A deep learning-based wireless communication interference signal open set identification method

Abstract

The application discloses a wireless communication open set identification method for interference signals based on deep learning, directly utilizes original I / Q data to realize open set interference identification of communication signals, fully utilizes complex value signals, so that the semantic space of interference signals can be accurately described, and accurate open set identification is realized through a self-adaptive threshold classifier.In addition, aiming at the case that the data distribution of a training set and a test set is different, an adversarial domain adaptation is introduced, known samples in a target domain and known samples in a source domain are aligned, domain invariance of the model is ensured, and accurate open set identification of interference modes can be realized in a real complex electromagnetic environment.

Description

Technical Field

[0001] This invention relates to the field of open set identification technology for communication interference, and specifically to an open set identification method for wireless communication interference signals based on deep learning. Background Technology

[0002] Due to the openness of wireless channels, wireless communication systems are easily susceptible to malicious interference. In today's complex electromagnetic environment, intelligent jamming devices equipped with advanced artificial intelligence and general-purpose software-defined radio technology further exacerbate this threat. To ensure reliable transmission of communication signals and frequency security for legitimate users, it is necessary to quickly and accurately identify interference signal patterns in the wireless communication environment, thereby effectively supporting subsequent anti-interference and other communication protection measures.

[0003] Traditional interference signal pattern recognition methods primarily rely on signal processing techniques and prior knowledge. They analyze the time and frequency domain characteristics of interference signals to design suitable classifiers, ensuring efficient and accurate interference detection. However, traditional interference recognition methods heavily depend on manual data preprocessing and feature extraction, requiring specialized prior knowledge and making them unsuitable for today's complex electromagnetic environments. With the development of deep learning technology, its ability to automatically extract signal features and reduce human intervention has led to its widespread application in interference recognition. Deep learning-based open-set interference recognition methods typically employ models such as Convolutional Neural Networks (CNNs), ResNets, and Autoencoders to achieve accurate interference pattern recognition through a data-driven approach. However, most existing deep learning-based interference pattern recognition work follows a closed-set scenario, assuming that the interference patterns in the test set are the same as those in the training set, and that all interference patterns are known in advance. However, in real-world systems, it is often impossible to collect training samples for all potential interference patterns, and new, unknown interference patterns may emerge during actual testing. Therefore, closed-set methods often incorrectly classify unknown interference patterns as one of the known patterns in the training dataset. Furthermore, in practical systems, due to dynamic factors such as the communication environment and user location, even for known signal patterns, the data distribution may differ between the source and target domains, further complicating interference identification. Currently, most algorithms convert I / Q data into time-frequency diagrams, constellation diagrams, etc., then use deep learning methods to extract interference signal features, employing a pre-designed classifier to reject unknown samples and identify known interference. However, it is worth noting that while converting the original I / Q signal into a time-frequency representation or a spectral waterfall plot aids feature extraction, this conversion may incur additional processing overhead and potentially lose useful information. Effectively utilizing the rich information in the original I / Q data to accurately identify communication interference patterns in open-set scenarios remains a significant challenge.

[0004] The invention disclosed in CN114239672A presents an open-set interference pattern recognition model and method based on zero-shot learning. In the supervised training phase, the encoder is trained to learn the latent feature representations of known interference patterns. In the unsupervised classification phase, known and unknown interference patterns are classified in the feature space according to a distance-based classification criterion. This effectively addresses unknown interference patterns and achieves high open-set recognition accuracy. While this invention effectively characterizes the open-set interference pattern recognition scenario in wireless communication anti-interference environments, it does not solve the problems of lost useful information and additional processing overhead. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-based open-set identification method for wireless communication interference signals, which can effectively utilize the rich information in the original I / Q data and accurately reject unknown interference modes while maintaining the performance of identifying known interference modes.

[0006] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:

[0007] In a first aspect, the present invention discloses a method for identifying open-set interference signals in wireless communication based on deep learning, the method comprising the following steps:

[0008] Step A: The interaction features between the I / Q channels are extracted using a complex-valued fully connected autoencoder, and then fused with the input I / Q data in parallel to form an enhanced data sample.

[0009] Step B: Construct an interference feature extraction module. The interference feature extraction module includes an encoder, a decoder, and a closed-set classifier. The encoder is used to extract the feature information of the input interference signal, the decoder is used to reconstruct the input signal from the extracted feature information, and the closed-set classifier uses the SoftMax function to classify known interference patterns.

[0010] Step C: Use the enhanced data samples as training data for the interference feature extraction module, constrain the feature space of the interference signal to minimize the intra-class distance in the feature space; use the mean of the feature vectors of each known sample output by the encoder as the center vector of each known pattern.

[0011] Step D: Use the trained encoder to extract the feature vector of the enhanced I / Q sample to be detected, use the open set classification module to calculate the minimum distance from the feature vector to each center vector, and compare the minimum distance with the set adaptive threshold. If it is greater than the adaptive threshold, it is determined to be an unknown class; if it is less than the threshold, the enhanced I / Q sample to be detected is identified as the class to which the nearest center vector belongs.

[0012] As a preferred example, the encoder includes an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, and a pooling layer connected in sequence; the encoder takes augmented data samples as input and maps them to the feature space through convolution and pooling operations; during the training process of the encoder, a center loss function is introduced to minimize the intra-class space to make the extracted feature space compact, and a boundary constraint function is introduced to further compress the feature space;

[0013] The decoder includes a first deconvolution layer, a second deconvolution layer, a third deconvolution layer, and a fourth deconvolution layer connected in sequence. The decoder reconstructs the extracted features into input. During the training process of the decoder, mean squared error is introduced as a reconstruction loss so that the feature vector extracted by the autoencoder contains features that represent the original data.

[0014] The closed-set classifier includes an input layer, a flattening layer, a first fully connected layer, a second fully connected layer, and a third fully connected layer. The closed-set classifier transforms the feature vector into a predicted probability distribution through multiple fully connected layers. A cross-entropy loss function is introduced during the training process of the closed-set classifier to improve the classification accuracy of known interference patterns.

[0015] As a preferred example, the interference feature extraction module clusters interference features by minimizing the joint loss function LOSS during training:

[0016] LOSS = λ cen L cen +λ rec L rec +λ cross L cross +λ b L b

[0017] Where λ cen , λ rec ,λc cross , λb b Represents the weighting coefficient;

[0018] The central loss is:

[0019]

[0020] in, This represents the augmented training sample, and d represents the Mahalanobis distance. This represents the feature vector extracted from the training samples. Represents y n The center vector of the class, where N2 represents the number of training samples;

[0021] The boundary constraints are:

[0022]

[0023] in It is a learnable parameter;

[0024] The reconstruction loss is:

[0025]

[0026] in, Represents the reconstructed output of the decoder;

[0027] The cross-entropy loss is:

[0028]

[0029] in, g represents the probability vector output by the closed-set classifier. n This represents the true probability distribution of the labels.

[0030] As a preferred example, the encoder and decoder incorporate a batch normalization layer; the closed-set classifier incorporates a random dropout layer.

[0031] As a preferred example, step D further includes:

[0032] The minimum Mahalanobis distance from the eigenvectors to each center vector is calculated using the following formula, expressed as:

[0033]

[0034] Where F(X) represents the feature vector extracted by the encoder, c k The center vector of class k, The type of interference is known to the receiver;

[0035] Considering different interference modes and interference-to-signal ratios, set an adaptive threshold;

[0036] If the minimum Mahalanobis distance is greater than the adaptive threshold, it is classified as an unknown class; if it is less than the adaptive threshold, the sample to be detected is identified as belonging to the class of its nearest center vector, as follows:

[0037]

[0038] Where y0 represents the category of the nearest center vector to the feature vector of the sample to be detected, expressed as:

[0039]

[0040] As a preferred example, in step D, the adaptive threshold is:

[0041]

[0042] in, It is the training dataset corresponding to the k-th sample pattern under the specified JSR, where Pr[·] represents the probability of the event occurring, and d k Represents the minimum Mahalanobis distance that satisfies the conditions, where η is the statistical confidence level. The type of interference is known to the receiver.

[0043] Secondly, this invention discloses a method for identifying open-set interference signals in wireless communication based on deep learning, the method comprising:

[0044] A complex-valued fully connected autoencoder is used to augment the source and target domain samples respectively. Specifically, the interaction features between the I / Q channels of the source and target domains are extracted respectively, and then fused with the input I / Q data of the source and target domains in parallel to form augmented source and target domain samples.

[0045] An interference feature extraction module is constructed, which includes an encoder, a decoder, and a closed-set classifier. The encoder is used to extract the feature information of the input interference signal, the decoder is used to reconstruct the input signal from the extracted feature information, and the closed-set classifier uses the SoftMax function to classify known interference patterns.

[0046] The interference feature extraction module is trained using source domain samples to obtain the optimal parameters of the encoder, thus completing the encoder training. The trained encoder is then used to extract feature vectors from target domain samples, and an open-set classification module is used to perform the extraction.

[0047] Select a predetermined proportion of target domain samples that have the smallest distance to each center vector, train an adversarial adaptation framework using source domain data and the selected target domain samples, optimize the encoder parameters, and align the feature distributions of known classes in the source and target domains.

[0048] The optimized encoder is used to extract the feature vector of the sample to be detected, and the minimum distance from the feature vector to each center vector is calculated. The minimum distance is compared with the set adaptive threshold. If it is greater than the adaptive threshold, it is determined to be an unknown class; if it is less than the adaptive threshold, the sample to be detected is identified as the class to which the nearest center vector belongs.

[0049] As a preferred example, the interference feature extraction module is trained using source domain samples to obtain the optimal parameters θ of the encoder. f :

[0050]

[0051] in, and These represent the reconstruction loss function, cross-entropy loss function, center loss function, and boundary constraint, respectively, when training using only source samples.

[0052] As a preferred example, the adversarial domain adaptation framework is composed of a feature extraction module and a domain discriminator. During the training process, the feature extraction module is used to extract similar feature vectors from signals with different data distributions to confuse the domain discriminator, which is used to distinguish samples from different data distributions.

[0053] The adversarial loss function of the adversarial domain adaptation framework is:

[0054]

[0055] Among them, L bce It is binary cross-entropy loss; α q Represents a given sample Is it from the source domain or the target domain, when α q When α = 0, the sample comes from the source domain; when α = 0, the sample comes from the source domain. q When = 1, the sample comes from the target domain. G represents the feature vector extracted from the training samples. d (·) indicates the output of the domain discriminator.

[0056] As a preferred example, the encoder parameters are updated according to the following formula. Parameters of the domain discriminator

[0057]

[0058] In the formula, L d This represents the adversarial loss function of the adversarial domain adaptation framework. L represents the cross-entropy loss function when training using only the source samples. rec This represents the reconstruction loss when training using source domain data and selected target domain samples.

[0059] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0060] First, the deep learning-based open-set identification method for wireless communication interference signals of the present invention directly utilizes the original I / Q data to realize open-set interference identification of communication signals, makes full use of complex-valued signals, thereby accurately describing the semantic space of interference signals, and achieves accurate open-set identification through an adaptive threshold classifier.

[0061] Second, the open set identification method for wireless communication interference signals based on deep learning in this invention introduces adversarial domain adaptation to address the situation where the data distributions of the training set and the test set are different. This aligns known samples in the target domain with known samples in the source domain, ensuring that the model has domain invariance and enabling accurate open set identification of interference patterns in real-world complex electromagnetic environments. Attached Figure Description

[0062] Figure 1 A model diagram of an open-set interference pattern recognition system for wireless communication;

[0063] Figure 2 A schematic diagram of a feature-enhanced open-set identification method for communication interference.

[0064] Figure 3 A schematic diagram of a complex-valued fully connected self-encoder;

[0065] Figure 4 A schematic diagram of an open-set interference identification method for communication based on adversarial domain adaptation;

[0066] Figure 5 This is a schematic diagram of the open-set recognition performance recognition curve as the interference-to-signal ratio varies; where Figure 5 In the figure, (a), (b), and (c) represent the performance of the proposed method on the JPR2024 dataset under different JSR conditions, namely NA, AKS, and AUS.

[0067] Figure 6 This is a schematic diagram illustrating the distribution of the extracted feature vectors; where Figure 6 Subgraphs a, c, and e in the diagram represent t-SNE embeddings obtained by directly reducing the dimensionality of the original data in JPR2024, JPR2016, and JPR2018, respectively. Figure 6 Subgraphs b, d, and f in the diagram represent the t-SNE embeddings of the feature vectors extracted from JPR2024, JPR2016, and JPR2018 using the trained model, respectively. Detailed Implementation

[0068] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0069] Example 1

[0070] like Figure 1As shown, this embodiment establishes a wireless communication open-set interference pattern recognition system to identify interference signals in the communication system. The system includes a jammer, a pair of legitimate transmitters and receivers, and an interference detector configured at the legitimate receiver. The jammer can transmit energy signals resembling noise to disrupt communication between the legitimate transmitters and receivers. The interference detector can determine whether the received signal is affected by interference; if so, it identifies a specific interference pattern. In practice, legitimate receivers cannot be aware of all interference patterns; therefore, it is assumed that the receiver is aware of the following interference categories: The remaining modes are unknown, denoted as K+1. For ease of description, 0 represents no interference, corresponding to a normal signal received by the receiver. The signal received by the receiver can be represented as:

[0071]

[0072] Where S0(t) represents the normal signal from a legitimate transmitter, S j w(t) represents the communication interference signal from the jammer, and w(t) represents additive white Gaussian noise.

[0073] This embodiment discloses a deep learning-based open-set identification method for wireless communication interference signals, the method comprising the following steps:

[0074] Step A: The interaction features between the I / Q channels are extracted using a complex-valued fully connected autoencoder, and then fused with the input I / Q data in parallel to form an enhanced data sample.

[0075] Step B: Construct an interference feature extraction module. The interference feature extraction module includes an encoder, a decoder, and a closed-set classifier. The encoder is used to extract the feature information of the input interference signal, the decoder is used to reconstruct the input signal from the extracted feature information, and the closed-set classifier uses the SoftMax function to classify known interference patterns.

[0076] Step C: Use the enhanced data samples as training data for the interference feature extraction module, constrain the feature space of the interference signal to minimize the intra-class distance in the feature space; use the mean of the feature vectors of each known sample output by the encoder as the center vector of each known pattern.

[0077] Step D: Use the trained encoder to extract the feature vector of the enhanced I / Q sample to be detected, use the open set classification module to calculate the minimum distance from the feature vector to each center vector, and compare the minimum distance with the set adaptive threshold. If it is greater than the adaptive threshold, it is determined to be an unknown class; if it is less than the threshold, the enhanced I / Q sample to be detected is identified as the class to which the nearest center vector belongs.

[0078] This embodiment is a feature-enhanced open-set recognition (FOSR) method for wireless communication interference signals, which specifically includes the following steps:

[0079] I / Q data feature enhancement steps: Use a complex-valued fully connected autoencoder to extract the interaction features between the I / Q channels, and fuse them with the I / Q data to form enhanced data samples.

[0080] Interference Feature Extraction Steps: Using a classification-reconstruction framework, the enhanced data samples are trained to extract interference features. Further, the interference feature extraction module includes an encoder, a decoder, and a closed-set classifier. The encoder extracts feature information from the input interference signal through operations such as convolution and pooling. The decoder reconstructs the input signal from the extracted feature information through operations such as deconvolution and unpooling. The closed-set classifier uses SoftMax to classify known interference patterns. Further, the enhanced I / Q samples are used as training data for the model, and center loss, boundary constraint loss, reconstruction loss, and cross-entropy loss are introduced to constrain the feature space of the interference signal to minimize the intra-class distance in the feature space. Finally, the mean of the feature vectors of each known sample output by the encoder is used as the center vector of each known pattern.

[0081] The open-set recognition step involves extracting feature vectors from the enhanced I / Q samples to be detected using a trained encoder, calculating the minimum distance from each feature vector to its nearest center vector, and comparing this minimum distance with a set adaptive threshold. If the distance is greater than the threshold, the sample is classified as an unknown class; otherwise, it is identified as belonging to the class of the nearest center vector. Furthermore, the open-set recognition module can adaptively set reasonable thresholds for different interference patterns and interference-to-signal ratios to reject unknown classes and accurately classify known classes.

[0082] A feature-enhanced open-set framework for communication interference identification, such as... Figure 2 As shown, firstly, the I / Q data feature enhancement module extracts the interaction features between the I / Q channels. Then, the interference feature extraction module extracts latent features from the enhanced data based on known patterns. Finally, a distance-based adaptive threshold open-set classifier is used to achieve interference pattern recognition. Specifically, the I / Q data feature enhancement module consists of a complex-valued fully connected autoencoder, the structure of which is shown below. Figure 3 As shown, the encoder and decoder have symmetrical structures, each consisting of two complex-valued fully connected layers. The model is trained using the mean squared error function as the loss function. The encoder output is concatenated in parallel with the input data to obtain the augmented data.

[0083] The interference feature extraction module was trained using enhanced I / Q data, and its structure is shown in Table 1. The interference feature extraction module consists of an encoder, a decoder, and a closed-set classifier.

[0084] Table 1 Structure of Interference Feature Extraction Module

[0085]

[0086] The encoder takes the augmented data as input and maps it to the feature space through convolutional and pooling layers. Introducing a center loss function during training minimizes the intra-class space, making the extracted feature space more compact.

[0087] The central loss is:

[0088]

[0089] in, Represents the augmented training samples, This represents the feature vector extracted from the training samples. Represents y n The center vector of the class (the average of the feature vectors of all known class samples), and d represent the Mahalanobis distance, as shown in the following formula:

[0090]

[0091] In addition, boundary constraint functions are introduced to further compress the feature space. The boundary constraints are:

[0092]

[0093] in It is the minimum distance learned by the encoder.

[0094] The decoder reconstructs the extracted features into input. During training, mean squared error is introduced as a reconstruction loss to ensure that the feature vector extracted by the autoencoder contains features that can characterize the original data.

[0095] The reconstruction loss is:

[0096]

[0097] in, This represents the reconstructed output of the decoder.

[0098] The closed-set classifier transforms feature vectors into predicted probability distributions through fully connected layers, and introduces a cross-entropy loss function during training to ensure accurate classification of known interference patterns.

[0099] The cross-entropy loss is:

[0100]

[0101] in, g represents the probability vector output by the closed-set classifier. n This represents the true label probability distribution. During training, clustering of interfering features is achieved by minimizing the joint loss function, ensuring the performance of subsequent open-set recognition. The joint loss function is shown below:

[0102] LOSS = λ cen L cen +λ rec L rec +λ cross L cross +λ b L b ,

[0103] Where λ cen , λ rec , λ cross and λ b This represents the weighting coefficient.

[0104] To stabilize the training process, batch normalization is introduced in both the encoder and decoder. To prevent overfitting during training, a dropout layer is introduced in the closed-set classifier.

[0105] Extract the enhanced I / Q sample X to be detected using the trained encoder. te,en The eigenvectors are calculated, and the minimum Mahalanobis distance from the eigenvectors to each center vector is expressed as:

[0106]

[0107] Where F(X) te,en The value represents the feature vector extracted by the encoder. This minimum distance is compared with a set adaptive threshold. Furthermore, considering different interference modes and interference-to-signal ratios, the adaptive threshold of the open-set recognition module is set as follows:

[0108]

[0109] in, This is the training dataset corresponding to the k-th sample pattern under a specified JSR, where η is the statistical confidence score. If the minimum distance is greater than this threshold, it is classified as an unknown class; if it is less than this threshold, the sample to be detected is identified as belonging to the class of its nearest center vector, which can be expressed as:

[0110]

[0111] Where y0 represents the category of the nearest center vector to the feature vector of the sample to be detected, which can be expressed as:

[0112]

[0113] Example 2

[0114] This embodiment discloses a deep learning-based open-set identification method for wireless communication interference signals, the method comprising:

[0115] A complex-valued fully connected autoencoder is used to augment the source and target domain samples respectively. Specifically, the interaction features between the I / Q channels of the source and target domains are extracted respectively, and then fused with the input I / Q data of the source and target domains in parallel to form augmented source and target domain samples.

[0116] An interference feature extraction module is constructed, which includes an encoder, a decoder, and a closed-set classifier. The encoder is used to extract the feature information of the input interference signal, the decoder is used to reconstruct the input signal from the extracted feature information, and the closed-set classifier uses the SoftMax function to classify known interference patterns.

[0117] The interference feature extraction module is trained using source domain samples to obtain the optimal parameters of the encoder, thus completing the encoder training. The trained encoder is then used to extract feature vectors from target domain samples, and an open-set classification module is used to perform the extraction.

[0118] Select a predetermined proportion of target domain samples that have the smallest distance to each center vector, train an adversarial adaptation framework using source domain data and the selected target domain samples, optimize the encoder parameters, and align the feature distributions of known classes in the source and target domains.

[0119] The optimized encoder is used to extract the feature vector of the sample to be detected, and the minimum distance from the feature vector to each center vector is calculated. The minimum distance is compared with the set adaptive threshold. If it is greater than the adaptive threshold, it is determined to be an unknown class; if it is less than the adaptive threshold, the sample to be detected is identified as the class to which the nearest center vector belongs.

[0120] To address the issue of different data distributions between the training and test sets, this embodiment proposes an open set recognition method for communication interference signals based on adversarial domain adaptation (FOSR-DA), comprising the following steps:

[0121] I / Q data feature enhancement steps: Use a complex-valued fully connected autoencoder to extract the interaction features between the I / Q channels, and fuse them with the I / Q data to form enhanced data samples.

[0122] Known / Unknown Class Separation Step: To prevent negative transfer that may be caused by unknown samples during domain adaptation, it is necessary to first separate the unknown samples. Further, the known / unknown separation step includes two steps: interference feature extraction and known class selection. These two steps are the same as those in the feature-enhanced open-set identification method for communication interference described earlier. Specifically, the feature extractor is trained using enhanced source domain I / Q samples. The trained encoder is used to extract feature vectors from the target domain samples, and an open-set classifier is used to identify the target domain samples, classifying them into known and unknown classes. The top γ% (e.g., 20%) of target domain samples closest to each center vector are selected for subsequent domain adaptation tasks.

[0123] The adversarial domain adaptation process involves training an adversarial domain adaptation framework using source domain data and selected target domain samples. This optimizes the encoder parameters and aligns the feature distributions of known classes in both the source and target domains. Furthermore, the adversarial domain adaptation framework consists of a feature extraction module connected to a domain discriminator. During training, the feature extraction module extracts similar feature vectors from signals with different data distributions to confuse the domain discriminator, which in turn distinguishes samples from different data distributions.

[0124] Open set recognition steps: The encoder trained by the adversarial domain is used to extract the feature vector of the sample to be detected, and the minimum distance from the feature vector to each center vector is calculated. The minimum distance is compared with the set adaptive threshold. If it is greater than the threshold, it is determined to be an unknown class; if it is less than the threshold, the sample to be detected is identified as the class to which the nearest center vector belongs.

[0125] A schematic diagram of open-set identification of communication interference based on adversarial domain adaptation is attached. Figure 4 As shown, the entire process can be divided into two stages: the known-unknown separation stage and the adversarial domain adaptation stage. In these two stages, iterative optimization is continuously performed to gradually adapt to known samples with different data distributions, while simultaneously eliminating unknown samples.

[0126] Specifically, the known / unknown separation stage consists of three modules: a feature enhancement module, an interference feature extraction module, and an open set classification module, with the same structure as in Table 1. First, the feature enhancement module is used to augment the source and target domain samples. Then, the enhanced source domain samples are used to train the feature extraction module, and the optimal parameters θ of the encoder can be obtained in the following way. f :

[0127]

[0128] in, and These represent the corresponding loss functions when training using only the source samples. Finally, the trained encoder extracts the feature vectors of the enhanced target samples, and the open-set classification module performs open-set recognition on the target samples. After classification, the minimum distance from each known target sample to its corresponding center vector is calculated. The top γ% (e.g., 20%) of samples with the smallest distance to each center vector are selected for subsequent adversarial domain adaptation.

[0129] Furthermore, in the adversarial domain adaptation stage, the initially trained feature extraction module is connected to a domain discriminator to perform adversarial training on the encoder. This aligns the known samples in the source and target domains, ensuring stable feature extraction capabilities across different data distributions. Adversarial domain adaptation training is then performed using source domain samples and filtered target domain samples. For a given sample... Using α q express It comes from the source domain (α) q =0) or target domain (α) q =1). During training, the feature extraction module extracts similar feature vectors from signals with different data distributions to confuse the domain discriminator, which in turn tries to distinguish samples from different data distributions. The adversarial loss function is defined as follows:

[0130]

[0131] Among them, L bce It is binary cross-entropy loss.

[0132] Update the encoder parameters according to the following formula. Parameters of the domain discriminator

[0133]

[0134] The encoder updated according to the above formula can retain knowledge of the source domain and has domain invariance, and can stably and effectively extract interference features in I / Q data.

[0135] The encoder, trained using the adversarial domain adaptation method, extracts the feature vectors of the samples to be detected and calculates the minimum distance from each feature vector to each center vector. This minimum distance is compared with a set adaptive threshold. If the distance is greater than the threshold, the sample is classified as an unknown class; if it is less than the threshold, the sample is identified as belonging to the class of its nearest center vector. The environment and hyperparameter configuration required for network training are shown in Table 2.

[0136] Table 2. Environment and Hyperparameter Configuration Required for Network Training

[0137] GPU NVIDIA GeForce GTX 4090 GPU Development tools PyCharmCommunityEdition2022.2.2 Python Python 3.9 I / Q data feature enhancement module learning rate 0.001 Training rounds of the I / Q data feature enhancement module 300 I / Q Data Feature Enhancement Module Batch size 400 Learning rate of interference feature extraction module 0.0008 Interference feature module extracts batch size 256 Number of training rounds for the interference feature extraction module 250 Learning rate during the adversarial domain adaptation phase 0.0001 Batch size during the adversarial domain adaptation phase 200 Number of training rounds in the adversarial domain adaptation phase 100

[0138] Open set recognition needs to consider the accuracy in recognizing unknown classes, making the evaluation metrics for closed set recognition no longer applicable. Therefore, we introduce Known Class Accuracy (AKS), Unknown Class Accuracy (AUS), and Normalized Accuracy (NA):

[0139]

[0140] NA = λ r AKS+(1-λ r AUS,

[0141] Among them, TP, FP, TN, and FN represent true positive, false positive, true negative, and false negative, respectively, and λ r A coefficient of 0.5 is set to balance the accuracy of known class and unknown class recognition. In addition, known class accuracy, unknown class accuracy, and F1-score are introduced.

[0142]

[0143] Where S correct U represents the total number of known samples that are correctly classified. correct This represents the total number of unknown samples correctly classified as unknown patterns. The two types of metrics have different focuses: accuracy emphasizes the discovery of unknown categories, while F1-score treats known and unknown classes equally. Combining both metrics provides a better evaluation of the model's overall performance.

[0144] Experimental verification

[0145] In the experimental verification process of this invention, BPSK was selected as the normal signal, and six typical interference suppression methods were chosen: single-tone interference, multi-tone interference, frequency sweep interference, comb spectrum interference, impulse interference, and noise sweep frequency interference. BPSK signal, single-tone interference, multi-tone interference, and impulse interference were set as known interference modes, while the remaining interference modes were set as unknown modes. All communication and interference signals were generated using MATLAB simulation. The main parameters of the interference modes are as follows: the carrier frequency range for each interference mode was set to kHz. Multi-tone interference consists of seven syllables. Frequency sweep interference has a sweep rate of 5000 kHz / s. Comb spectrum interference consists of two signals with fixed frequencies of 200-500 kHz. The frequency modulation coefficient for noise sweep frequency interference was set to 5. For each known interference mode, 2500 samples were generated every 5 dB between JSR values ​​(-10 dB, 15 dB), with 80% used for training and 20% for testing. For each type of unknown interference, 500 samples were generated for testing, and the number of I / Q sampling points in the dataset was set to 128. In addition, the simulation-generated dataset was integrated with two public datasets, RML2016.10a and RML2018.01a, to create two datasets, JPR2016 and JPR2018, for experimental verification.

[0146] The performance of the feature-enhanced open-set communication interference identification method in this invention is compared with that of IIAE-JPR, SR2CNN, ZSL-JPR, and OpenMax. Figure 5 To compare performance across different JSRs on the JPR2024 dataset, both training and testing used I / Q samples with the same JSR. Under different JSRs, our method outperformed the other four methods in terms of NA, AKS, and AUS. Specifically, when JSR = 10 dB, our method improved NA by 7%, 13%, 14%, and 21% compared to IIAE-JPR, ZSL-JPR, SR2CNN, and OpenMax-JPR, respectively. Furthermore, AKS, AUS, and NA also increased with increasing JSR. This is because as JSR increases, the intensity of the interfering signal becomes stronger than the normal signal, allowing the network to better recognize and learn its features.

[0147] Table 3 shows the recognition accuracy of the feature-enhanced open-set communication interference identification method on three datasets under different signal-to-noise ratios (SNRs). I / Q samples with the same SNR were used in both training and testing. The recognition performance gradually decreases as the SNR decreases. This is because the intensity of background noise increases with decreasing SNR, making it more difficult to capture discriminative features in the interference signal.

[0148] Table 3 compares the performance of the feature-enhanced open-set interference identification method on three datasets under different signal-to-noise ratios.

[0149]

[0150] The trained encoder is used to extract feature vectors of interference signals from three datasets, and t-SNE is used to reduce their dimensionality, resulting in the feature vector distribution map shown below. Figure 6 As shown. From Figure 6 As can be seen from a, c, and e in the dataset, the perturbation features of the original samples overlap and cannot be separated in each dataset. After feature enhancement and feature extraction, features belonging to the same perturbation pattern will cluster together, while features of unknown patterns will become outliers.

[0151] To evaluate the performance of the feature-enhanced open-set interference identification method on different open-set tasks, the method was tested on JPR2024 with different combinations of unknown patterns, as shown in Table 4. Comparison of experiments ②, ③, and ④ shows that the proposed method exhibits stable identification performance across different combinations of unknown patterns. However, experiments ①, ③, and ⑤ reveal that the NA (Accuracy Requirement) decreases as the number of unknown patterns increases. This is because an increase in the number of unknown patterns reduces the number of features learned by the model, thus limiting its ability to represent the signal feature space and consequently leading to a decrease in identification performance.

[0152] Table 4 shows the performance of this method on JPR2024 for different combinations of unknown modes.

[0153]

[0154]

[0155] To verify the open-set identification method for communication interference based on adversarial domain adaptation, two cases were further considered in JPR2024: different channel conditions and different JSRs. Table 5 shows the maximum NA decrease rate for each method when the source and target domains are generated under Gaussian and Rayleigh fading channel conditions, respectively. It can be seen that the open-set identification method for communication interference based on adversarial domain adaptation exhibits the best robustness, with an average decrease rate of 19.1% on the test set, which is 1%, 5%, and 19% higher than FOSR, SR2CNN, and OpenMax-JPR, respectively.

[0156] Table 5. Maximum NA decrease rate when target domain samples are generated under Rayleigh fading channel.

[0157]

[0158] For the second scenario, the experiment was set up so that the source domain contained samples with a specific JSR, while the target domain contained samples with multiple JSRs. Table 6 shows the NA values ​​of each method on the JPR2024 dataset, when the JSR of the source domain samples was -5 dB and the JSR of the target domain samples was in the range of -10 dB to 15 dB. Notably, within the considered JSR range, FOSR-DA consistently outperformed all baseline methods. FOSR-DA achieved a maximum accuracy of 81%. This performance is 3%, 6%, 9%, 21%, and 41% higher than the baseline methods, respectively. These results highlight the superior performance and robustness of FOSR-DA in handling signals with domain offset issues.

[0159] Table 6 shows the NA values ​​of six methods when the JSRs of the source and target samples are different.

[0160]

[0161]

[0162] All the experimental results above demonstrate that the feature-enhanced open-set interference identification method not only has excellent identification accuracy, but also maintains stable NA when facing different open-set interference pattern identification tasks. Furthermore, when the channel conditions of the training and test samples differ from those of the JSR, the adversarial domain-adaptive open-set interference identification method exhibits superior open-set interference signal identification performance and strong robustness.

[0163] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0164] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, produce instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0165] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0166] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment, causing a series of operational steps to be executed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that run on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0167] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0168] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for open set identification of wireless communication jamming signals based on deep learning, characterized in that, The method includes the following steps: Step A: The interaction features between the I / Q channels are extracted using a complex-valued fully connected autoencoder, and then fused with the input I / Q data in parallel to form an enhanced data sample. Step B: Construct an interference feature extraction module. The interference feature extraction module includes an encoder, a decoder, and a closed-set classifier. The encoder is used to extract the feature information of the input interference signal, the decoder is used to reconstruct the input signal from the extracted feature information, and the closed-set classifier uses the SoftMax function to classify known interference patterns. Step C: Use the enhanced data samples as training data for the interference feature extraction module, constrain the feature space of the interference signal to minimize the intra-class distance in the feature space; use the mean of the feature vectors of each known sample output by the encoder as the center vector of each known pattern. Step D: Use the trained encoder to extract the feature vector of the enhanced I / Q sample to be detected, use the open set classification module to calculate the minimum distance from the feature vector to each center vector, and compare the minimum distance with the set adaptive threshold. If it is greater than the adaptive threshold, it is determined to be an unknown class; if it is less than the threshold, the enhanced I / Q sample to be detected is identified as the class to which the nearest center vector belongs. The encoder includes an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, and a pooling layer connected in sequence. The encoder takes augmented data samples as input and maps them to the feature space through convolution and pooling operations. During the training process of the encoder, a center loss function is introduced to minimize the intra-class space to make the extracted feature space compact, while a boundary constraint function is introduced to further compress the feature space. The decoder includes a first deconvolution layer, a second deconvolution layer, a third deconvolution layer, and a fourth deconvolution layer connected in sequence. The decoder reconstructs the extracted features into input. During the training process of the decoder, mean squared error is introduced as a reconstruction loss so that the feature vector extracted by the autoencoder contains features that represent the original data. The closed-set classifier includes an input layer, a flattening layer, a first fully connected layer, a second fully connected layer, and a third fully connected layer. The closed-set classifier transforms the feature vector into a predicted probability distribution through multiple fully connected layers. A cross-entropy loss function is introduced during the training process of the closed-set classifier to improve the classification accuracy of known interference patterns. The interference feature extraction module minimizes a joint loss function during training to cluster the interference features: ； wherein , , , represent weight coefficients; The central loss is: ； wherein, represents an enhanced training sample, denotes a Mahalanobis distance, represents a feature vector extracted from a training sample, represents a center vector of a class, denotes a number of training samples; The boundary constraints are: ； wherein is the minimum distance learned by the encoder; The reconstruction loss is: ； wherein, represents the reconstructed output of the decoder; The cross-entropy loss is: ； wherein, a probability vector representing the closed set classifier output, denotes the true label probability distribution. 2.The deep learning based wireless communication out-of-set jamming signal identification method of claim 1, wherein, The encoder and decoder incorporate batch normalization layers; the closed-set classifier incorporates random dropout layers. 3.The deep learning based wireless communication out-of-set jamming signal identification method of claim 1, wherein, Step D further includes: The minimum Mahalanobis distance from the eigenvectors to each center vector is calculated using the following formula, expressed as: ； wherein represents a feature vector extracted by the encoder, represents a center vector of the class, is an interference class known to the receiver; Considering different interference modes and interference-to-signal ratios, set an adaptive threshold; If the minimum Mahalanobis distance is greater than the adaptive threshold, it is classified as an unknown class; if it is less than the adaptive threshold, the sample to be detected is identified as belonging to the class of its nearest center vector, as follows: ； wherein, represents the class to which the nearest center vector to the sample feature vector to be detected belongs, and is represented as: 。 4.The deep learning based wireless communication out-of-set jamming signal identification method according to claim 1 or 3, characterized in that, In step D, the adaptive threshold is: ； wherein, is a training dataset corresponding to the th sample pattern under a designated JSR, is a statistical confidence, denotes a probability of an event occurrence, denotes a minimum Mahalanobis distance satisfying a condition, is an interference class known to the receiver.

5. A method for open set identification of wireless communication jamming signals based on deep learning, characterized in that, The method includes: A complex-valued fully connected autoencoder is used to augment the source and target domain samples respectively. Specifically, the interaction features between the I / Q channels of the source and target domains are extracted respectively, and then fused with the input I / Q data of the source and target domains in parallel to form augmented source and target domain samples. An interference feature extraction module is constructed, which includes an encoder, a decoder, and a closed-set classifier. The encoder is used to extract the feature information of the input interference signal, the decoder is used to reconstruct the input signal from the extracted feature information, and the closed-set classifier uses the SoftMax function to classify known interference patterns. The encoder includes an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, and a pooling layer connected in sequence. The encoder takes augmented data samples as input and maps them to the feature space through convolution and pooling operations. During the training process of the encoder, a center loss function is introduced to minimize the intra-class space to make the extracted feature space compact, while a boundary constraint function is introduced to further compress the feature space. The decoder includes a first deconvolution layer, a second deconvolution layer, a third deconvolution layer, and a fourth deconvolution layer connected in sequence. The decoder reconstructs the extracted features into input. During the training process of the decoder, mean squared error is introduced as a reconstruction loss so that the feature vector extracted by the autoencoder contains features that represent the original data. The closed-set classifier includes an input layer, a flattening layer, a first fully connected layer, a second fully connected layer, and a third fully connected layer. The closed-set classifier transforms the feature vector into a predicted probability distribution through multiple fully connected layers. A cross-entropy loss function is introduced during the training process of the closed-set classifier to improve the classification accuracy of known interference patterns. The interference feature extraction module minimizes a joint loss function during training to cluster the interference features: ； wherein , , , represent weight coefficients; The central loss is: ； wherein, represents an enhanced training sample, denotes a Mahalanobis distance, represents a feature vector extracted from a training sample, represents a center vector of a class, denotes a number of training samples; The boundary constraints are: ； wherein is the minimum distance learned by the encoder; The reconstruction loss is: ； wherein, representing a reconstructed output of the decoder; The cross-entropy loss is: ； wherein, a probability vector representing the closed set classifier output, denotes the true label probability distribution; The interference feature extraction module is trained using source domain samples to obtain the optimal parameters of the encoder, thus completing the training of the encoder. The trained encoder is used to extract the feature vectors of the target domain samples, and the open set classification module is used to perform open set recognition on the target domain samples, classifying the target domain samples into known classes and unknown classes. Select a predetermined proportion of target domain samples that have the smallest distance to each center vector, train an adversarial adaptation framework using source domain data and the selected target domain samples, optimize the encoder parameters, and align the feature distributions of known classes in the source and target domains. The optimized encoder is used to extract the feature vector of the sample to be detected, and the minimum distance from the feature vector to each center vector is calculated. The minimum distance is compared with the set adaptive threshold. If it is greater than the adaptive threshold, it is determined to be an unknown class; if it is less than the adaptive threshold, the sample to be detected is identified as the class to which the nearest center vector belongs. 6.The deep learning based wireless communication interference signal open set identification method of claim 5, wherein, The interference feature extraction module is trained by using source domain samples to obtain optimal parameters of the encoder : ； wherein, , , and represent the reconstruction loss function, the cross-entropy loss function, the center loss function and the boundary constraint respectively when training using only source samples. 7.The deep learning based wireless communication interference signal open set identification method of claim 5, wherein, The adversarial domain adaptation framework consists of a feature extraction module and a domain discriminator. During training, the feature extraction module is used to extract similar feature vectors from signals with different data distributions to confuse the domain discriminator, which is used to distinguish samples from different data distributions. The adversarial loss function of the adversarial domain adaptation framework is: ； wherein, is the binary cross-entropy loss; denotes a given sample is from the source domain or the target domain, when the sample is from the source domain, when the sample is from the target domain, represents a feature vector extracted from the training sample, denotes the output of the domain discriminator. 8.The deep learning based wireless communication interference signal open set identification method of claim 5, wherein, The parameters of the encoder are updated according to the following formula and the parameters of the domain discriminator : ； ； In the formula, represents the adversarial loss function of the adversarial domain adaptation framework, represents the cross-entropy loss function when training only using source samples, represents the reconstruction loss when training using source domain samples and screened target domain samples.

Images