A method for identifying multi-component interference signals based on multi-label classification networks

Abstract

This invention provides a method for identifying multi-component interference signals based on a multi-label classification network. It converts one-dimensional interference signals into time-frequency images through time-frequency transformation and utilizes the global feature extraction capability of the MobileViT-v2 module for efficient classification and identification of multi-component interference signals. This invention solves the problem that current multi-class classification networks cannot identify signal combinations not present in the training set, has lower requirements for the number of training samples, and overcomes the application limitations of existing radar active interference identification methods.

Description

Technical Field

[0001] This invention relates to a method for identifying multi-component interference signals based on a multi-label classification network, belonging to the field of radar interference signal identification. Background Technology

[0002] As the number of different radiation sources increases, the electromagnetic environment becomes increasingly complex, with high signal density, high pulse sequence interleaving, and even intense competition and confrontation in the frequency spectrum. Radar may receive multiple intentional or unintentional interference signals, which alias in the time or frequency domain to form multi-component interference signals. Because the types and quantities of interference in the received mixed signals are unknown, they often have a significant impact on radar detection. To take more effective and targeted anti-interference measures, it is necessary to accurately and quickly identify the patterns of any unknown interference signals that may exist in the radar echo, thereby ensuring the radar's operational performance.

[0003] A review of existing technical literature reveals that Qu Qizhe et al.'s paper, "JRNet: Jamming Recognition Networks for Radar Compound Suppression Jamming Signals," published in IEEE Transactions on Vehicular Technology (2020, 69(12):15035-45), utilizes the significant power spectrum characteristics of jamming signals to propose a robust power spectrum feature-based jamming recognition network, improving the ability to identify subtle features. However, this method only considers a few fixed combinations of jamming types, limiting the types of jamming discussed. Zhu Mengtao et al.'s paper, "Automatic modulation recognition of compound signals using a deep multi-label classifier: A case study with radar jamming signals," published in Signal Processing, proposes a new framework for compound signals based on deep multi-label learning, applying multi-label classification methods to the identification of compound radar jamming signals. Although it has good scalability for combinations of invisible signals, its performance is somewhat lacking due to the lack of in-depth network mining.

[0004] Existing literature review results indicate that current algorithms for multi-component interference pattern recognition mainly rely on multi-class classification methods. While achieving some results in certain areas, these methods suffer from poor scalability, can only identify fixed types of multi-component interference signals, and their accuracy is affected by the increase in interference types. Furthermore, existing multi-label classification methods lack sufficient depth in feature extraction. Therefore, this paper proposes a multi-component interference signal recognition method based on a multi-label classification network. This method converts the interference signal into a two-dimensional time-frequency image and extracts global features of the interference signal for classification using the ML-SGMViTv2 network. This approach offers higher accuracy and stronger scalability, overcoming the current limitations in radar multi-component interference signal recognition applications. Summary of the Invention

[0005] To address the shortcomings and deficiencies of existing interference signal identification algorithms, this invention designs a multi-component interference signal identification method based on the ML-SGMViTv2 multi-label classification network. The method converts the one-dimensional interference signal into a time-frequency image through time-frequency transformation, and utilizes the global feature extraction capability of the MobileViT-v2 module to efficiently classify and identify the multi-component interference signal. This invention solves the problem that current multi-class classification networks cannot identify signal combinations not present in the training set, has lower requirements for the number of training samples, and overcomes the application limitations of existing radar active interference identification methods.

[0006] The objective of this invention is achieved by the following main steps:

[0007] Step 1: Obtain the multi-component interference signal formed by the linear superposition of radar echo signal and multiple interference signals. J represents the number of components that constitute a single interference signal, which can be composed of at most q types of interference components. i (n) represents the received interference signal, A i Let awgn(n) be the amplitude coefficient modulated when propagating through the environment, and let x be the additive white Gaussian noise. t (n) represents the actual target echo signal.

[0008] Step 2: Preprocessing the multi-component interference signal. First, perform time-frequency transformation on the multi-component interference signal to obtain its time-frequency joint distribution image X(t,ω).

[0009]

[0010] Where ω represents the frequency and h(τ-t) is the analysis window function.

[0011] Step 3: Input the training set into the multi-label classification network for training:

[0012] The multi-label classification network includes a MobileViT-v2 module, a SandGlass module, a global average pooling module, a fully connected layer, and a multi-threshold optimization module. The loss function is set as a binary cross-entropy function.

[0013] Step 31: Input the time-frequency images of the training set into the proposed ML-SGMViTv2 network, use the MobileViT-v2 module to extract the global features of the signal, and use the SandGlass module to remove redundant information, reduce model parameters, and obtain the global feature matrix F;

[0014] Step 32: Input the global feature matrix F into the global average pooling module layer to obtain AvgPool(F); input AvgPool(F) into the fully connected layer to obtain the global feature vector G of the signal. q

[0015] G q =FC(AvgPool(F))

[0016]

[0017] Where FC represents the fully connected layer operation, H and W are the row and column values ​​of the feature matrix F, and F(i,j) represents the element in the i-th row and j-th column of the feature matrix F.

[0018] Step 33: Use the Sigmoid function as the activation function, which transforms a real-valued output z into a probability output in the range (0,1). Then, use the global feature vector G... q The output p of the Sigmoid layer for the j-th output label is input into the Sigmoid function classification layer. j Calculated as

[0019]

[0020] Among them, g j q is the j-th element of the global feature vector, and q is the number of output nodes of the fully connected layer.

[0021] Steps 3 and 4: Using binary cross-entropy as the loss function, calculate the loss independently for each of the q output labels:

[0022] V(p j , t j )=-t j log(p j )-(1-t j log(1-P) j ), j = 1, 2, ..., q

[0023] Where p jt represents the output of the ML-SGMViTv2 model for the j-th label. j This indicates the corresponding real label.

[0024] Step 35: Optimize the ML-SGMViTv2 network using gradient descent and cosine annealing algorithms.

[0025] Step 4: Optimize the decision threshold based on the output results:

[0026] Thresholding strategies significantly impact the performance of multi-label classifiers; therefore, a multi-threshold strategy is employed. This step treats threshold calibration as an optimization problem, based on the input label vector T of the training sample X and the output probability vector P obtained after feature extraction, optimizing the multiple threshold vector M. The threshold induced by l training samples for the j-th label is as follows:

[0027]

[0028] In the formula, y ij , t ij ∈(0,1) represent training samples s respectively i The decision label and the true label of the j-th label.

[0029] Step 5: Input the test set into the trained multi-label classification network to obtain the interference signal components contained in the input signal.

[0030] Compared with existing technologies, the advantages of this invention are as follows: Addressing the problems of fixed identification types and poor robustness in existing methods, this invention designs a more intelligent multi-component interference signal identification method. By combining and modeling multiple interference signals, and extracting global signal features from the perspective of the signal's time-frequency image, the ML-SGMViTv2 network is used for identification. Simulation experiments demonstrate the effectiveness of the proposed ML-SGMViTv2 network in identifying multi-component interference signals. Compared to the CNN method of extracting local signal features, the method using global signal features achieves higher identification accuracy. Compared to multi-class classification methods, multi-label classification methods can still maintain a certain recognition rate for signals with unknown combinations, and have stronger scalability. Attached Figure Description

[0031] Figure 1 This is the overall flowchart of the present invention;

[0032] Figure 2 This is a schematic diagram of the structure of the ML-SGMViTv2 network in this invention;

[0033] Figure 3 This is the performance curve of the ML-SGMViTv2 network as a function of the number of iterations in this invention;

[0034] Figure 4 It is a measure of the recognition performance of the ML-SGMViTv2 network for unknown combination signals in the training set in this invention. Detailed Implementation

[0035] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0036] Step 1: Establish a mathematical model for multi-component radar jamming signals.

[0037] The simulation generates multi-component radar jamming signals, which include radar echo signals, suppression jamming, smart noise jamming, and deception jamming. The input signal model is as follows: k∈(1, q) represents the number of individual interference signal components, J i (n) represents the received interference signal, A i Let awgn(n) be the amplitude coefficient modulated when propagating through the environment, and let x be the Gaussian white noise. t (n) represents the actual target echo signal.

[0038] Step 2: Preprocess the multi-component interference signal.

[0039] A time-frequency transformation is performed on the multi-component interference signal to obtain the joint time-frequency distribution image of the signal. A short-time Fourier transform (STFT) is used to perform the time-frequency transformation on the interference signal, converting the one-dimensional interference signal into a two-dimensional time-frequency image X(t, ω).

[0040]

[0041] Where ω represents the frequency and h(τ-t) is the analysis window function.

[0042] (2) For each signal, several time-frequency images are generated at each drying ratio. Considering the variation of given q and k, the number of all possible signal combinations N can be defined as:

[0043]

[0044] (3) Preprocess the time-frequency images, normalize the scale of the images and standardize the data to improve the generalization ability of the model.

[0045] Step 3: Construct and train the ML-SGMViTv2 multi-label classification network

[0046] The ML-SGMViTv2 multi-label classification network includes a MobileViT-v2 module, a SandGlass module, a global average pooling module, a fully connected layer, and a multi-threshold optimization module. The loss function is set as a binary cross-entropy function.

[0047] Step 31: Construct the MobileViT-v2 module. The MobileViT-v2 module is the core module of the ML-SGMViTv2 network. It uses a separable self-attention method, replacing the quadratic multi-headed self-attention (MHA) mechanism with two separate linear computations to encode global information, thereby addressing the limitation of MHA in efficient inference on resource-constrained devices.

[0048] The separable self-attention structure is inspired by the multi-head self-attention mechanism. Similar to MHA, it processes the input x using three branches: input I, key K, and value V. The input branch I uses W... I ∈R d The linear layer maps each d-dimensional token in x to a scalar. This linear projection is an inner product operation that computes the distance between the latent token L and x, resulting in a k-dimensional vector. A softmax operation is then performed on this k-dimensional vector to produce a context score c. s =R k

[0049] Context score c s Used to calculate the context vector c v Specifically, the weight value is W. K ∈R d×d The key branch K linearly projects the input x into a d-dimensional space, producing the output x. k ∈R k×d Context vector c v ∈R d Then it is calculated as x k Weighted sum:

[0050]

[0051] c v The encoded context information is shared with all tokens in x. Therefore, the input x uses a weight of W. V ∈R d×d The value branch V is linearly projected onto d-dimensional space, and then the ReLU activation function produces the output x. v ∈R k×d Then, c v Contextual information in the broadcast element is propagated to x through the multiplication operation of the broadcast element. V The generated output is then input into another input with weight W.O ∈R d×d A linear layer is used to produce the final output y.

[0052]

[0053] Here, * and ∑ are broadcastable element-wise multiplication and summation operations, respectively.

[0054] Step 32: Construct the SandGlass module. The SandGlass module is the bottleneck module, which performs identity mapping and spatial transformation in a higher dimension, effectively mitigating information loss and gradient confusion.

[0055] The Sandglass module performs dimensionality increase followed by dimensionality reduction in the original inverse residual module, while also adjusting the order of the two 1×1 convolutions. To address the difficulty of capturing spatial information with 1×1 convolutions, a depthwise convolution is introduced to encode spatial information. This depthwise convolution is placed outside the two 1×1 convolutions, and short connections are established between high-dimensional features, which helps to transfer more information from the input to the output, thus enabling more gradient backpropagation.

[0056] Assume S i Let S represent the input vector. o The output vector is represented by the following formula:

[0057]

[0058]

[0059] in, These represent 1×1 convolution and depthwise convolution, respectively, ensuring that depthwise convolution processes data in a high-dimensional space and yields richer feature representations.

[0060] Step 33: Input the time-frequency images of the training set into the proposed ML-SGMViTv2 network. Use the MobileViT-v2 module to extract global features of the signal, and use the SandGlass module to remove redundant information and reduce model parameters to obtain the global feature matrix F. Then, input the global feature matrix F into the global average pooling module layer to obtain AvgPool(F). Input AvgPool(F) into the fully connected layer to obtain the global feature vector G of the signal. q

[0061] G q =FC(AvgPool(F))

[0062]

[0063] Where FC represents the fully connected layer operation, H and W are the row and column values ​​of the feature matrix F, and F(i,j) represents the element in the i-th row and j-th column of the feature matrix F.

[0064] Steps 3 and 4: Use the Sigmoid function as the activation function, which transforms a real-valued output z into a probability output in the range (0, 1). Then, apply the global feature G... q The output p of the Sigmoid layer for the j-th output label is input into the Sigmoid function classification layer. j Calculated as

[0065]

[0066] Among them, g j q is the j-th element of the global feature vector, and q is the number of output nodes of the fully connected layer.

[0067] Step 35: Using binary cross-entropy as the loss function, calculate the loss independently for each of the q output labels:

[0068] V(p j , t j )=-t j log(p j )-(1-t j log(1-p) j ), j = 1, 2, ..., q

[0069] Where p j t represents the output of the ML-SGMViTv2 model for the j-th label. j This represents the corresponding true label. The gradient descent method is used to optimize the ML-SGMViTv2 network.

[0070] Step 36: Optimize the decision threshold based on the output results:

[0071] Based on the input label vector T of the training sample X and the output probability vector P obtained after feature extraction, threshold calibration is considered as an optimization problem, and optimization is performed on the multiple threshold vector M. The threshold induced by l training samples for the j-th label is as follows:

[0072]

[0073] In the formula, y ij , t ij ∈(0,1) represent training samples s respectively i The decision label and the true label of the j-th label.

[0074] Step 4: Input the test set into the trained multi-label classification network to obtain the interference signal components contained in the input signal.

[0075] A multi-component interference signal dataset was constructed by combining seven types of signals: amplitude-modulated noise interference, frequency-modulated noise interference, white noise convolution interference, white noise product interference, spectral dispersion interference, intermittent sampling and forwarding interference, and echo signals, resulting in 127 possible combinations. For each combination, 20 time-frequency images were generated at 10 different interference-to-noise ratios (IRRs): -8dB, -6dB, -4dB, -2dB, 0dB, 2dB, 4dB, 6dB, 8dB, and 10dB, resulting in a total of 25,400 video images forming the dataset. The dataset was then divided into training and testing sets in a 7:3 ratio. Each sample was assigned a corresponding multi-label vector.

[0076] Figure 3 The figure shows the performance curve of the training network as the number of iterations increases. As can be seen from the simulation graph, the method designed in this invention has a high recognition accuracy, with an overall recognition rate of up to 99.7% on the test set, which is better than similar networks based on multi-class classification and multi-label convolutional neural networks (MLCNN) based on local features.

[0077] Figure 4 The figure shows the performance measurement of the proposed method for recognizing unknown combination signals in the training set. The experiment divided the original dataset into known signal combinations and unknown signal combinations according to the maximum number of single interference components constituting the multi-component interference signal. 80% of the known combination samples were used for training, and 20% of the known combination samples and all unknown combination samples were used for testing. The test results show that the recognition performance of the proposed method for unknown combinations increases rapidly with the number of training combinations, reaching maximum accuracy at the maximum number of known combinations (126). At this point, the recognition accuracy for unknown multi-component interference signals composed of 7 single interference components is 98.5%. When training is used with multi-component interference signals composed of a maximum of 3 single interference components, the recognition accuracy for 64 unknown combination signals can reach 93.98%, which can greatly reduce the data cost required for training.

Claims

1. A method for identifying multi-component interference signals based on a multi-label classification network, characterized in that, The steps are as follows: Step 1: Obtain the multi-component interference signal formed by the linear superposition of radar echo signal and multiple interference signals. k∈(1,Q) represents the number of components constituting a single interference signal, which can be composed of at most Q types of interference components, J m (n) represents the received interference signal, A m Let awgn(n) be the amplitude coefficient modulated when propagating through the environment, and let x be the additive white Gaussian noise. t (n) represents the actual target echo signal; Step 2: Preprocess the multi-component interference signal; perform time-frequency transformation on the multi-component interference signal to obtain its time-frequency joint distribution image X(t,ω): Where ω represents the frequency, and h(τ-t) is the analysis window function; Step 3: Input the training set into the multi-label classification network for training: The multi-label classification network includes the MobileViT-v2 module, SandGlass module, global average pooling module, fully connected layer and multi-threshold optimization module, and the loss function is set as the binary cross-entropy function; Step 4: Optimize the decision threshold based on the output results: Based on the input label vector T of the training sample X and the output probability vector P obtained after feature extraction, the threshold calibration is regarded as an optimization problem, and the multi-threshold vector M is optimized; the threshold induced by l training samples for the j-th label is as follows: In the formula, y ij ,t ij ∈(0,1) represent training samples s respectively i The decision label and the true label of the j-th label; q is the total number of labels; Step 5: Input the test set into the trained multi-label classification network to obtain the interference signal components contained in the input signal.

2. The method for identifying multi-component interference signals based on a multi-label classification network according to claim 1, characterized in that, Step three includes: Step 31: Input the time-frequency images of the training set into the proposed ML-SGMViTv2 network, use the MobileViT-v2 module to extract the global features of the signal, and use the SandGlass module to remove redundant information, reduce model parameters, and obtain the global feature matrix F; Step 32: Input the global feature matrix F into the global average pooling module layer to obtain AvgPool(F); input AvgPool(F) into the fully connected layer to obtain the global feature vector G of the signal. q The number of output nodes in the fully connected layer is q: G q =FC(AvgPool(F)) Where FC represents the fully connected layer operation, H and W are the row and column values ​​of the feature matrix F, and F(h,w) represents the element in the h-th row and w-th column of the feature matrix F; Step 33: Use the Sigmoid function as the activation function, which transforms a real-valued output z into a probability output in the range (0,1); then convert the global feature vector G... q The output p of the Sigmoid layer for the j-th label is input into the Sigmoid function classification layer. j The calculation is as follows: Where, q j For the global feature vector G q The j-th element; Steps 3 and 4: Using binary cross-entropy as the loss function, calculate the loss independently for each of the q labels: V(p j ,t j )=-t j log(p j )-(1-t j )log(1-p j ),j=1,2,...,q Where, p j t represents the output of the ML-SGMViTv2 model for the j-th label. j This indicates the corresponding real label; Step 35: Optimize the ML-SGMViTv2 network using gradient descent and cosine annealing algorithms.

Images