Search

CN-121980336-A - Model identification method for low-interception probability radar

CN121980336ACN 121980336 ACN121980336 ACN 121980336ACN-121980336-A

Abstract

The invention discloses a model identification method for a radar with low interception probability, which comprises the steps of inputting a pulse description word PDW original data set obtained by radar detection, analyzing and format standardization of the original data set to obtain full pulse sequence data, sorting the full pulse sequence data by adopting a signal sorting algorithm to separate pulse sequences from different radars, constructing a two-way feature extraction network based on a time sequence convolution and an attention mechanism, extracting signal features of each sorted radar pulse sequence to obtain a high-dimensional feature vector, finely dividing all comprehensive features extracted by the radar with the same model into a plurality of sub-class clusters by using a K-Means clustering algorithm to represent different possible working modes of the radar with the same model, carrying out equalization processing on all sub-class cluster samples after division, constructing a training sample set with balanced distribution between the sub-class clusters, carrying out supervision training on a preset radar model classification model by using an equalized training sample set to obtain a complete model identification model. The invention improves the data fitting capacity, generalization capacity and robustness of the data-driven artificial intelligent recognition method.

Inventors

  • LIU JIAN
  • WU LIANHUI
  • XU PENGTAO
  • LIANG BING

Assignees

  • 中国船舶集团有限公司第七二三研究所

Dates

Publication Date
20260505
Application Date
20251231

Claims (8)

  1. 1. The model identification method for the low interception probability radar is characterized by comprising the following steps of: Step 1, inputting a pulse description word PDW original data set obtained by radar detection; step 2, analyzing and format standardization is carried out on the original data set to obtain full pulse sequence data; Step 3, sorting the whole pulse sequence data by adopting a signal sorting algorithm, and separating pulse sequences from different radars; step 4, constructing a two-way feature extraction network based on a time sequence convolution and an attention mechanism, and extracting signal features of each sorted radar pulse sequence to obtain a high-dimensional feature vector, wherein: The second path of characteristic extraction path, input pulse sequence into a attention weighting module, the module carries out channel attention calculation and space attention calculation sequentially, carry on the adaptive weight distribution to different characteristic dimension and different time points of the pulse sequence, then send its output into a time sequence convolution network TCN to carry on deep characteristic extraction; step 5, performing intra-class fine division on all comprehensive features extracted from the radar of the same model by using a K-Means clustering algorithm to form a plurality of sub-class clusters so as to represent different possible working modes of the radar of the same model; Step 6, carrying out equalization treatment on all the divided sub-class cluster samples, constructing a training sample set with balanced distribution in the class and among the classes, and for a few class clusters with the number of samples lower than a set threshold, wherein: For a plurality of clusters with the sample number higher than a set threshold value, calculating the distance between the clusters and a few samples by adopting a distance-based heuristic sampling method, and reserving the most representative samples to reduce the number of the clusters; step 7, performing supervision training on a preset radar model classification model by using an balanced training sample set to obtain a model identification model with complete training; And 8, inputting the comprehensive characteristics obtained after the pulse sequences of the radar to be identified are processed in the steps 2 to 4 into a model identification model with complete training, and outputting the radar model identification result to which the model identification model belongs.
  2. 2. The model identification method for the low-interception-probability radar according to claim 1, wherein the time sequence convolution network TCN adopts a hole causal convolution structure, hole convolution is introduced on the basis of causal convolution, wherein the convolution result of the causal convolution is only related to the current moment and sequences before the current moment so as to ensure the causality of a time sequence, and the hole factors are increased layer by layer along with the increase of the network layer number, so that the network obtains a larger time sequence receptive field under the condition of not increasing depth so as to capture history dependent features of a longer range in a radar pulse sequence.
  3. 3. The method for model identification for low-probability-of-interception radar according to claim 2, wherein for a sample sequence And convolution kernel The hole causal convolution is expressed as: ; Wherein, the Representing the result of s time after one hole causal convolution, and obtaining the sequence after the convolution of the sequence and n convolution kernels D represents a cavity factor, the power of 2 increases along with the change of a convolution layer, T is a time sequence length, C is a characteristic channel number, and then a receptive field calculation formula is feld = (dim-1) d+1, dim is a characteristic dimension.
  4. 4. The method for identifying a model for a radar with low probability of interception according to claim 1, wherein the process of channel attention calculation is: For input sample sequences Wherein T is the time sequence length, C is the number of characteristic channels, and global maximum pooling and global average pooling are respectively carried out on the space dimension to obtain two characteristic vectors; Respectively inputting two feature vectors into a multi-layer perceptron MLP with shared parameters, wherein the MLP consists of two full-connection layers and a ReLU activation function, firstly compressing the channel number to be 1/r, wherein r is a compression rate larger than 1, and then expanding the channel number back to the original channel number C; after adding the two output feature vectors of the MLP, generating the channel attention weight through the Sigmoid activation function Expressed as: ; multiplying the channel attention weight with the original input feature map channel by channel to obtain a channel weighted feature map expressed as: 。
  5. 5. The model identification method for the low-interception-probability radar according to claim 1, wherein the process of spatial attention calculation is as follows: Weighting channels As input, carrying out maximum pooling and average pooling on the dimension of the characteristic channel respectively to obtain two characteristic diagrams; splicing the two feature graphs in channel dimension to obtain a feature graph with dimension of 2 xT x1, inputting the feature graph into a one-dimensional convolution layer with convolution kernel size of 1 xk for feature fusion and coding, wherein k is an odd number greater than 1, and finally generating space attention weight through Sigmoid activation function Expressed as: ; Multiplying the spatial attention weight with the channel weighted feature map to obtain a final attention weighted feature map: 。
  6. 6. A model identification system for a low-probability-of-interception radar, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the model identification method for a low-probability-of-interception radar according to any one of claims 1 to 5 when executing the program.
  7. 7. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements a model identification method for a low interception probability radar according to any one of claims 1 to 5 when executing the program.
  8. 8. A non-transitory computer readable storage medium having stored thereon a computer program, which when executed by a processor, implements the model identification method for a low interception probability radar according to any one of claims 1 to 5.

Description

Model identification method for low-interception probability radar Technical Field The invention relates to a signal processing technology, in particular to a model identification method aiming at a radar with low interception probability. Background In recent years, along with the continuous upgrading and updating of radar technology systems, the characteristics of various electromagnetic signals, overlapped density, dynamic overlapping and rapid style change are achieved. The development of big data artificial intelligence technology brings new ideas for radar radiation source identification, and based on a large amount of detection data, the extraction of radar features and the construction of identification space are automatically completed through a machine learning training method, so that the accurate identification of a new system radar is solved. However, the signal facing the low-interception radar radiation source identification is a small sample space, and the corresponding data has the defects of insufficient completeness, poor continuity and unbalanced category due to the limitation of actual conditions, and the artificial intelligence identification method based on data driving has the problems of over fitting of training data, poor generalization capability and insufficient robustness. Disclosure of Invention The invention aims to provide a model identification method aiming at a radar with low interception probability, so as to solve the problems of insufficient data completeness, poor continuity and unbalanced category. The technical scheme for realizing the purpose of the invention is that the model identification method aiming at the radar with low interception probability comprises the following steps: Step 1, inputting a pulse description word PDW original data set obtained by radar detection; step 2, analyzing and format standardization is carried out on the original data set to obtain full pulse sequence data; Step 3, sorting the whole pulse sequence data by adopting a signal sorting algorithm, and separating pulse sequences from different radars; step 4, constructing a two-way feature extraction network based on a time sequence convolution and an attention mechanism, and extracting signal features of each sorted radar pulse sequence to obtain a high-dimensional feature vector, wherein: The second path of characteristic extraction path, input pulse sequence into a attention weighting module, the module carries out channel attention calculation and space attention calculation sequentially, carry on the adaptive weight distribution to different characteristic dimension and different time points of the pulse sequence, then send its output into a time sequence convolution network TCN to carry on deep characteristic extraction; step 5, performing intra-class fine division on all comprehensive features extracted from the radar of the same model by using a K-Means clustering algorithm to form a plurality of sub-class clusters so as to represent different possible working modes of the radar of the same model; Step 6, carrying out equalization treatment on all the divided sub-class cluster samples, constructing a training sample set with balanced distribution in the class and among the classes, and for a few class clusters with the number of samples lower than a set threshold, wherein: For a plurality of clusters with the sample number higher than a set threshold value, calculating the distance between the clusters and a few samples by adopting a distance-based heuristic sampling method, and reserving the most representative samples to reduce the number of the clusters; step 7, performing supervision training on a preset radar model classification model by using an balanced training sample set to obtain a model identification model with complete training; And 8, inputting the comprehensive characteristics obtained after the pulse sequences of the radar to be identified are processed in the steps 2 to 4 into a model identification model with complete training, and outputting the radar model identification result to which the model identification model belongs. Furthermore, the time sequence convolution network TCN adopts a hole causal convolution structure, and hole convolution is introduced on the basis of causal convolution, wherein the convolution result of the causal convolution is only related to the current moment and the sequence before the current moment so as to ensure the causality of the time sequence, and the hole factor is increased layer by layer along with the increase of the network layer number, so that the network obtains a larger time sequence receptive field under the condition of not increasing the depth so as to capture the history dependent characteristics of a longer range in the radar pulse sequence. Further, for sample sequencesAnd convolution kernelThe hole causal convolution is expressed as: , Wherein, the Representing the result of s time after one hole causal con