Radar radiation source model identification method based on multivariable time sequence Transform

Abstract

The invention discloses a radar radiation source model identification method based on multivariable time sequence Transform, and relates to radar radiation source identification. Comprising the following steps: 1) constructing a self-supervised contrast learning architecture by using PDW and IF data of the same radar signal, zooming in common features of the PDW and IF data, zooming out common features and specific features of the same domain, and performing cross-domain translation before inputting a model to ensure that a domain missing problem can be handled; 2) constructing a multivariable time sequence Transform model for processing the multivariable time sequence data, not modifying any structure of the Transform, and only redefining the function of each module, that is, learning the relationship between variables through an attention mechanism, and learning the long-time relationship through layer normalization and a feedforward network; data of different domains of radar signals are fully utilized, the degree of dependence of the model on the data is reduced, and the robustness of the model is improved; and the model identification effect of the model on the non-cooperative radar radiation source with limited data is improved.

Description

Technical Field

[0001] The invention belongs to the field of radar radiation source identification, and specifically relates to a radar radiation source model identification method based on a multivariate time series Transformer. Background Art

[0002] Radar emitter model identification refers to the process of intercepting radar signals through reconnaissance systems, analyzing their operating parameters and signal characteristic parameters, and classifying them into corresponding radar models. Non-cooperative emitters refer to radars or other electronic devices that do not work in conjunction with the identification system. They usually belong to the enemy or unknown parties. However, since the modulation method of non-cooperative radar emitters is more complex and the signal is difficult to obtain, it is more challenging to identify their models.

[0003] Most of the current radar emitter model recognition models are still based on single-domain data (Liu Fuyue. Research on Passive Radar Identification and Sorting Technology [D]. Harbin Engineering University, 2022). However, this method of extracting only single-domain parameters cannot fully describe the information contained in the radar signal, resulting in limited performance in related tasks. A few models (Zhang Z, Shi X, Guo X, et al. Tr-ragcn-aff-ress: A method for radar emittersignal sorting [J]. Remote Sensing, 2024, 16(7): 1121; Jing Bojun. Research on Radar Emitter Identification Technology Based on Deep Learning [D]. Xi'an: Xi'an University of Electronic Science and Technology, 2017) extract parameters from different domains separately and treat them as a sample. However, this scheme does not utilize the common information between parameters in different domains, so its performance still has room for improvement. The pulse descriptive word (PDW) data of the radar signal is composed of extracting some parameters from the intermediate frequency (IF) and adding parameters such as position. Therefore, when conducting comparative learning on PDW data and IF data, it is necessary to capture not only the common information in each domain, but also the unique information between different domains, so as to generate richer and meaningful feature representations and make the feature representations more discriminative and generalizable.

[0004] Traditional Transformerm models usually encode variables at the same timestamp in multivariate time series data into a token, and use the attention mechanism to model the time series correlation of different timestamps. However, different variables at the same timestamp in radar signals do not correspond exactly to the timestamps, but have different degrees of time lag. This type of data is multivariate time lag data, which is a type of multivariate time series data with time lag. The processing method of the traditional Transformer model ignores the time lag between variables, and also causes the correlation between variables to be eliminated. In addition, as the length of the historical window increases, it will also lead to an explosion in the amount of model calculation.

[0005] Based on this, the present invention proposes a Multivariate Time Series Transformer (MTS-Trans) model for PDW and IF data, and constructs a special contrastive learning architecture based on PDW and IF data to bring the common information of different domains closer and the common information and unique information of the same domain farther apart. Without changing the network architecture of the Transformer, the role of the attention mechanism and the feedforward network are transformed. The model regards a time-varying signal of each variable as an independent token, so that the attention mechanism models the correlation of the change of the variable over a period of time, and the feedforward network models the temporal correlation of the variable. Summary of the invention

[0006] The purpose of the present invention is to improve the radar emitter model recognition effect from the idea of ​​improving the representation learning ability, improve its model robustness and scene generalization ability, especially for non-cooperative radar emitters that are limited by the difficulty in obtaining data. By constructing a suitable self-supervised pre-training method for pulse descriptor words (PDW) and intermediate frequency (IF) data, the characteristics of radar signals are learned from a large amount of data from cooperative radar emitters, and then fine-tuned on a small amount of data from non-cooperative radar emitters, finally obtaining good generalization performance. In addition, compatible information mining means are constructed for PDW and IF data, and a radar emitter model recognition method based on multivariate time series Transformer is proposed.

[0007] In order to achieve the above-mentioned object of the invention, the present invention provides the following technical solutions.

[0008] A radar emitter model identification method based on multivariate time series Transformer includes the following steps:

[0009] 1) Construct a comparative self-supervised learning architecture based on pulse descriptors and IF data. Use the pulse descriptors and IF data of the same radar signal to construct a self-supervised comparative learning architecture, bring the common features of pulse descriptors and IF data closer, and pull the common features and unique features of the same domain further apart, so as to effectively utilize the information in pulse descriptors and IF data. Before inputting the model, cross-domain translation is also performed to ensure that the domain missing problem can be handled.

[0010] 2) Construct a multivariate time series Transformer model that is compatible with pulse descriptors and intermediate frequency data: By modifying the way Tokens are constructed, the roles of the attention mechanism and the feedforward network are transformed so that the attention mechanism models the correlation between time series variables, and the long-term relationships are learned through layer normalization and the feedforward network, and the feedforward network models the temporal correlation of variables.

[0011] In step 1), the contrastive self-supervised learning architecture based on pulse descriptors and intermediate frequency data separates the feature embedding of each domain into a domain-specific part and a domain-common part to retain domain-specific information.

[0012] In step 1), the construction of a comparative self-supervised learning architecture based on pulse description words and intermediate frequency data includes:

[0013] (1) Input coding: Encode inputs from different domains and process the pulse descriptor data and intermediate frequency data of radar signals in different ways. Unify the dimensions of the intermediate frequency data through discrete cosine transform (DCT), and increase the dimension of the pulse descriptor data through a fully connected layer with ReLU nonlinearity for subsequent representation learning.

[0014] (2) Cross-domain translation: First, use the module composed of MLP to unify the representations of different domains into the same dimension. At the same time, cross-domain translation is achieved through MLP, and the out-of-domain translation loss is constructed to solve the problem of the model missing in the generalization time domain.

[0015] (3) Feature segmentation: The feature embedding of each domain is split into a domain-specific part and a domain-common part through a multivariate temporal transformer (MTS-Trans), ensuring that the complementary information of different domains is preserved while constructing a dissimilarity loss.

[0016] In feature segmentation, the dimensions of domain-common features and domain-specific features are consistent.

[0017] (4) Construct a loss function to perform comparative learning on the common features of the domains and construct a similarity loss. The final loss is the weighted sum of the cross-domain translation loss, dissimilarity loss, and similarity loss.

[0018] In step 1), the construction is based on a comparative self-supervised learning architecture of pulse descriptors and intermediate frequency data. The feature extractor of radar signals is obtained by comparative self-supervised learning of pulse descriptors and intermediate frequency data of a large number of unlabeled cooperative radar emitter data. The final pre-training loss is composed of cross-domain loss, common feature similarity loss and dissimilarity between unique features and common features:

[0019]

[0020] in, and is a hyperparameter that can control the relative importance in contrastive learning; is the cross-domain translation loss; is a similar loss; is the dissimilarity loss; and then a small amount of labeled non-cooperative radar emitter data is used for transfer learning to obtain good generalization performance.

[0021] In the self-supervised learning stage, a cross-domain translation module is introduced:

[0022]

[0023]

[0024] The problem of missing the model in the generalization time domain is solved by translating the data from one domain to another to ensure that the model input has both the representation of the pulse descriptor data and the representation of the intermediate frequency data.

[0025] In step 2), the multivariate time series Transformer model compatible with pulse description words and intermediate frequency data adopts the pure encoder architecture of Transformer, including an embedding layer, The Transformer block consists of a self-attention layer, a layer normalization layer, and a feed-forward network. The output of the projection layer is the final output. The projection layer, like the embedding layer, is composed of a single-layer MLP.

[0026] The multivariate time series Transformer treats a variable over a period of time as an independent process. By extracting representations for each period of time, the self-attention module uses linear projection to obtain the query ( ),key( ) and value ( ), ,in is the projection dimension. Representing a unique query and key of a Token, each attention matrix element of the pre-Softmax score is expressed as ; Since each Token is normalized on its variable dimension before, the Toke construction method implies the variability of variables over time, and the attention mechanism mainly focuses on the correlation between variables that change over a period of time. Shows the multivariate correlation between paired variable tokens. The feature transfer process in the multivariate time series Transformer is:

[0027]

[0028] in, Indicates The representation of the layer, represents the self-attention layer.

[0029] In a multivariate time series Transformer, layer normalization operates on a sequence of a single variable over time. , the formula is as follows:

[0030]

[0031] in, Representatives include The hidden layer representation of a single variable time series with values, express No. values, express The mean of express The variance of .

[0032] In the multivariate time series Transformer, the Token vector is formed by a radar signal of the same variable. Based on the universal representation theorem of the multi-layer perceptron, it has a large enough model capacity to extract the time features shared in historical observations and future predictions. The feedforward network acts on a sequence and can extract the intrinsic properties of the sequence, such as amplitude, periodicity, frequency spectrum, etc., thereby improving the generalization on other sequences.

[0033] Compared with the prior art, the outstanding advantages and technical effects of the present invention include:

[0034] The present invention introduces information from different domains into the model by introducing PDW and IF data. A new comparative learning method is designed for PDW data and IF data to bring their common features closer and pull the common features and unique features of the same domain further apart. Before inputting the model, cross-domain translation is also performed to ensure that the domain missing problem can be handled. In addition, in order to mine the long-term relationship of the data, by constructing Tokens along the time-varying direction of each parameter, without modifying any structure of the Transformer model, limited Tokens can better capture the time dependency of parameters in the PDW data, overcoming the deficiency of the classic Transformer solution that it is difficult to capture long-term dependencies due to the limitation of Token length. Due to the modification of the Token construction method, the functions of the functional components of the Transformer are changed, that is, the self-attention module focuses on the correlation between parameters, while the fully connected layer focuses on the time correlation of parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 Schematic diagram of the self-supervised pre-training architecture.

[0036] Figure 2 Schematic diagram of the MTS-Trans model.

[0037] Figure 3 It is the self-attention layer mechanism of the MTS-Trans model.

[0038] Figure 4 It is the feed-forward network of the MTS-Trans model. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solution and advantages of the present invention more clear, the embodiments of the present invention are further described in detail below with reference to the accompanying drawings:

[0040] like Figure 1 and 2 The embodiment of the present invention includes a self-supervised pre-training architecture for PDW data and IF data, and a multivariate time series Transformer model compatible with PDW and IF data included therein, and the specific steps are as follows:

[0041] 1) Construct a self-supervised pre-training architecture for PDW data and IF data: This architecture compares PDW data and IF data to narrow the distance between their shared information and distance between shared information and unique information in the same domain, thereby effectively utilizing the information in PDW data and IF data; use this self-supervised pre-training architecture to process PDW data and IF data to obtain processed data;

[0042] 2) Construct a multivariate time series Transformer model compatible with PDW and IF data: Input the data processed in step 1) into the multivariate time series Transformer model for further analysis and processing. By modifying the way Token is constructed and transforming the roles of the attention mechanism and the feedforward network, the attention mechanism models the correlation between time series variables, and the feedforward network models the temporal correlation of variables.

[0043] In step 1), MTS-Trans uses PDW data and IF data for comparative learning. The framework is as follows: Figure 1 As shown in the figure, it is divided into four parts: (1) Input encoding, encoding inputs from different domains for subsequent representation learning; (2) Cross-domain translation, unifying the representations of different domains into the same dimension, and constructing a domain translation loss; (3) Feature segmentation, splitting the feature embedding of each domain into a domain-specific part and a domain-common part through MTS-Trans, ensuring that the complementary information of different domains is preserved, and constructing a dissimilar loss; (4) Constructing a loss function, performing comparative learning on the common features of the domains, and constructing a similarity loss. The weighted sum of the above losses is the final loss.

[0044] The data formats of the radar signal's PDW and IF data are quite different, so different processing methods should be adopted in the input encoding stage. In the example, the dimensions of each sample are not uniform and the span is large, so it is necessary to unify the IF data of different dimensions to the same specified dimension. Discrete Cosine Transform (DCT) is a commonly used signal processing tool. It is a linear reversible function. ,in is the set of real numbers, or equivalently, a DCT is based on the following formula real number Transform to another real number Operation:

[0045]

[0046] So far, The main information is compressed into In the front dimension, whether it is Pad with zeros or cut off The later dimensions will not lose much information. Then, we can fill in zeros or trim the dimensions of different Unify to the same dimensions The above steps correspond to Figure 1 In , and finally get the same dimension , and the amount of information contained does not change much after the dimension changes.

[0047] PDW data for radar signals For , its dimension is generally single digit, and needs to be increased to be close to the dimension of IF data. The present invention uses a separate fully connected layer with ReLU nonlinearity for each variable, thereby projecting the one-dimensional input to dimensional space. The above steps correspond to Figure 1 In , and finally get the upgraded dimension .

[0048] In the process of cross-domain translation, we first use the MLP , the representation of different domains Representations unified to the same dimension Then, MLP is used to achieve cross-domain translation, using the implicit relationship and approximate mapping between the two domains in the same radar signal to obtain feature embedding .For example, Means through Embedding PDW data Mapping to embedding of IF data To ensure that the translated embedding is a meaningful representation of the IF data. True embedding of distance and IF data To minimize, the cross-domain translation loss is defined as:

[0049]

[0050] Embedding IF data Mapping to embedding of PDW data Similar to:

[0051]

[0052] Finally, the total cross-domain translation loss is:

[0053]

[0054] In addition, cross-domain translation can also solve the problem of missing domains in the generalization of the model. By translating data from one domain to another, it ensures that the model input has both PDW data representation and IF data representation.

[0055] This paper proposes a novel contrastive learning model, which separates the feature embedding of each domain into a domain-specific part and a domain-common part. For example, given an embedding , the present invention expresses it as ,in is a domain-specific feature, is a domain-wide feature, and and dimensions. and To distinguish them as much as possible, we need to construct different loss functions. First, we construct contrastive learning only for the common features of the domains. As close as possible, so as to maintain the consistency of domain common features. You need to maintain uniqueness, which requires domain-specific features of the same domain Common features with domains As far away as possible.

[0056] The present invention hopes that the domain has common characteristics As close as possible in the embedding space, in order to achieve this goal, the present invention utilizes the pairing information in different domains of each data instance and adopts the contrast loss method for effective guidance. The radar radiation source signals of samples are each composed of PDW data and IF data. The paired data pairs are called positive sample pairs in contrastive learning, and the opposite are called negative sample pairs. The similarity of two vectors is defined as , expressed as and of norm (i.e. cosine similarity). Then a pair of positive samples The loss function is defined as:

[0057]

[0058] in, is an indicator function if and only if and When it represents the temperature parameter, it is The final loss is calculated for all positive sample pairs, including and .

[0059] To ensure domain-specific features Can carry unique complementary information, hoping to have unique features in the same domain and common features As far as possible. To achieve this goal, use distance, and and The distance loss is calculated as:

[0060]

[0061] in, and are the unique and common features of PDW data. is the loss corresponding to the PDW data. The loss corresponding to the IF data is similar, and the sum of the two is the loss of the corresponding sample. .

[0062] The final loss is the weighted sum of all the above losses:

[0063]

[0064] in, and is a hyperparameter that can control the relative importance in contrastive learning.

[0065] In step 2), the overall structure of the multivariate time series Transformer is as follows Figure 2 As shown in Figure 2, the pure encoder architecture of Transformer is adopted, including the embedding layer, The Transformer block consists of a self-attention layer, a layer normalization layer, and a feed-forward network. The output of the projection layer is the final output, which, like the embedding layer, is composed of a single-layer MLP.

[0066] The multivariate time series Transformer treats the sequence of a variable over a period of time as an independent process, such as Figure 3 Specifically, by extracting representations for each time series, the self-attention module uses linear projection to obtain the query ( ),key( ) and value ( ), ,in is the projection dimension. Representing a unique query and key of a Token, each attention matrix element of the pre-Softmax score is expressed as Since each Token is normalized on its variable dimension, the Toke construction method implies the variability of variables over time. The attention mechanism focuses on the correlation between variables that change over a period of time. Shows the multivariate correlation between paired variable tokens. The feature transfer process in the multivariate time series Transformer is:

[0067]

[0068] in, Indicates The representation of the layer, represents the self-attention layer.

[0069] In a multivariate time series Transformer, layer normalization operates on a sequence of a single variable over time. , the formula is as follows. This method makes the characteristics of all variables under a relatively uniform distribution, reducing the difference in measurement units. This method can also effectively deal with the non-stationary problem of time series. In addition, since the characteristic representations of all variables are normalized to a normal distribution, the differences caused by different variable value ranges can be reduced.

[0070]

[0071] in Representatives include The hidden layer representation of a single variable time series with values, express No. values, express The mean of express The variance of .

[0072] In the multivariate time series transformer, the Token vector is formed by a radar signal of the same variable, such as Figure 4 As shown. Based on the universal representation theorem of multi-layer perceptron, it has a large enough model capacity to extract time features shared in historical observations and future predictions. The feedforward network acts on a sequence and can extract the intrinsic properties of the sequence, such as amplitude, periodicity, frequency spectrum, etc., thereby improving the generalization on other sequences.

[0073] The dimensions of IF data in MTS-Trans vary from 103 to 1900. The DCT module is used to unify the IF data into The MLP that encodes the input of each variable in the PDW data consists of a single hidden layer with 128 dimensions, and the output is Dimension. Cross-domain translation uses an MLP with two hidden layers, each containing 2048 dimensions, and finally outputs the PDW data and IF data uniformly to 512 dimensions. After obtaining the feature embedding from MTS-Trans, the embedding is divided into domain-common features (the first half of the embedding, 128 dimensions) and domain-specific features (the remaining half, 128 dimensions). The projection network in contrastive learning is also a multilayer perceptron, which consists of two hidden layers, each containing 1024 neurons, and the output vector is 128 dimensions. In addition, the present invention sets , . MTS-Trans model, random dropout rate is , the activation function uses ReLu. In the training phase, the present invention uses the Adam optimizer with a learning rate of , a scalar temperature parameter Set as , the network is trained in total wheel.

[0074] In the task of identifying non-cooperative radar emitter models, the test data set contains not only a small number of known non-cooperative emitters with labels, but also unknown non-cooperative emitters that have never been seen before. How to identify unknown emitters is also a very important part. In the field of machine learning, this is a classic open set recognition task. The model directly adopts the classic OpenMax (Bendale A, Boult T E. Towards open set deep networks[C] / / Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 1563-1572) open set recognition method (Yang C, Liu H, Yang S, et al. Open-set radar emitter recognition via deep metric auto-encoder[J]. IEEE Internetof Things Journal, 2024) in open set recognition, and realizes the recognition of unknown emitters by processing the features of the previous layer of Softmax, that is, the processing of the fully connected layer. The core goal of OpenMax is set as follows: if the input belongs to a predefined set of categories, OpenMax clearly identifies the specific known category corresponding to the input; conversely, when encountering input of unknown categories during the testing phase, its design ensures that it can output a label of the "unknown" category.

[0075] After the unknown non-cooperative radiation source is identified by the open set identification method, the unknown non-cooperative radiation source data has not been seen before, and there is no training process. The direct clustering method is directly adopted to cluster the unknown non-cooperative radiation source data by the elbow method. The OpenMax method is a plug-and-play module that is compatible with various classifiers. Among the hyperparameters of OpenMax, the number of tail samples used to fit the Weibull distribution is , the confidence threshold is .

[0076] The purpose of the present invention is to improve the recognition effect of non-cooperative radar emitter models. In actual situations, it is difficult to obtain data on non-cooperative radar emitters, and only a small amount of training data can be obtained, or even no training data can be obtained. To this end, the present invention randomly samples the training sets of the simulation data set, competition data set, and mixed data set used. , , The data of known non-cooperative radar emitters are used for training, and the remaining training set data are used as validation sets. The complete test set is used to test the model effect. The data composition of the above data sets is described in Tables 1 to 3 respectively.

[0077] Table 1 Data composition of the data set obtained by sampling the simulation data set

[0078]

[0079] Table 2 Data composition of the dataset obtained by sampling the competition dataset

[0080]

[0081] Table 3 Data composition of the dataset obtained by sampling the mixed dataset

[0082]

[0083] During testing, the open set recognition method is used so that the model can output the recognition results of unknown radar emitter models, but all unknown radar emitters can only be identified as unknown, and the specific unknown emitter cannot be identified. Since the unknown non-cooperative emitter data has not been seen during model fine-tuning or training, the direct clustering model can only be used to divide the data identified as unknown emitters. For the data identified as unknown emitters, the elbow method K-Means clustering is used, and the optimal clustering result is the recognition result of the unknown non-cooperative emitter. Here, the input of the K-Means model is not the original features of the sample, but the new representation of the sample obtained by the model. This representation is better than the original feature representation and is conducive to the subsequent clustering results. For models that do not have the ability to learn representations, only the original features can be used for clustering. Therefore, the evaluation indicators for the recognition of known and unknown non-cooperative radar emitter models should be displayed separately.

[0084] For known non-cooperative radar emitters, the ratio of the number of correctly identified known non-cooperative radar emitters to the total known non-cooperative radar emitters is used as the evaluation index, which is the recall rate in machine learning:

[0085]

[0086] Among them, TP (True Positives) is the true positive example, that is, the number of samples correctly predicted as positive; FN (False Negatives) is the false negative example, that is, the number of samples incorrectly predicted as negative; the sum of the two is the number of positive samples in reality.

[0087] For unknown non-cooperative radar emitters, the evaluation index is slightly different from that of known non-cooperative radar emitters. The final output of the unknown non-cooperative radar emitter model recognition result is a cluster label, which does not correspond to the real label one by one. Therefore, the cluster label needs to be relabeled first (Zhou Zhihua. Machine Learning [M]. Qing hua da xue chu banshe, 2016). Specifically, the label with the highest proportion of real labels in the data clustered with the same label is used as the new identification label. Then, the evaluation index of the unknown non-cooperative radar emitter is calculated according to the above recall rate.

[0088] For fair comparison, the random seeds of all schemes are the same. The CPU of the device used in this implementation plan is 2 Intel Xeon Gold 6240, the memory is 128G, and the GPU is 4 NVIDIA RTX 3090.

[0089] Specific embodiments are given below:

[0090] 1. Identification of known radar emitter models:

[0091] 1) Enter the fenxuanshibie directory;

[0092] 2) Use VSCode to open the test file test_fewshot_recog.py, set the model path parameter finetune_ckpt to 'checkpoints / finetune / fewshot_recog_model.pt', which means that the trained model will be loaded from this path for testing. Set the test data path parameter test_data_path to 'dataset / pdw_data / recog_task / fewshot / fewshot_test.csv', which means that the model will read the test data from the file in this path, and set the random seed parameter seed to any integer between 0-2^32-1, which means that part of the data from the test set is sampled from a random starting point as the test data for this test;

[0093] 3) Run the following command in the command line: python test_fewshot_recog.py and wait for the result output;

[0094] 4) When the prompt "Program running is finished" appears in the command line terminal, it means that the test is completed. Open the result file generated by this run to view the results;

[0095] 5) This part of the test needs to record the following results in the result file:

[0096] ①Total number of test data samples;

[0097] ②The total number of samples correctly identified;

[0098] ③Recognition rate of known radar radiation source models;

[0099] ④The random seed used in this test.

[0100] 2. Identification of unknown radar emitter models:

[0101] 1) Enter the fenxuanshibie directory;

[0102] 2) Use VSCode to open the test file test_zeroshot_recog.py, set the model path parameter finetune_ckpt to 'checkpoints / finetune / fewshot_separate_model.pt', which means that the trained model will be loaded from this path for testing. Set the test data path parameter test_data_path to 'dataset / pdw_data / recog_task / zeroshot / zeroshot_test.csv', which means that the model will read the test data from the file in this path, and set the random seed parameter seed to any integer between 0-2^32-1, which means that part of the data from the test set is sampled from a random starting point as the test data for this test;

[0103] 3) Run the following command in the command line: python test_zeroshot_recog.py and wait for the result output;

[0104] 4) When the prompt "Program running is finished" appears in the command line terminal, it means that the test is completed. Open the result file generated by this run to view the results;

[0105] 5) This part of the test needs to record the following results in the result file:

[0106] ①Total number of test data sample pairs;

[0107] ②The number of correct logarithms of the relationship among the sample pairs;

[0108] ③ Identification rate of unknown radar radiation source models;

[0109] ④The random seed used in this test.

[0110] Finally, Table 4 describes the MTS-Trans model in each type of sampling. , , The recognition effect of known and unknown radar emitters on simulation datasets, competition datasets, and mixed datasets.

[0111] Table 4 Results of MTS-Trans model

[0112]

[0113] As can be seen from Table 4, in the three data sets, as the number of samples increases from 100 to 10,000, the recognition accuracy of the model for known radar emitters generally increases; for unknown radar emitters, the simulation and competition data sets show an upward trend, while the mixed data set increases first and then decreases. Among different data sets, under the same number of samples, the simulation data set is usually more accurate than the competition data set in recognizing known and unknown radar emitters, and the mixed data set performs in the middle.

[0114] The present invention introduces PDW and IF data to provide the model with information from different domains, and through a new comparative learning method, effectively brings common features closer, distances common and unique features in the same domain, and enhances information utilization efficiency. Cross-domain translation is performed before inputting the model to ensure that the model can handle domain missing situations and improve the adaptability of the model. Tokens are constructed along the time-varying direction so that limited tokens can better capture the time dependency of parameters in PDW data, overcoming the problem that traditional Transformers are limited by the length of tokens and are difficult to capture long-term dependencies. The self-attention module focuses on parameter correlation, and the fully connected layer focuses on parameter time correlation to improve model performance. The MTS-Trans model of the present invention is particularly suitable for non-cooperative radar emitter model identification that is limited by the difficulty in obtaining data. By learning a large amount of data on cooperative radar emitters and fine-tuning on a small amount of data on non-cooperative radar emitters, good generalization performance is obtained, and it can be widely used in radar signal processing, electronic countermeasures and other fields.

[0115] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A radar emitter model recognition method based on multivariate time series Transformer, characterized in that The specific steps are as follows: 1) Construct a comparative self-supervised learning architecture based on pulse descriptors and IF data. Use the pulse descriptors and IF data of the same radar signal to construct a self-supervised comparative learning architecture, bring the common features of pulse descriptors and IF data closer, and pull the common features and unique features of the same domain further apart, so as to effectively utilize the information in pulse descriptors and IF data. 2) Construct a multivariate time series Transformer model that is compatible with pulse descriptors and intermediate frequency data: By modifying the way Tokens are constructed, the roles of the attention mechanism and the feedforward network are transformed so that the attention mechanism models the correlation between time series variables, and the long-term relationships are learned through layer normalization and the feedforward network, and the feedforward network models the temporal correlation of variables.

2. The radar emitter model identification method based on multivariate time series Transformer as claimed in claim 1, characterized in that: In step 1), the construction of a comparative self-supervised learning architecture based on pulse description words and intermediate frequency data includes: (1) Input encoding: Encode inputs from different domains, and process the pulse descriptor data and intermediate frequency data of radar signals in different ways. The intermediate frequency data is unified in dimension through discrete cosine transform, and the pulse descriptor data is dimensionally upgraded through a fully connected layer with ReLU nonlinearity to facilitate subsequent representation learning. (2) Cross-domain translation: First, use the module composed of MLP to unify the representations of different domains into the same dimension. At the same time, cross-domain translation is achieved through MLP, and the out-of-domain translation loss is constructed to solve the problem of the model missing in the generalization time domain. (3) Feature segmentation: The feature embedding of each domain is split into a domain-specific part and a domain-common part through a multivariate temporal Transformer to ensure that the complementary information of different domains is preserved while constructing a dissimilarity loss. In feature segmentation, the dimensions of domain-common features and domain-specific features are consistent. (4) Construct a loss function to perform comparative learning on the common features of the domains and construct a similarity loss. The final loss is the weighted sum of the cross-domain translation loss, dissimilarity loss, and similarity loss.

3. The radar emitter model identification method based on multivariate time series Transformer as claimed in claim 1, characterized in that: In step 1), based on the contrastive self-supervised learning architecture of pulse descriptors and intermediate frequency data, the feature extractor of radar signals is obtained by contrastive self-supervised learning of pulse descriptors and intermediate frequency data of a large number of unlabeled cooperative radar emitter data. The final pre-training loss is composed of cross-domain loss, common feature similarity loss and dissimilarity between unique features and common features: in, and is a hyperparameter that controls the relative importance in contrastive learning; is the cross-domain translation loss; is a similar loss; is the dissimilarity loss; and then a small amount of labeled non-cooperative radar emitter data is used for transfer learning to obtain good generalization performance.

4. The radar emitter model identification method based on multivariate time series Transformer as claimed in claim 3, characterized in that: In the self-supervised learning stage, a cross-domain translation module is introduced: in, and They are the MLPs that realize the embedding mapping from PDW to IF and from IF to PDW respectively; they solve the problem of missing the model in the generalization time domain by translating data from one domain to another, ensuring that the model input has both the representation of the pulse descriptor data and the representation of the intermediate frequency data.

5. The radar emitter model identification method based on multivariate time series Transformer as claimed in claim 1, characterized in that: In step 1), a contrastive self-supervised learning architecture based on pulse descriptors and IF data separates the feature embedding of each domain into a domain-specific part and a domain-common part to preserve domain-specific information.

6. The radar emitter model identification method based on multivariate time series Transformer as claimed in claim 1, characterized in that: In step 2), the multivariate time series Transformer model compatible with pulse description words and intermediate frequency data adopts the pure encoder architecture of Transformer, including an embedding layer, The Transformer block consists of a self-attention layer, a layer normalization layer, and a feed-forward network. The output of the projection layer is the final output. Both the projection layer and the embedding layer are composed of a single-layer MLP.

7. The radar emitter model identification method based on multivariate time series Transformer as claimed in claim 1, characterized in that: In step 2), the multivariate time series Transformer treats a variable over a period of time as an independent process. By extracting representations for each period of time, the self-attention module uses linear projection to obtain the query ( ),key( ) and value ( ), ,in is the projection dimension; Representing a unique query and key of a Token, each attention matrix element of the pre-Softmax score is expressed as ; Since each Token is normalized on its variable dimension before, the Toke construction method implies the variability of variables over time, and the attention mechanism mainly focuses on the correlation between variables that change over a period of time. Displays the multivariate correlation between paired variables Token; The feature transfer process in the multivariate time series Transformer is: in, Indicates The representation of the layer, represents the self-attention layer.

8. The radar emitter model identification method based on multivariate time series Transformer as claimed in claim 1, characterized in that: In step 2), in the multivariate time series Transformer, layer normalization operates on a sequence of a single variable over time. , the formula is as follows: in, Representatives include The hidden layer representation of a single variable time series with values, express No. values, express The mean of express The variance of .

9. The radar emitter model identification method based on multivariate time series Transformer as claimed in claim 1, characterized in that: In step 2), in the multivariate time series Transformer, the Token vector is formed by a radar signal of the same variable. Based on the universal representation theorem of the multi-layer perceptron, it has a large enough model capacity to extract the time features shared in historical observations and future predictions; the feedforward network acts on a sequence and can extract the intrinsic properties of the sequence, including amplitude, periodicity, and frequency spectrum, thereby improving the generalization to other sequences.

Images