Radiation source signal identification method based on feature matching and mapping learning

Abstract

The invention provides a radiation source signal identification method based on feature matching and mapping learning, and relates to the technical field of mode identification, intelligent information processing and signal identification. In a training stage, firstly, source domain and target domain signals are extracted from a collected radiation source signal full pulse sequence to construct a radiation source signal data set, and time-frequency characteristics Vsf and Vtf are extracted from source domain signal Vs and target domain signal Vt data in the data set respectively; and finally, inputting the P into a signal identification model, learning consistent representation of the P, and realizing a radiation source signal identification task. When a radiation source signal is identified, whether input signal data is empty or not is firstly judged, if the input signal data is empty, the whole process is terminated, if the input signal data is not empty, the input signal is preprocessed, then time-frequency characteristics of the preprocessed signal are extracted, and finally a category prediction result is given through a signal identification model. And jumping to radiation source signal input after the current input signal identification is completed, and ending the whole process until the input data is empty.

Description

Technical Field

[0001] The present invention relates to the technical fields of pattern recognition, intelligent information processing and signal recognition. Background Art

[0002] The identification of radiation source signals is the basis and guarantee of electronic countermeasures. During the use of the transmitter, due to some external and internal factors, such as electromagnetic interference, high temperature, humidity and hardware aging, the data distribution of the transmitter signal will change over time. This phenomenon can be regarded as domain shift, that is, the data distribution shifts from the source domain (data of the previous period) to the target domain (data of the next period), which usually leads to a significant decrease in the individual identification performance of the transmitter. Researching radiation source signal identification methods based on feature matching and mapping learning to improve the accuracy of target domain signal identification is of great significance to the realization of electronic countermeasures.

[0003] At present, most of the research on radiation source signal recognition is concentrated in the same domain, while there are relatively few cross-domain studies that consider domain offset. In the same domain, the signal data distribution and fingerprint features are relatively stable, and the radiation source signal is usually identified based on the extracted fingerprint features. Wang Huanhuan et al. proposed a communication signal feature extraction method based on wavelet theory by analyzing the correlation between the nonlinear harmonics and spectrum of the radiation source signal and published it at IAEAC in 2017. This method extracts wavelet entropy as a feature and identifies the communication signal based on support vector machine (SVM). Satija et al. proposed a transmitter identification method based on variational mode decomposition and spectral features (VMD-SF) and published it in IEEE Trans. Inf. Forensics Secur in 2018. The received signal is decomposed into various time and spectrum modes by empirical mode decomposition (EMD), and the recognition accuracy is improved by selecting different spectrum features such as spectrum flatness and spectrum brightness. Roy et al. proposed an adversarial learning method called Rfal in the IEEE Trans. Cognit. Commun. Networking journal in 2019 for the identification of radio frequency transmitters. This method trains a generative adversarial network (GAN) to enable it to learn effective feature representations from radio frequency signals, thereby improving recognition accuracy. Chen Haiyong and others used the fact that afterglow images contain more information to transform the signal recognition problem into a target detection problem, achieving the task of identifying specific radiation source signals and published the results in the Modern Radar Journal in 2023.

[0004] Most of the research on cross-domain identification of radiator signals is based on transfer learning, which transfers signal knowledge in one scenario to another scenario for identification. In the 2021 B Po1 Acad Sci-tech journal, Huang Keju et al. proposed a method based on continuous wavelet transform (CWT) combined with domain adversarial neural network (DANN) to improve the performance of radiator signal identification at different frequencies. This method uses labeled training data and unlabeled test data for training to learn the discriminative features of each transmitter at different frequencies. When the carrier frequencies of the training data and the test data are different, the proposed method has superior performance. Wang Jian et al. proposed a radiator signal identification method based on deep adversarial domain adaptation (DADA) at the 2021 IICSP conference. DADA integrates deep neural networks and adversarial methods into the domain adaptation problem of transfer learning. It trains the feature extractor by maximizing the discriminant loss to extract high-quality deep RF fingerprints under weak interference conditions. Finally, by minimizing the classifier loss, accurate classification and identification of unlabeled transmitters are achieved. All of the above methods are based on the DANN model, which can transfer knowledge across different frequencies and channels, thereby significantly improving performance. However, with the complexity of the electromagnetic environment and the significant increase in the types of transmitters, these methods can only achieve domain-level mapping alignment in cross-domain scenarios, and they cannot map and align individuals in the target domain with individuals in the source domain. Therefore, domain mapping alignment by maximizing the discrimination loss between the source and target domains is difficult to adapt to situations with multiple types of transmitters. How to map individuals in the target domain and individuals in the source domain to the same domain is the key to solving the problem of cross-domain radiation source signal recognition in complex electromagnetic environments. Summary of the invention

[0005] The purpose of the present invention is to provide a radiation source signal recognition method based on feature matching and mapping learning, which can effectively solve the technical problem of cross-domain radiation source signal recognition in complex electromagnetic environments.

[0006] The object of the present invention is achieved by the following steps:

[0007] Step 1: Construction of radiation source signal dataset

[0008] First, the single pulse signal of each radiation source is extracted from the full pulse sequence of the collected radiation source signal according to the pulse address and its length; then the source domain signal of the previous period and the target domain signal of the next period are used to construct data sets respectively, where the source domain signal has a corresponding category label and the target domain signal has no category label; finally, the source domain signal and the target domain signal are processed by removing outliers, removing the mean and standardizing, so that the source domain signal and the target domain signal data have the same length;

[0009] Step 2: Signal feature extraction and feature set construction

[0010] In feature extraction, the input radiation source signal is first decomposed by EMD, and then the first component after decomposition is input into the wavelet scattering network to extract the time-frequency characteristics of the radiation source signal; the wavelet scattering network is set to three layers, and for the signal x(t), its scattering coefficient characteristic calculation is as follows:

[0011] Zero-order scattering coefficient S 0 x(t):

[0012] S 0 x(t)=x(t)*φ t

[0013] First-order scattering coefficient S 1 x(t):

[0014]

[0015] Second-order scattering coefficient S 2 x(t):

[0016]

[0017] The third-order scattering coefficient S 3 x(t):

[0018]

[0019] in, The center frequency is λ i The wavelet filter bank, φ t It represents the operation of averaging the values ​​in a window with a time length of t; after the radiation source signal data is decomposed by EMD, the first component is passed through the wavelet scattering network to generate the scattering coefficient characteristic matrix [S 0 x(t), S 1 x(t), S 2 x(t), S 3 x(t)]; In order to better perform feature matching, each column in the wavelet scattering coefficient matrix is ​​spliced ​​together in order, which represents the characteristic information combination of the input signal at different scales and frequencies;

[0020] Step 3: Source domain classifier construction and training

[0021] Source domain signal data Where N s is the amount of source domain signal data, V s Extract its time-frequency features through step 2 V sf Used to train the source domain classifier and obtain weights Among them, the source domain classifier is composed of ResNet-18 network, f represents features, and s represents source domain;

[0022] Step 4: Feature matching set construction

[0023] First, the target domain signal data Extract its time-frequency features through step 2 Where N t is the target domain signal data volume; then based on the distance measurement module, the target domain signal feature V tf Match the source domain signal feature V with its nearest neighbor sf Forming characteristic pair P ST , P ST The calculation formula is as follows:

[0024]

[0025] in, yes and The Wasserstein distance between is the label corresponding to the source domain signal data, and the target domain signal label generated by the distance measurement module is recorded as Y s In order to improve the accuracy of feature matching, source domain knowledge is introduced, that is, the target domain signal is obtained through the source domain classifier to obtain its predicted label The feature pair P is formed by the target domain signal feature and its predicted label T , P T The calculation formula is as follows:

[0026]

[0027] Finally, the target domain signal label Y generated by the distance metric module is s The labels based on the output of the source domain classifier Compare them, if they are consistent, retain the generated feature pair P, that is in is the matched source domain signal data, is the matched target domain signal data, for The corresponding label, N tr is the amount of source domain signal and target domain signal data after matching;

[0028] Step 5: Signal recognition model construction and training

[0029] The signal recognition model consists of a domain mapping module G, a domain distribution calculation module D, and a domain adaptation classifier C; the domain mapping module G consists of a multi-layer perceptron; the domain distribution calculation module D consists of a one-dimensional convolution, an activation function, and batch normalization, and the domain adaptation classifier C consists of ResNet-18;

[0030] In the signal recognition model, first, based on the domain mapping module G, V in the feature pair P is transformed into str and V ttr Mapping is performed to generate source domain signal mapping features g s and target domain signal mapping feature g t Then, in the domain distribution calculation module D, the source domain signal mapping feature g is calculated s and target domain signal mapping feature g t The Wasserstein distance loss between d , by minimizing the distance loss l d Reduce the source domain signal mapping feature g s and target domain signal mapping feature g t The distance between individual features of the same category in the domain adaptation classifier C is minimized by the classification cross entropy loss l c Learning to obtain source domain signal mapping features g s The difference between the individual features of different categories in the training process; by jointly training the distance loss l d and categorical cross entropy loss l c , so that the source domain signal mapping feature g generated by the domain mapping module G s and target domain signal mapping feature g t Domain alignment is achieved in the feature space, and their classification information is learned in the domain adaptation classifier C, with a distance loss l d and categorical cross entropy loss l c The calculation of is as follows:

[0031]

[0032] Among them, N c represents the number of source domain signal categories, express Belongs to category n c probability;

[0033] Step 6: Radiation source signal input

[0034] Extracting the single pulse signal of each radiation source from the full pulse sequence of the radiation source signal collected by the signal receiver according to the pulse address and its length as the signal input to be identified; if the input signal is empty, the whole process is terminated, otherwise, jump to step seven;

[0035] Step 7: Signal preprocessing

[0036] The input signal is detected and removed from outliers based on the Hampel filter method, and then the original data of each dimension of the signal is subtracted from the average value of each dimension of the signal to achieve the purpose of removing the mean value; finally, the signal data after removing outliers and the mean value is standardized to make the signal length consistent;

[0037] Step 8: Signal feature extraction

[0038] Signal feature extraction uses the same feature extraction method as step 2, and uses wavelet scattering features as signal features to be identified; first, the preprocessed signal is decomposed by EMD to obtain multiple components, then the wavelet scattering features are extracted from the first component after decomposition, and finally each column in the wavelet scattering coefficient matrix is ​​spliced ​​together in order as the time-frequency features of the input signal;

[0039] Step 9: Signal Identification

[0040] The signal time-frequency features are input into the signal recognition model obtained by training in step five to infer the recognition results; in the signal recognition model, the input signal features are first mapped to the same feature space as the source domain through the domain mapping module G, and then the prediction results are output through the last layer of the domain adaptation classifier C network. After completing the current input signal recognition process, jump to step six.

[0041] The present invention proposes a radiation source signal recognition method of feature matching and mapping learning, which enables source domain signal data and target domain signal data to achieve domain alignment in feature space, and can effectively improve the recognition accuracy of target domain signals. The feature matching proposed by the method provides the signal recognition model with source domain and spatial distance knowledge, thereby effectively guiding the individual alignment in the source domain and the target domain, and has significant advantages in complex electromagnetic environments where there are multiple types of radiation source signals. The introduction of Wasserstein distance in the signal recognition model avoids the collapse of the model during training and accelerates the convergence of the model. The method can be used not only for the recognition of radiation source signals, but also for the recognition of signal modulation methods, and can also be modified to realize the recognition of unknown radiation source signals. In addition, the method can be widely applied to various types of radiation source signals, including radar radiation source signals, communication radiation source signals, etc., and the data dimension will not limit the application of the method.

[0042] The advantages and effects of the radiation source signal recognition method of feature matching and mapping learning proposed in the present invention compared with the prior art. The method first extracts the time-frequency features of the signal through a wavelet scattering network, then generates feature pairs that fuse source domain knowledge and spatial distance knowledge based on feature matching, and finally uses the feature pairs to train the signal recognition model, learns to obtain consistent representations of signal features, and realizes the radiation source signal recognition task. In addition, the method of the present invention can process different types of signals. In the actual application process, it only needs to modify the input form according to the specific situation and appropriately configure the relevant parameters to realize the radiation source signal recognition task, and has a strong scene adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 The feature matching diagram of the method of the present invention

[0044] Figure 2 The signal recognition model diagram of the method of the present invention is

[0045] Figure 3 The technical flow chart of the method of the present invention is Example

[0046] The technical flow chart of the method of the present invention is as follows Figure 3 In the training phase, we first extract the source domain and target domain signals from the full pulse sequence of the collected radiation source signal to construct a radiation source signal dataset. s and the target domain signal V t The time-frequency features V are extracted from the data sf and V tf Then set V sf Used to train the source domain classifier to learn and obtain its weight Then, based on feature matching, V sf and V tf Generate feature pairs P, where P is composed of source domain training features V str , target domain training features V ttr and source domain training labels y str Finally, P is input into the signal recognition model to learn their consistent representations and realize the radiation source signal recognition task. When identifying the radiation source signal, first determine whether the input signal data is empty. If it is empty, the whole process is terminated. If it is not empty, the input signal is preprocessed, and then the time-frequency features of the preprocessed signal are extracted. Finally, the category prediction result is given through the signal recognition model. After completing the recognition of the current input signal, jump to the radiation source signal input until the input data is empty and the whole process ends.

[0047] Taking the radar radiation source signal as an example, the method proposed in the present invention can be used to improve the accuracy of radar radiation source signal recognition in the presence of signal characteristic offset. Specifically, firstly, the wavelet scattering features of the source domain and target domain signals of the radar radiation source are extracted respectively, and then the source domain classifier is trained based on the source domain signal features, and then the source domain features and the target domain features are matched to generate feature pairs, and the signal recognition model is trained based on the feature pairs. Finally, the signal to be recognized is preprocessed and feature extracted, and the obtained signal features are used to give the predicted category through the domain mapping module and domain adaptation classifier in the signal recognition model.

[0048] The method of the present invention can also be used in other application occasions, such as individual identification of communication radiation source signals and modulation mode identification.

[0049] The method of the present invention can be implemented by programming in any computer programming language (such as Python language), and the signal recognition model based on the method can be implemented in any PC or embedded system.

Claims

1. A method for identifying radiation source signals by feature matching and mapping learning, characterized in that: The method is implemented by the following steps: Step 1: Constructing the radiation source signal dataset; First, the single pulse signal of each radiation source is extracted from the full pulse sequence of the collected radiation source signal according to the pulse address and its length; then the source domain signal of the previous period and the target domain signal of the next period are used to construct data sets respectively, where the source domain signal has a corresponding category label and the target domain signal has no category label; finally, the source domain signal and the target domain signal are processed by removing outliers, removing the mean and standardizing, so that the source domain signal and the target domain signal data have the same length; Step 2: Signal feature extraction and feature set construction; In feature extraction, the input radiation source signal is first decomposed by EMD, and then the first component after decomposition is input into the wavelet scattering network to extract the time-frequency characteristics of the radiation source signal; the wavelet scattering network is set to three layers, and for the signal x(t), its scattering coefficient characteristic calculation is as follows: Zero-order scattering coefficient S0x(t): S0x(t)=x(t)*φ t First-order scattering coefficient S1x(t): Second-order scattering coefficient S2x(t): Third-order scattering coefficient S3x(t): in, The center frequency is λ i The wavelet filter bank, φ t It represents the operation of averaging the values ​​in a window with a time length of t. After the radiation source signal data is decomposed by EMD, the first component is passed through the wavelet scattering network to generate a scattering coefficient feature matrix [S0x(t), S1x(t), S2x(t), S3x(t)]. In order to better perform feature matching, each column in the wavelet scattering coefficient matrix is ​​spliced ​​together in sequence, which represents the characteristic information combination of the input signal at different scales and frequencies. Step 3: Source domain classifier construction and training; Source domain signal data Where N s is the amount of source domain signal data, V s Extract its time-frequency features through step 2 V sf Used to train the source domain classifier and obtain weights Among them, the source domain classifier is composed of ResNet-18 network, f represents features, and s represents source domain; Step 4: Construct feature matching set; First, the target domain signal data Extract its time-frequency features through step 2 Where N t is the target domain signal data volume; then based on the distance measurement module, the target domain signal feature V tf Match the source domain signal feature V with its nearest neighbor sf Forming characteristic pair P ST , P ST The calculation formula is as follows: in, yes and The Wasserstein distance between is the label corresponding to the source domain signal data, and the target domain signal label generated by the distance measurement module is recorded as Y s In order to improve the accuracy of feature matching, source domain knowledge is introduced, that is, the target domain signal is obtained through the source domain classifier to obtain its predicted label The feature pair P is formed by the target domain signal feature and its predicted label T , P T The calculation formula is as follows: Finally, the target domain signal label Y generated by the distance metric module is s Based on the labels output by the source domain classifier Compare them, if they are consistent, retain the generated feature pair P, that is in is the source domain signal data after matching, is the matched target domain signal data, for The corresponding label, N tr is the amount of source domain signal and target domain signal data after matching; Step 5: Signal recognition model construction and training; The signal recognition model consists of a domain mapping module G, a domain distribution calculation module D, and a domain adaptation classifier C; the domain mapping module G consists of a multi-layer perceptron; the domain distribution calculation module D consists of a one-dimensional convolution, an activation function, and batch normalization, and the domain adaptation classifier C consists of ResNet-18; In the signal recognition model, first, based on the domain mapping module G, V in the feature pair P is transformed into str and V ttr Mapping is performed to generate source domain signal mapping features g s and target domain signal mapping feature g t ; Then, in the domain distribution calculation module D, the source domain signal mapping feature g is calculated s and target domain signal mapping feature g t The Wasserstein distance loss between d , by minimizing the distance loss l d Reduce the source domain signal mapping feature g s and target domain signal mapping feature g t The distance between individual features of the same category in the domain adaptation classifier C is minimized by the classification cross entropy loss l c Learning to obtain source domain signal mapping features g s The difference between the individual features of different categories in the training process; by jointly training the distance loss l d and categorical cross entropy loss l c , so that the source domain signal mapping feature g generated by the domain mapping module G s and target domain signal mapping feature g t Domain alignment is achieved in the feature space, and their classification information is learned in the domain adaptation classifier C, with a distance loss l d and categorical cross entropy loss l c The calculation of is as follows: Among them, N c represents the number of source domain signal categories, express Belongs to category n c probability; Step 6: Input radiation source signal; Extracting the single pulse signal of each radiation source from the full pulse sequence of the radiation source signal collected by the signal receiver according to the pulse address and its length as the signal input to be identified; if the input signal is empty, the whole process is terminated, otherwise, jump to step seven; Step 7: Signal preprocessing; The input signal is detected and removed from outliers based on the Hampel filter method, and then the original data of each dimension of the signal is subtracted from the average value of each dimension of the signal to achieve the purpose of removing the mean value; finally, the signal data after removing outliers and the mean value is standardized to make the signal length consistent; Step 8: Signal feature extraction; Signal feature extraction uses the same feature extraction method as step 2, and uses wavelet scattering features as signal features to be identified; first, the preprocessed signal is decomposed by EMD to obtain multiple components, then the wavelet scattering features are extracted from the first component after decomposition, and finally each column in the wavelet scattering coefficient matrix is ​​spliced ​​together in order as the time-frequency features of the input signal; Step nine: signal identification; The signal time-frequency features are input into the signal recognition model obtained by training in step five to infer the recognition results; in the signal recognition model, the input signal features are first mapped to the same feature space as the source domain through the domain mapping module G, and then the prediction results are output through the last layer of the domain adaptation classifier C network. After completing the current input signal recognition process, jump to step six.

Images