A method and device for extracting multi-level features of radar signals

Abstract

This invention relates to the field of radar signal processing, and provides a method and apparatus for extracting multi-level features of radar signals. The method involves convolving the time-frequency map of the original DRFM relay signal at multiple feature extraction scales to extract local detail features at various locations, obtaining a first distinguishing feature; modeling the global information in the first distinguishing feature to obtain a first global feature; fusing the first distinguishing feature and the first global feature to obtain the time-frequency map feature; processing the bispectral map of the original DRFM relay signal using the same method to obtain the bispectral map feature; and fusing the time-frequency map feature and the bispectral map feature to obtain the multi-level features of the original DRFM relay signal. This method solves the problem that the accuracy of DRFM relay signal category recognition drops sharply due to severe aliasing and the lack of obvious artificial features in DRFM relay signals.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of radar signal processing, in particular to a method and device for extracting multi-level features of radar signals. BACKGROUND

[0002] With the continuous development of digital radio frequency memory (DRFM) technology, various DRFM retransmission signal patterns are emerging, the electromagnetic environment faced by radars is becoming increasingly complex, and fast and effective DRFM retransmission signals pose a serious threat to the detection capability of radars, which requires radars to have stronger environmental perception capability.

[0003] In order to realize the stronger environmental perception capability of radars, it is first necessary to accurately identify the category to which the DRFM retransmission signal belongs, and only by extracting the effective features of the DRFM retransmission signal can a better recognition effect be achieved.

[0004] In a complex electromagnetic environment, due to the severe aliasing of DRFM retransmission signals, artificial features are not obvious, and the feature extraction capability of the traditional method of the prior art is greatly reduced, resulting in a sharp decline in the recognition accuracy of the category of DRFM retransmission signals, which leads to the fact that in a complex electromagnetic environment, the radar system cannot accurately identify and filter out unnecessary DRFM retransmission signals. Among them, artificial features refer to the features of DRFM retransmission signals manually extracted by traditional methods, that is, based on prior knowledge or experience of DRFM retransmission signals, some characteristics of DRFM retransmission signals are described.

[0005] In view of this, overcoming the defects of the prior art is a problem to be solved in the technical field. SUMMARY

[0006] The technical problem to be solved by the present application is to provide a method and device for extracting multi-level features of radar signals, which aims to extract local detail features of original DRFM retransmission signals on multiple feature extraction scales, model corresponding global information, and extract more levels of feature information to obtain multi-level features of original DRFM retransmission signals, thereby solving the problem of sharp decline in recognition accuracy of signal categories due to severe aliasing of signals and non-obvious artificial features.

[0007] The present application adopts the following technical solutions:

[0008] In a first aspect, the present application provides a method for extracting multi-level features of radar signals, comprising:

[0009] obtaining a to-be-extracted time-frequency graph and a to-be-extracted bispectrum graph of an original DRFM retransmission signal;

[0010] convolve the to-be-extracted time-frequency graph on multiple feature extraction scales to extract local detail features of each position in the to-be-extracted time-frequency graph, to obtain first distinguishing features; model global information in the to-be-extracted time-frequency graph to obtain first global features; and fuse the first distinguishing features and the first global features to obtain time-frequency graph features;

[0011] convolve the to-be-extracted bispectrum graph on multiple feature extraction scales to extract local detail features of each position in the to-be-extracted bispectrum graph, to obtain second distinguishing features; model global information in the to-be-extracted bispectrum graph to obtain second global features; and fuse the second distinguishing features and the second global features to obtain bispectrum graph features;

[0012] fuse the time-frequency graph features and the bispectrum graph features to obtain multi-level features of the original DRFM retransmission signal.

[0013] Further, the convolving the to-be-extracted time-frequency graph on multiple feature extraction scales to extract local detail features of each position in the to-be-extracted time-frequency graph, to obtain first distinguishing features includes:

[0014] extracting local detail features in parallel on multiple spatial resolutions of the to-be-extracted time-frequency graph to obtain multiple intermediate features of different scales;

[0015] performing feature stitching on the multiple intermediate features of different scales to extract signal category information of the first to-be-processed feature under each spatial resolution, to obtain first distinguishing features.

[0016] Further, the extracting local detail features in parallel on multiple spatial resolutions of the to-be-extracted time-frequency graph to obtain multiple intermediate features of different scales includes:

[0017] convolving the to-be-extracted time-frequency graph using a first-size convolution kernel to obtain intermediate features of a first scale;

[0018] convolving the to-be-extracted time-frequency graph using a second-size convolution kernel to obtain first features, and convolving the first features using the first-size convolution kernel to obtain intermediate features of a second scale;

[0019] convolving the to-be-extracted time-frequency graph using a first-size convolution kernel to obtain second features, and convolving the second features using a third-size convolution kernel to obtain intermediate features of a third scale;

[0020] convolve the to-be-extracted time-frequency graph using a convolution kernel of a first size to obtain a third feature; convolve the third feature using a convolution kernel of a second size to obtain a fourth feature; and convolve the fourth feature using a convolution kernel of the second size to obtain an intermediate feature of a fourth scale.

[0021] Further, the modeling of the global information in the to-be-extracted time-frequency graph to obtain a first global feature comprises:

[0022] segmenting the to-be-extracted time-frequency graph into fixed-size tiles; and converting each of the tiles into a fixed-dimension to-be-processed vector;

[0023] determining an attention weight of the to-be-processed vector, and performing attention calculation on the to-be-processed vector using the attention weight to obtain a first encoding feature;

[0024] performing linear transformation and nonlinear activation on the first encoding feature to learn the first encoding feature using a feedforward neural network to obtain a second encoding feature;

[0025] performing an addition operation on the first encoding feature and the second encoding feature to obtain a first global feature.

[0026] Further, the determining of the attention weight of the to-be-processed vector and the attention calculation on the to-be-processed vector using the attention weight to obtain a first encoding feature comprises:

[0027] calculating a product of the to-be-processed vector and a query weight parameter to obtain a query vector; calculating a product of the to-be-processed vector and a key weight parameter to obtain an original key vector; and calculating a product of the to-be-processed vector and a value weight parameter to obtain an original value vector;

[0028] performing convolution operation on the original key vector in the spatial dimension according to a reduction ratio to obtain a first reduction feature; and performing linear projection operation on the first reduction feature using the reduction ratio to compress the channel dimension of the first reduction feature to obtain a target key vector;

[0029] performing convolution operation on the original value vector in the spatial dimension according to a reduction ratio to obtain a second reduction feature; and performing linear projection operation on the second reduction feature using the reduction ratio to compress the channel dimension of the second reduction feature to obtain a target value vector;

[0030] calculating a dot product of the query vector and the target key vector to obtain an attention weight;

[0031] using the attention weight for weighted summation on the target value vector to generate a first encoding feature.

[0032] Further, the fusing the first distinguishing feature and the first global feature to obtain a time-frequency map feature comprises:

[0033] extracting local detail features of the first distinguishing feature in the channel dimension and the spatial dimension respectively to obtain local attention weights;

[0034] extracting global dependency relationships between different positions in the first global feature to obtain channel attention features;

[0035] performing feature splicing on the local attention weights and the channel attention features to obtain a time-frequency map feature.

[0036] Further, the extracting local detail features of the first distinguishing feature in the channel dimension and the spatial dimension respectively to obtain local attention weights comprises:

[0037] performing average pooling on the first distinguishing feature in the channel dimension to obtain a first channel global feature; performing maximum pooling on the first distinguishing feature in the channel dimension to obtain a second channel global feature; inputting the first channel global feature and the second channel global feature into a shared fully connected layer to generate channel attention weights;

[0038] performing average pooling on the first distinguishing feature in the spatial dimension to obtain a first spatial global feature; performing maximum pooling on the first distinguishing feature in the spatial dimension to obtain a second spatial global feature; inputting the first spatial global feature and the second spatial global feature into a shared fully connected layer to generate spatial attention weights;

[0039] performing an addition operation on the channel attention weights and the spatial attention weights to obtain local attention weights.

[0040] Further, the obtaining a to-be-extracted time-frequency map and a to-be-extracted bispectrum map of the original DRFM forwarding signal comprises:

[0041] converting an original time-frequency map of the original DRFM forwarding signal into a gray-scale map to obtain a to-be-extracted time-frequency map;

[0042] converting an original bispectrum map of the original DRFM forwarding signal into a gray-scale map to obtain a to-be-extracted bispectrum map.

[0043] In a second aspect, the present application further provides a device for extracting multi-level features of a radar signal, comprising:

[0044] at least one processor; and a memory communicatively connected with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the processor to execute the method for extracting multi-level features of a radar signal according to the first aspect.

[0045] In a third aspect, the present application provides a non-volatile computer storage medium storing computer executable instructions, which, when executed by one or more processors, perform the method for extracting multi-level features of a radar signal according to the first aspect.

[0046] In a fourth aspect, a computer program product containing instructions, which, when executed on a computer or processor, cause the computer or processor to perform the method for extracting multi-level features of a radar signal according to the first aspect.

[0047] In a fifth aspect, the present application provides a system for extracting multi-level features of a radar signal, which comprises the device for extracting multi-level features of a radar signal according to the second aspect, and uses the method for extracting multi-level features of a radar signal according to the first aspect to complete the interaction of the device for extracting multi-level features of a radar signal according to the second aspect.

[0048] Compared with the prior art, the present application has at least the following beneficial effects:

[0049] The present application uses the time-frequency graph and the bispectrum graph of the DRFM retransmitted signal to extract the local detailed features of the DRFM retransmitted signal in multiple feature extraction scales, and models the global information of the DRFM retransmitted signal, captures the features of the original DRFM retransmitted signal at a deeper level by fusing the local detailed features and the global information, and finally obtains the multi-level features, thereby improving the robustness and precision of signal recognition in a complex electromagnetic environment, and solving the problem of sharp decline in recognition accuracy of signal categories due to serious signal aliasing and unobvious artificial features. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments of the present application. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can also be obtained according to these drawings without creative labor for those skilled in the art.

[0051] Figure 1 is a flowchart of a method for extracting multi-level features of a radar signal provided by an embodiment of the present application;

[0052] Figure 2 is a specific example diagram of a neural network structure of an embodiment of the present application;

[0053] Figure 3 is a flowchart of the first step 20 provided by an embodiment of the present application;

[0054] Figure 4 is a flowchart of the first step 201a provided by an embodiment of the present application;

[0055] Figure 5 is a specific example schematic diagram of a multi-scale module provided by an embodiment of the present application;

[0056] Figure 6 is a specific example schematic diagram of another neural network structure provided by an embodiment of the present application;

[0057] Figure 7 is a flowchart of the second step 20 provided by an embodiment of the present application;

[0058] Figure 8 is a flowchart of the step 202b provided by an embodiment of the present application;

[0059] Figure 9 is a flowchart of the third step 20 provided by an embodiment of the present application;

[0060] Figure 10 is a specific example schematic diagram of a device for extracting first distinguishing features in channel dimension and spatial dimension respectively provided by an embodiment of the present application;

[0061] Figure 11 is a specific example schematic diagram of a device for obtaining channel attention features provided by an embodiment of the present application;

[0062] Figure 12 is a schematic diagram of a device for extracting multi-level features of radar signals provided by an embodiment of the present application. DETAILED DESCRIPTION

[0063] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and should not be used to limit the present application.

[0064] Unless the context otherwise requires, throughout the specification and claims, the term "comprising" is interpreted as openly inclusive, meaning "including, but not limited to." In the description of the specification, terms such as "one embodiment," "some embodiments," "exemplary embodiment," "example," "specific example," or "some examples" are intended to indicate that a particular feature, structure, material, or characteristic associated with that embodiment or example is included in at least one embodiment or example of this disclosure. The illustrative representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics mentioned may be included in any suitable manner in any one or more embodiments or examples; that is, although they may be incorporated into embodiments or examples using the above terms for reasons such as order and position, it does not limit them to be incorporated in combination by a single embodiment or example.

[0065] In the description of this invention, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this disclosure and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this disclosure.

[0066] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of this disclosure, unless otherwise stated, "a plurality of" means two or more. Furthermore, for example, the description may use the prefix "A" or "B" to describe the same type of nouns as two independent entities. In this case, the corresponding features defined with "A" and "B" are used only to distinguish between similar entities and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features.

[0067] In describing some embodiments, the use of "coupled" or "coupling" and "connected" or "connecting" along with their derivatives, can be used. For example, some embodiments can be described as being "connected" where directly connected or physically connected with the two components that are "connected" directly physically touching one another. As another example, some embodiments can be described as being "coupled" where directly coupled or electrically coupled with the two components that are "coupled" directly electrically touching one another. However, it should be understood that "connected" or "coupled" can also mean that two or more components are not directly physically or electrically touching one another, but nonetheless still co-operate or interact with one another. For example, "optically coupled", "wirelessly connected", etc. The embodiments disclosed herein are not necessarily limited to the inventive concept.

[0068] In the description of the present application, the expression "A and / or B" (wherein A and B represent specific features) includes the following three combinations: A alone, B alone, and a combination of A and B.

[0069] As used herein, "about," "approximately," or "around" includes the recited value and the average value within an acceptable range of deviation from the recited value, as determined by one of ordinary skill in the art considering the measurement in question and the error in measuring the particular quantity (i.e., the limitations of the measurement system).

[0070] Furthermore, the technical features involved in each of the embodiments of the present application described below can be combined with each other as long as there is no conflict.

[0071] Embodiment 1:

[0072] In a complex electromagnetic environment, due to the development of modulation technology, multiple DRFM retransmission signals can enter the radar receiver at the same time, and the received composite DRFM retransmission signal increases the difficulty of identifying the DRFM retransmission signal. In the strong overlapping DRFM retransmission signal, the characteristics of the DRFM retransmission signal cover each other, and useless noise will pollute a large part of the feature image area. The ordinary convolutional neural network cannot capture useful features, and it is an urgent need to improve the feature extraction and characterization ability of the convolutional neural network.

[0073] Short-time Fourier transform captures instantaneous frequency changes through localized time-frequency windows. Since the frequency distribution of different types of DRFM retransmission signals at different time points is different, different DRFM retransmission signals can be better displayed in the time-frequency domain, and the difference characteristics between DRFM retransmission signals are intuitive.

[0074] However, for several types of DRFM signal such as range gate pull-off (RGPO), range-velocity gate pull-off (RGPO-VGPO), velocity gate pull-off (VGPO) and azimuth gate pull-off (AGPO), the frequency components of the corresponding signals evolve similarly over time, making it difficult to distinguish them in the time-frequency domain. For this type of signal, it is particularly important to design a method for analyzing and extracting effective features to achieve signal recognition.

[0075] In some cases, the difference between the DRFM signal and the real echo is reflected in the modulation information carried by the DRFM signal itself, which is generally derived from the phase quantization or amplitude quantization fingerprint characteristics left by the DRFM during modulation. In other cases, the DRFM signal and the real echo will exhibit differences in the time domain, frequency domain, time-frequency domain or other transform domains. Based on these subtle differences, some effective features can be extracted in each domain to improve the recognition effect of the DRFM signal.

[0076] In some solutions of the prior art, due to the phase quantization, a delay quantization effect is introduced, and the frequency modulation rate of the received signal is detected by using the fractional Fourier transform, the difference between the DRFM repeating signal and the real echo is compared, and the features are extracted to realize the identification of the DRFM repeating signal. Some solutions propose a digital radio frequency memory speed modulation DRFM repeating signal identification algorithm based on singular spectrum analysis, the difference between the harmonic components reflected on the singular value distribution of the real echo and the DRFM repeating signal is used, the singular spectrum analysis is performed on the received signal, the corresponding singular value statistical histogram is obtained by decomposition, and the related features are extracted from the singular value statistical histogram to realize the identification and classification of the DRFM repeating signal. Some solutions perform feature extraction on the DRFM repeating signal in the time-frequency domain, for example, the time-frequency image of the received signal is obtained by using the smoothed pseudo-Wigner-Ville distribution, the texture features of the time-frequency image are extracted by using the image local binary pattern algorithm, and the identification and classification of the spectrum dispersion DRFM repeating signal and the slice combination DRFM repeating signal are realized. Some solutions propose an identification algorithm based on the bispectrum of the range modulation signal, the speed modulation signal and the angle modulation signal, the bispectrum of the three kinds of DRFM repeating signals is estimated, the variance and the information entropy of the corresponding bispectrum distribution diagonal slice are taken as the feature parameters, and the classification of the three signals is completed. Some solutions perform bispectrum transformation on the real target echo and the three kinds of trailing DRFM repeating signals, use the dimension reduction method of Y-axis vertical projection and normalization to process the DRFM repeating signals, extract the Renyi entropy, exponential entropy and box dimension to form a feature set, put the feature set into a v-support vector classifier, obtain the corresponding feature vector and perform identification.

[0077] However, the above method can only extract the shallow features of the DRFM repeating signal in the time domain, the frequency domain, the time-frequency domain or other transform domains, and the deep features of the signal are not fully utilized, so that in a complex electromagnetic environment, the radar system cannot accurately identify and filter out unnecessary signals according to the extracted features.

[0078] In order to solve the above problems, as shown in Figure 1 The present application provides a method for extracting multi-level features of radar signals, comprising:

[0079] Step 10: obtaining the to-be-extracted time-frequency image and the to-be-extracted bispectrum image of the original DRFM repeating signal.

[0080] The to-be-extracted time-frequency image is obtained by preprocessing the time-frequency image of the original DRFM repeating signal, and the time-frequency image is the time-frequency domain image of the original DRFM repeating signal. The to-be-extracted bispectrum image is obtained by preprocessing the time-frequency image of the original DRFM repeating signal, and the bispectrum image is the bispectrum domain image of the original DRFM repeating signal.

[0081] In an embodiment, a gray image of the original time-frequency map of the original DRFM retransmission signal can be converted to obtain a time-frequency map to be extracted; a gray image of the original bispectrum map of the original DRFM retransmission signal can be converted to obtain a bispectrum map to be extracted.

[0082] In a complex electromagnetic environment, the received signal characteristics of a radar are various and complex, and a single dimension feature is difficult to distinguish the differences between composite DRFM retransmission signals. Therefore, multi-dimensional and multi-field features need to be extracted to describe the DRFM retransmission signal, which is helpful for subsequent identification of the type of the DRFM retransmission signal. The composite DRFM retransmission signal refers to a composite signal composed of multiple types of DRFM retransmission signals, that is, the original DRFM retransmission signal of the embodiment.

[0083] The embodiment uses the time-frequency map and the bispectrum map of the DRFM retransmission signal to obtain the features of the signal. The projection of the bispectrum map of the DRFM retransmission signal to two frequency components will form a plane, and the corresponding plane can reflect the phase coupling relationship of different frequency components in the DRFM retransmission signal. The brightness or color represents the phase coupling degree of the frequency pair, and the nonlinear interaction between different frequency components is revealed through the phase relationship. Since the bispectrum map can effectively suppress Gaussian noise while retaining the amplitude and phase information of the signal, for the DRFM retransmission signal with low distinguishability in the time-frequency domain, the embodiment introduces the corresponding bispectrum map to capture the nonlinear features of the DRFM retransmission signal. Through multi-level feature learning and enhancement, the effective complement and deep fusion of cross-domain (i.e., time-frequency domain and bispectrum domain) features are realized.

[0084] Step 20: convolve the to-be-extracted time-frequency map on multiple feature extraction scales to extract local detail features of each position in the to-be-extracted time-frequency map to obtain first distinguishing features; model global information in the to-be-extracted time-frequency map to obtain first global features; and fuse the first distinguishing features and the first global features to obtain time-frequency map features.

[0085] On the one hand, although different types of DRFM retransmission signals are difficult to distinguish in the time-frequency domain, there are essential differences in the subtle features and energy distribution in the time-frequency domain. Therefore, the embodiment adopts multi-scale convolution to extract local detail features of the to-be-extracted time-frequency map or the to-be-extracted bispectrum map on multiple feature extraction scales to enhance the sensitivity of the neural network of the embodiment to the local subtle features of the DRFM retransmission signal. In an embodiment, as shown in FIG. 8, after the to-be-extracted time-frequency map is input into the multi-scale module, the multi-scale module convolves the to-be-extracted time-frequency map on multiple feature extraction scales to realize extraction of local detail features of each position in the to-be-extracted time-frequency map to obtain first distinguishing features. Figure 2

[0086] ​On the other hand, the prior art cannot better extract features conducive to identifying DRFM-repeated signals by only extracting artificial features from signals, because the prior art cannot know the location of features important to the task of identifying DRFM-repeated signals in signals, and in DRFM-repeated signals, context information representing the nature of different signal types occupies a larger area than the signal itself. The embodiment of the present application models global information in the time-frequency graph to be extracted, and then extracts key features conducive to identifying DRFM-repeated signals, i.e., first global features; and uses the key features as context to combine local detailed features (i.e., first distinguishing features or second distinguishing features in the following context) to obtain features (i.e., time-frequency domain features or bispectrum domain features in the following context) conducive to identifying DRFM-repeated signals from the time-frequency graph to be extracted or the bispectrum graph to be extracted. In one embodiment, as shown in Figure 2 the global information in the time-frequency graph to be extracted is modeled by a global feature module after the time-frequency graph to be extracted is input into a multi-scale module, and first global features are obtained.

[0087] It should be noted that the multi-scale module and the global feature module are two independent neural network modules, and do not share network parameters.

[0088] Step 30: convolve the bispectrum graph to be extracted at multiple feature extraction scales to extract local detailed features of each position in the bispectrum graph to be extracted, obtain second distinguishing features, model global information in the bispectrum graph to be extracted, obtain second global features, and fuse the second distinguishing features and the second global features to obtain bispectrum graph features.

[0089] As shown in Figure 2 the processing flow of the time-frequency graph to be extracted and the bispectrum graph to be extracted is the same in the embodiment of the present application.

[0090] Step 40: fuse the time-frequency graph features and the bispectrum graph features to obtain multi-level features of the original DRFM-repeated signal.

[0091] The present application uses the time-frequency graph and the bispectrum graph of the original DRFM-repeated signal, extracts local detailed features of the signal at multiple feature extraction scales, models global information of the signal, captures features of the original DRFM-repeated signal at a deeper level by fusing local detailed features and global information, and finally obtains multi-level features, thereby improving the robustness and accuracy of signal recognition in a complex electromagnetic environment, and solving the problem of a sharp decline in the accuracy of signal classification due to severe signal aliasing and unobvious artificial features.

[0092] The following example illustrates the method for extracting multi-level features of radar signals according to an embodiment of the present invention. The process of obtaining bispectral features is the same as that of obtaining time-frequency map features, and will not be repeated here.

[0093] The following describes the processing flow of the multi-scale module for extracting local detail features of the signal at multiple feature extraction scales, such as... Figure 3 As shown, in step 20, the convolution of the time-frequency image to be extracted at multiple feature extraction scales to extract local detail features at various locations in the time-frequency image to obtain the first distinguishing feature includes:

[0094] Step 201a: Extract local detail features in parallel at multiple spatial resolutions of the time-frequency map to be extracted, and obtain multiple intermediate features at different scales.

[0095] like Figure 5 and Figure 6 As shown, by using convolutional kernels of different sizes to perform convolution operations on the time-frequency graph to be processed, different time and frequency resolutions can be obtained. Therefore, compared with traditional frequency domain or time domain analysis methods, the embodiments of the present invention can extract information of the original DRFM forwarding signal at different scales.

[0096] In one embodiment, such as Figure 6 As shown, the methods for extracting local detail features in parallel at multiple spatial resolutions of the time-frequency map to be extracted can be as follows:

[0097] (1) A 1×1 two-dimensional convolution kernel is used to perform a convolution operation on the time-frequency image to be extracted (i.e., Figure 6 The intermediate features of the first scale are obtained by using “1×1Conv”.

[0098] (2) First, a 3×3 two-dimensional convolution kernel is used to perform a convolution operation on the time-frequency image to be extracted (i.e., Figure 6 The "3×3Conv" in the diagram is used to obtain the initial extracted feature map; then, a 1×1 two-dimensional convolution kernel is used to perform a convolution operation on the initial extracted feature map (i.e., Figure 6 The intermediate features of the second scale are obtained by using “1×1Conv” in the model.

[0099] (3) First, perform a convolution operation on the time-frequency image to be extracted using a 1×1 two-dimensional convolution kernel (i.e., Figure 6 The "1×1Conv" in the code is used to obtain the initial extracted feature map; then, a 5×5 two-dimensional convolution kernel is used to perform a convolution operation on the initial extracted feature map (i.e., ...). Figure 6 The intermediate features of the third scale are obtained by using “5×5Conv”.

[0100] (4) first use a two-dimensional convolution kernel with a size of 1x1 to perform convolution operation on the to-be-extracted time-frequency graph (i.e., "1x1Conv" in Figure 6 ), to obtain a preliminary extracted feature map; then use a two-dimensional convolution kernel with a size of 3x3 to perform convolution operation on the preliminary extracted feature map (i.e., "3x3Conv" in Figure 6 ), to obtain a further extracted feature map; and finally use a two-dimensional convolution kernel with a size of 3x3 again to perform convolution operation on the further extracted feature map (i.e., "3x3Conv" in Figure 6 ), to obtain an intermediate feature of a fourth scale.

[0101] It should be noted that the above-mentioned convolution operation using a two-dimensional convolution kernel refers to: performing standard convolution operation on the corresponding feature map, then performing batch normalization (Batch Normalization, abbreviated as BN) on the features obtained by convolution, and using a rectified linear unit (Rectified Linear Unit, abbreviated as ReLu) activation function. The specific manner of standard convolution operation is selected by a person skilled in the art in combination with experience according to the specific use scene, which is not limited here.

[0102] Step 202a: performing feature splicing on the intermediate features of the plurality of different scales to extract signal category information of the first to-be-processed feature at each spatial resolution, to obtain a first distinguishing feature.

[0103] In an embodiment, as shown in Figure 6 , the feature splicing is performed on the intermediate features of the four different scales to obtain a first to-be-processed feature, and then a maximum pooling operation is performed on the first to-be-processed feature to obtain a corresponding first distinguishing feature.

[0104] In an optional embodiment, for the plurality of different scale convolution kernels used above, an asymmetric convolution manner can also be used for a single convolution kernel, that is, by decomposing a traditional large-size convolution kernel (for example, a convolution kernel with a size of 7x7) into two smaller asymmetric convolution kernels (1x7 and 7x1), to effectively enhance the feature expression ability and diversity of the first distinguishing feature, and significantly reduce the model parameter amount in the extraction process, thereby improving the training and inference efficiency.

[0105] The embodiment of the present application extracts features in parallel on different spatial resolutions of the to-be-extracted time-frequency graph, combines different sizes of convolution kernels and maximum pooling operations, and captures local detailed features of the DRFM forwarded signal. Specifically, as shown in Figure 7 , the step 201a includes:

[0106] Step 2011: convolve the to-be-extracted time-frequency graph using a convolution kernel of a first size to obtain an intermediate feature of a first scale.

[0107] The first size is selected by a person skilled in the art according to a specific use scenario; for example, as shown in the figure, in an embodiment, the first size can be 1x1, that is, a 1x1 convolution kernel is used for convolution to obtain an intermediate feature of a size of Figure 5 .

[0108] Step 2012: convolve the to-be-extracted time-frequency graph using a convolution kernel of a second size to obtain a first feature; and convolve the first feature using a convolution kernel of the first size to obtain an intermediate feature of a second scale.

[0109] The second size is selected by a person skilled in the art according to a specific use scenario; for example, as shown in the figure, in an embodiment, the second size can be 3x3, that is, a 3x3 convolution kernel and a 1x1 convolution kernel are used in sequence for convolution to obtain an intermediate feature of a size of Figure 5 .

[0110] Step 2013: convolve the to-be-extracted time-frequency graph using a convolution kernel of the first size to obtain a second feature; and convolve the second feature using a convolution kernel of a third size to obtain an intermediate feature of a third scale.

[0111] The third size is selected by a person skilled in the art according to a specific use scenario; for example, as shown in the figure, in an embodiment, the third size can be 5x5, that is, a 1x1 convolution kernel and a 5x5 convolution kernel are used in sequence for convolution to obtain an intermediate feature of a size of Figure 5 .

[0112] Step 2014: convolve the to-be-extracted time-frequency graph using a convolution kernel of the first size to obtain a third feature; convolve the third feature using a convolution kernel of the second size to obtain a fourth feature; and convolve the fourth feature using a convolution kernel of the second size to obtain an intermediate feature of a fourth scale.

[0113] As shown in the figure, in an embodiment, a 1x1 convolution kernel, a 3x3 convolution kernel, and a 3x3 convolution kernel are used in sequence for convolution to obtain an intermediate feature of a size of Figure 5 .

[0114] ​​​​The embodiment of the present application captures rich features with signal class discriminativeness from fine textures to larger structures through multi-scale convolution, so that the neural network model of the embodiment of the present application can efficiently integrate multiple scale information without significantly increasing the computational burden. However, convolution is limited by its local characteristics, and the receptive field of the convolution kernel of different scales is limited in the local neighborhood, which cannot capture the global context information in the signal, and the global context information is often necessary for better identifying complex signal patterns. Therefore, the embodiment of the present application also captures local detail features of the signal and models global context information, so that the neural network model of the embodiment of the present application uses the global context information to guide the learning of each local detail feature, and then better learns the features that help to identify the original DRFM forwarding signal. The processing flow of the global feature module modeling the global information of the signal is described below, as shown in Figure 7 The step 20 of modeling the global information in the to-be-extracted time-frequency graph includes:

[0115] Step 201b: dividing the to-be-extracted time-frequency graph into fixed-size graph blocks; converting each graph block into a fixed-dimension to-be-processed vector.

[0116] In an embodiment, as shown in Figure 6 After the input image (i.e., the to-be-extracted time-frequency graph or the to-be-extracted bispectrum graph) is divided into fixed-size graph blocks by a linear layer, each graph block is converted into a fixed-dimension to-be-processed vector (i.e., token) by a line normalization layer, and position encoding representing the relative position of each graph block in the input image is added.

[0117] Step 202b: determining the attention weight of the to-be-processed vector, and performing attention calculation on the to-be-processed vector using the attention weight to obtain the first encoded feature.

[0118] Then the features of the to-be-processed vector are learned by the encoder, which will be described below.

[0119] Step 203b: performing linear transformation and nonlinear activation on the first encoded feature to learn the first encoded feature using a feedforward neural network to obtain a second encoded feature.

[0120] The specific way of learning the first encoded feature using the feedforward neural network is determined by those skilled in the art according to the specific use scenario and reference to the prior art, which is not limited here.

[0121] Step 204b: performing addition operation on the first encoded feature and the second encoded feature to obtain the first global feature.

[0122] The global feature module of the embodiment of the present application can model the global relationship of the entire input image, can capture long-range dependencies and global information, and the global features extracted by the global feature module supplement the distinguishing features captured by the multi-scale module.

[0123] As shown in Figure 6 , the embodiment of the present application captures the global dependency relationship between different positions in the sequence of the to-be-processed vector by calculating the weighted sum of the query vector, the key vector and the value vector using the multi-head attention mechanism, and specifically, as shown in Figure 8 , the step 202b includes:

[0124] Step 2021: calculating the product of the to-be-processed vector and the query weight parameter to obtain a query vector; calculating the product of the to-be-processed vector and the key weight parameter to obtain an original key vector; calculating the product of the to-be-processed vector and the value weight parameter to obtain an original value vector.

[0125] Wherein, the query weight parameter, the key weight parameter and the value weight parameter are weight matrices.

[0126] Step 2022: performing convolution operation on the original key vector in the spatial dimension according to the reduction ratio to obtain a first reduced feature; performing linear projection operation on the first reduced feature using the reduction ratio to compress the channel dimension of the first reduced feature to obtain a target key vector.

[0127] Wherein, the reduction ratio is selected by a person skilled in the art according to the specific use scenario. In an embodiment, the original key vector is reduced to 1 / R of the original size, wherein 1 / R is the reduction ratio.

[0128] Step 2023: performing convolution operation on the original value vector in the spatial dimension according to the reduction ratio to obtain a second reduced feature; performing linear projection operation on the second reduced feature using the reduction ratio to compress the channel dimension of the second reduced feature to obtain a target value vector.

[0129] For example, for a feature map with a size of HxW, the spatial dimensions of K and V are reduced to Wherein, H is the length of the feature map, and W is the width of the feature map.

[0130] Step 2024: calculating the dot product of the query vector and the target key vector to obtain an attention weight.

[0131] Step 2025: using the attention weight for weighted sum of the target value vector to generate a first encoding feature.

[0132] In the calculation of the weighted sum of the query vector, the key vector and the value vector to obtain the first encoding feature, the dot product of the query vector and the target key vector is calculated first to obtain the attention weight, and then the weighted sum of the attention weight and the value vector is calculated, which reduces the spatial dimension of the calculation, effectively alleviates the problems of calculation complexity and memory consumption caused by high-dimensional data, and can greatly reduce the calculation memory overhead; so that the embodiment of the application can use the multi-head attention mechanism to capture long-distance interaction without increasing the number of parameters, and further enable the neural network model of the embodiment of the application to more efficiently process large-scale DRFM forwarding signal data, greatly improve the efficiency of extracting global features in the training and reasoning process, and not sacrifice the representation ability of the model.

[0133] As shown in Figure 6 , the embodiment of the application further effectively merges the local detail features captured by the multi-scale module and the global information extracted by the global feature module through the fusion module to further enhance the overall feature representation, specifically, as shown in Figure 9 , in step 20, the first local detail feature and the first global feature are fused to obtain the time-frequency graph feature, which includes:

[0134] Step 201c: Extracting local detail features of the first distinguishing feature in the channel dimension and the spatial dimension respectively to obtain local attention weights.

[0135] As shown in Figure 10 , in one embodiment, the first distinguishing feature is averaged and pooled in the channel dimension to obtain a first channel global feature; the first distinguishing feature is maximally pooled in the channel dimension to obtain a second channel global feature; and the first channel global feature and the second channel global feature are input into a shared fully connected layer to generate channel attention weights. The shared fully connected layer is used to fuse the first channel global feature and the second channel global feature, and the specific implementation of the shared fully connected layer is selected by a person skilled in the art according to the specific use scenario, and in one optional embodiment, a multilayer perceptron (MLP) can be used for implementation. Similarly, the first distinguishing feature is averaged and pooled in the spatial dimension to obtain a first spatial global feature; the first distinguishing feature is maximally pooled in the spatial dimension to obtain a second spatial global feature; and the first spatial global feature and the second spatial global feature are input into a shared fully connected layer to generate spatial attention weights. Finally, the channel attention weights and the spatial attention weights are added to obtain local attention weights.

[0136] Since the local detail features captured by the scaling module (i.e., the first or second distinguishing features) contain shallow features of the DRFM forwarding signal, and these shallow features often contain noise, this embodiment of the invention utilizes spatial attention as a spatial filter to enhance local details and suppress irrelevant regions. Appropriate attention weights are calculated in both spatial and channel dimensions to adaptively enhance key features that help identify the DRFM forwarding signal category, thereby guiding the neural network model of this embodiment to learn more discriminative features during training and inference.

[0137] Step 202c: Extract the global dependencies between different positions in the first global feature to obtain the channel attention feature.

[0138] In one embodiment, such as Figure 11 As shown, the first global feature is first subjected to average pooling (i.e., Figure 11 The Avg-Pool method in the feature map uses average pooling to average the local regions of each channel to extract global information from the feature map and reduce its spatial dimension, thereby reducing the spatial resolution of the feature map and the computational cost of subsequent operations. Then, a fully connected layer (i.e., ...) is used... Figure 11 The Fc function in the fully connected layer performs a linear transformation on the compressed feature map, mapping the input features to another space. Then, a non-linear activation function (e.g., Fc) is applied after the fully connected layer. Figure 11 The circle containing "S" represents the Sigmoid function. By introducing nonlinearity through a nonlinear activation function, the expressive power of the neural network model in this embodiment of the invention is enhanced.

[0139] The channel attention feature in this embodiment of the invention represents the importance of different channels in the feature map, thereby enhancing the neural network model's attention to important features.

[0140] Step 203c: Perform feature concatenation on the local attention weights and the channel attention features to obtain time-frequency map features.

[0141] At the same time, by focusing on the relationship between channels in the feature map through channel attention weights, we can ensure that the most important channels captured by the global feature module are highlighted and global information is effectively presented.

[0142] The present application can extract more deep-level features of the DRFM repeated signal in the complex electromagnetic environment, improve the accuracy and diversity of feature extraction, and further improve the recognition accuracy of the signal category. In addition, the adaptability of the neural network model of the embodiment of the present application in the complex electromagnetic environment can be greatly improved, especially when facing strong signal overlap, effective information can be extracted from the complex background.

[0143] Embodiment 2:

[0144] As Figure 12 shown in FIG. 1, which is a schematic diagram of an architecture of the device for extracting multi-level features of radar signals according to an embodiment of the present application. The device for extracting multi-level features of radar signals according to the embodiment of the present application comprises one or more processors 21 and a memory 22. Among them, Figure 12 The processor 21 is taken as an example in the embodiment.

[0145] The processor 21 and the memory 22 can be connected through a bus or other means, Figure 12 The connection through the bus is taken as an example in the embodiment.

[0146] The memory 22 is a kind of non-volatile computer readable storage medium, which can be used to store non-volatile software programs and non-volatile computer executable programs, such as the method for extracting multi-level features of radar signals in the embodiment. The processor 21 executes the non-volatile software programs and instructions stored in the memory 22, thereby executing the method for extracting multi-level features of radar signals.

[0147] The memory 22 can include a high-speed random access memory, and can also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 22 can optionally include a memory remotely arranged with respect to the processor 21, and these remote memories can be connected to the processor 21 through a network. Examples of the above-mentioned network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0148] The program instructions / modules are stored in the memory 22, and when executed by the one or more processors 21, the method for extracting multi-level features of radar signals in the above-mentioned embodiment is executed, for example, each step of the method for extracting multi-level features of radar signals according to the embodiment of the present application described above is executed.

[0149] The embodiment of the present application also provides a non-volatile computer storage medium, which stores computer executable instructions, and the computer executable instructions are executed by one or more processors, for example Figure 12one processor 21, so that the one or more processors can execute the method for extracting multi-level features of radar signals in the embodiments of the present application, for example, execute each step of the method for extracting multi-level features of radar signals in the embodiments of the present application described above; and also can realize Figure 12 the various modules, units, or execute the method for extracting multi-level features of radar signals in the embodiments of the present application, for example, execute each step of the method for extracting multi-level features of radar signals in the embodiments of the present application described above; and also can realize Figure 12 the various modules, units.

[0150] It is worth noting that the information interaction, execution process and the like between the modules and units in the above apparatus and system are based on the same concept as the processing method embodiments of the present application, and the specific content can be referred to the description in the method embodiments of the present application, which will not be described here.

[0151] Those skilled in the art can understand that all or part of the steps in the various methods of the embodiments can be completed by a program instructing the related hardware, and the program can be stored in a computer readable storage medium, which can include read only memory (ROM), random access memory (RAM), magnetic disk or optical disk, etc.

[0152] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for extracting multi-level features of radar signals, characterized in that, include: Obtain the time-frequency plot and bispectral plot of the original DRFM forwarding signal to be extracted; Local detail features are extracted in parallel at multiple spatial resolutions of the time-frequency map to be extracted, resulting in multiple intermediate features at different scales. The intermediate features at multiple different scales are concatenated to extract the signal category information of the first feature to be processed at each spatial resolution, thereby obtaining the first distinguishing feature; The time-frequency image to be extracted is segmented into fixed-size patches; each patch is converted into a fixed-dimensional vector to be processed; the attention weights of the vectors to be processed are determined, and attention calculation is performed on the vectors to be processed using the attention weights to obtain a first encoded feature; the first encoded feature is subjected to linear transformation and nonlinear activation to learn the first encoded feature using a feedforward neural network to obtain a second encoded feature; the first encoded feature and the second encoded feature are added together to obtain a first global feature. Local detail features of the first distinguishing feature are extracted in both the channel dimension and the spatial dimension to obtain local attention weights; global dependencies between different positions in the first global feature are extracted to obtain channel attention features; and the local attention weights and the channel attention features are concatenated to obtain time-frequency map features. The bispectral image to be extracted is convolved at multiple feature extraction scales to extract local detail features at various locations in the bispectral image to obtain the second distinguishing feature; The global information in the bispectral image to be extracted is modeled to obtain the second global feature; By fusing the second distinguishing feature and the second global feature, a dual-spectral feature is obtained; By fusing the time-frequency plot features and the bispectral plot features, the multi-level features of the original DRFM forwarding signal are obtained.

2. The method for extracting multi-level features of radar signals according to claim 1, characterized in that, The parallel extraction of local detail features at multiple spatial resolutions of the time-frequency map to be extracted, resulting in multiple intermediate features at different scales, includes: The time-frequency image to be extracted is convolved using a convolution kernel of the first size to obtain intermediate features of the first scale; The time-frequency image to be extracted is convolved using a convolution kernel of the second size to obtain a first feature; the first feature is then convolved using a convolution kernel of the first size to obtain an intermediate feature of the second scale. The time-frequency image to be extracted is convolved using a convolution kernel of the first size to obtain a second feature; the second feature is then convolved using a convolution kernel of the third size to obtain an intermediate feature of the third scale. The time-frequency image to be extracted is convolved using a convolution kernel of the first size to obtain a third feature; the third feature is convolved using a convolution kernel of the second size to obtain a fourth feature; and the fourth feature is convolved using a convolution kernel of the second size to obtain an intermediate feature of the fourth scale.

3. The method for extracting multi-level features of radar signals according to claim 1, characterized in that, The step of determining the attention weights of the vector to be processed, and using the attention weights to perform attention calculations on the vector to be processed to obtain the first encoded feature includes: Calculate the product of the vector to be processed and the query weight parameter to obtain the query vector; calculate the product of the vector to be processed and the key weight parameter to obtain the original key vector; calculate the product of the vector to be processed and the value weight parameter to obtain the original value vector. The original key vector is convolved in the spatial dimension according to the reduction ratio to obtain the first reduced feature; the first reduced feature is then linearly projected using the reduction ratio to compress the channel dimension of the first reduced feature to obtain the target key vector. The original value vector is convolved in the spatial dimension according to the reduction ratio to obtain the second reduced feature; the second reduced feature is then linearly projected using the reduction ratio to compress the channel dimension of the second reduced feature to obtain the target value vector. Calculate the dot product between the query vector and the target key vector to obtain the attention weights; The attention weights are used to perform a weighted summation of the target value vector to generate a first encoded feature.

4. The method for extracting multi-level features of radar signals according to claim 1, characterized in that, The step of extracting local detail features of the first distinguishing feature in both the channel dimension and the spatial dimension to obtain local attention weights includes: The first discriminative feature is average pooled along the channel dimension to obtain the first channel global feature; the first discriminative feature is max pooled along the channel dimension to obtain the second channel global feature; the first channel global feature and the second channel global feature are input into a shared fully connected layer to generate channel attention weights. The first distinguishing feature is average pooled in the spatial dimension to obtain the first spatial global feature; the first distinguishing feature is max pooled in the spatial dimension to obtain the second spatial global feature; the first spatial global feature and the second spatial global feature are input into a shared fully connected layer to generate spatial attention weights. The local attention weight is obtained by adding the channel attention weight and the spatial attention weight.

5. The method for extracting multi-level features of radar signals according to any one of claims 1-4, characterized in that, The process of obtaining the time-frequency diagram and bispectral diagram to be extracted from the original DRFM forwarding signal includes: The original time-frequency diagram of the original DRFM forwarding signal is converted into a grayscale image to obtain the time-frequency diagram to be extracted; The original bispectral image of the original DRFM forwarding signal is converted into a grayscale image to obtain the bispectral image to be extracted.

6. A device for extracting multi-level features of radar signals, characterized in that, The apparatus for extracting multi-level features of radar signals includes at least one processor and a memory, which are connected via a data bus. The memory stores instructions that can be executed by the at least one processor. After being executed by the processor, the instructions are used to implement the method for extracting multi-level features of radar signals according to any one of claims 1-5.

7. A non-volatile computer storage medium, characterized in that, The computer storage medium stores computer-executable instructions, which are executed by one or more processors to perform the method for extracting multi-level features of radar signals as described in any one of claims 1-5.

Images