the final recognition result, wherein the final recognition result includes at least one of the drone model, manufacturer, and operation type.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images