Unmanned aerial vehicle signal open set identification method under interference condition

Abstract

This invention discloses an open-set UAV signal recognition method under interference conditions, relating to the field of UAV radio frequency signal recognition. This invention achieves open-set UAV RF signal recognition under interference conditions by embedding a coordinate attention mechanism and a designed rejection model into a residual network structure. Our proposed RF-based open-set UAV recognition method mainly includes three steps: (i) RF signal acquisition; (ii) MVDR feature extraction; and (iii) open-set UAV RF signal recognition based on a deep learning network. For the joint loss rejection model for unknown RF signal categories, this model comprehensively considers cross-entropy loss and open-set regularization loss. A novel CARJ-OSR deep learning network is designed, combining the coordinate attention mechanism and the aforementioned rejection model with a residual network for open-set UAV RF signal recognition.

Description

Technical Field

[0001] This invention relates to the field of radio frequency signal identification for unmanned aerial vehicles (UAVs). Background Technology

[0002] Existing open set recognition methods can be roughly divided into three categories: (1) confidence-based open set recognition methods; (2) reconstruction-based open set recognition methods; and (3) distance learning-based open set recognition methods.

[0003] The core idea of ​​confidence-based OSR methods is to assume that signals from unidentified drones typically result in low prediction confidence from the classifier. Therefore, the unknown class is detected by comparing the output probability with a preset threshold. Based on this idea, the paper A. Bendale and TE Boult, “Towards open set deepnetworks,” in Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016, pp. 1563–1572, proposes the OpenMax layer for estimating the probability that an input signal belongs to an unknown class in open-set image recognition. The paper X. Yin, B. Cao, Q. Hu, and Q. Wang, “RD-OpenMax: Rethinking OpenMax for robust realistic open-set recognition,” IEEE Transactions on Neural Networks and Learning Systems, vol. 36, no. 4, pp. 7565 7579, April 2025, introduces a covariance pooling module based on covariance attention as a global feature aggregation step and proposes a regularized discriminative OpenMax (RD-OpenMax) layer to enhance classification robustness. The paper X. Zhang, Y. Huang, M. Lin, Y. Tian, ​​and J. An, “Transmitter identification with contrastive learning in incremental open-set recognition,” IEEE Internet of Things Journal, vol. 11, no. 3, pp. 4693–4711, February 2024, proposes an open-set recognizer based on Weibull calibration, which establishes a suitable open-set decision boundary through output calibration for radio frequency fingerprint recognition.

[0004] The reconstruction-based OSR method is based on the following assumption: only UAV signals of known categories can be accurately reconstructed by autoencoders or generative models, while UAV signals of unknown categories will produce large reconstruction errors and thus be rejected. The literature P. Oza and VM Patel, “C2AE: Class conditioned auto-encoder for open-set recognition,” in Proceedings of IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR), vol. June, April 2019, pp. 2302–2311, proposes an open-set recognition algorithm based on a class-conditioned autoencoder and designs a novel training and testing strategy. The paper R. Yoshihashi, W. Shao, R. Kawakami, S. You, M. Iida, and T. Naemura, “Classification-reconstruction learning for open-set recognition,” in Proceedings of IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019, pp. 4011–4020, proposes a classification-reconstruction joint learning (CROSR) method, which enhances the ability to detect unknown classes without reducing the accuracy of known class classification by utilizing latent representations for reconstruction.

[0005] Distance-based OSR methods assume that unknown class samples are typically far from the cluster centers of known class samples in the feature space. Therefore, they identify unknown class UAV signals by constructing decision boundaries in the feature space. The paper Y. Dong, X. Jiang, H. Zhou, Y. Lin, and Q. Shi, “SR2CNN: Zero shot learning for signal recognition,” IEEE Transactions on Signal Processing, vol. 69, pp.2316–2329, March 2021, proposes a Signal Recognition and Reconstruction Convolutional Neural Network (SR2CNN), which can learn the representation of the signal semantic feature space even when training data is lacking for certain signal categories. The paper C. Wu, S. Chen, G. Sun, and H. Fang, “Open set RF fingerprint identification for wireless communication devices,” IEEE Wireless Communications Letters, vol.14, no.3, pp.776–780, March 2025, proposes an Open Set Support Vector Data Description (OpenSVDD) model based on feature distribution optimization and classification boundary fitting for RF fingerprint recognition.

[0006] In summary, all existing open-set identification schemes only refer to "unknown category" as the type of UAV signal, failing to consider "interference signals." In real-world environments, LLS UAVs typically communicate with their controllers via the Industrial, Scientific, and Medical (ISM) band, which contains numerous congested coexisting signals such as Wi-Fi, Bluetooth, and ZigBee. However, existing literature data generally does not include these coexisting ISM signals. Besides additive white Gaussian noise (AWGN), these coexisting ISM signals (regardless of whether they originate from a known or unknown category) should be considered interference with UAV radio frequency (RF) signals during testing, a problem that has not been adequately addressed in current research. Therefore, conducting open-set UAV RF signal identification in the presence of such interference is particularly important and challenging. Summary of the Invention

[0007] To address the aforementioned problems, this invention proposes a novel open-set UAV identification method based on radio frequency (RF) and designs a novel deep learning (DL) network. Minimum variance distortionless response (MVDR) spectral analysis is performed on the RF signals acquired within the UAV's operating frequency band to extract MVDR features. Based on this, a novel open-set rejection model based on joint loss is designed.

[0008] This invention achieves open-set UAV RF signal recognition under interference conditions by embedding a coordinate attention mechanism and a designed rejection model into a residual network structure. Our proposed RF-based open-set UAV recognition method mainly includes three steps: (i) RF signal acquisition; (ii) MVDR feature extraction; and (iii) open-set UAV RF signal recognition based on a deep learning network.

[0009] The technical solution of this invention includes: a method for open-set identification of UAV signals under interference conditions, the method comprising:

[0010] Step 1: Acquire RF signals;

[0011] A superheterodyne receiver is used to convert the signal acquired in the ISM band to the intermediate frequency band, and sampling is performed at a fixed sampling rate; discrete-time RF signal sampling sequence. Represented as:

[0012] ;

[0013] in, Indicates the first frequency band in the ISM band A discrete-time RF signal sequence; This indicates that the mean is zero and the variance is... Additive white Gaussian noise; and These represent the acquired RF signal sequences. The quantity and corresponding sampling length;

[0014] Step 2: Extract MVDR features, where MVDR represents the minimum variance distortion-free response;

[0015] Once the monitoring receiver acquires the RF signal sampling sequence Therefore, the MVDR spectral analysis method was applied to it; firstly, the received signal was represented as a column vector. :

[0016] ;

[0017] Next, using a length of A sliding window with a step size of 1 receives the signal sequence. Divided into The data sample matrix is ​​constructed by dividing the data into overlapping sub-segments and converting each sub-segment into a column vector.

[0018] ;

[0019] in Therefore, the corresponding correlation matrix Represented as:

[0020] ;

[0021] Let the normalized frequency be ,in , The index of discrete frequency points is represented by the number of indexes. MVDR spectral analysis adaptively adjusts the weight vector. The optimization problem, which aims to minimize other frequency components while maintaining a distortion-free target frequency response, can be expressed as:

[0022] ;

[0023] Where the direction vector for:

[0024] ;

[0025] It is the standard representation of the imaginary unit;

[0026] Optimal weight vector for:

[0027] ;

[0028] Corresponding MVDR power spectrum for:

[0029] ;

[0030] To improve the resolution of spectral estimation, the correlation matrix is ​​diagonally loaded to obtain the loaded correlation matrix. :

[0031] ;

[0032] in , For trace operation, for The identity matrix; at this time, the MVDR power spectrum is updated as follows:

[0033] ;

[0034] The normalized MVDR power spectrum is defined as:

[0035] ;

[0036] in:

[0037] ;

[0038] Next, this The vectors are plotted as curves and converted into grayscale matrices;

[0039] First, adopt Bit-bit quantizer Calculate quantized elements :

[0040] ;

[0041] in This represents the ceiling operation, which returns a value greater than or equal to the number it contains; correspondingly... Grayscale image matrix It means that, among them:

[0042] ;

[0043] Reconstruct the grayscale image matrix as follows: A square matrix; finally normalized and converted into a third-order tensor; ,in

[0044] ;

[0045] ;

[0046] The MVDR feature tensor represents the standardized grayscale value. This will be used as the input to the open set recognition network proposed in this paper;

[0047] Step 3: Use an open-set deep learning network to identify open-set UAV RF signals;

[0048] The open-set deep learning network includes: a convolutional unit, a feature extraction unit based on coordinate attention mechanism, a known UAV RF signal classification unit, and a rejection unit based on joint loss;

[0049] The convolutional unit sequentially comprises: a convolutional layer, a batch normalization (BN) layer, a ReLU activation layer, and a max pooling layer. The input to the convolutional unit is the MVDR feature tensor. , The data is processed sequentially through BN, ReLU, and max pooling layers, and finally the output of the convolutional unit is... Represented as:

[0050] ;

[0051] in Indicates the convolution kernel weights. Indicate the bias term; then, output the tensor. It is fed into the feature extraction unit based on the coordinate attention mechanism;

[0052] The feature extraction unit based on the coordinate attention (CA) mechanism sequentially includes eight residual blocks with identical structures, a CA mechanism, and a global average pooling layer; wherein, each residual block sequentially includes: a first convolutional layer Conv1, a first batch normalization (BN) layer BN1, a first ReLU layer ReLU1, a second convolutional layer Conv2, a second batch normalization (BN) layer BN2, and then concatenates it with the input of the current input residual block and inputs it into the second ReLU layer ReLU2;

[0053] For the Each residual block The input tensor is The output tensor is , among which when hour, ,otherwise, This is the output of the previous residual block; the first convolutional layer processes the input sequentially through Conv1→BN1→ReLU, and the intermediate results are:

[0054] ;

[0055] in For the first The convolution kernel of the first convolutional layer for each residual block The bias term is used; the second convolutional layer is connected to the residual, and the intermediate result continues to pass through Conv2→BN2→shortcut addition with the input→ReLU, and the output is:

[0056] ;

[0057] in The kernel of the second convolutional layer. For bias; output of the 8th residual block To simplify the representation ;

[0058] The specific methods of the CA mechanism are as follows:

[0059] S1: Height direction encoding is achieved by averaging the channel along the width direction.

[0060] ;

[0061] in , Construct vectors The corresponding output tensor is ;

[0062] S2: Width-direction encoding is achieved by averaging the channels along the height direction.

[0063] ;

[0064] in , Construct vectors The corresponding output tensor is Next, the formula Given Japanese style Given Aggregation is performed through a concatenation layer, which enables:

[0065] ;

[0066] S3: After passing through the first convolutional layer, the BN layer, and the ReLU activation function layer in sequence, we obtain:

[0067] ;

[0068] in, and These represent the convolution kernel and bias term involved in the first convolutional layer of the CA mechanism, respectively.

[0069] S3: It is divided into two subtensors, denoted as follows: and ;right and Apply convolutional layers and the Sigmoid activation function respectively, such that:

[0070] ;

[0071] ;

[0072] in, and These represent the convolution kernels used in the last two parallel convolutional layers of the CA mechanism; and These represent the bias terms involved in the last two parallel convolutional layers of the CA mechanism; for the c-th channel, the output of the CA mechanism is:

[0073] ;

[0074] Therefore, the overall output tensor of the CA mechanism is Finally, a global average pooling layer, denoted here as "GAP", is applied to... Thus, the output of the CA-based feature extraction unit is obtained:

[0075] ;

[0076] This output tensor serves as the input to both the rejection unit and the classification unit.

[0077] The classification unit for the known UAV radio frequency signals:

[0078] Output of the feature extraction unit based on CA The data is directly input into the classification unit and then passes through a fully connected layer and a softmax activation function layer to classify the received UAV radio frequency signals. This classification is based on a pre-trained candidate set. The process is carried out; the pre-trained candidate set of UAV RF signals is represented as... in, Indicates the first Given known candidate categories of drone RF signals, and The output of a fully connected layer is defined as:

[0079] ;

[0080] in, , , Represents a set The cardinality; subsequently, the received signal belongs to the set. Radio frequency signals of the k-th type of UAV The probability is defined as:

[0081] ;

[0082] Therefore, the type of UAV radio frequency signal can be estimated as follows:

[0083] ;

[0084] in,

[0085] ;

[0086] The rejection unit under the joint loss constraint:

[0087] The rejection unit uses feature vectors As input, unknown radio frequency signals are identified based on the extracted features; here, the joint loss function of the rejection unit combines cross-entropy loss and open set regularization loss to measure the feature differences between known and unknown radio frequency signals and enhance the distinguishability of known UAV radio frequency signal features.

[0088] (1) Establish cross-entropy loss:

[0089] During the training phase, each training batch contains M radio frequency signal sample sequences. These samples were randomly selected from the training dataset; according to the formula – Each has a length of Received signal vector Generate the corresponding feature vector From the M received signal vectors selected in the training set, the corresponding M feature vectors are obtained: The actual types of drones are respectively ,Mode and Rewritten as:

[0090] ;

[0091] ;

[0092] Cross-entropy loss is defined as:

[0093] ;

[0094] in Represents the learnable parameters of the model; This is an indicator function used to determine the first... Does each sample belong to the k-th class?

[0095] ;

[0096] (2) Establish open set regularized loss:

[0097] To enhance intra-class compactness, the feature vectors are normalized:

[0098] ;

[0099] Define the feature set of the k-th class:

[0100] ;

[0101] Further expressed as:

[0102] ;

[0103] in , The corresponding mean feature vector of the k-th class is:

[0104] ;

[0105] The open set regularization loss is defined as:

[0106] ;

[0107] (3) Obtain the joint loss function;

[0108] The joint loss function is defined as:

[0109] ;

[0110] Where ϱ represents the weight coefficients, and the model parameters are updated iteratively using the Adam optimizer:

[0111] ;

[0112] Where α is the learning rate; This is used to avoid overflow or numerical divergence in numerical calculations. and express Deviation correction estimates for the first and second moments of the gradient;

[0113] (4) Feature statistical modeling;

[0114] For different signal-to-noise ratio conditions: The normalized eigenvectors are:

[0115] ;

[0116] Calculate the mean:

[0117] ;

[0118] Covariance matrix:

[0119] ;

[0120] (5) Rejection decision during the testing phase;

[0121] During the testing phase, the feature vector of the acquired RF signal can be obtained according to the above process, denoted as . ; can be calculated Mahalanobis distances between the mean eigenvectors and the mean eigenvectors:

[0122]

[0123] The minimum distance is:

[0124] ;

[0125] (6) Criteria for refusing to recognize judgments;

[0126] The rejection criterion used in our novel deep learning network is given by the following equation:

[0127] ;

[0128] Where τ is the rejection threshold; if the signal belongs to a known class, it is classified according to the classification unit.

[0129] This invention addresses the problem of open-set UAV RF signal recognition under interference conditions. It proposes a joint loss rejection model for unknown RF signal categories, which comprehensively considers cross-entropy loss and open-set regularization loss. A novel CARJ-OSR deep learning network is designed, which combines the coordinate attention mechanism and the aforementioned rejection model with a residual network for open-set UAV RF signal recognition. Attached Figure Description

[0130] Figure 1 Block diagram of an unmanned open set recognition system.

[0131] Figure 2 This is the CARJ-OSR deep learning network architecture. Detailed Implementation

[0132] To address the problem of drone identification under interference conditions, this invention proposes a novel radio frequency-based open-set drone identification method and designs a novel deep learning network. For example... Figure 1 As shown, our proposed RF-based open-set UAV recognition method mainly includes three steps: (i) RF signal acquisition; (ii) MVDR feature extraction; and (iii) open-set UAV RF signal recognition based on a deep learning network. These three parts will be explained in detail below:

[0133] RF signal acquisition

[0134] Typically, a superheterodyne (monitoring) receiver is used to convert signals acquired in the ISM band to the intermediate frequency band and sample them at a fixed sampling rate. Since the receiver captures all signals within the entire ISM band, the discrete-time received RF signal samples inevitably contain RF signals from drones and interference signals emitted by other devices coexisting in the ISM band (e.g., Wi-Fi hotspots and Bluetooth transmitters). In this invention, signals generated by other ISM devices are considered "interference," and the drone RF signal (i.e., drone type) must be accurately identified in the presence of both interference and noise. The discrete-time RF signal sampling sequence can be represented as:

[0135] ;

[0136] in, Indicates the first frequency band in the ISM band A discrete-time RF signal sequence; This indicates that the mean is zero and the variance is... Additive white Gaussian noise (AWGN); and These represent the acquired RF signal sequences. The quantity and corresponding sampling length. This invention employs... All are 10000.

[0137] MVDR Feature Extraction

[0138] Once the monitoring receiver acquires the RF signal sampling sequence Therefore, the Minimum Variance Distortionless Response (MVDR) spectral analysis method is applied to it. First, the received signal is represented as a column vector:

[0139] ;

[0140] Next, using a length of (In this invention) A sliding window with a step size of 1 is used to receive the signal sequence. Divided into The data sample matrix is ​​constructed by dividing the data into overlapping sub-segments and converting each sub-segment into a column vector.

[0141] ;

[0142] in Therefore, the corresponding correlation matrix is ​​expressed as:

[0143] ;

[0144] Let the normalized frequency be ,in (In this invention) According to the literature, MVDR spectral analysis adaptively adjusts the weight vector. The optimization problem, which aims to minimize other frequency components while maintaining a distortion-free target frequency response, can be expressed as:

[0145] ;

[0146] Where the direction vector for:

[0147] ;

[0148] Therefore, the optimal weight vector is:

[0149] ;

[0150] The corresponding MVDR power spectrum is:

[0151] ;

[0152] Furthermore, to improve the resolution of spectral estimation, the correlation matrix is ​​diagonally loaded:

[0153] ;

[0154] in , For trace operation, for The identity matrix. The MVDR power spectrum is then updated as follows:

[0155] ;

[0156] Furthermore, the normalized MVDR power spectrum is defined as:

[0157] ;

[0158] in:

[0159] ;

[0160] Next, this The vector is converted into a grayscale matrix. First, using... Bit-bit quantizer ( Quantization of elements

[0161] ;

[0162] in This indicates a ceiling operation, which returns a value greater than or equal to the contained number. Correspondingly... Grayscale image matrix can be used It means that among them

[0163] ;

[0164] Furthermore, reconstruct it into (in A square matrix. Finally, normalize and convert it into a third-order tensor. ,in

[0165] ;

[0166] ;

[0167] The MVDR feature tensor obtained above This will be used as the input to the open set recognition network proposed in this paper.

[0168] Step 3: Apply the proposed novel deep learning network to identify open-set UAV RF signals.

[0169] like Figure 2 As shown, our proposed novel open-set deep learning network consists of the following components: a convolutional unit; a feature extraction unit based on coordinate attention (CA); a known UAV RF signal classification unit; and a rejection unit based on joint loss.

[0170] 1) Convolutional unit

[0171] A convolutional unit consists of the following components: a convolutional layer; a batch normalization (BN) layer; a ReLU activation layer; and a max pooling layer. The input to the convolutional unit is the MVDR feature tensor. The tensor is fed into a convolutional layer containing 64 7×7 convolutional kernels with a stride of 1 and padding of 3. The convolution result is then processed sequentially through a BN layer, a ReLU layer, and a max-pooling layer. The output of the convolutional unit can be represented as:

[0172] ;

[0173] in Indicates the convolution kernel weights. This represents the bias term. Then, the output tensor is used. It is fed into the CA-based feature extraction unit.

[0174] 2) CA-based feature extraction unit

[0175] This unit is used to further extract from the tensor. Key features are extracted from the data. This unit contains eight structurally identical residual blocks, a CA mechanism, and a global average pooling layer. Each residual block contains two convolutional layers (Conv1 and Conv2), two batch normalization (BN1 and BN2) layers, and two ReLU layers (ReLU1 and ReLU2).

[0176] For the Each residual block The input tensor is The output tensor is , among which when hour, ,otherwise, This is the output of the previous residual block. The first convolutional layer processes the input sequentially through Conv1→BN1→ReLU, and the intermediate results are:

[0177] ;

[0178] in For the first The convolution kernel of the first convolutional layer for each residual block This is the bias term. The second convolutional layer is connected to the residual, and the intermediate result continues to pass through Conv2→BN2→shortcut addition with the input→ReLU, and the output is:

[0179] ;

[0180] in The kernel of the second convolutional layer. For bias. Output of the 8th residual block. To simplify the representation .

[0181] The CA mechanism includes: two global average pooling layers (along the H and W directions), three convolutional layers, one batch normalization (BN) layer, one ReLU activation function, and two sigmoid activation functions. Height-direction encoding averages the channels along the width direction.

[0182] ;

[0183] in , Construct vectors The corresponding output tensor is .

[0184] Width-direction encoding is achieved by averaging the channel along the height direction:

[0185] ;

[0186] in , Construct vectors The corresponding output tensor is Next, the formula will be... Given Japanese style Given Aggregation is achieved through a concatenation layer, making...

[0187] ;

[0188] Subsequently, After passing through the first convolutional layer, the BN layer, and the ReLU activation function layer in sequence, the result is obtained.

[0189] ;

[0190] in, and These represent the convolution kernel and bias term involved in the first convolutional layer of the CA mechanism, respectively.

[0191] To further extract attention from spatial features, a third-order tensor... It is divided into two subtensors, denoted as follows: and Subsequently, regarding and Apply convolutional layers and the sigmoid activation function respectively, so that

[0192] ;

[0193] ;

[0194] in, and These represent the convolution kernels used in the last two parallel convolutional layers of the CA mechanism; and These represent the bias terms involved in the last two parallel convolutional layers of the CA mechanism. For the c-th channel, the output of the CA mechanism is:

[0195] ;

[0196] Therefore, the overall output tensor of the CA mechanism is Finally, a global average pooling layer, denoted here as "GAP", is applied to... Thus, the output of the CA-based feature extraction unit is obtained:

[0197] ;

[0198] The output is simultaneously sent to both the rejection unit and the classification unit.

[0199] 3) Classification units of known UAV radio frequency signals:

[0200] Output of the feature extraction unit based on CA The signal is directly input into the classification unit and then passes through a fully connected layer and a softmax activation function layer to classify the received UAV radio frequency signal. This classification is based on a known candidate set. The process is carried out in this way. Without loss of generality, the pre-trained candidate set of UAV RF signals can be represented as... in, Indicates the first Given known candidate categories of drone RF signals, and .

[0201] Specifically, the output of a fully connected layer (FC) is defined as

[0202] ;

[0203] in, , , Represents a set The cardinality (i.e., the number of known drone radio frequency signal categories). Subsequently, the received signal belongs to the set. Radio frequency signals of the k-th type of UAV The probability is defined as

[0204] ;

[0205] Therefore, the type of UAV radio frequency signal can be estimated as follows:

[0206] ;

[0207] in,

[0208] ;

[0209] 4) Rejection unit under joint loss constraint:

[0210] The rejection unit uses feature vectors As input, unknown radio frequency signals are identified based on the extracted features. Here, the present invention proposes a novel joint loss function that combines cross-entropy loss and open-set regularization loss to measure the feature differences between known and unknown radio frequency signals and enhance the discriminability of known UAV radio frequency signal features.

[0211] (1) Cross-entropy loss

[0212] During the training phase, each training batch contains M radio frequency signal sample sequences. These samples were randomly selected from the training dataset. According to the formula... – Each has a length of Received signal vector It can generate the corresponding feature vector. From the M received signal vectors selected in the training set, we can obtain the corresponding M feature vectors: The actual types of drones are respectively Therefore, the formula and It can be rewritten as:

[0213] ;

[0214] ;

[0215] Cross-entropy loss is defined as:

[0216] ;

[0217] in Represents the learnable parameters of the model; This is an indicator function used to determine the first... Does each sample belong to the k-th class?

[0218] ;

[0219] (2) Open set regularization loss

[0220] To enhance intra-class compactness, the feature vectors are normalized:

[0221] ;

[0222] Define the feature set of the k-th class:

[0223] ;

[0224] Further expressed as:

[0225] ;

[0226] in , The corresponding mean eigenvector is:

[0227] ;

[0228] The open set regularization loss is defined as:

[0229] ;

[0230] (3) Joint loss function

[0231] The joint loss function is defined as:

[0232] ;

[0233] Where ϱ represents the weight coefficients. The model parameters are updated iteratively using the Adam optimizer:

[0234] ;

[0235] Where α is the learning rate; It is a preset, very small positive constant used to avoid overflow or numerical divergence in numerical calculations; and express Deviation correction estimation of the first and second moments of the gradient.

[0236] (4) Feature statistical modeling

[0237] For different signal-to-noise ratio conditions: The normalized eigenvectors are:

[0238] ;

[0239] Furthermore, we can see that the mean is:

[0240] ;

[0241] Covariance matrix:

[0242] ;

[0243] (5) Rejection decision during the testing phase

[0244] During the testing phase, the feature vector of the acquired RF signal can be obtained according to the above process, denoted as . It can be calculated. Mahalanobis Distance (MD) between the mean eigenvectors and the mean eigenvectors:

[0245]

[0246] The minimum distance is:

[0247] ;

[0248] (6) Criteria for Denial of Recognition

[0249] The rejection criterion used in our novel deep learning network is given by the following equation:

[0250] ;

[0251] Where τ is the rejection threshold. If the signal belongs to a known class, it is classified according to the classification unit.

[0252] This invention proposes a modeling method for open-set UAV RF signal recognition in complex electromagnetic environments. Addressing the interference from multiple wireless signals such as Bluetooth and WiFi in real-world electromagnetic environments, an open-set recognition model is constructed that includes both known UAV signal recognition and unknown interference signal rejection capabilities, enabling the recognition of open-set UAV RF signals under interference conditions.

[0253] This invention constructs a joint loss rejection model for unknown RF signals. Based on the traditional cross-entropy loss function, an open-set regularization loss term is introduced to improve the model's rejection capability for unknown RF signals, enhancing open-set recognition performance while maintaining the accuracy of known category recognition.

[0254] This invention designs a deep learning network structure for open-set UAV RF signal recognition. Based on a residual network framework, a coordinate attention mechanism is introduced, and combined with the proposed joint loss rejection model, a CARJ-OSR deep learning network is constructed to effectively extract key features of UAV RF signals, thereby improving the recognition accuracy of UAV RF signals in interference environments and the ability to reject unknown signals.

Claims

1. A method for open-set identification of UAV signals under interference conditions, the method comprising: Step 1: Acquire RF signals; A superheterodyne receiver is used to convert the signal acquired in the ISM band to the intermediate frequency band and sample it at a fixed sampling rate; Discrete-time RF signal sampling sequence Represented as: ； in, Indicates the first frequency band in the ISM band A discrete-time RF signal sequence; This indicates that the mean is zero and the variance is... Additive white Gaussian noise; and These represent the acquired RF signal sequences. The quantity and corresponding sampling length; Step 2: Extract MVDR features, where MVDR represents the minimum variance distortion-free response; Once the monitoring receiver has acquired the sequence of RF signal samples , it applies the MVDR spectral analysis method to them; first, the received signal is represented in the form of a column vector : ； Next, using a length of A sliding window with a step size of 1 receives the signal sequence. Divided into The data sample matrix is ​​constructed by dividing the data into overlapping sub-segments and converting each sub-segment into a column vector. ； in Therefore, the corresponding correlation matrix Represented as: ； Let the normalized frequency be ,in , The index of discrete frequency points is represented by the number of indexes. MVDR spectral analysis adaptively adjusts the weight vector. The optimization problem, which aims to minimize other frequency components while maintaining a distortion-free target frequency response, can be expressed as: ； where the direction vector is: ； is the standard notation for imaginary units; optimal weight vector is: ； corresponding MVDR power spectrum is: ； To improve the resolution of the spectrum estimation, the correlation matrix is diagonally loaded to obtain a loaded correlation matrix : ； where , is the trace operation, is the identity matrix; in this case the MVDR power spectrum update is: ； The normalized MVDR power spectrum is defined as: ； in: ； Next, the vectors of the are plotted as a graph and converted into a gray scale matrix; First, a bit quantizer , the quantized elements are calculated ； in This represents the ceiling operation, which returns a value greater than or equal to the number it contains; correspondingly... Grayscale image matrix It means that, among them: ； Reconstruct the grayscale image matrix as follows: A square matrix; finally normalized and converted into a third-order tensor; ,in ； ； The MVDR feature tensor represents the standardized grayscale value. This will be used as the input to the open set recognition network proposed in this paper; Step 3: Use an open-set deep learning network to identify open-set UAV RF signals; The open-set deep learning network includes: a convolutional unit, a feature extraction unit based on coordinate attention mechanism, a known UAV RF signal classification unit, and a rejection unit based on joint loss.

2. The method for open-set identification of UAV signals under interference conditions as described in claim 1, characterized in that, In step 3, the convolutional unit sequentially includes: a convolutional layer, a batch normalization (BN) layer, a ReLU activation layer, and a max pooling layer. The input to the convolutional unit is the MVDR feature tensor. , The data is processed sequentially through BN, ReLU, and max pooling layers, and finally the output of the convolutional unit is... Represented as: ； in Indicates the convolution kernel weights. Indicate the bias term; then, output the tensor. It is fed into the feature extraction unit based on the coordinate attention mechanism.

3. The method for open-set identification of UAV signals under interference conditions as described in claim 1, characterized in that, The feature extraction unit based on the coordinate attention mechanism (CA) in step 3 sequentially includes 8 residual blocks with the same structure, a CA mechanism, and a global average pooling layer; wherein, each residual block sequentially includes: a first convolutional layer Conv1, a first batch normalization (BN) layer BN1, a first ReLU layer ReLU1, a second convolutional layer Conv2, a second batch normalization (BN) layer BN2, and then concatenates it with the input of the current input residual block and inputs it into the second ReLU layer ReLU2; For the Each residual block The input tensor is The output tensor is , among which when hour, ,otherwise, This is the output of the previous residual block; the first convolutional layer processes the input sequentially through Conv1→BN1→ReLU, and the intermediate results are: ； in For the first The convolution kernel of the first convolutional layer of each residual block The bias term is used; the second convolutional layer is connected to the residual, and the intermediate result continues to pass through Conv2→BN2→shortcut addition with the input→ReLU, and the output is: ； in The kernel of the second convolutional layer. For bias; output of the 8th residual block To simplify the representation .

4. The method for open-set identification of UAV signals under interference conditions as described in claim 1, characterized in that, The specific method for classifying the UAV RF signal in step 3 is as follows: Output of the feature extraction unit based on CA The data is directly input into the classification unit and then passes through a fully connected layer and a softmax activation function layer to classify the received UAV radio frequency signals. This classification is based on a pre-trained candidate set. The process is carried out; the pre-trained candidate set of UAV RF signals is represented as... in, Indicates the first Given known candidate categories of drone RF signals, and The output of a fully connected layer is defined as: ； in, , , Represents a set The cardinality; subsequently, the received signal belongs to the set. Radio frequency signals of the k-th type of UAV The probability is defined as: ； Therefore, the type of UAV radio frequency signal can be estimated as follows: ； in, 。 5. The method for open-set identification of UAV signals under interference conditions as described in claim 1, characterized in that, The rejection unit under the joint loss constraint in step 3: The rejection unit uses feature vectors As input, unknown radio frequency signals are identified based on the extracted features; here, the joint loss function of the rejection unit combines cross-entropy loss and open set regularization loss to measure the feature differences between known and unknown radio frequency signals and enhance the distinguishability of known UAV radio frequency signal features. (1) Establish cross-entropy loss: During the training phase, each training batch contains M radio frequency signal sample sequences. These samples were randomly selected from the training dataset; according to the formula – Each has a length of Received signal vector Generate the corresponding feature vector From the M received signal vectors selected in the training set, the corresponding M feature vectors are obtained: ; The actual types of drones are respectively ,Mode and Rewritten as: ； ； Cross-entropy loss is defined as: ； in Represents the learnable parameters of the model; This is an indicator function used to determine the first... Does each sample belong to the k-th class? ； (2) Establish open set regularized loss: To enhance intra-class compactness, the feature vectors are normalized: ； Define the feature set of the k-th class: ； Further expressed as: ； in , The corresponding mean feature vector of the k-th class is: ； The open set regularization loss is defined as: ； (3) Obtain the joint loss function; The joint loss function is defined as: ； Where ϱ represents the weight coefficients, and the model parameters are updated iteratively using the Adam optimizer: ； Where α is the learning rate; This is used to avoid overflow or numerical divergence in numerical calculations. and express Deviation correction estimates for the first and second moments of the gradient; (4) Feature statistical modeling; For different signal-to-noise ratio conditions: The normalized eigenvectors are: ； Calculate the mean: ； Covariance matrix: ； (5) Rejection decision during the testing phase; During the testing phase, the feature vector of the acquired RF signal can be obtained according to the above process, denoted as . ; can be calculated Mahalanobis distances between the mean eigenvectors and the mean eigenvectors: The minimum distance is: ； (6) Criteria for refusing to recognize judgments; The rejection criterion used in our novel deep learning network is given by the following equation: ； Where τ is the rejection threshold; if the signal belongs to a known class, it is classified according to the classification unit.

6. The method for open-set identification of UAV signals under interference conditions as described in claim 3, characterized in that, The specific method of the CA mechanism is as follows: S1: Height direction encoding is achieved by averaging the channel along the width direction. ； in , Construct vectors The corresponding output tensor is ; S2: Width-direction encoding is achieved by averaging the channels along the height direction. ； in , Construct vectors The corresponding output tensor is Next, the formula Given Japanese style Given Aggregation is performed through a concatenation layer, which enables: ； S3: After passing through the first convolutional layer, the BN layer, and the ReLU activation function layer in sequence, we obtain: ； in, and These represent the convolution kernel and bias term involved in the first convolutional layer of the CA mechanism, respectively. S3: It is divided into two sub-tensors, denoted as follows: and ;right and Apply convolutional layers and the Sigmoid activation function respectively, such that: ； ； in, and These represent the convolution kernels used in the last two parallel convolutional layers of the CA mechanism; and These represent the bias terms involved in the last two parallel convolutional layers of the CA mechanism; for the c-th channel, the output of the CA mechanism is: ； Therefore, the overall output tensor of the CA mechanism is Finally, a global average pooling layer, denoted here as "GAP", is applied to... Thus, the output of the CA-based feature extraction unit is obtained: ； This output tensor serves as the input to both the rejection unit and the classification unit.

Images