A SAR Interference Recognition Method Based on Feature Space Distance Measurement

Abstract

The present invention discloses a SAR interference identification method based on feature space distance measurement, comprising: obtaining at least one interference signal data to be identified; performing short-time Fourier transform on at least one interference signal data to be identified by using preset short-time Fourier transform parameters to obtain at least one time-frequency diagram; inputting at least one time-frequency diagram into a well-trained interference feature vector extraction model, and outputting at least one feature vector; the well-trained interference feature vector extraction model is obtained by training with a plurality of training data pairs; each training data pair comprises a time-frequency diagram and an interference type label of the time-frequency diagram; determining the interference type of the interference signal data to be identified corresponding to the feature vector according to the distance between each feature vector and each interference feature vector with known interference type in an interference feature vector library; the interference feature vector library is constructed by extracting feature vectors from the time-frequency diagram with known interference type according to the well-trained interference feature vector extraction model.

Description

Technical Field

[0001] The invention belongs to the field of radar technology, and in particular relates to a SAR interference identification method based on feature space distance measurement. Background Art

[0002] With the development of electronic warfare, the endless jamming measures have brought severe challenges to the survival and operation of Synthetic Aperture Radar (SAR). Most of the existing anti-jamming technologies are aimed at specific types of jamming. Therefore, it is of great research significance and practical value to accurately identify radar jamming through design methods.

[0003] At present, there are many methods for accurate identification of SAR radar interference, mainly including likelihood-based methods and multi-domain feature extraction-based methods. The likelihood-based method mainly achieves interference detection by modeling the echo signal and using the generalized likelihood ratio test (GLRT). The multi-domain feature extraction-based method uses professional knowledge to extract features from different domains such as time domain, frequency domain, time-frequency domain and polarization domain, and on this basis uses machine learning or other algorithms to design classifiers to achieve interference identification.

[0004] However, this method of interference detection by modeling the echo signal and using the generalized likelihood ratio test requires prior information of known interference, and its detection threshold needs to be set manually, so it is only applicable to a limited number of interference types. In addition, when this type of method faces new types of interference data, it is necessary to manually adjust the detection threshold, which has the problems of poor adaptability and low generalization. Manual extraction of multi-domain features of interference based on multi-domain feature extraction methods is prone to large computational complexity, which affects the efficiency of interference identification and is difficult to meet the real-time requirements in specific interference identification scenarios. In addition, the interference identification performance is directly determined by the quality of the proposed features, and manually extracted features are difficult to guarantee the accuracy of the classification results.

[0005] That is to say, the interference type identified by the current SAR radar interference identification method is not accurate enough, and the generalization and identification efficiency are low. Summary of the invention

[0006] In order to solve the above problems existing in the related art, the present invention provides a SAR interference identification method based on feature space distance measurement. The technical problem to be solved by the present invention is achieved by the following technical solutions:

[0007] The present invention provides a SAR interference identification method based on feature space distance measurement, comprising:

[0008] Obtain at least one piece of interference signal data to be identified;

[0009] Performing short-time Fourier transform processing on the at least one piece of interference signal data to be identified using preset short-time Fourier transform parameters to obtain at least one time-frequency graph corresponding to the at least one piece of interference signal data to be identified;

[0010] The at least one time-frequency graph is input into a well-trained interference feature vector extraction model, and an interference feature vector corresponding to the at least one time-frequency graph is output. Figure 1 At least one feature vector corresponding to one; the well-trained interference feature vector extraction model is obtained by training with multiple training data pairs; each training data pair includes a sample time-frequency graph and an interference type label of the sample time-frequency graph;

[0011] According to the distance between each feature vector and each interference feature vector with known interference type in the interference feature vector library, the interference type of an interference signal data to be identified corresponding to the feature vector is determined; the interference feature vector library is constructed based on the feature vector extracted from the time-frequency diagram with known interference type according to the trained interference feature vector extraction model.

[0012] In some embodiments, the at least one time-frequency graph is input into a well-trained interference feature vector extraction model, and the at least one time-frequency graph is output. Figure 1 A corresponding at least one feature vector, including:

[0013] After inputting the at least one time-frequency graph into a well-trained interference feature vector extraction model, the well-trained interference feature vector extraction model performs a plurality of different convolution pooling operations on each time-frequency graph to obtain a first feature graph of the time-frequency graph;

[0014] Performing global average pooling, global maximum pooling, activation and attention layer processing on the first feature map in sequence to obtain a second feature map;

[0015] Performing feature fusion processing on the first feature map and the second feature map to obtain a fused feature map;

[0016] Perform global maximum pooling processing on the fused feature map to obtain a feature vector corresponding to the time-frequency map.

[0017] In some embodiments, determining the interference type of a piece of interference signal data to be identified corresponding to each feature vector according to the distance between each feature vector and each interference feature vector with known interference type in the interference feature vector library includes:

[0018] For each feature vector in the at least one feature vector, calculating the Euclidean distance between the feature vector and each interference feature vector with a known interference type in the interference feature vector library, to obtain a plurality of Euclidean distances corresponding one to one to each interference feature vector;

[0019] The interference type of the interference feature vector corresponding to the smallest Euclidean distance among the multiple Euclidean distances is used as the interference type of a piece of interference signal data to be identified corresponding to the feature vector.

[0020] In some embodiments, the fully trained interference feature vector extraction model includes, in sequence: a first convolution pooling layer, a second convolution pooling layer, a third convolution pooling layer and a fourth convolution pooling layer with the same convolution kernel size and pooling step size, as well as an attention module and a first global maximum pooling layer; the attention module includes, in sequence: a global average pooling layer, a second global maximum pooling layer, a sigmoid layer and an attention layer; the number of channels of the first convolution pooling layer, the second convolution pooling layer, the third convolution pooling layer and the fourth convolution pooling layer increases in sequence by the same multiple; the number of channels of the fourth convolution pooling layer, the first global maximum pooling layer, the global average pooling layer, the second global maximum pooling layer, the sigmoid layer and the attention layer are all the same; the step size of the first global maximum pooling layer, the global average pooling layer and the second global maximum pooling layer is the same.

[0021] In some embodiments, the convolution kernel sizes of the first convolution pooling layer, the second convolution pooling layer, the third convolution pooling layer, and the fourth convolution pooling layer are all , the pooling step size is 2; the pooling step size of the first global maximum pooling layer, the global average pooling layer and the second global maximum pooling layer is 8.

[0022] In some embodiments, the at least one time-frequency graph is input into a well-trained interference feature vector extraction model, and the at least one time-frequency graph is output. Figure 1 Before the at least one corresponding feature vector, the method further comprises:

[0023] Get Signal data of known interference types; each signal data of known interference type includes Signal data; and All are positive integers;

[0024] Using the preset short-time Fourier transform parameter The signal data is processed by short-time Fourier transform, and the corresponding Sample frequency spectrum;

[0025] For the Each sample time-frequency diagram in the sample time-frequency diagram Add corresponding interference type labels ; ;

[0026] By As a training data pair, we get training data pairs;

[0027] Adopt the The initial interference feature vector extraction model is iteratively trained multiple times based on a pair of training data, a preset loss function, a preset learning rate and a preset momentum value to obtain the fully trained interference feature vector extraction model.

[0028] In some embodiments, the use of The initial interference feature vector extraction model is trained multiple times with a training data pair, a preset loss function, a preset learning rate and a preset momentum value to obtain the fully trained interference feature vector extraction model, including:

[0029] In the During the training, get The interference feature vector extraction model and the times the momentum value, and from the Training data pairs are selected training data pairs; and are all positive integers; when When The interference feature vector extraction model of the second time is the initial interference feature vector extraction model, and the The momentum value of the second is the preset momentum value;

[0030] The The training data is input into the The interference feature vector extraction model is obtained The feature vector of each time-frequency graph in the training data pair;

[0031] According to the feature vectors of each time-frequency graph, the interference type labels of each time-frequency graph and the preset loss function, the first The loss value of times;

[0032] According to the said The loss value of the first The interference feature vector is used to extract the gradient corresponding to each network parameter in the model;

[0033] According to the gradient, the preset learning rate and the The momentum value of the Momentum value of times;

[0034] According to the said The momentum value of the first update The interference feature vector is used to extract each network parameter in the model, and the The interference feature vector extraction model is repeated until the first When extracting the interference feature vector model for the second time, The interference feature vector extraction model obtained this time is used as the fully trained interference feature vector extraction model; is the default value.

[0035] In some embodiments, the The expression of the momentum value is as follows:

[0036] ;

[0037] in, For the said The momentum value of times, For the said The momentum value of times, is the preset learning rate, is the gradient.

[0038] In some embodiments, the expression of the preset loss function is as follows:

[0039] ;

[0040] in, represents the preset loss function, is the number of training data pairs used in each training, For each training session The first The feature vector of the time-frequency graph in the training data pairs; For each training session The training data is matched with the The feature vector of the time-frequency graph in a training data pair with the same interference type; For each training session The training data is matched with the The feature vector of the time-frequency graph in a training data pair with different interference types, ; Indicates the maximum value operation. is the preset distance threshold, Represents the Euclidean distance between vectors.

[0041] In some embodiments, before determining the interference type of a piece of interference signal data to be identified corresponding to each feature vector according to the distance between each feature vector and each interference feature vector with known interference type in the interference feature vector library, the method further includes:

[0042] The The sample time-frequency graph is input into the well-trained interference feature vector extraction model to obtain a one-to-one correspondence interference feature vectors;

[0043] The The interference type label of the sample time-frequency diagram corresponds to the The interference type label of the interference feature vector is obtained interference characteristic vectors with known interference types; The interference feature vector library is composed of interference feature vectors with known interference types.

[0044] The present invention has the following beneficial technical effects:

[0045] The present invention achieves the acquisition of identifiable features (i.e., feature vectors) of interference data by training a complete interference feature vector extraction model based on a convolutional neural network. No artificial parameters are required in the feature extraction process, thereby greatly improving the efficiency of interference feature extraction, which is beneficial to improving the efficiency of type recognition. In addition, the present invention calculates the distance between the extracted identifiable features and the vectors in the interference feature vector library, and identifies the type of the interference feature vector according to the distance, thereby having strong generalization for different interference types. In addition, the present invention only needs to update the interference feature vector library to complete the identification of new interference types, without the need to retrain the model, and can therefore be directly deployed in the identification scenario of new interference types.

[0046] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A schematic flow chart of a SAR interference identification method based on feature space distance measurement provided by an embodiment of the present invention;

[0048] Figure 2 A schematic diagram of the principle of extracting feature vectors of a time-frequency graph using an exemplary fully trained interference feature vector extraction model provided in an embodiment of the present invention;

[0049] Figure 3 A schematic diagram of the principle of an exemplary attention module provided in an embodiment of the present invention generating a second feature map according to an input first feature map;

[0050] Figure 4 An exemplary schematic diagram of the principle of identifying the interference type of a feature vector according to an interference feature vector library provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0051] The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0052] In the description of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0053] 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 invention. 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.

[0054] Although the present invention is described herein in conjunction with various embodiments, in the process of implementing the claimed invention, those skilled in the art may understand and implement other variations of the disclosed embodiments by viewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "one" or "an" does not exclude multiple situations. A single processor or other unit may implement several functions listed in the claims. Certain measures are recorded in different dependent claims, but this does not mean that these measures cannot be combined to produce good results.

[0055] Figure 1 FIG. 1 is a flow chart of a SAR interference identification method based on feature space distance measurement provided by an embodiment of the present invention. Figure 1 As shown, the method comprises the following steps:

[0056] S101. Obtain at least one piece of interference signal data to be identified.

[0057] Exemplarily, the interference signal data to be identified may be SAR radar echo data to be identified.

[0058] S102: Perform short-time Fourier transform processing on at least one piece of interference signal data to be identified using preset short-time Fourier transform parameters to obtain at least one time-frequency diagram corresponding to the at least one piece of interference signal data to be identified.

[0059] Here, the short-time Fourier transform parameters may include: window length, number of Fourier transform points, type of window function and sliding step of window function. Exemplarily, the preset short-time Fourier transform parameters include: window length is 31, number of Fourier transform points is 256, window function is Hamming window, and sliding step of window function is 1.

[0060] S103, inputting at least one time-frequency graph into a well-trained interference feature vector extraction model, and outputting Figure 1 One corresponds to at least one feature vector; the fully trained interference feature vector extraction model is obtained by training with multiple training data pairs; each training data pair includes a sample time-frequency diagram and an interference type label of the sample time-frequency diagram.

[0061] Here, the fully trained interference feature vector extraction model includes: convolution pooling layer 1, convolution pooling layer 2, convolution pooling layer 3 and convolution pooling layer 4, which have the same convolution kernel size and pooling step size, as well as attention module 1 and global maximum pooling layer 2; attention module 1 includes: global average pooling layer 1, global maximum pooling layer 1, sigmoid layer 1 and attention layer 1. The number of channels of convolution pooling layer 1, convolution pooling layer 2, convolution pooling layer 3 and convolution pooling layer 4 increases in the same multiple. The number of channels of convolution pooling layer 4, global average pooling layer 1, global maximum pooling layer 1, global maximum pooling layer 2, sigmoid layer 1 and attention layer 1 are all the same. The step size of global average pooling layer 1, global maximum pooling layer 1 and global maximum pooling layer 2 is the same. Each of the convolutional pooling layers 1, 2, 3, and 4 has an activation function.

[0062] For example, as shown in Table 1 below, the convolution kernel sizes of convolution pooling layer 1, convolution pooling layer 2, convolution pooling layer 3, and convolution pooling layer 4 are all , the pooling step size is 2, and the activation function is ReLU, and the number of channels of convolution pooling layer 1, convolution pooling layer 2, convolution pooling layer 3 and convolution pooling layer 4 are 4, 8, 16 and 32 respectively. The number of channels of global average pooling layer 1, global maximum pooling layer 1, global maximum pooling layer 2, sigmoid layer 1 and attention layer 1 is 32, and the pooling step size of global average pooling layer 1, global maximum pooling layer 1 and global maximum pooling layer 2 is 8. Specifically, each convolution pooling layer is a network layer composed of a convolution layer and a pooling layer.

[0063] Table 1:

[0064]

[0065] In some embodiments, S103 may be implemented by the following steps:

[0066] S1031. After inputting at least one time-frequency graph into a well-trained interference feature vector extraction model, the well-trained interference feature vector extraction model performs multiple different convolution pooling operations on each time-frequency graph to obtain a first feature graph of the time-frequency graph.

[0067] Specifically, Figure 2 As shown, for a time-frequency diagram ( Figure 2 After the interference time-frequency diagram is input into the well-trained interference feature vector extraction model, the time-frequency diagram is firstly subjected to a convolution pooling process through the convolution pooling layer 1 to obtain the feature vector. Figure 1 After that, convolutional pooling layer 2 is used to perform feature Figure 1 Perform a convolution pooling process to obtain the features Figure 2 After that, convolutional pooling layer 3 is used to perform feature Figure 2 Perform a convolution pooling process to obtain the features Figure 3 After that, convolutional pooling layer 4 is used to perform feature Figure 3 Perform a convolution pooling process to obtain the features Figure 4 (called the first feature map).

[0068] S1032. Perform global average pooling, global maximum pooling, activation and attention layer processing on the first feature map in sequence to obtain a second feature map.

[0069] like Figure 3As shown in the figure, after obtaining the first feature map, the global average pooling layer 1 is first used to perform global average pooling processing on the first feature map to obtain the feature map after average pooling. Then, the global maximum pooling layer 1 is used to perform global maximum pooling processing on the feature map after average pooling to obtain the feature map after maximum pooling. Then, the sigmoid layer 1 (i.e., the sigmoid function) is used to activate the feature map after maximum pooling to obtain the activated feature map. Then, the attention layer 1 is used to perform attention layer processing on the activated feature map to obtain the attention feature map (called the second feature map).

[0070] S1033: Perform feature fusion processing on the first feature map and the second feature map to obtain a fused feature map.

[0071] Continue with Figure 3 Taking the example of FIG. 1 as an example, after obtaining the first feature map and the second feature map, the first feature map and the second feature map are subjected to feature fusion processing ( Figure 3 in ” indicates feature fusion operation), and we get Figure 2 The feature map 5 (i.e., the fused feature map) is shown.

[0072] S1034. Perform global maximum pooling processing on the fused feature map to obtain a feature vector corresponding to the time-frequency map.

[0073] Continue with Figure 2 Taking the feature map 5 as an example, after obtaining the feature map 5, the global maximum pooling layer 2 is used to perform global maximum pooling on the feature map 5, and a feature vector corresponding to the time-frequency map can be obtained. .

[0074] S104. Determine the interference type of a piece of interference signal data to be identified corresponding to the feature vector based on the distance between each feature vector and each interference feature vector with known interference type in the interference feature vector library; the interference feature vector library is constructed based on the feature vector extracted from the time-frequency diagram with known interference type according to a well-trained interference feature vector extraction model.

[0075] In some embodiments, S104 can be implemented by the following steps:

[0076] S1041. For each feature vector in the at least one feature vector, calculate the Euclidean distance between the feature vector and each interference feature vector in the interference feature vector library to obtain a plurality of Euclidean distances corresponding to each interference feature vector.

[0077] S1042: Use the interference type of the interference feature vector corresponding to the smallest Euclidean distance among the multiple Euclidean distances as the interference type of a piece of interference signal data to be identified corresponding to the feature vector.

[0078] For example, Figure 4 As shown, when the interference feature vector library contains multiple interference feature vectors with known interference types ( Figure 4 Known Interference Feature Vector Library , ) specifically includes interference feature vectors of known interference types, and these The interference characteristic vector of the known interference type belongs to When there are interference types, for a feature vector , we can calculate the eigenvector Interference Feature Vector Library The Euclidean distance between each interference feature vector in the The corresponding interference feature vector The interference type is taken as the feature vector type of interference.

[0079] In some embodiments, before S103, the method further includes:

[0080] S201, Acquisition Signal data of known interference types; each signal data of known interference type includes Signal data; and All are positive integers.

[0081] For example, you can get Various types of interference SAR radar echo data.

[0082] S202, using preset short-time Fourier transform parameters The signal data is processed by short-time Fourier transform, and the corresponding Sample frequency diagram.

[0083] Here, the preset short-time Fourier transform parameters are used to obtain Various types of interference The SAR radar echo data are short-time Fourier transformed to obtain A time-frequency diagram, The time-frequency diagram is Sample frequency diagram.

[0084] S203, for Each sample time-frequency diagram in the sample time-frequency diagram Add corresponding interference type labels ; .

[0085] S204, by As a training data pair, we get training data pairs.

[0086] S205, use training data pairs, preset loss function, preset learning rate and preset momentum values , the initial interference feature vector extraction model is iteratively trained multiple times to obtain a fully trained interference feature vector extraction model.

[0087] Specifically, step S205 is implemented by the following steps:

[0088] S2051, in During the training, get The interference feature vector extraction model and the times the momentum value, and from Training data pairs are selected training data pairs; and are all positive integers; when At that time, The interference feature vector extraction model is the initial interference feature vector extraction model. The momentum value of the times is the preset momentum value .

[0089] Here, the initial interference feature vector extraction model is an interference feature vector extraction model obtained by initialization.

[0090] S2052, will The training data is input to The interference feature vector extraction model is obtained The feature vector of each time-frequency graph in the training data pair.

[0091] S2053, according to the characteristic vector of each time-frequency graph, the interference type label of each time-frequency graph and the preset loss function, calculate the first The loss value .

[0092] Specifically, the expression of the preset loss function is as follows:

[0093] ;in, represents the preset loss function, For each training session The first The feature vector of the time-frequency graph in the training data pairs; For each training session The training data is aligned with the The feature vector of the time-frequency graph in a training data pair with the same interference type; For each training session The training data is aligned with the The feature vector of the time-frequency graph in a training data pair with different interference types, ; Indicates the maximum value operation. is the preset distance threshold, Represents the Euclidean distance between vectors.

[0094] S2054, according to The loss value , calculate the The gradient corresponding to each network parameter in the model is extracted by the interference feature vector .

[0095] Specifically, , For the The interference feature vector is used to extract each network parameter in the model. is the derivative symbol.

[0096] S2055, according to the gradient, preset learning rate and Momentum value , calculate the Momentum value .

[0097] Specifically, .

[0098] S2056, according to The momentum value is updated The interference feature vector is used to extract each network parameter in the model, and the The interference feature vector extraction model is repeated until the first When extracting the interference feature vector model, The interference feature vector extraction model of this time is taken as the fully trained interference feature vector extraction model; is the default value.

[0099] Specifically, , For the The interference feature vector is extracted for each network parameter in the model.

[0100] Here, in the In the next training, the same principles as in S2051~S2056 are used for training, and this is repeated until When The interference feature vector extraction model is taken as the fully trained interference feature vector extraction model.

[0101] In some embodiments, after S2056 and before S103, the method further includes:

[0102] S301, The sample time-frequency graph is input into the well-trained interference feature vector extraction model to obtain a one-to-one correspondence interference feature vector.

[0103] S302, The interference type label of the sample time-frequency diagram corresponds to The interference type label of the interference feature vector is obtained interference feature vectors with known interference types; The interference feature vector library is composed of interference feature vectors with known interference types.

[0104] The present invention constructs an interference feature extraction model based on a neural network to extract deep identifiable features of interference data, thereby overcoming the shortcomings of the prior art of manually extracting interference features, which leads to large amount of calculation, poor adaptability, and reduced interference recognition accuracy; the present invention uses deep learning technology to train the interference feature vector extraction model, combines distance measurement to calculate the distance between the extracted feature vector and the vector in the interference feature vector library, and completes accurate identification of the interference type according to the minimum distance. Under the technical framework proposed by the present invention, the type of interference data used in training can be different from the interference type of the test set. Therefore, compared with the traditional interference recognition algorithm, the interference recognition system of the present invention trained offline can be directly deployed in the application scenario of new interference type recognition; the present invention uses distance measurement to determine the category of the interference feature vector. When there is a new interference type, only the interference feature vector library needs to be updated without retraining the entire model, thereby overcoming the problem that the existing interference recognition technology based on classification ideas has poor generalization and low recognition accuracy for interference types outside the training set.

[0105] The present invention can be applied to SAR systems of various platforms such as airborne and satellite-borne, and can efficiently and accurately identify various types of interference. At present, it has been tested in the measured SAR interference echo data of airborne flight. The present invention can accurately identify different types of interference, provide prior knowledge for the selection of subsequent SAR anti-interference measures, and prove the practicality of the method of the present invention. The following Table 2 shows the recognition accuracy of the present invention for three types of interference, which are 91.35%, 87.62% and 90.24% respectively. According to Table 2, it can be seen that the present invention can accurately and efficiently complete the recognition of different types of interference.

[0106] Table 2:

[0107]

[0108] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A SAR interference identification method based on feature space distance measurement, characterized in that: include: Obtain at least one piece of interference signal data to be identified; Performing short-time Fourier transform processing on the at least one piece of interference signal data to be identified using preset short-time Fourier transform parameters to obtain at least one time-frequency graph corresponding to the at least one piece of interference signal data to be identified; Inputting the at least one time-frequency graph into a well-trained interference feature vector extraction model, and outputting at least one feature vector corresponding to the at least one time-frequency graph in one-to-one correspondence; said inputting the at least one time-frequency graph into the well-trained interference feature vector extraction model, and outputting at least one feature vector corresponding to the at least one time-frequency graph in one-to-one correspondence, comprises: After inputting the at least one time-frequency graph into a well-trained interference feature vector extraction model, the well-trained interference feature vector extraction model performs a plurality of different convolution pooling operations on each time-frequency graph to obtain a first feature graph of the time-frequency graph; Performing global average pooling, global maximum pooling, activation and attention layer processing on the first feature map in sequence to obtain a second feature map; Performing feature fusion processing on the first feature map and the second feature map to obtain a fused feature map; Performing global maximum pooling processing on the fused feature map to obtain a feature vector corresponding to the time-frequency map; The fully trained interference feature vector extraction model is obtained by training with a plurality of training data pairs; each training data pair comprises a sample time-frequency graph and an interference type label of the sample time-frequency graph; According to the distance between each feature vector and each interference feature vector with known interference type in the interference feature vector library, the interference type of an interference signal data to be identified corresponding to the feature vector is determined; the interference feature vector library is constructed based on the feature vector extracted from the time-frequency diagram with known interference type according to the trained interference feature vector extraction model.

2. The SAR interference identification method based on feature space distance measurement according to claim 1 is characterized in that: The step of determining the interference type of a piece of interference signal data to be identified corresponding to each feature vector according to the distance between each feature vector and each interference feature vector with known interference type in the interference feature vector library comprises: For each feature vector in the at least one feature vector, calculating the Euclidean distance between the feature vector and each interference feature vector with a known interference type in the interference feature vector library, to obtain a plurality of Euclidean distances corresponding one to one to each interference feature vector; The interference type of the interference feature vector corresponding to the smallest Euclidean distance among the multiple Euclidean distances is used as the interference type of a piece of interference signal data to be identified corresponding to the feature vector.

3. The SAR interference identification method based on feature space distance measurement according to claim 1 is characterized in that: The fully trained interference feature vector extraction model includes, in sequence: a first convolution pooling layer, a second convolution pooling layer, a third convolution pooling layer and a fourth convolution pooling layer with the same convolution kernel size and pooling step size, as well as an attention module and a first global maximum pooling layer; the attention module includes, in sequence: a global average pooling layer, a second global maximum pooling layer, a sigmoid layer and an attention layer; the number of channels of the first convolution pooling layer, the second convolution pooling layer, the third convolution pooling layer and the fourth convolution pooling layer increases in sequence by the same multiple; the number of channels of the fourth convolution pooling layer, the first global maximum pooling layer, the global average pooling layer, the second global maximum pooling layer, the sigmoid layer and the attention layer are all the same; the step size of the first global maximum pooling layer, the global average pooling layer and the second global maximum pooling layer is the same.

4. The SAR interference identification method based on feature space distance measurement according to claim 3 is characterized in that: The convolution kernel sizes of the first convolution pooling layer, the second convolution pooling layer, the third convolution pooling layer and the fourth convolution pooling layer are all 3*3, and the pooling steps are all 2; the pooling steps of the first global maximum pooling layer, the global average pooling layer and the second global maximum pooling layer are all 8.

5. The SAR interference identification method based on feature space distance measurement according to claim 1 is characterized in that: Before inputting the at least one time-frequency graph into a well-trained interference feature vector extraction model and outputting at least one feature vector corresponding to the at least one time-frequency graph, the method further includes: Acquire signal data of N known interference types; each signal data of known interference type includes M pieces of signal data; N and M are both positive integers; Using the preset short-time Fourier transform parameters to perform short-time Fourier transform processing on N×M signal data, and correspondingly obtain N×M sample time-frequency diagrams; For each sample time-frequency graph x in the N×M sample time-frequency graphs i Add the corresponding interference type label l i ; i = 1, 2, ..., M × N; By adding [x i ,l i ] as a training data pair, and obtain N×M training data pairs; The N×M training data pairs, the preset loss function, the preset learning rate and the preset momentum value are used to perform multiple iterative training on the initial interference feature vector extraction model to obtain the fully trained interference feature vector extraction model.

6. The SAR interference identification method based on feature space distance measurement according to claim 5 is characterized in that: The N×M training data pairs, the preset loss function, the preset learning rate and the preset momentum value are used to perform multiple iterations of training on the initial interference feature vector extraction model to obtain the fully trained interference feature vector extraction model, including: During the k-th training, the k-1-th interference feature vector extraction model and the k-1-th momentum value are obtained, and B training data pairs are selected from the N×M training data pairs; k and B are both positive integers; when k=1, the k-1-th interference feature vector extraction model is the initial interference feature vector extraction model, and the k-1-th momentum value is the preset momentum value; Inputting the B training data pairs into the k-1th interference feature vector extraction model to obtain feature vectors of each time-frequency graph in the B training data pairs; Calculate the k-th loss value according to the feature vectors of each time-frequency graph, the interference type label of each time-frequency graph and the preset loss function; Calculate the gradient corresponding to each network parameter in the k-1th interference feature vector extraction model according to the k-th loss value; Calculate the k-th momentum value according to the gradient, the preset learning rate and the k-1-th momentum value; Update each network parameter in the k-1th interference feature vector extraction model according to the kth momentum value to obtain the kth interference feature vector extraction model, and iterate in this way until the k+nth interference feature vector extraction model is obtained, and use the k+nth interference feature vector extraction model as the fully trained interference feature vector extraction model; n is a preset value.

7. The SAR interference identification method based on feature space distance measurement according to claim 6 is characterized in that: The expression of the k-th momentum value is as follows: m k =0.9×m k-1 +p·grad; Among them, m k-1 is the k-1th momentum value, m k is the k-th momentum value, ρ is the preset learning rate, and grad is the gradient.

8. The SAR interference identification method based on feature space distance measurement according to claim 5, characterized in that: The expression of the preset loss function is as follows: in, represents the preset loss function, B is the number of training data pairs used in each training, is the feature vector of the time-frequency graph of the jth training data pair among the B training data pairs used in each training; is a feature vector of a time-frequency graph in a training data pair of the same interference type as the j-th training data pair among the B training data pairs used in each training; is the feature vector of the time-frequency graph of a training data pair with a different interference type from the j-th training data pair among the B training data pairs used in each training, j=1,2,...,B; [·] + represents the maximum value operation, α is the preset distance threshold, Represents the Euclidean distance between vectors.

9. The SAR interference identification method based on feature space distance measurement according to claim 5, characterized in that: Before determining the interference type of a piece of interference signal data to be identified corresponding to each feature vector according to the distance between each feature vector and each interference feature vector with known interference type in the interference feature vector library, the method further includes: Inputting the N×M sample time-frequency graphs into the fully trained interference feature vector extraction model to obtain one-to-one corresponding N×M interference feature vectors; The interference type labels of the N×M sample time-frequency diagrams are corresponded to the interference type labels of the N×M interference feature vectors to obtain N×M interference feature vectors with known interference types; the N×M interference feature vectors with known interference types constitute the interference feature vector library.

Images