Radiation source classification method based on polarization domain feature learning extraction
Through polarization domain feature learning, polarized signal characteristics of communication radiation sources are extracted, which solves the problems of high signal analysis complexity and poor recognition effect in the prior art, and realizes high-precision classification of communication radiation sources.
Patent Information
- Application Number
- CN202411004358.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-25
- Publication Date
- 2024-11-19
AI Technical Summary
The existing communication radiation source classification method is greatly affected by noise when analyzing non-Gaussian and non-stationary signals, has poor generalization ability, and has high computational complexity, making it difficult to achieve accurate identification under conditions similar time and frequency information.
Using a method based on polarization domain feature learning, the polarization amplitude ratio and phase difference of the polarization signal are measured, combined with representation learning and feature clustering, the polarization domain characteristics of the communication radiation source are extracted to realize the classification of different radiation sources.
It effectively improves the classification accuracy of communication radiation sources and improves the recognition ability of polarized signals, especially under low signal-to-noise ratio and variable modulation style conditions, showing good classification effects.
Smart Images

Figure CN118981660A_ABST
Abstract
Description
Technical Field
[0001] The embodiments of the present disclosure relate to a communication radiation source classification method, and more particularly to a radiation source classification method based on polarization domain feature learning and extraction. Background Art
[0002] At present, the classification and identification of communication radiation sources mainly uses specific radiation source identification technology, namely radio frequency fingerprint identification technology, which utilizes the hardware differences of communication radiation source equipment, namely the reflection of radio frequency fingerprints in spatial radiation signals to achieve the distinction and identification of radiation sources.
[0003] The existing communication radiation source classification based on expert experience and deep learning has different degrees of limitations in the analysis and use of RF signal parameter statistical characteristics and transform domain characteristics. The signal parameter statistical characteristics recognition method is greatly affected by noise, has weak analysis capabilities for non-Gaussian and non-stationary signals, and has poor generalization capabilities. The signal change domain recognition method has difficulty in selecting basis functions, is prone to cross-term interference, has high computational complexity, and is inefficient. Under the condition that the space, time, frequency, and energy domain characteristics of the communication radiation source are similar, the difference in time-frequency information (time domain characteristics, frequency domain characteristics, high-order spectrum characteristics, and time-frequency transform domain characteristics) decreases, and the difficulty of analysis and extraction using direct or indirect methods is further increased. From the principle analysis, it can be seen that the recognition effects of the above two methods will deteriorate to varying degrees, and even some methods are difficult to support the recognition of such individual communication radiation sources.
[0004] Therefore, it is necessary to improve one or more problems existing in the above-mentioned related technical solutions.
[0005] It should be noted that this section is intended to provide background or context for the technical solutions of the present disclosure stated in the claims. The description herein is not admitted to be prior art by virtue of being included in this section. Summary of the invention
[0006] The purpose of the embodiments of the present invention is to provide a radiation source classification method based on polarization domain feature learning and extraction, thereby overcoming one or more problems caused by limitations and defects of related technologies at least to a certain extent.
[0007] According to an embodiment of the present invention, a radiation source classification method based on polarization domain feature learning and extraction is provided, the method comprising:
[0008] Receiving polarization signals from different communication radiation sources, and obtaining polarization characteristic representations of the polarization signals by measuring polarization amplitude ratios and polarization phase differences of the polarization signals;
[0009] Combining the original data of the polarization signal with the polarization feature representation, and obtaining a training data set and a test data set through data preprocessing;
[0010] The training data set outputs features extracted by learning through representation learning phase;
[0011] The features extracted by the representation learning and the test data set complete the classification of communication radiation sources through the feature clustering stage.
[0012] In one embodiment of the present disclosure, the data preprocessing includes the following steps:
[0013] Splitting the received radio frequency I / Q complex valued data of the main polarization and cross polarization signals into two groups of I-channel and Q-channel real valued data;
[0014] The amplitude and phase of the main polarization and cross-polarization signals are calculated using two sets of corresponding I-channel and Q-channel data, and then the polarization amplitude ratio and polarization phase difference of the main polarization and cross-polarization signals are calculated using the two sets of amplitude and phase values obtained;
[0015] The polarization amplitude ratio and polarization phase difference in the main polarization and cross-polarization signals are aligned and then spliced and cut with the original four-way I / Q data to obtain M×6×128-dimensional training and test data sets.
[0016] In one embodiment of the present disclosure, a contrastive learning model is used in the representation learning stage, and the contrastive learning model is composed of data enhancement, a basic coding network, a projection head, and a contrastive loss.
[0017] In one embodiment of the present disclosure, in the feature clustering stage, a sub-cluster is created for each cluster, and the number of clusters is changed by splitting clusters and merging clusters.
[0018] In one embodiment of the present disclosure, when splitting the cluster, each cluster is split into two sub-clusters with probability P1, and the k value is changed adaptively. The cluster will copy the kth unit in the last layer of the clustering network and the weight connected to the previous hidden layer, and use the parameters learned by the sub-clustering network to initialize two new clusters. The expression for splitting the cluster includes:
[0019] P1=Min(1,H s ) (1)
[0020]
[0021] Where P1 is the cluster split probability, H s The Hastings ratio is the ratio of the marginal likelihood of the data under the two sub-clusters to the marginal likelihood under the cluster, Γ is the gamma function, represents the data in cluster K, N k =|c k |, Represents the data N in sub-cluster j (j∈{1,2}) k,j =|c k,j |, f x (g,l) is the marginal likelihood function, and l=(m,k,y,v) is the normal inverse wishart hyperparameter.
[0022] In one embodiment of the present disclosure, when merging clusters, each sub-cluster is merged with its three nearest sub-clusters with probability P2 each time it is aggregated. After merging, a new sub-cluster network is initialized, and the last layer unit of the merged cluster and the weights of its connection with the previous hidden layer are removed from the cluster network, and the weighted maximum a posteriori estimation is used to initialize the parameters and weights of the newly generated clusters. The weighting is based on the output of the deep network (r i ,k) completed; the expressions for merging clusters include:
[0023]
[0024] Where P2 is the cluster merging probability, H m is the inverse of the Hastings ratio, m is the prior mean of the normal inverse wishart hyperparameter distribution, is the updated mean of the normal inverse wishart hyperparameter distribution, k is the initial precision parameter of the normal inverse wishart hyperparameter distribution, is the updated precision parameter of the normal inverse wishart hyperparameter distribution, the prior scale matrix of the covariance matrix of the y normal inverse wishart hyperparameter distribution, The update scale matrix of the covariance matrix of the normal inverse wishart hyperparameter distribution, v is the initial degrees of freedom of the normal inverse wishart hyperparameter distribution, is the updated degrees of freedom for the normal inverse wishart hyperparameter distribution.
[0025] In one embodiment of the present disclosure, a first network update optimization is performed on the representation learning phase, and the expression of the first network update optimization includes:
[0026]
[0027] Among them, q is the representation of the query sample, k + is the representation of the positive sample, is the representation of negative samples, t is the temperature hyperparameter, f(g) is the similarity calculation of dot product, and in the contrastive learning architecture, the backbone network uses the standard ResNet-18 network. The projection head added to the backbone network is a two-layer MLP network with 2048 neurons in the first layer and 8 neurons in the second layer for output feature dimension, and the ReLU function is used for activation in the output layer. The other hyperparameters of the network are set to batch size batchsize = 128 and temperature hyperparameter t = 0.2.
[0028] In one embodiment of the present disclosure, a second network update optimization is performed on the feature clustering stage, and the expression of the second network update optimization includes:
[0029]
[0030] Among them, r i,k is the soft assignment of feature representation, i.e., the mapping from feature representation to clustering.
[0031] In one embodiment of the present disclosure, a third network update optimization is performed on the feature clustering stage, and the third network update optimization expression includes:
[0032]
[0033] z i = argmax k (r i,k ) (13)
[0034] Among them, x i is the characteristic representation of the sample, is the mean of subcluster j of cluster k.
[0035] The present disclosure proposes a radiation source classification method based on polarization domain feature learning and extraction, which can include the following beneficial effects: the method obtains subtle differences between polarization domain features of communication radiation sources through measurement and analysis of cross-polarization discrimination of different communication radiation source antennas, obtains dual-polarization feature representation of communication radiation sources, splits and merges clusters through representation learning stage and feature clustering stage, and finally completes autonomous learning and extraction of polarization domain features to achieve classification of communication radiation sources. The method effectively improves the classification effect of communication radiation sources and improves the classification accuracy of communication radiation sources. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] The accompanying drawings herein are incorporated into the specification and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the specification are used to explain the principles of the present disclosure. Obviously, the accompanying drawings described below are only some embodiments of the present disclosure, and for ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without creative work.
[0037] Figure 1 A step diagram of a radiation source classification method based on polarization domain feature learning and extraction in an exemplary embodiment of the present disclosure is shown;
[0038] Figure 2 A diagram showing a deep clustering network structure representing learning in an exemplary embodiment of the present disclosure is shown;
[0039] Figure 3 An example graph of network training loss in Experiment 1 in an exemplary embodiment of the present disclosure is shown;
[0040] Figure 4 An example diagram of a classification confusion matrix of four communication radiation source individuals at a signal-to-noise ratio of 0-20 dB in Experiment 1 in an exemplary embodiment of the present disclosure is shown;
[0041] Figure 5 An example diagram showing classification effects under different signal-to-noise ratios in Experiment 1 in an exemplary embodiment of the present disclosure is shown;
[0042] Figure 6 An example diagram of feature dimensionality reduction extracted by a learning network under a signal-to-noise ratio of 0-20 dB in Experiment 1 in an exemplary embodiment of the present disclosure is shown;
[0043] Figure 7 An example diagram of the classification confusion matrix of six communication radiation source individuals at 5dB, 10dB, and 15dB signal-to-noise ratios in Experiment 2 in an exemplary embodiment of the present disclosure is shown;
[0044] Figure 8 An example diagram showing classification effects under different signal-to-noise ratios in Experiment 2 in an exemplary embodiment of the present disclosure is shown;
[0045] Fig. 9 An example diagram of dimensionality reduction of features extracted by the representation learning network at 5dB, 10dB, and 15dB signal-to-noise ratios in Experiment 2 of the exemplary embodiment of the present disclosure is shown. DETAILED DESCRIPTION
[0046] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that the disclosure will be more comprehensive and complete and to fully convey the concepts of the example embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.
[0047] In addition, the accompanying drawings are only schematic illustrations of the embodiments of the present disclosure and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated descriptions will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities.
[0048] This example embodiment provides a radiation source classification method based on polarization domain feature learning and extraction. Figure 1 As shown in , the radiation source classification method based on polarization domain feature learning and extraction may include: steps S101 to S104.
[0049] Step S101: receiving polarization signals from different communication radiation sources, and obtaining polarization characteristic representations of the polarization signals by measuring polarization amplitude ratios and polarization phase differences of the polarization signals;
[0050] Step S102: combining the original data of the polarization signal with the polarization feature representation, and obtaining a training data set and a test data set through data preprocessing;
[0051] Step S103: The training data set outputs features extracted by representation learning in the representation learning phase;
[0052] Step S104: the features extracted by the representation learning and the test data set are used to complete the classification of communication radiation sources through a feature clustering stage.
[0053] Through the above-mentioned radiation source classification method based on polarization domain feature learning and extraction, the polarization amplitude ratio and polarization phase difference of different communication radiation source antennas are measured to obtain the characteristic representation of the polarization signal of the communication radiation source. The clustering clusters are split and merged through the representation learning stage and the feature clustering stage, and finally the autonomous learning and extraction of polarization domain features are completed to realize the classification of communication radiation sources. This method effectively improves the classification effect of communication radiation sources and improves the classification accuracy of communication radiation sources.
[0054] Next, we will refer to Figures 1 to 9 The various steps of the above-mentioned radiation source classification method based on polarization domain feature learning and extraction in this example implementation are described in more detail.
[0055] The main part of the electromagnetic wave radiated by the communication antenna is the main polarization component, and a small amount of polarization components orthogonal to it are cross-polarization components or orthogonal to it. According to the antenna design principle, the ratio of the power level of the main polarization component of the electromagnetic wave radiated by the transmitting system to the total power level is called cross-polarization discrimination. As an important indicator to measure the quality of the system, the better the cross-polarization discrimination, the more ideal the system performance. However, in practical applications, under the influence of factors such as the inherent error of the device and the machining error, there are different degrees of deviation in the cross-polarization discrimination of the antenna, which is reflected in the polarization signal as the polarization domain characteristics of the communication radiation source. The reception of polarization signals requires a polarization reconnaissance receiver, which usually adopts the polarization-pattern method, the amplitude-phase method and the multi-antenna method. The amplitude-phase method is currently widely used in modern receiving systems.
[0056] Step S101: receiving polarization signals from different communication radiation sources, and obtaining polarization characteristic representations of the polarization signals by measuring polarization amplitude ratios and polarization phase differences of the polarization signals.
[0057] Specifically, the polarization state can be described by the shape and rotation of the space trajectory formed by the change of the electric field vector endpoint over time, which can be divided into three types: fully polarized wave, partially polarized wave and non-polarized wave. Assuming that the electromagnetic wave emitted by the communication radiation source is a monochromatic fully polarized electromagnetic wave propagating along the positive direction of the z-axis, the vector form of the instantaneous value of the electromagnetic wave is
[0058]
[0059] Where e x 、e y are the unit basis vectors in the x-axis and y-axis directions respectively. x 、E y for
[0060] E x =E 0x cos(2pf0×t-kz+f x ) (2)
[0061] E y =E 0y cos(2pf0×t-kz+f y ) (3)
[0062] Where: f0 is the frequency of the electromagnetic wave; K is the space propagation constant; E 0x and f x are the electric field amplitude and phase in the positive direction of the x-axis respectively; E 0y and f y are the electric field amplitude and phase in the positive direction of the y-axis respectively.
[0063] Write equation (1) in complex vector form
[0064]
[0065] Where: e is the complex electric field vector; e 0x and f x are the electric field amplitude and phase in the positive direction of the x-axis respectively; e 0y and f y are the electric field amplitude and phase in the positive direction of the y-axis, respectively. From equations (4) and (5), it can be concluded that the polarization information is reflected in the amplitude ratio and phase difference of the electric field components in two orthogonal directions. Ignoring the absolute phase information, the polarization ratio is defined according to the linear polarization ratio of polarization.
[0066]
[0067] Therefore, for a given electromagnetic wave, its polarization state can be completely described by the polarization ratio, where: tang = e 0y / e 0x ,g∈[0,2p);f=f y -f x ,f∈[0,2p].
[0068] The (g, f) given above is the phase descriptor of the polarization state of the electromagnetic wave, which can be used to characterize the polarization domain characteristics of the radiation source. In order to simplify the calculation amount and more intuitively characterize the polarization domain characteristics of the RF signal of the communication radiation source, this paper directly uses tang=e y / e x ,g∈[0,2p) is taken as the amplitude ratio of the main polarization signal to the cross-polarization signal, i.e., the polarization amplitude ratio, and f=f y -f x ,f∈[0,2p] is taken as the phase difference between the main polarization and cross polarization signals, i.e., the polarization phase difference. By calculating the amplitude and phase of the main polarization and cross polarization signals, the polarization domain characteristics of the communication radiation source defined above can be obtained by formula [6]: polarization amplitude ratio tang and polarization phase difference f.
[0069] Step S102: combining the original data of the polarization signal with the polarization feature representation, and obtaining a training data set and a test data set through data preprocessing.
[0070] Specifically, the polarization signal collected by the polarization reconnaissance receiver contains two sets of I / Q data of the main polarization channel and the cross-polarization channel. When the four-channel I / Q data is directly processed using a deep clustering network, the network is easily affected by noise and errors, making it difficult to quickly capture the polarization features in the signal. Therefore, in order to enable the deep clustering network to extract polarization features quickly and stably, a training data set and a test data set are constructed through preprocessing to guide the network to adapt to the polarization features. The preprocessing is divided into three steps. First, the received RF I / Q complex data of the main polarization and cross-polarization signals are split into two groups of I-channel and Q-channel real-valued data; then, the amplitude and phase of the main polarization and cross-polarization signals are calculated using the two groups of corresponding I-channel and Q-channel data, and then the polarization amplitude ratio and polarization phase difference of the main polarization and cross-polarization signals are calculated using the two groups of amplitude and phase values obtained; finally, the polarization amplitude ratio and polarization phase difference in the main polarization and cross-polarization signals are aligned and then spliced and cut with the original four-channel I / Q data to obtain M×6×128-dimensional training data sets and test data sets.
[0071] Step S103: The training data set outputs features extracted by representation learning in the representation learning phase.
[0072] Specifically, the representation learning stage adopts a contrastive learning model, which consists of four parts: data enhancement, basic encoding network, projection head and contrastive loss. The strategy adopted for data enhancement is random segmentation including row and column flipping and length adjustment, as well as center value smoothing. Through data enhancement, the generalization ability and robustness of the model can be improved, and the effectiveness of extracting features from the original data can be enhanced. The basic encoding network adopts Momentum Contrast (MoCo), which includes sequence and momentum encoders. The sequence ensures that the current small batch of training data enters the sequence and squeezes the earliest small batch of training data out of the sequence, realizing stack-like processing. The momentum encoder establishes an association between the key encoder at time t and the key encoder at time t-1, ensuring that all key encoders have similar structures. The above two structures realize the encoding of the enhanced data, that is, mapping the data to the representation in the latent space. The projection head is a multi-layer perceptron. Compared with the traditional fully connected network, the multi-layer perceptron can learn more complex nonlinear mappings, so as to better capture the connection between features. Through further encoding of the projection head, the representation of the latent space is more distinctive, which can improve the contrast performance. The contrast loss adopts the classic contrast learning loss function Noise Contrast Estimation (Noise Contrast Estimation). Contrastive Estimation, InfoNCE), whose goal is to maximize the consistency between two related features, that is, to make the representation of positive samples closer and the representation of negative samples farther away, which can be well trained for the basic encoder network and the projection head.
[0073] What needs to be understood is that the learning algorithm process is as shown in Figure 1
[0074]
[0075]
[0076] Figure 1
[0077] Step S104: the features extracted by the representation learning and the test data set are used to complete the classification of communication radiation sources through a feature clustering stage.
[0078] Specifically, based on the high-dimensional discrete data characteristics of polarization signals and the high-dimensional data representation capabilities of deep neural networks, a Bayesian nonparametric (BNP) type method deep nonparametric clustering method was adopted for unknown number of clusters, which is adaptive to data and mathematically inferable. This method is a Dirichlet Process Mixture (DPM) algorithm similar to the Dirichlet process. It creates a subcluster for each cluster cluster and uses splitting and merging to change the number of clusters.
[0079] It is important to understand that the clustering algorithm process is as shown in Figure 2:
[0080]
[0081]
[0082] Figure 2
[0083] When splitting clusters, each cluster is split into two sub-clusters with probability P1. At the same time, for adaptive k value changes, the cluster will copy the kth unit in the last layer of the clustering network and the weights connected to the previous hidden layer, and use the parameters learned by the sub-clustering network to initialize two new clusters. The expression of the above process is:
[0084] P1=Min(1,H s ) (7)
[0085]
[0086] Where P1 is the cluster split probability, H s The Hastings ratio is the ratio of the marginal likelihood of the data under the two sub-clusters to the marginal likelihood under the cluster, Γ is the gamma function, represents the data in cluster K, N k =|c k |, Represents the data N in sub-cluster j (j∈{1,2}) k,j =|ck ,j|,f x (g,l) is the marginal likelihood function, and l=(m,k,y,v) is the normal inverse wishart hyperparameter.
[0087] When merging clusters, in order to avoid merging two groups of sub-clusters at the same time and causing multi-class aggregation errors, and to maximize the efficiency of sub-cluster merging, each sub-cluster is merged with its three nearest sub-clusters with probability P2 each time it is aggregated. After the merger, a new sub-cluster network will be initialized, and the last layer unit of the merged cluster and its weights connected to the previous hidden layer will be removed from the clustering network. The weighted maximum a posteriori (MAP) estimate is used to initialize the parameters and weights of the newly generated clusters. The weighting is based on the output of the deep network (r i ,k) is completed. The expression of the above process is
[0088]
[0089]
[0090] Where P2 is the cluster merging probability, H m is the inverse of the Hastings ratio, m is the prior mean of the normal inverse wishart hyperparameter distribution, is the updated mean of the normal inverse wishart hyperparameter distribution, k is the initial precision parameter of the normal inverse wishart hyperparameter distribution, is the updated precision parameter of the normal inverse wishart hyperparameter distribution, the prior scale matrix of the covariance matrix of the y normal inverse wishart hyperparameter distribution, The update scale matrix of the covariance matrix of the normal inverse wishart hyperparameter distribution, v is the initial degrees of freedom of the normal inverse wishart hyperparameter distribution, is the updated degrees of freedom of the normal inverse wishart hyperparameter distribution
[0091] The network structure is as follows Figure 2 As shown, the network consists of two main parts. The first part is a clustering network, and the second part is K sub-clustering networks (k is the number of clusters) and is trained through the EM process in the hybrid model.
[0092] It should be understood that the update optimization of the deep clustering network based on contrastive representation learning is divided into two steps. The first step is to represent the network update optimization in the learning phase. The InfoNCE loss function expression used is
[0093]
[0094] Among them, q is the representation of the query sample, k + is the representation of the positive sample, is the representation of negative samples, t is the temperature hyperparameter, and f(g) is the similarity calculation of dot product. In the contrastive learning architecture set in this application, the backbone network uses the standard ResNet-18 network. The projection head added to the backbone network is a two-layer MLP network. The number of neurons in the first layer is 2048, and the number of neurons in the second layer is the output feature dimension 8. The ReLU function is used for activation in the output layer. The other hyperparameters of the network are set to batch size batchsize = 128 and temperature hyperparameter t = 0.2.
[0095] The second step is the network update optimization in the feature clustering stage. Two optimization methods are used. The first one is the KL divergence network optimization for the E step of the EM process in the hybrid model, and its expression is:
[0096]
[0097] Among them, r i,k is the soft assignment of feature representation, i.e., the mapping from feature representation to clustering.
[0098] The second is the network optimization of the M step of the EM process in the hybrid model with homogeneous loss, which is expressed as
[0099]
[0100] z i = arg max k (r i,k ) (19)
[0101] Among them, x i is the characteristic representation of the sample, It is the mean of sub-cluster j of cluster k. In the clustering architecture set in the present application, MLP has an input layer, a single hidden layer and an output layer. The number of neurons in the input layer is 8 (the input data dimension of the clustering module), the number of neurons in the hidden layer is 50 (can be set to other numbers with little impact on the clustering results), and the number of neurons in the output layer is k (varies with the clustering results).
[0102] The other hyperparameters of the network are set as follows: batch size batchsize = 128, clustering network learning rate lr = 0.0005, and sub-clustering network learning rate lr = 0.005.
[0103] Three parameters are used to measure network performance evaluation, namely accuracy (ACC), mutual information entropy (NMI) and modified Rand index (ARI). Accuracy (ACC) refers to the ratio of the correct match between the true label of the sample and the predicted cluster label. The calculation formula is:
[0104]
[0105] Where N is the number of samples, y i is the true label of the sample, z i is the label of the cluster assignment predicted by the algorithm, l(g) is the indicator function, and m is the number of matches between the cluster assignment label of the predicted sample and the true label.
[0106] Mutual information entropy (NMI) represents the entropy of the mutual information of the predicted clustering label and the mutual information of the true label of the sample, and its calculation formula is
[0107]
[0108] in, H(g) is the entropy function and I(g) is the mutual information function.
[0109] The modified Rand index (ARI) represents the percentage of correct assignments for each pair of samples and is calculated as
[0110]
[0111] Among them, TP represents the number of correct classifications of similar samples into one category, TN represents the number of correct classifications of different samples into different categories, FP represents the number of incorrect classifications of dissimilar samples into the same category, and FN represents the number of incorrect classifications of similar samples into different categories.
[0112] The larger the values of the above three parameters are, the more consistent the clustering results are with the actual sample classification, which means that the clustering effect of the network is more ideal.
[0113] In order to further verify the effect of the radiation source classification method based on polarization domain feature learning and extraction proposed in this disclosure, the following simulation experiments were conducted:
[0114] According to the dual-polarization antenna structure of current engineering applications, two groups of modulation signals of four main polarization and cross-polarization channels are simulated and generated in Matlab, including two types of signals: digital modulation (BPSK, 2FSK, 2ASK) and analog modulation (PM, FM, AM). They are then modulated into RF signals for transmission by the communication RF system, constructing two groups of I / Q signals used in the experiment.
[0115] Experiment 1: Classification and recognition under variable signal-to-noise ratio conditions
[0116] Since polarization characteristics are susceptible to noise pollution and aliasing, it is necessary to verify the effect of the method of this application on low signal-to-noise ratio data. Four communication radiation source individuals with polarization amplitude ratios and polarization phase differences of (1 / 10, 1.5°) (1 / 9, 3°) (1 / 7, 5°) (1 / 6, 8°) are set according to the simulation data set settings in Table 1 to generate experimental data sets. The representation learning network is trained using the generated data with a signal-to-noise ratio of 20dB, and the network parameters are fixed. Cluster analysis is performed on the data of each signal-to-noise ratio of 0-20dB.
[0117] Table 3 Simulation data set parameters
[0118]
[0119] like Figure 3 As shown, Figure 3 (a) represents the learning network training loss, Figure 3 (b) is the feature clustering network loss. Figure 3 (a) shows that the learning network gradually converges in 50 rounds of training. Overall, it can be seen that the network maintains good stability when extracting features, and the training effect is good; Figure 3 (b) The loss of the available feature clustering network changes significantly each time the clustering cluster is split, and gradually decreases and tends to be stable with subsequent training.
[0120] like Figure 4 As shown, the classification confusion matrix of four communication radiation source individuals under 0-20dB signal-to-noise ratio is shown. It can be seen from the figure that when it is above 0dB, the method of the present application can clearly classify 4 communication radiation source individuals, and as the signal-to-noise ratio increases, the classification effect improves rapidly. When the signal-to-noise ratio reaches 10dB, the classification recognition accuracy reaches 100%.
[0121] like Figure 5 As shown, Figure 5 (a) shows the classification accuracy of the four communication radiation sources at 0-20 dB. It can be seen from the figure that when the signal-to-noise ratio is 0 dB, the classification accuracy of each individual is above 80%. As the signal-to-noise ratio gradually increases, the classification accuracy gradually approaches 1. Figure 5 (b) shows the three evaluation indicators of the method of the present application. As can be seen from the figure, as the signal-to-noise ratio increases, the values of the three evaluation indicators gradually approach 1, which means that the closer the values of the three indicators are to 1, the better the classification and recognition effect of the method of the present application.
[0122] like Figure 6As shown in the figure, it can be seen that the learning network is effective in extracting features from the original high-dimensional I / Q data, and as the signal-to-noise ratio increases, the extracted features gradually stabilize and tend to the true polarization amplitude ratio and polarization phase difference features, proving the method of the present application's ability to adaptively learn and extract polarization domain features.
[0123] The above experiments as a whole verified the effectiveness and stability of the method of the present application under different signal-to-noise ratio conditions, and proved that it also has a good classification and recognition effect for individual communication radiation sources under low signal-to-noise ratio conditions.
[0124] Experiment 2: Classification and recognition under variable modulation pattern conditions
[0125] Since the modulation styles of communication radiation sources are diverse, the method of the present application is required to be unaffected by the modulation style, adaptively learn to extract polarization features and perform cluster analysis. Six communication radiation source individuals with polarization amplitude ratios and polarization phase differences of (1 / 10, 1°) (1 / 9, 2.5°) (1 / 8, 4.5°) (1 / 7, 6°) (1 / 6, 7.5°) (1 / 5, 8.5°) are modulated in BPSK, BFSK, BASK, PM, FM, and AM modulation modes in turn, and then set according to the data settings except the modulation mode in Table 1 to generate an experimental data set with a signal-to-noise ratio of -5dB-20dB, and use the representation learning network trained in Experiment 1 for feature learning and cluster analysis, and the classification recognition effect is as follows.
[0126] like Figure 7 As shown, the classification confusion matrix of six communication radiation source individuals under 5dB, 10dB and 15dB signal-to-noise ratios is shown. It can be seen from the figure that when the signal-to-noise ratio is above 5dB, the method of the present application can clearly classify 6 communication radiation source individuals, among which individual 0, individual 1, individual 2 and individual 3 have higher classification accuracy, and the classification accuracy of the remaining 2 is lower. However, as the signal-to-noise ratio increases, the classification effect improves rapidly. When the signal-to-noise ratio reaches 10dB, the overall classification and recognition accuracy reaches more than 90%, which proves that the method of the present application has a good classification and recognition effect for communication radiation source individuals under variable modulation patterns.
[0127] like Figure 8 As shown, Figure 8 (a) shows the classification accuracy of the six communication radiation sources at 0-20 dB. It can be seen from the figure that at 10 dB SNR, the classification accuracy of the six different modulation patterns is basically above 70%. As the SNR gradually increases, the classification accuracy gradually approaches 1. Figure 8(b) shows the three evaluation indicators of the method of the present application. It can be seen from the figure that with the improvement of the signal-to-noise ratio, the values of the three evaluation indicators of the communication radiation source individuals with 6 different modulation styles of the method of the present application gradually approach 1, indicating that the closer the values of the three indicators are to 1, the better the classification and recognition effect of the method of the present application.
[0128] like Fig. 9 As shown, the feature dimensionality reduction graph extracted by the representation learning network under 5dB, 10dB, and 15dB signal-to-noise ratios is displayed. It can be seen from the figure that the representation learning network is effective in extracting features from original high-dimensional I / Q data with different modulation styles, and as the signal-to-noise ratio increases, the extracted features gradually stabilize and tend to the true polarization amplitude ratio and polarization phase difference features, proving the method of the present application's ability to adaptively learn and extract polarization domain features.
[0129] The above experiments as a whole verified the effectiveness and stability of the method of the present application under different modulation patterns, and proved that it also has good classification and recognition effects for individual communication radiation sources with variable modulation patterns under lower signal-to-noise ratios.
[0130] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification.
[0131] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. This application is intended to cover any modification, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present disclosure are indicated by the appended claims.
Claims
1. A radiation source classification method based on polarization domain feature learning and extraction, characterized in that: The following steps are involved: Receiving polarization signals from different communication radiation sources, and obtaining polarization characteristic representations of the polarization signals by measuring polarization amplitude ratios and polarization phase differences of the polarization signals; Combining the original data of the polarization signal with the polarization feature representation, and obtaining a training data set and a test data set through data preprocessing; The training data set outputs features extracted by learning through representation learning phase; The features extracted by the representation learning and the test data set complete the classification of communication radiation sources through the feature clustering stage.
2. The radiation source classification method based on polarization domain feature learning and extraction according to claim 1 is characterized in that: The data preprocessing comprises the following steps: Splitting the received radio frequency I / Q complex valued data of the main polarization and cross polarization signals into two groups of I-channel and Q-channel real valued data; The amplitude and phase of the main polarization and cross-polarization signals are calculated using two sets of corresponding I-channel and Q-channel data, and then the polarization amplitude ratio and polarization phase difference of the main polarization and cross-polarization signals are calculated using the two sets of amplitude and phase values obtained; The polarization amplitude ratio and polarization phase difference in the main polarization and cross-polarization signals are aligned and then spliced and cut with the original four-way I / Q data to obtain M×6×128-dimensional training and test data sets.
3. The radiation source classification method based on polarization domain feature learning and extraction according to claim 1 is characterized in that: A contrastive learning model is used in the representation learning stage, and the contrastive learning model is composed of data enhancement, a basic encoding network, a projection head and a contrast loss.
4. The radiation source classification method based on polarization domain feature learning and extraction according to claim 3 is characterized in that: In the feature clustering stage, a subcluster is created for each cluster, and the number of clusters is changed by splitting clusters and merging clusters.
5. The radiation source classification method based on polarization domain feature learning and extraction according to claim 4 is characterized in that: When splitting the cluster, each cluster is split into two sub-clusters with probability P1. At the same time, the k value changes adaptively. The cluster will copy the kth unit in the last layer of the clustering network and the weight connected to the previous hidden layer, and use the parameters learned by the sub-clustering network to initialize two new clusters. The expression for splitting the cluster includes: P1=Min(1,H s ) (1) Where P1 is the cluster split probability, H s The Hastings ratio is the ratio of the marginal likelihood of the data under the two sub-clusters to the marginal likelihood under the cluster, Γ is the gamma function, represents the data in cluster K, N k =|χ k |, Represents the data N in sub-cluster j (j∈{1,2}) k,j =|χ k,j |, f x (·,λ) is the marginal likelihood function, and λ=(m,κ,ψ,v) is the normal inverse wishart hyperparameter.
6. The radiation source classification method based on polarization domain feature learning and extraction according to claim 4 is characterized in that: When merging clusters, each sub-cluster is merged with its three nearest sub-clusters with probability P2 each time it is aggregated. After merging, a new sub-cluster network is initialized, and the last layer unit of the merged cluster and the weights of its connection with the previous hidden layer are removed from the cluster network. The weighted maximum a posteriori estimation is used to initialize the parameters and weights of the newly generated cluster. The weighting is based on the output of the deep network (r i ,k) completed; Expressions for merging clusters include: Where P2 is the probability of cluster merging, H m is the inverse of the Hastings ratio, m is the prior mean of the normal inverse wishart hyperparameter distribution, is the updated mean of the normal inverse wishart hyperparameter distribution, κ is the initial precision parameter of the normal inverse wishart hyperparameter distribution, is the updated precision parameter of the normal inverse wishart hyperparameter distribution, ψ is the prior scale matrix of the covariance matrix of the normal inverse wishart hyperparameter distribution, The update scale matrix of the covariance matrix of the normal inverse wishart hyperparameter distribution, v is the initial degrees of freedom of the normal inverse wishart hyperparameter distribution, is the updated degrees of freedom for the normal inverse wishart hyperparameter distribution.
7. The radiation source classification method based on polarization domain feature learning and extraction according to claim 1 is characterized in that: The representation learning phase is subjected to a first network update optimization, wherein the expression of the first network update optimization includes: Among them, q is the representation of the query sample, k + is the representation of the positive sample, is the representation of negative samples, τ is the temperature hyperparameter, f(·) represents the similarity calculation of dot products, and in the contrastive learning architecture, the backbone network uses the standard ResNet-18 network. The projection head added to the backbone network is a two-layer MLP network with 2048 neurons in the first layer and 8 neurons in the second layer for the output feature dimension. The ReLU function is used for activation in the output layer. The other hyperparameters of the network are set to batch size batchsize = 128 and temperature hyperparameter τ = 0.
2.
8. The radiation source classification method based on polarization domain feature learning and extraction according to claim 1 is characterized in that: The feature clustering stage is subjected to a second network update optimization, and the expression of the second network update optimization includes: Among them, r i,k is the soft assignment of feature representation, i.e., the mapping from feature representation to clustering.
9. The radiation source classification method based on polarization domain feature learning and extraction according to claim 8 is characterized in that: The third network update optimization is performed on the feature clustering stage, and the third network update optimization expression includes: z i =argmax k (r i,k ) (13) Among them, x i is the characteristic representation of the sample, is the mean of subcluster j of cluster k.
