Wireless communication interference signal open set identification method based on deep learning

Abstract

The invention discloses a wireless communication interference signal open set identification method based on deep learning, which directly utilizes original I / Q data to realize open set interference identification of communication signals and fully utilizes complex value signals so as to accurately describe semantic space of interference signals and realize accurate open set identification through an adaptive threshold classifier. Besides, according to the situation that data distribution of a training set and a test set is different, adversarial domain adaptation is introduced, a known sample in a target domain is aligned with a known sample in a source domain, it is ensured that the model has domain invariance, and accurate open-set recognition of the interference mode can be achieved in the real complex electromagnetic environment.

Description

Technical Field

[0001] The present invention relates to the technical field of open-set identification of communication interference, and in particular to an open-set identification method for wireless communication interference signals based on deep learning. Background Art

[0002] Due to the openness of wireless channels, wireless communication systems are vulnerable to malicious interference. In today's complex electromagnetic environment, intelligent jammers equipped with advanced artificial intelligence and general software-defined radio technology further exacerbate this threat. In order to ensure the reliable transmission of communication signals and the frequency security of legitimate users, it is necessary to quickly and accurately identify the interference signal pattern in the wireless communication environment, so as to effectively support subsequent anti-interference and other communication protection measures.

[0003] Traditional interference signal pattern recognition methods mainly rely on signal processing technology and prior knowledge. By analyzing the time domain, frequency domain and other characteristics of the interference signal, a suitable classifier is designed to ensure efficient and accurate detection of interference. However, traditional interference identification methods rely heavily on manual data preprocessing and feature extraction, which requires professional prior knowledge and is difficult to apply to today's complex electromagnetic environment. With the development of deep learning technology, deep learning has been widely used in the field of interference identification because it can automatically extract signal features and reduce manual intervention. The open set identification method of communication interference based on deep learning generally adopts models such as convolutional neural network (CNN), residual neural network (ResNet), and autoencoder to achieve accurate identification of interference patterns in a data-driven manner. However, most of the existing interference pattern recognition work based on deep learning follows the closed set scenario, that is, it is assumed that the interference pattern in the test set is the same as the interference pattern in the training set, and all interference patterns are known in advance. However, it is often impossible to collect training samples of all potential interference patterns in actual systems, and new and unknown interference patterns may appear in the actual testing process. Therefore, the closed set method will often mistakenly classify the unknown interference pattern as one of the known patterns in the training data set. In addition, in actual systems, due to dynamic factors such as the communication environment and user location, even for known signal patterns, the data distribution between the source domain and the target domain may be different, which further exacerbates the difficulty of interference identification. At present, most algorithms convert I / Q data into time-frequency diagrams, constellation diagrams, etc., and then use deep learning methods to extract interference signal features, and use designed classifiers to reject unknown samples and identify known interference. However, it is worth noting that although converting the original I / Q signal into a time-frequency representation or a spectrum waterfall diagram helps feature extraction, this conversion may bring additional processing overhead and may lose useful information. How to effectively use the rich information in the original I / Q data to accurately identify communication interference patterns in open set scenarios still faces severe challenges.

[0004] The invention with publication number CN114239672A discloses an interference pattern open set recognition model and method based on zero-sample learning. In the supervised training stage, the encoder is trained to learn the potential feature representation of the known interference pattern. In the unsupervised classification stage, the known and unknown interference patterns are classified in the feature space according to the distance-based classification criteria, which can effectively deal with unknown interference patterns and obtain a high open set recognition accuracy. The invention can well characterize the interference pattern open set recognition scenario in the wireless communication anti-interference scenario, but it cannot solve the problem of losing useful information and additional processing overhead. Summary of the invention

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

[0006] In order to achieve the above technical objectives, the technical solution adopted by the present invention is:

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

[0008] Step A: A complex-valued fully connected autoencoder is used to extract the interaction features between I / Q channels, and then it is connected in parallel with the input I / Q data and fused into enhanced data samples;

[0009] Step B, constructing 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 feature information of the input interference signal, the decoder is used to reconstruct the extracted feature information into the input signal, and the closed set classifier uses the SoftMax function to classify the known interference pattern;

[0010] Step C, using the enhanced data samples as training data for the interference feature extraction module, constraining the feature space of the interference signal to minimize the intra-class distance in the feature space; using 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, using the trained encoder to extract the feature vector of the enhanced I / Q sample to be detected, using the open set classification module to calculate the minimum distance from the feature vector to each center vector, and comparing 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 category to which the center vector closest to it belongs.

[0012] As a preferred example, the encoder includes an input layer, a first convolution layer, a second convolution layer, a third convolution layer and a pooling layer connected in sequence; the encoder takes the enhanced data sample as input and maps it 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 so that the extracted feature space becomes compact, and a boundary constraint function is introduced to further compress the feature space;

[0013] The decoder comprises 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 as input, and introduces a mean square error as a reconstruction loss during the training process of the decoder so that the feature vector extracted by the autoencoder contains features representing 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 converts the feature vector into a predicted probability distribution through multiple fully connected layers, and introduces a cross entropy loss function in 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 the interference features by minimizing the joint loss function LOSS during the training process:

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

[0017] where λ cen , rec ,λc cross , λb b represents the weight coefficient;

[0018] Among them, the center loss is:

[0019]

[0020] in, represents the enhanced training sample, d represents the Mahalanobis distance, represents the feature vector extracted from the training sample, Represents y n The center vector of the class, N 2 Indicates the number of training samples;

[0021] The boundary constraints are:

[0022]

[0023] in 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, represents the probability vector output by the closed set classifier, g n Represents the true label probability distribution.

[0030] As a preferred example, a batch normalization layer is introduced into the encoder and the decoder; and a random dropout layer is introduced into the closed set classifier.

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

[0032] The following formula is used to calculate the minimum Mahalanobis distance from the feature vector to each center vector, expressed as:

[0033]

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

[0035] Consider different interference patterns and interference-to-signal ratios and set adaptive thresholds;

[0036] If the minimum Mahalanobis distance 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, expressed as:

[0037]

[0038] Among them, y 0 Indicates the category to which the center vector closest to the feature vector of the sample to be detected belongs, expressed as:

[0039]

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

[0041]

[0042] in, is the training data set corresponding to the kth sample pattern under the specified JSR, Pr[·] represents the probability of the event occurring, d k represents the minimum Mahalanobis distance that satisfies the conditions, η is the statistical confidence, is the type of interference known to the receiver.

[0043] In a second aspect, the present invention discloses a method for open set identification of wireless communication interference signals based on deep learning, the method comprising:

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

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

[0046] The source domain samples are used to train the interference feature extraction module to obtain the optimal parameters of the encoder and complete the training of the encoder; the trained encoder is used to extract the feature vector of the target domain samples and the open set classification module is used to extract the feature vector of the target domain samples.

[0047] Filter out a preset proportion of target domain samples with the smallest distance to each center vector, use the source domain data and the filtered target domain samples to train the adversarial domain adaptation framework, optimize the encoder parameters, and align the feature distribution of known classes in the source domain and target domain;

[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 judged as an unknown class; if it is less than the adaptive threshold, the sample to be detected is identified as the category to which the center vector closest to it belongs.

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

[0050]

[0051] in, and They represent the reconstruction loss function, cross entropy loss function, center loss function and boundary constraint when only source samples are used for training.

[0052] As a preferred example, the adversarial domain adaptation framework is formed by connecting a feature extraction module and a domain discriminator. During the training process, the feature extraction module is used to extract close feature vectors from signals with different data distributions to confuse the domain discriminator, and the domain discriminator 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 is the binary cross entropy loss; α q Represents a given sample Is it from the source domain or the target domain? q = 0, the sample comes from the source domain, when α q = 1, the sample comes from the target domain, represents the feature vector extracted from the training sample, G d (·) denotes the output of the domain discriminator.

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

[0057]

[0058] Where, L d represents the adversarial loss function of the adversarial domain adaptation framework, represents the cross entropy loss function when only source samples are used for training, L rec represents the reconstruction loss when using source domain data and filtered target domain samples for training.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] First, the open-set identification method of wireless communication interference signals based on deep learning of the present invention directly uses 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 realizes accurate open-set identification through an adaptive threshold classifier.

[0061] Second, the open set identification method of wireless communication interference signals based on deep learning of the present invention introduces adversarial domain adaptation to address the situation where the data distribution of the training set and the test set are different. It aligns the known samples in the target domain with the known samples in the source domain to ensure that the model has domain invariance and can achieve accurate open set identification of interference patterns in real complex electromagnetic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a model diagram of the open set interference pattern recognition system for wireless communication;

[0063] Figure 2 Schematic diagram of the open set identification method for communication interference based on feature enhancement

[0064] Figure 3 Schematic diagram of a complex-valued fully connected autoencoder;

[0065] Figure 4 Schematic diagram of a communication open set interference identification method based on adversarial domain adaptation;

[0066] Figure 5 It is a schematic diagram of the open set recognition performance recognition curve with the change of interference signal ratio; Figure 5 (a), (b), and (c) respectively represent the NA, AKS, and AUS performance of the proposed method on the JPR2024 dataset under different JSR conditions.

[0067] Figure 6 is a schematic diagram of the extracted feature vector distribution; Figure 6 Sub-figures a, c, and e in the figure represent the t-SNE embedding representations obtained by directly reducing the dimensionality of the original data in JPR2024, JPR2016, and JPR2018, respectively; Figure 6 Sub-figures b, d, and f in the figure represent the t-SNE embedding representations of the feature vectors extracted from JPR2024, JPR2016, and JPR2018 using the trained models, respectively. DETAILED DESCRIPTION

[0068] The embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings.

[0069] Example 1

[0070] like Figure 1As shown, this embodiment establishes an open set interference pattern recognition system for wireless communication to identify interference signals in the communication system. Consider a system including a jammer, a pair of legal transmitters and receivers, and an interference detector configured at the legal receiving end. The jammer can send noise-like energy signals to disrupt the communication between the legal transmitter and the receiver. The interference detector can determine whether the received signal is affected by interference, and if so, identify a specific interference pattern. In practice, it is impossible for a legal receiver to know all interference patterns, so it is assumed that the interference categories known to the receiver are The rest are unknown modes, recorded as K+1. For the convenience of description, 0 is used to represent no interference, and the signal received by the corresponding receiver is a normal signal. The signal received by the receiver can be expressed as:

[0071]

[0072] Where S 0 (t) represents the normal signal from the legal transmitter, S j (t) represents the communication interference signal from the jammer, and w(t) represents the additive white Gaussian noise.

[0073] This embodiment discloses a method for identifying an open set of wireless communication interference signals based on deep learning, and the method comprises the following steps:

[0074] Step A: A complex-valued fully connected autoencoder is used to extract the interaction features between I / Q channels, and then it is connected in parallel with the input I / Q data and fused into enhanced data samples;

[0075] Step B, constructing 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 feature information of the input interference signal, the decoder is used to reconstruct the extracted feature information into the input signal, and the closed set classifier uses the SoftMax function to classify the known interference pattern;

[0076] Step C, using the enhanced data samples as training data for the interference feature extraction module, constraining the feature space of the interference signal to minimize the intra-class distance in the feature space; using 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, using the trained encoder to extract the feature vector of the enhanced I / Q sample to be detected, using the open set classification module to calculate the minimum distance from the feature vector to each center vector, and comparing 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 category to which the center vector closest to it belongs.

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

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

[0080] Interference feature extraction step: Use the classification-reconstruction framework to train the enhanced data samples and extract interference features. Furthermore, the interference feature extraction module includes an encoder, a decoder, and a closed set classifier, wherein the encoder extracts feature information of the input interference signal through operations such as convolution and pooling, and the decoder reconstructs the extracted feature information into the input signal through operations such as deconvolution and depooling. The closed set classifier uses SoftMax to classify known interference patterns. Furthermore, the enhanced I / Q samples are used as training data for the model, and the 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. Furthermore, the mean of the feature vectors of each type of known samples output by the encoder is used as the center vector of each known pattern.

[0081] Open set identification step: Use the trained encoder to extract the feature vector of the enhanced I / Q sample to be detected, and calculate the minimum distance from the feature vector to each center vector, compare the minimum distance 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. Furthermore, for different interference patterns and interference-to-signal ratios, the open set identification module can adaptively set a reasonable threshold to reject unknown categories and accurately classify known categories.

[0082] The communication interference open set recognition framework based on feature enhancement is as follows Figure 2 As shown in Figure 1, the I / Q data feature enhancement module is first used to extract the interaction features between the I / Q channels. Then, the interference feature extraction module extracts the potential features of the enhanced data for the known pattern. Finally, the distance-based adaptive threshold open set classifier is used to realize the interference pattern recognition. Specifically, the I / Q data feature enhancement module is composed of a complex-valued fully connected autoencoder. The complex-valued fully connected autoencoder structure is shown in Figure 1. Figure 3 As shown in the figure, the encoder and decoder have symmetrical structures, both consisting of two layers of complex-valued fully connected layers, and the mean square error function is used as the loss function to train the model. The output of the encoder is connected in parallel with the input data to obtain the enhanced data.

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

[0084] Table 1 Interference feature extraction module structure table

[0085]

[0086] The encoder takes the augmented data as input and maps it to the feature space through convolution and pooling layers. Introducing the center loss function during training can minimize the intra-class space and make the extracted feature space compact.

[0087] The center loss is:

[0088]

[0089] in, represents the enhanced training samples, represents the feature vector extracted from the training sample, Represents y n The center vector of the class (the average of the feature vectors of samples of each known class), d represents the Mahalanobis distance, as shown in the following formula:

[0090]

[0091] In addition, the boundary constraint function is introduced to further compress the feature space. The boundary constraint is:

[0092]

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

[0094] The decoder reconstructs the extracted features as input, and introduces the mean square error as the reconstruction loss during the training process to ensure that the feature vector extracted by the autoencoder contains features that can represent the original data.

[0095] The reconstruction loss is:

[0096]

[0097] in, Represents the reconstructed output of the decoder.

[0098] The closed set classifier converts the feature vector into a predicted probability distribution through a fully connected layer, and introduces the cross entropy loss function during the training process to ensure accurate classification of known interference patterns.

[0099] The cross entropy loss is:

[0100]

[0101] in, Represents the probability vector of the closed set classifier output, g n Represents the true label probability distribution. During the training process, the clustering of interference features is achieved by minimizing the joint loss function to ensure the subsequent open set recognition performance. The joint loss function is as follows:

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

[0103] where λ cen , rec , cross and λ b Represents the weight coefficient.

[0104] In order to stabilize the training process, batch normalization is introduced in the encoder and decoder. In order to prevent overfitting during training, a random dropout layer (Dropout) is introduced in the closed set classifier.

[0105] Use the trained encoder to extract the enhanced I / Q sample X to be detected te,en The eigenvector of , and calculate the minimum Mahalanobis distance from the eigenvector to each center vector, expressed as:

[0106]

[0107] Where F(X te,en ) represents the feature vector extracted by the encoder. The minimum distance is compared with the set adaptive threshold. Further, 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, is the training data set corresponding to the kth sample pattern under the specified JSR, and η is the statistical confidence. If the minimum distance 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, which can be expressed as:

[0110]

[0111] Among them, y 0The category to which the center vector closest to the feature vector of the sample to be detected belongs can be expressed as:

[0112]

[0113] Example 2

[0114] This embodiment discloses a method for identifying an open set of wireless communication interference signals based on deep learning, the method comprising:

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

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

[0117] The source domain samples are used to train the interference feature extraction module to obtain the optimal parameters of the encoder and complete the training of the encoder; the trained encoder is used to extract the feature vector of the target domain samples and the open set classification module is used to extract the feature vector of the target domain samples.

[0118] Filter out a preset proportion of target domain samples with the smallest distance to each center vector, use the source domain data and the filtered target domain samples to train the adversarial domain adaptation framework, optimize the encoder parameters, and align the feature distribution of known classes in the source domain and target domain;

[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 judged as an unknown class; if it is less than the adaptive threshold, the sample to be detected is identified as the category to which the center vector closest to it belongs.

[0120] In view of the different data distributions of the training set and the test set, this embodiment proposes an open set identification method for communication interference signals based on adversarial domain adaptation (FOSR-DA), comprising the steps of:

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

[0122] Known and unknown class separation step: In order to prevent the negative transfer that may be caused by unknown samples during the domain adaptation process, it is necessary to separate the unknown samples first. Furthermore, the known and unknown separation step includes two steps: interference feature extraction and known class screening. These two steps are the same as the steps in the open set identification method of communication interference based on feature enhancement in the previous article. Specifically, the feature extractor is trained using the enhanced source domain I / Q samples. The feature vector of the target domain sample is extracted using the trained encoder, and the target domain sample is identified using an open set classifier to classify the target domain samples into known classes and unknown classes. The top γ% (for example, 20%) of target domain samples closest to each center vector are selected for subsequent domain adaptation tasks.

[0123] Adversarial domain adaptation step: Use source domain data and selected target domain samples to train the adversarial domain adaptation framework, optimize the encoder parameters, and align the feature distributions of known classes in the source and target domains. Furthermore, the adversarial domain adaptation framework is composed of a feature extraction module connected to a domain discriminator. During the training process, the feature extraction module extracts close feature vectors from signals with different data distributions as much as possible to confuse the domain discriminator, and the domain discriminator distinguishes samples from different data distributions as much as possible.

[0124] Open set recognition step: Use the encoder trained after adversarial domain adaptation to extract the feature vector of the sample to be detected, and calculate the minimum distance from the feature vector to each center vector. Compare the minimum distance with the set adaptive threshold. If it is greater than the threshold, it is judged as an unknown class; if it is less than the threshold, the sample to be detected is identified as the category to which the nearest center vector belongs.

[0125] The schematic diagram of communication interference open set recognition based on adversarial domain adaptation is shown in the attached figure. Figure 4 As shown in Figure 2, the whole process can be divided into two stages: the known and 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 and eliminate unknown samples.

[0126] Specifically, the known / unknown separation stage consists of three modules: feature enhancement module, interference feature extraction module and open set classification module, and its structure is the same as Table 1. First, the feature enhancement module is used to perform data enhancement on the source domain and target domain samples. Then, the feature extraction module is trained using the enhanced source domain samples. The optimal parameter θ of the encoder can be obtained in the following way. f :

[0127]

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

[0129] Furthermore, in the adversarial domain adaptation stage, the preliminarily trained feature extraction module is connected to a domain discriminator, and the encoder is adversarially trained to align known samples in the source and target domains, so that it can maintain stable feature extraction capabilities when facing different data distributions. Adversarial domain adaptation training is performed using source domain samples and filtered target domain samples. For a given sample Using alpha q express is from the source domain (α q =0) or the target domain (α q =1). During the training process, the feature extraction module extracts close feature vectors from signals of different data distributions as much as possible to confuse the domain discriminator, and the domain discriminator distinguishes samples from different data distributions as much as possible. The adversarial loss function is defined as follows:

[0130]

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

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

[0133]

[0134] The encoder updated according to the above formula can not only retain the knowledge of the source domain but also have domain invariance, and can stably and effectively extract interference features in I / Q data.

[0135] The encoder trained by adversarial domain adaptation 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 category to which the nearest center vector belongs. The environment and hyperparameter configuration required for network training are shown in Table 2:

[0136] Table 2 Network training environment and hyperparameter configuration

[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 Number of training rounds for I / Q data feature enhancement module 300 I / Q data feature enhancement module Batch size 400 Interference feature extraction module learning rate 0.0008 Interference feature module extraction Batch size 256 Interference feature extraction module training rounds 250 Learning rate in adversarial domain adaptation phase 0.0001 Batch size in the adversarial domain adaptation stage 200 Number of training rounds in the adversarial domain adaptation phase 100

[0138] Open set recognition needs to consider the recognition accuracy of unknown classes, and the evaluation indicators of closed set recognition are no longer applicable. Therefore, the accuracy of known classes (AKS), the accuracy of unknown classes (AUS), and the normalized accuracy (NA) are introduced:

[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, λ r In order to balance the coefficient of recognition accuracy between known and unknown classes, it is set to 0.5. In addition, the accuracy of known classes, the accuracy of unknown classes and F1-score are also introduced:

[0142]

[0143] Where S correct Indicates the total number of known samples that are correctly classified, U correct Indicates the total number of unknown samples correctly classified as unknown patterns. The above two types of indicators have different focuses. Accuracy focuses more on the discovery of unknown categories, while F1-score treats known and unknown categories equally. Combining the two indicators for comprehensive evaluation can better evaluate the overall performance of the model.

[0144] Experimental verification

[0145] The present invention selects BPSK as a normal signal during the experimental verification process, and selects six typical suppression interferences: single-tone interference, multi-tone interference, swept frequency interference, comb spectrum interference, pulse interference, and noise swept frequency interference. BPSK signal, single-tone interference, multi-tone interference, and pulse interference are set as known interference modes, and the remaining interference modes are set as unknown modes. The above communication signals and interference signals are all generated by MATLAB simulation. The main parameters of the above interference modes are as follows: the carrier frequency range of each interference is set to kHz. Multi-tone interference consists of seven syllables. The sweep frequency speed of the swept frequency interference is 5000kHz / s. The comb spectrum interference consists of two fixed frequency signals with a frequency of 200-500kHz. The frequency modulation coefficient of the noise frequency modulation interference is set to 5. For each known interference mode, 2500 samples are generated every 5dB when the JSR is between (-10dB, 15dB), of which 80% are used for training and 20% are used for testing. For each type of unknown interference, 500 samples are generated for testing, and the number of I / Q sampling points in the data set is set to 128. In addition, the simulation-generated data set is integrated with two public data sets RML2016.10a and RML2018.01a to create the JPR2016 and JPR2018 data sets for experimental verification.

[0146] The communication interference open set identification method based on feature enhancement in the present invention is compared with IIAE-JPR, SR2CNN, ZSL-JPR, and OpenMax. Figure 5 To compare the performance under different JSRs on the dataset JPR2024, I / Q samples with the same JSR are used in both training and testing. Under different JSRs, our method outperforms the other four methods in terms of NA, AKS, and AUS. Specifically, when JSR = $10$dB, the method of this patent improves NA by 7%, 13%, 14%, and 21% compared with IIAE-JPR, ZSL-JPR, SR2CNN, and OpenMax-JPR, respectively. In addition, as JSR increases, AKS, AUS, and NA also increase. This is because when JSR rises, the strength of the interference signal becomes stronger than the normal signal, allowing the network to better identify and learn its features.

[0147] Table 3 shows the recognition accuracy of the communication interference open set recognition method based on feature enhancement under different signal-to-noise ratios on three data sets. I / Q samples with the same signal-to-noise ratio are used in both training and testing. As the signal-to-noise ratio decreases, the recognition performance gradually decreases. This is because as the signal-to-noise ratio decreases, the intensity of the background noise increases. In this case, it becomes more difficult to capture the identification features in the interference signal.

[0148] Table 3 Performance comparison of communication interference open set recognition method based on feature enhancement under different signal-to-noise ratios on three data sets

[0149]

[0150] The trained encoder is used to extract the feature vectors of the interference signals in the three data sets, and the feature vector distribution diagram obtained by t-SNE is shown in the figure below. Figure 6 As shown. Figure 6 From a, c, and e in Figure 1, we can see that the interference features of the original samples in each data set are overlapping and cannot be separated. After feature enhancement and feature extraction, features belonging to the same interference pattern are clustered together, while features of unknown patterns become outliers.

[0151] In order to evaluate the performance of the communication interference open set recognition method based on feature enhancement in different open set tasks, the performance of this method in the face of different unknown pattern combinations on JPR2024 was tested, as shown in Table 4. By comparing the three experiments ②, ③, and ④, it was found that this method showed stable recognition performance in different unknown pattern combinations. Through ①, ③, and ⑤, it was found that as the number of unknown patterns increased, the NA also decreased. The reason is that as the number of unknown patterns increased, the number of features learned by the model would decrease, thereby limiting the model's ability to represent the signal feature space, which in turn led to a decrease in recognition performance.

[0152] Table 4 Performance of this method on JPR2024 for different unknown pattern combinations

[0153]

[0154]

[0155] In order to verify the communication interference open set identification method based on adversarial domain adaptation, two cases of different channel conditions and different JSRs in JPR2024 are further considered. Table 5 shows the maximum NA reduction rate of each method when the source domain and target domain are generated under Gaussian and Rayleigh fading channel conditions, respectively. It can be seen that the communication interference open set identification method based on adversarial domain adaptation has the best robustness, with an average reduction 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 drop rate when target domain samples are generated under Rayleigh fading channel

[0157]

[0158] For the second case, the experiment sets the source domain to contain samples with a specific JSR, while the target domain contains 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 is -5dB and the JSR of the target domain samples is in the range of -10dB to 15dB. It is worth noting that FOSR-DA consistently outperforms the baseline methods in the considered JSR range. The highest accuracy of FOSR-DA reaches 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 when processing signals with domain offset problems.

[0159] Table 6 NA values ​​of six methods when the JSR of source domain samples and target samples are different

[0160]

[0161]

[0162] All the above experimental results prove that the communication interference open set recognition method based on feature enhancement not only has excellent recognition accuracy, but also can maintain a stable NA when facing different open set interference pattern recognition tasks. In addition, when the channel conditions and JSRs of training samples and test samples are different, the communication interference open set recognition method based on adversarial domain adaptation can show excellent interference signal open set recognition performance and strong robustness.

[0163] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. Moreover, the present application can adopt 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.) that contain computer-usable program code. The scheme in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0164] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0165] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0166] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0167] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0168] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for open set identification of wireless communication interference signals based on deep learning, characterized in that: The method comprises the following steps: Step A: A complex-valued fully connected autoencoder is used to extract the interaction features between I / Q channels, and then it is connected in parallel with the input I / Q data and fused into enhanced data samples; Step B, constructing 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 feature information of the input interference signal, the decoder is used to reconstruct the extracted feature information into the input signal, and the closed set classifier uses the SoftMax function to classify the known interference pattern; Step C, using the enhanced data samples as training data for the interference feature extraction module, constraining the feature space of the interference signal to minimize the intra-class distance in the feature space; using the mean of the feature vectors of each known sample output by the encoder as the center vector of each known pattern; Step D, using the trained encoder to extract the feature vector of the enhanced I / Q sample to be detected, using the open set classification module to calculate the minimum distance from the feature vector to each center vector, and comparing 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 category to which the center vector closest to it belongs.

2. The method for identifying wireless communication interference signals based on deep learning according to claim 1, characterized in that: The encoder comprises an input layer, a first convolution layer, a second convolution layer, a third convolution layer and a pooling layer connected in sequence; the encoder takes the enhanced data sample as input and maps it 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 so that the extracted feature space becomes compact, and a boundary constraint function is introduced to further compress the feature space; The decoder comprises 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 as input, and introduces a mean square error as a reconstruction loss during the training process of the decoder so that the feature vector extracted by the autoencoder contains features representing 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 converts the feature vector into a predicted probability distribution through multiple fully connected layers, and introduces a cross entropy loss function in the training process of the closed set classifier to improve the classification accuracy of known interference patterns.

3. The method for identifying wireless communication interference signals based on deep learning according to claim 2, characterized in that: The interference feature extraction module clusters the interference features by minimizing the joint loss function LOSS during the training process: LOSS=λ cen L cen +λ rec L rec +λ cross L cross +λ b L b where λ cen , rec , cross , b represents the weight coefficient; Among them, the center loss is: in, represents the enhanced training sample, d(·) represents the Mahalanobis distance, represents the feature vector extracted from the training sample, Represents y n The center vector of the class, N2 represents the number of training samples; The boundary constraints are: in is the minimum distance learned by the encoder; The reconstruction loss is: in, represents the reconstructed output of the decoder; The cross entropy loss is: in, Represents the probability vector of the closed set classifier output, g n Represents the true label probability distribution.

4. The method for identifying wireless communication interference signals based on deep learning according to claim 2, characterized in that: A batch normalization layer is introduced into the encoder and the decoder; and a random dropout layer is introduced into the closed set classifier.

5. The method for identifying open sets of wireless communication interference signals based on deep learning according to claim 1, characterized in that: Step D further comprises: The following formula is used to calculate the minimum Mahalanobis distance from the feature vector to each center vector, expressed as: Where F(X) represents the feature vector extracted by the encoder, c k represents the center vector of class k, The type of interference known to the receiver; Consider different interference patterns and interference-to-signal ratios and set adaptive thresholds; If the minimum Mahalanobis distance 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, expressed as: Among them, y0 represents the category to which the center vector closest to the feature vector of the sample to be detected belongs, which is expressed as:

6. The method for identifying open sets of wireless communication interference signals based on deep learning according to claim 1 or 5, characterized in that: In step D, the adaptive threshold is: in, is the training data set corresponding to the kth sample pattern under the specified JSR, η is the statistical confidence, Pr[·] represents the probability of the event, d k represents the minimum Mahalanobis distance that satisfies the conditions, is the type of interference known to the receiver.

7. A method for open set identification of wireless communication interference signals based on deep learning, characterized in that: The method comprises: A complex-valued fully connected autoencoder is used to perform data enhancement on source domain and target domain samples respectively. Specifically, the interaction features between the I / Q channels of the source domain and the target domain are extracted respectively, and then connected in parallel with the input I / Q data of the source domain and the target domain to form enhanced source domain samples and target domain samples. Construct an interference feature extraction module, which includes an encoder, a decoder and a closed set classifier. The encoder is used to extract feature information of the input interference signal, the decoder is used to reconstruct the extracted feature information into the input signal, and the closed set classifier uses the SoftMax function to classify the known interference patterns. The interference feature extraction module is trained using source domain samples to obtain the optimal parameters of the encoder and complete 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 to classify the target domain samples into known classes and unknown classes. Filter out a preset proportion of target domain samples with the smallest distance to each center vector, use the source domain data and the filtered target domain samples to train the adversarial domain adaptation framework, optimize the encoder parameters, and align the feature distribution of known classes in the source domain and target domain; 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 judged as an unknown class; if it is less than the adaptive threshold, the sample to be detected is identified as the category to which the center vector closest to it belongs.

8. The method for identifying open sets of wireless communication interference signals based on deep learning according to claim 7, characterized in that: Use source domain samples to train the interference feature extraction module and obtain the optimal parameter θ of the encoder f : in, and They represent the reconstruction loss function, cross entropy loss function, center loss function and boundary constraint when only source samples are used for training.

9. The method for identifying open sets of wireless communication interference signals based on deep learning according to claim 7, characterized in that: The adversarial domain adaptation framework is composed of a feature extraction module connected with a domain discriminator. During the training process, the feature extraction module is used to extract close feature vectors from signals with different data distributions to confuse the domain discriminator, and the domain discriminator is used to distinguish samples from different data distributions. The adversarial loss function of the adversarial domain adaptation framework is: Among them, L bce is the binary cross entropy loss; α q Represents a given sample Is it from the source domain or the target domain? q = 0, the sample comes from the source domain, when α q = 1, the sample comes from the target domain, represents the feature vector extracted from the training sample, G d (·) denotes the output of the domain discriminator.

10. The method for open set identification of wireless communication interference signals based on deep learning according to claim 7, characterized in that: Update the encoder parameters according to the following formula and the parameters of the domain discriminator Where, L d represents the adversarial loss function of the adversarial domain adaptation framework, represents the cross entropy loss function when only source samples are used for training, L rec represents the reconstruction loss when using source domain samples and filtered target domain samples for training.

Images