Feature extraction method and device for composite signal

Abstract

The invention relates to the field of radar signal processing, and provides a composite signal feature extraction method and device. The method comprises the following steps: respectively extracting local detail features and global context information of a to-be-extracted time-frequency graph and a to-be-extracted double-spectrum graph of a composite signal to obtain time-frequency graph features and double-spectrum graph features, carrying out channel exchange on the time-frequency graph features and the double-spectrum graph features, and determining effective channel features corresponding to each channel; obtaining a first fusion feature according to all the effective channel features; respectively extracting local detail features of each position in the to-be-extracted time-frequency graph and the to-be-extracted bispectral graph to obtain a first distinguishing feature and a second distinguishing feature, and performing feature fusion on the first distinguishing feature and the second distinguishing feature based on the attention weight to obtain a second fusion feature; and performing feature splicing on the first fusion feature and the second fusion feature to obtain a target feature, thereby solving the problem of relatively low recognition rate of the composite signal under the condition of relatively low signal-to-noise ratio.

Description

Technical Field

[0001] The present invention relates to the field of radar signal processing technology, and in particular to a method and device for extracting features of a composite signal. Background Art

[0002] The forwarding signal of the Digital Radio Frequency Memory (DRFM) is used to reduce the radar's ability to detect target signals, or even make the radar completely ineffective. In modern radar systems, the characteristics of the DRFM forwarding signal are often highly overlapping and complex, that is, the radar often detects a composite signal; in this scenario, the radar is likely to be unable to detect the target signal.

[0003] Traditional time domain or frequency domain analysis methods often cannot effectively handle this problem, especially when the DRFM forwarding signal has structural features of multiple scales. The difficulty of feature extraction will increase with the increase of noise; when the noise power increases to a certain extent, the noise will completely submerge the DRFM forwarding signal. In this case, it is difficult to extract effective features from the composite signal detected by the radar. Under the conditions of low signal-to-noise ratio and complex signals, the signal quality of the composite signal that the radar can detect is poor, resulting in low recognition ability of the composite signal, which seriously affects the normal operation of the radar.

[0004] In view of this, overcoming the defects of the prior art is an urgent problem to be solved in the field of this technology. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a feature extraction method and device for a composite signal, the purpose of which is to perform inter-domain fusion of the time-frequency diagram features of the composite signal in the time-frequency domain and the dual-spectrum diagram features in the dual-spectrum domain, so as to better extract effective features, improve the recognition ability of the composite signal, and solve the problem that when the signal-to-noise ratio is low, the signal quality of the composite signal that can be detected by the radar is poor, resulting in a low recognition rate of the composite signal.

[0006] The present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a method for extracting features of a composite signal, comprising:

[0008] Obtain a time-frequency graph to be extracted and a bispectrum graph to be extracted of the composite signal; extract local detail features and global context information of the time-frequency graph to be extracted to obtain time-frequency graph features; extract local detail features and global context information of the bispectrum graph to be extracted to obtain bispectrum features;

[0009] Perform channel swapping on the time-frequency map features and the bispectrum map features to determine the effective channel features corresponding to each channel in the time-frequency map features and the bispectrum map features; obtain the first fusion feature based on all the effective channel features;

[0010] Extract the local detail features at each position in the time-frequency map to be extracted to obtain the first discrimination feature; extract the local detail features at each position in the bispectrum map to be extracted to obtain the second discrimination feature; perform feature fusion on the first discrimination feature and the second discrimination feature based on the attention weights to obtain the second fusion feature;

[0011] Perform feature splicing on the first fusion feature and the second fusion feature to obtain the target feature of the composite signal.

[0012] Further, the performing channel swapping on the time-frequency map features and the bispectrum map features to determine the effective channel features corresponding to each channel in the time-frequency map features and the bispectrum map features; obtaining the first fusion feature based on all the effective channel features includes:

[0013] Perform batch normalization on the time-frequency map features to obtain the first feature map to be evaluated; perform batch normalization on the bispectrum map features to obtain the second feature map to be evaluated;

[0014] If the channel features in the first feature map to be evaluated overlap with the channel features at the corresponding positions in the second feature map to be evaluated, then determine the overlapping channel features as the effective channel features in the first feature map to be evaluated and the effective channel features in the second feature map to be evaluated;

[0015] If the channel features in the first feature map to be evaluated do not overlap with the channel features at the corresponding positions in the second feature map to be evaluated, then determine the non-overlapping channel features in the first feature map to be evaluated as the first original channel features, and determine the non-overlapping channel features in the second feature map to be evaluated as the second original channel features; determine whether the first original channel features are effective channel features, determine whether the second original channel features are effective channel features, and perform channel swapping on the original channel features that are not effective channel features to obtain the corresponding effective channel features;

[0016] Determine all the effective channel features in the first feature map to be evaluated as the time-frequency intermediate features; determine all the effective channel features in the second feature map to be evaluated as the bispectrum intermediate features; perform feature splicing on the time-frequency intermediate features and the bispectrum intermediate features to obtain the first fusion feature.

[0017] Further, the performing channel swapping on the original channel features that are not effective channel features to obtain the corresponding effective channel features includes:

[0018] When the first original channel feature is not a valid channel feature, calculate the mean of the channel features of all channels in the current second feature map to be evaluated, obtain a first feature mean, and determine the first feature mean as the valid channel feature of the channel in the current first feature map to be evaluated, so as to perform channel swapping on the first original channel feature;

[0019] When the second original channel feature is not a valid channel feature, calculate the mean of the channel features of all channels in the current first feature map to be evaluated, obtain a second feature mean, and determine the second feature mean as the valid channel feature of the channel in the current second feature map to be evaluated, so as to perform channel swapping on the first feature map to be evaluated and the second feature map to be evaluated, so as to perform channel swapping on the second original channel feature.

[0020] Further, the determining whether the first original channel feature is a valid channel feature includes:

[0021] When performing batch normalization on the time-frequency map feature, set a scaling factor parameter for each channel in advance;

[0022] Add a penalty function to the total loss function to use the penalty function to constrain the sparsity of the scaling factor parameter; wherein, the total loss function is used to extract the target feature;

[0023] When the scaling factor parameter of the channel is lower than the threshold, the first original channel feature is not a valid channel feature;

[0024] When the scaling factor parameter of the channel is not lower than the threshold, the first original channel feature is a valid channel feature.

[0025] Further, the expression of the valid channel feature is:

[0026]

[0027] wherein, f c represents the first original channel feature of the c-th channel of the time-frequency map feature, f c ′ represents the valid channel feature corresponding to the c-th channel, μ c represents the mean of the feature values of all channels in the current second feature map to be evaluated, σ c represents the standard deviation of the feature values of all channels in the current second feature map to be evaluated, γ c represents the scaling factor parameter, β c represents the bias parameter, and ε is a constant.

[0028] Further, the feature fusion of the first distinguishing feature and the second distinguishing feature based on the attention weight to obtain the second fusion feature includes:

[0029] Calculate the dot product of each position in the first distinguishing feature using multiple asymmetric convolution kernels to extract the spatial features of the first distinguishing feature and obtain the first deep feature; calculate the dot product of each position in the second distinguishing feature using multiple asymmetric convolution kernels to extract the spatial features of the second distinguishing feature and obtain the second deep feature;

[0030] Perform feature splicing on the first deep feature and the second deep feature to obtain the feature to be processed; divide the feature to be processed into tiles of a fixed size; convert each tile into a vector to be processed with a fixed dimension;

[0031] Determine the attention weight of the vector to be processed, and perform attention calculation on the vector to be processed using the attention weight to obtain the first encoded feature;

[0032] Perform linear transformation and non-linear activation on the first encoded feature to learn the first encoded feature using a feed-forward neural network and obtain the second encoded feature;

[0033] Perform an addition operation on the first encoded feature and the second encoded feature to obtain the second fusion feature.

[0034] Further, the determining the attention weight of the vector to be processed, and performing attention calculation on the vector to be processed using the attention weight to obtain the first encoded feature includes:

[0035] 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;

[0036] Perform a convolution operation on the original key vector in the spatial dimension according to the reduction ratio to obtain the first reduced feature; perform a linear projection operation on the first reduced feature using the reduction ratio to compress the channel dimension of the first reduced feature and obtain the target key vector;

[0037] Perform a convolution operation on the original value vector in the spatial dimension according to the reduction ratio to obtain the second reduced feature; perform a linear projection operation on the second reduced feature using the reduction ratio to compress the channel dimension of the second reduced feature and obtain the target value vector;

[0038] Calculate the dot product of the query vector and the target key vector to obtain the attention weight;

[0039] Weighted sum of the target value vectors is performed using the attention weights to generate a first encoded feature.

[0040] Further, obtaining the to-be-extracted time-frequency diagram and to-be-extracted bispectrum diagram of the composite signal; extracting local detail features and global context information of the to-be-extracted time-frequency diagram to obtain time-frequency diagram features; extracting local detail features and global context information of the to-be-extracted bispectrum diagram to obtain bispectrum diagram features includes:

[0041] Converting the original time-frequency diagram of the composite signal into a grayscale image to obtain the to-be-extracted time-frequency diagram; performing convolution on the to-be-extracted time-frequency diagram at multiple feature extraction scales to determine first discriminative features; modeling the global context information in the to-be-extracted time-frequency diagram to obtain first global features; fusing the first discriminative features and the first global features to obtain time-frequency diagram features;

[0042] Converting the original bispectrum diagram of the composite signal into a grayscale image to obtain the to-be-extracted bispectrum diagram; performing convolution on the to-be-extracted bispectrum diagram at multiple feature extraction scales to determine second discriminative features; modeling the global context information in the to-be-extracted bispectrum diagram to obtain second global features; fusing the second discriminative features and the second global features to obtain bispectrum diagram features.

[0043] In a second aspect, the present invention further provides a feature extraction device for a composite signal, including:

[0044] At least one processor; and a memory communicatively connected to 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 for performing the composite signal feature extraction method described in the first aspect.

[0045] In a third aspect, the present invention further provides a non-volatile computer storage medium, the computer storage medium stores computer-executable instructions, and the computer-executable instructions are executed by one or more processors for completing the composite signal feature extraction method described in the first aspect.

[0046] In a fourth aspect, there is provided a computer program product containing instructions, which when run on a computer or a processor, causes the computer or the processor to execute the composite signal feature extraction method as described in the first aspect.

[0047] In a fifth aspect, the present invention further provides a feature extraction system for a composite signal, including the feature extraction device for a composite signal as described in the second aspect, and using the composite signal feature extraction method as described in the first aspect to complete the interaction of the feature extraction device for a composite signal described in the second aspect.

[0048] Different from the prior art, the present invention has at least the following beneficial effects:

[0049] The present invention respectively extracts the local detail features and global context information of the time-frequency diagram to be extracted and the bispectrum diagram to be extracted. Since the distinguishability of different types of DRFM forwarding signals in the time-frequency domain is not high, the information of the composite signal in the bispectrum domain is used to capture the non-linear features to supplement the information in its time-frequency domain, and the corresponding time-frequency diagram features and bispectrum diagram features are obtained; by performing channel exchange on the time-frequency diagram features and bispectrum diagram features, after the channel exchange, the corresponding feature maps in the time-frequency domain and the bispectrum domain are all effective channel features, realizing the deep cross-domain fusion of the time-frequency diagram features and the bispectrum diagram features, and obtaining the key information in the local dimension of the time-frequency domain and the bispectrum domain; through the attention weight, the local detail features in the time-frequency diagram features and the bispectrum diagram features are feature-fused to obtain the key information in the global context dimension of the time-frequency domain and the bispectrum domain; finally, by performing feature splicing on the first fusion feature and the second fusion feature, integrating the key information in the local dimension and the global context dimension, and using the global context information to guide the attention to the local detail features, it is possible to better extract the effective features in the composite signal under the condition of low signal-to-noise ratio, so as to improve the recognition ability of the composite signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required to be used in the embodiments of the present invention. Obviously, the following described drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0051] Figure 1 is a schematic flow chart of a method for extracting features of a composite signal provided by an embodiment of the present invention;

[0052] Figure 2 is a schematic diagram of a specific example of generating time-frequency diagram features provided by an embodiment of the present invention;

[0053] Figure 3 is a schematic diagram of a specific example of channel exchange provided by an embodiment of the present invention;

[0054] Figure 4 is a schematic flow chart of step 10 provided by an embodiment of the present invention;

[0055] Figure 5 is a schematic diagram of the network structure of a neural network model of an embodiment of the present invention;

[0056] Figure 6 is a schematic flow chart of step 20 provided by an embodiment of the present invention;

[0057] Figure 7 It is a schematic flowchart of step 203 provided by an embodiment of the present invention;

[0058] Figure 8 It is a schematic flowchart of step 30 provided by an embodiment of the present invention;

[0059] Figure 9 It is a schematic flowchart of step 303 provided by an embodiment of the present invention;

[0060] Figure 10 It is a schematic architecture diagram of a feature extraction device for composite signals provided by an embodiment of the present invention. Detailed implementation manners

[0061] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be 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 invention, but not to limit the present invention.

[0062] Unless otherwise required by the context, in the entire specification and claims, the term "comprising" is interpreted in an open, inclusive sense, that is, "including, but not limited to". In the description of the specification, the terms "one embodiment", "some embodiments", "exemplary embodiments", "examples", "specific examples" or "some examples", etc., are intended to indicate that specific features, structures, materials or characteristics related to the embodiment or example are included in at least one embodiment or example of the present disclosure. The schematic representations of the above terms are not necessarily referring to the same embodiment or example. In addition, the specific features, structures, materials or characteristics may be included in any one or more embodiments or examples in any appropriate manner, that is, although they may be carried in the above-mentioned embodiments or examples due to reasons such as the order and position of appearance, they are not limited to being carried in a combined manner by one embodiment or example.

[0063] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present disclosure and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation to the present disclosure.

[0064] In the description of the present invention, the terms "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the embodiments of the present disclosure, unless otherwise specified, the meaning of "a plurality" is two or more. In addition, for example, in the description, for the same type of nouns, the method of adding "A" and "B" at the end is used to describe them as two independent individuals. In this case, the features defined with "A" and "B" are only used for the purpose of distinguishing similar individuals and should not be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features.

[0065] In the description of some embodiments, the expressions "coupled", "coupled to" and "connected" and their derivatives may be used. For example, in the description of some embodiments, the term "connected" may be used to indicate that two or more components have direct physical contact or electrical contact with each other. Another example is that in the description of some embodiments, the term "coupled to" may be used to indicate that two or more components have direct physical contact or electrical contact. However, the terms "connected" or "coupled" may also mean that two or more components do not have direct contact with each other, but still cooperate or interact with each other, such as "optical path coupling", "wireless connection", etc. The embodiments disclosed herein are not necessarily limited to the content of the present invention.

[0066] In the description of the present invention, the expression "A and / or B" (where A and B are used to formally represent specific feature contents) includes the following three combinations: only A, only B, and the combination of A and B.

[0067] As used in the present invention, "about", "substantially" or "approximately" includes the stated value and the average value within an acceptable deviation range of the specific value, where the acceptable deviation range is determined by those of ordinary skill in the art considering the measurement being discussed and the errors associated with the measurement of the specific quantity (i.e., the limitations of the measurement system).

[0068] In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0069] Embodiment 1:

[0070] To solve the above problems, as Figure 1 shown, an embodiment of the present invention provides a method for extracting features of a composite signal, including:

[0071] Step 10: Obtain the time-frequency diagram to be extracted and the bispectrum diagram to be extracted of the composite signal; extract the local detail features and global context information of the time-frequency diagram to be extracted to obtain time-frequency diagram features; extract the local detail features and global context information of the bispectrum diagram to be extracted to obtain bispectrum diagram features.

[0072] Among them, the time-frequency diagram to be extracted is obtained after preprocessing the time-frequency diagram of the composite signal, and the time-frequency diagram is the time-frequency domain image of the composite signal; the bispectrum diagram to be extracted is obtained after preprocessing the time-frequency diagram of the composite signal, and the bispectrum diagram is the bispectrum domain image of the composite signal.

[0073] It should be noted that the operation steps for processing the time-frequency diagram to be extracted to obtain time-frequency diagram features are the same as those for processing the bispectrum diagram to be extracted to obtain bispectrum diagram features, and this process will be described below.

[0074] In a complex electromagnetic environment, the features of DRFM retransmitted signals are diverse and complex, and it is difficult to distinguish the differences between composite DRFM retransmitted signals with features in a single dimension. Therefore, it is necessary to extract multi-dimensional and multi-domain features to describe DRFM retransmitted signals, which is helpful for subsequent identification of the types of DRFM retransmitted signals; among them, the composite DRFM retransmitted signal refers to: a signal composed of the superposition of multiple DRFM retransmitted signals, that is, the composite signal in the embodiments of the present invention.

[0075] Since the frequency distributions of different types of DRFM retransmitted signals at different time points are different, different DRFM retransmitted signals can be better displayed in the time-frequency domain, and the differential features between DRFM retransmitted signals are intuitively visible. Therefore, the embodiments of the present invention extract the features of the composite signal based on the time-frequency domain image.

[0076] However, for several types of DRFM retransmitted signals with gate pulling, such as Range Gate Pull-Off (abbreviated as RGPO), Range-Velocity Gate Pull-Off (abbreviated as RGPO-VGPO), Velocity Gate Pull-Off (abbreviated as VGPO), and Azimuth Gate Pull-Off (abbreviated as AGPO), the processes of the frequency components of the corresponding DRFM retransmitted signals evolving over time are similar, resulting in difficulty in distinguishing in the time-frequency domain.

[0077] Therefore, based on the time-frequency domain image, the embodiments of the present invention also combine the bispectrum domain image of the composite signal to obtain the characteristics of the DRFM forwarding signal. Projecting the bispectrum of the DRFM forwarding signal onto two frequency components forms a plane, and the corresponding plane can reflect the phase coupling relationship of different frequency components in the composite signal. The brightness or color represents the phase coupling degree of the frequency pair, and the non-linear interaction between different frequency components is revealed through the phase relationship. Since the bispectrum can effectively suppress Gaussian noise while retaining the amplitude and phase information of the signal, for the DRFM forwarding signal with low distinguishability in the time-frequency domain, the embodiments of the present invention introduce the corresponding bispectrum to capture the non-linear characteristics of the DRFM forwarding signal, and realize the effective complementarity and deep fusion of cross-domain (i.e., time-frequency domain and bispectrum domain) characteristics through multi-level feature learning and enhancement.

[0078] In the case of poor quality and low signal-to-noise ratio of the composite signal, the embodiments of the present invention first extract the local detail features and global context information of the time-frequency diagram to be extracted to obtain the time-frequency diagram features; and extract the local detail features and global context information of the bispectrum diagram to be extracted to obtain the bispectrum diagram features; enhancing through the global context information helps to identify the local detail features of the DRFM forwarding signal, and then extract the effective features in the time-frequency diagram to be extracted and the bispectrum diagram to be extracted.

[0079] In one embodiment, as Figure 2 shown, after inputting the time-frequency diagram to be extracted into the multi-scale module, the multi-scale module performs convolution on the time-frequency diagram to be extracted at multiple feature extraction scales to extract the local detail features at each position in the time-frequency diagram to be extracted, and obtain the first discrimination feature. After inputting the bispectrum diagram to be extracted into the multi-scale module, the global feature module models the global information in the time-frequency diagram to be extracted to obtain the first global feature. Fusing the first discrimination feature and the first global feature to obtain the time-frequency diagram feature. The processing flow of obtaining the time-frequency diagram feature is the same as that of obtaining the bispectrum diagram feature, and will not be elaborated here.

[0080] It should be noted that the multi-scale module and the global feature module are two independent neural network modules, and they do not share network parameters.

[0081] Step 20: Perform channel exchange on the time-frequency diagram feature and the bispectrum diagram feature, determine the effective channel features corresponding to each channel in the time-frequency diagram feature and the bispectrum diagram feature; and obtain the first fusion feature according to all the effective channel features.

[0082] Among them, the channel is a dimension in the feature map (i.e., the time-frequency diagram feature or the bispectrum diagram feature); as Figure 3As shown, the feature map of the embodiment of the present invention is a three-dimensional tensor with a shape of (C, H, W), where C represents the number of channels, H represents the height of the feature map, and W represents the width of the feature map.

[0083] In order to achieve effective complementarity and deep fusion of time-frequency domain features and bispectrum domain features, the embodiments of the present invention fuse cross-domain features in stages. First, the channel features in the time-frequency map features and the bispectrum map features are exchanged between different domains (i.e., the time-frequency domain and the bispectrum domain) to perform cross-domain feature fusion. In one embodiment, since there may be overlapping regions with the same features in the time-frequency map features and the bispectrum map features of the composite signal, it is necessary to compare the time-frequency map features and the bispectrum map features to determine the non-overlapping regions with different features, and determine the effective features that help identify the DRFM forwarding signal from the non-overlapping regions of the two. As Figure 3 shown, in the process of determining the effective features from the non-overlapping regions of the two, first, it is respectively determined whether the features in the time-frequency map features and the bispectrum map features are effective features; in one embodiment, when they are not effective features, the feature mean of the other domain is used as the effective feature and replaces the feature, so that the non-overlapping regions of the two are both effective features through channel exchange, achieving the purpose of deep fusion of the time-frequency map features and the bispectrum map features. The channel exchange process will be further described below.

[0084] Step 30: Extract the local detail features at each position in the to-be-extracted time-frequency map to obtain the first discrimination feature; extract the local detail features at each position in the to-be-extracted bispectrum map to obtain the second discrimination feature; perform feature fusion on the first discrimination feature and the second discrimination feature based on the attention weight to obtain the second fusion feature.

[0085] The process of obtaining the first discrimination feature will be described below. In one embodiment, the dot product at each position in the to-be-extracted time-frequency map is calculated using a convolution kernel to extract the DRFM forwarding category information of the to-be-extracted time-frequency map at each spatial resolution, obtaining the first discrimination feature.

[0086] Among them, the second fusion feature is: a feature map with channel attention.

[0087] Compared with the fusion process through channel exchange in the above text, the embodiments of the present invention enhance and aggregate the key information across spatial distributions and channels in the time-frequency map features and the bispectrum map features based on the attention mechanism, and promote their interaction and information complementarity by effectively capturing the complex correlation between the time-frequency map features and the bispectrum map features.

[0088] Step 40: Perform feature splicing on the first fusion feature and the second fusion feature to obtain the target feature of the composite signal.

[0089] In the embodiment of the present invention, after extracting the features in the time-frequency domain and the bispectrum domain (i.e., the first fusion feature and the second fusion feature) through two fusion methods, finally, the two are spliced to obtain the target feature that fuses the time-frequency domain and the bispectrum domain. It should be noted that, hereinafter Figure 5 the circles containing "C" represent feature splicing of the first fusion feature and the second fusion feature.

[0090] The present invention separately extracts the local detailed features and global context information of the time-frequency diagram to be extracted and the bispectrum diagram to be extracted. Since the distinguishability of different types of DRFM forwarding signals in the time-frequency domain is not high, the information of the composite signal in the bispectrum domain is used to capture non-linear features to supplement its information in the time-frequency domain, so as to obtain the corresponding time-frequency diagram features and bispectrum diagram features; by performing channel exchange on the time-frequency diagram features and the bispectrum diagram features, after the channel exchange, the corresponding feature maps in the time-frequency domain and the bispectrum domain are all effective channel features, realizing deep cross-domain fusion of the time-frequency diagram features and the bispectrum diagram features, and obtaining the key information in the local dimension of the time-frequency domain and the bispectrum domain; through attention weights, feature fusion is performed on the local detailed features in the time-frequency diagram features and the bispectrum diagram features to obtain the key information in the global context dimension of the time-frequency domain and the bispectrum domain; finally, by performing feature splicing on the first fusion feature and the second fusion feature, the key information in the local dimension and the global context dimension is integrated, and the global context information is used to guide the attention to the local detailed features, so that in the case of low signal-to-noise ratio, the effective features in the composite signal can be better extracted, so as to improve the recognition ability of the composite signal.

[0091] In order to illustrate the process of processing the time-frequency diagram to be extracted to obtain time-frequency diagram features and processing the bispectrum diagram to be extracted to obtain bispectrum diagram features, as Figure 4 shown, step 10 includes:

[0092] Step 101: Convert the original time-frequency diagram of the composite signal into a grayscale diagram to obtain the time-frequency diagram to be extracted; perform convolution on the time-frequency diagram to be extracted at multiple feature extraction scales to determine the first discrimination feature; model the global context information in the time-frequency diagram to be extracted to obtain the first global feature; fuse the first discrimination feature and the first global feature to obtain the time-frequency diagram features.

[0093] On the one hand, although DRFM forwarding signals are difficult to distinguish in the time-frequency domain, there are essential differences in the subtle features and energy distributions in the time-frequency domain. Therefore, the embodiment of the present invention adopts multi-scale convolution, at multiple feature extraction scales (i.e., using convolution kernels of different sizes), to extract the local detailed features of the time-frequency diagram to be extracted or the bispectrum diagram to be extracted, so as to enhance the sensitivity of the neural network in the embodiment of the present invention to the local subtle features of DRFM forwarding signals. In one embodiment, as Figure 5As shown, the multi-scale module convolves the time-frequency map to be extracted at multiple feature extraction scales to obtain the first discriminative feature, and convolves the bispectrum map to be extracted at multiple feature extraction scales to obtain the second discriminative feature.

[0094] On the other hand, since the prior art cannot know the position of the features important for the DRFM forwarding signal recognition task in the signal, and in the signal, the essential context information characterizing the type of the DRFM forwarding signal is larger than the area occupied by the DRFM forwarding signal itself, the features helpful for recognizing the DRFM forwarding signal cannot be well extracted by only extracting artificial features from the DRFM forwarding signal. The embodiments of the present invention model the global information in the time-frequency map to be extracted, and then extract the key features helpful for recognizing DRFM forwarding, that is, the first global feature; and use the key feature as the context, combine with the local detail features (that is, the first discriminative feature or the second discriminative feature in the following text), and obtain the features helpful for recognizing the type of the DRFM forwarding signal (that is, the time-frequency domain feature or the bispectrum domain feature in the following text) in the time-frequency map to be extracted or the bispectrum map to be extracted from the DRFM forwarding signal. In one embodiment, as Figure 5 shown, the global feature module models the global context information in the time-frequency map to be extracted to obtain the first global feature; the global feature module models the global context information in the bispectrum map to be extracted to obtain the second global feature.

[0095] Finally, the fusion module fuses the local detail features and the global information to capture the features of the composite signal at a deeper level, so as to improve the robustness and accuracy of recognizing the type of the DRFM forwarding signal for the composite signal composed of multiple types of DRFM forwarding signals.

[0096] Step 102: Convert the original bispectrum map of the composite signal into a grayscale map to obtain the bispectrum map to be extracted; convolve the bispectrum map to be extracted at multiple feature extraction scales to determine the second discriminative feature; model the global context information in the bispectrum map to be extracted to obtain the second global feature; fuse the second discriminative feature and the second global feature to obtain the bispectrum map feature.

[0097] The processing flow of the original time-frequency map and the original bispectrum map in the embodiments of the present invention is the same, which will not be elaborated here.

[0098] In this embodiment, it is necessary to first obtain the effective channel features, and obtain the first fusion feature according to all the effective channel features. Among them, the determination process of the effective channel features is divided into two stages: (1) For the overlapping channel features, the overlapping channel features can be directly used as the effective channel features; (2) For the non-overlapping channel features, it is necessary to determine whether they are effective channel features according to the scaling factor parameters of the channels. If they are not effective channel features, it is necessary to further perform channel swapping to obtain the effective channel features. The specific implementation process is as follows Figure 6 As shown, step 20 includes:

[0099] Step 201: Perform batch normalization on the time-frequency map features to obtain the first feature map to be evaluated; perform batch normalization on the bispectrum map features to obtain the second feature map to be evaluated.

[0100] In one embodiment, as Figure 5 shown, input the time-frequency map features and bispectrum map features output by the fusion module into the channel swapping module to perform channel swapping on the non-overlapping regions of the two. Before comparing the overlapping regions and non-overlapping regions of the two, first make the feature dimensions consistent through batch normalization.

[0101] Step 202: If the channel features in the first feature map to be evaluated overlap with the channel features at the corresponding positions in the second feature map to be evaluated, then determine the overlapping channel features as the effective channel features in the first feature map to be evaluated and the effective channel features in the second feature map to be evaluated.

[0102] Among them, channel feature overlap means that the feature values of the channel features are equal.

[0103] In one embodiment, as Figure 2 shown, for the first feature map to be evaluated in the time-frequency domain, the leftmost channel feature overlaps with the leftmost channel feature in the second feature map to be evaluated in the bispectrum domain; then the leftmost channel feature in the first feature map to be evaluated is the effective channel feature in the first feature map to be evaluated, and the leftmost channel feature in the second feature map to be evaluated is the effective channel feature in the second feature map to be evaluated.

[0104] Step 203: If the channel features in the first feature map to be evaluated do not overlap with the channel features at the corresponding positions in the second feature map to be evaluated, then determine the non-overlapping channel features in the first feature map to be evaluated as the first original channel features, and determine the non-overlapping channel features in the second feature map to be evaluated as the second original channel features; determine whether the first original channel features are effective channel features, determine whether the second original channel features are effective channel features, and perform channel swapping on the original channel features that are not effective channel features to obtain the corresponding effective channel features.

[0105] The judgment criteria for the effective channel features will be described below.

[0106] In one embodiment, when the first original channel feature is not an effective channel feature, calculate the mean of the channel features of all channels in the current second feature map to be evaluated, obtain a first feature mean, and determine the first feature mean as the effective channel feature of the channel in the current first feature map to be evaluated, so as to perform channel swapping on the first original channel feature. When the second original channel feature is not an effective channel feature, calculate the mean of the channel features of all channels in the current first feature map to be evaluated, obtain a second feature mean, and determine the second feature mean as the effective channel feature of the channel in the current second feature map to be evaluated, so as to perform channel swapping on the first feature map to be evaluated and the second feature map to be evaluated, so as to perform channel swapping on the second original channel feature.

[0107] For example, as Figure 2 shown, when the first original channel feature is the feature of the channel indicated by the light-colored square in the time-frequency domain, during the process of channel swapping between the first feature map to be evaluated and the second feature map to be evaluated, the channel features of each channel are likely to be swapped to the corresponding means; in the embodiment of the present invention, it is judged one by one whether the first original channel feature in the first feature map to be evaluated is an effective channel feature. When it is not an effective channel feature, calculate the mean of all channel feature values in the current second feature map to be evaluated during the channel swapping process as the first feature mean, and determine it as the effective channel feature of the channel indicated by the light-colored square.

[0108] For example, as Figure 2 shown, when the second original channel feature is the feature of the channel indicated by the dark-colored square in the bispectrum domain, judge one by one whether the second original channel feature in the second feature map to be evaluated is an effective channel feature. When it is not an effective channel feature, calculate the mean of all channel feature values in the current first feature map to be evaluated during the channel swapping process as the second feature mean, and determine it as the effective channel feature of the channel indicated by the dark-colored square.

[0109] Step 204: Determine all the effective channel features in the first feature map to be evaluated as time-frequency intermediate features; determine all the effective channel features in the second feature map to be evaluated as bispectrum intermediate features; perform feature splicing on the time-frequency intermediate features and the bispectrum intermediate features to obtain a first fusion feature.

[0110] In the embodiment of the present invention, by comparing the features of each channel in the time-frequency map features with the features of each channel in the bispectrum map features, and through channel swapping, the channel features that are not effective channel features in the non-overlapping regions of the two are replaced with the feature means of the other domain. On the one hand, through selective channel swapping, feature fusion between the time-frequency domain and the bispectrum domain is achieved; on the other hand, since during the process of channel swapping, the key information in the first feature map to be evaluated and the second feature map to be evaluated that helps identify the composite signal is continuously increasing, compared with the channel features that are not effective channel features, the feature mean of the feature map to be evaluated is very likely to have more key information. By replacing it with the feature mean of the corresponding feature map to be evaluated, more likely effective feature information is added.

[0111] To illustrate the judgment criterion for effective channel features, in one embodiment, as Figure 7 shown, in step 203, the judgment of whether the first original channel feature is an effective channel feature includes:

[0112] Step 2031: When batch-normalizing the time-frequency map features, a scaling factor parameter is set for each channel in advance.

[0113] Among them, the specific values of the scaling factor parameters set for each channel in advance are determined by those skilled in the art according to the specific usage scenario.

[0114] A penalty function is added to the total loss function to use the penalty function to constrain the sparsity of the scaling factor parameters; wherein, the total loss function is used to extract the target features.

[0115] In the channel swapping module, the scaling factor parameters are used in the normalization layer to evaluate the importance of each channel; in an alternative embodiment, as Figure 2 shown, a 1-norm penalty is imposed on the scaling factor parameters, and the filters that meet the sparsity criterion are explicitly pruned, thereby enabling direct fusion of the swapped channels between the time-frequency domain and the bispectrum domain.

[0116] In one embodiment, taking the case where the first original channel feature is not an effective channel feature as an example, the expression for the effective channel feature is:

[0117]

[0118] where f c represents the first original channel feature of the c-th channel of the time-frequency map features, f c ′ represents the effective channel feature corresponding to the c-th channel, μ c represents the mean of the feature values of all channels in the current second feature map to be evaluated, and σ crepresents the standard deviation of the feature values of all channels in the current second feature map to be evaluated, γ c represents the scaling factor parameter, β c represents the bias parameter, and ε is a constant.

[0119] Among them, the bias parameter and the constant are selected by those skilled in the art according to the specific usage scenario, and are not limited herein.

[0120] Step 2032: When the scaling factor parameter of the channel is lower than the threshold, the first original channel feature is not a valid channel feature.

[0121] Among them, the threshold is selected by those skilled in the art according to the specific usage scenario, and the threshold is as close to zero as possible.

[0122] Step 2033: When the scaling factor parameter of the channel is not lower than the threshold, the first original channel feature is a valid channel feature.

[0123] Specifically, applying the sparse constraint of the scaling factor parameter to the non-overlapping regions in the time-frequency domain and the bispectrum domain. If the scaling factor parameter of a certain channel in the feature map of one domain is lower than a certain threshold, it means that it has little impact on the final result and is redundant. Therefore, in the embodiments of the present invention, the feature mean of the other domain is used to replace it.

[0124] To illustrate the process of obtaining the second fusion feature, as Figure 8 shown, in step 30, the feature fusion of the first discriminative feature and the second discriminative feature based on the attention weight to obtain the second fusion feature includes:

[0125] Step 301: Use multiple asymmetric convolution kernels to calculate the dot product at each position in the first discriminative feature to extract the spatial features of the first discriminative feature, and obtain the first deep feature; use multiple asymmetric convolution kernels to calculate the dot product at each position in the second discriminative feature to extract the spatial features of the second discriminative feature, and obtain the second deep feature.

[0126] Since the initially extracted shallow features often have the problem of limited receptive fields, resulting in the neural network model in the embodiments of the present invention being insufficient to fully capture the complex and potential DRFM forwarding signal features in the image. Therefore, in the embodiments of the present invention, asymmetric convolution (i.e., Figure 5The asymmetric convolution module) further deepens the feature extraction processes in the time-frequency domain and the bispectrum domain independently and separately, enhancing the expression ability and diversity of features. Specifically, after the first discriminative feature is input into the asymmetric convolution module for the time-frequency domain, the asymmetric convolution module convolves the first discriminative feature at multiple feature extraction scales to extract the local detailed features at each position in the first discriminative feature, obtaining the first deep feature; after the second discriminative feature is input into the asymmetric convolution module for the bispectrum domain, the asymmetric convolution module convolves the second discriminative feature at multiple feature extraction scales to extract the local detailed features at each position in the second discriminative feature, obtaining the second deep feature. Among them, in the embodiment of the present invention, the traditional single large-size (such as n×n) convolution kernel is decomposed into two smaller and asymmetric convolution kernels (such as 1×n and n×1) to significantly reduce the number of parameters of the neural network model of the embodiment of the present invention. When the asymmetric convolution module performs convolution at multiple feature extraction scales, multiple asymmetric convolution kernels are used to calculate the dot product at each position in the first discriminative feature or the second discriminative feature to extract spatial features, thereby improving the efficiency of training and inference and effectively alleviating the risk of overfitting. More importantly, due to the addition of an additional non-linear transformation layer (i.e., Figure 5 the asymmetric convolution module), the neural network model of the embodiment of the present invention can capture richer and more detailed spatial features, thereby enhancing the representation ability for complex DRFM forwarding signals.

[0127] It should be noted that both the asymmetric convolution module and the multi-scale convolution module use multiple convolution kernels of different sizes for local detailed feature extraction. The core idea of the asymmetric convolution module and the multi-scale convolution module in feature extraction is the same, which is to use convolution kernels of different sizes to extract features in parallel, only the sizes and distribution rules of the convolution kernels used by the asymmetric convolution module and the multi-scale convolution module are different.

[0128] Step 302: Perform feature splicing on the first deep feature and the second deep feature to obtain a feature to be processed; divide the feature to be processed into tiles of a fixed size; convert each tile into a vector to be processed with a fixed dimension.

[0129] In one embodiment, after the input image (i.e., the feature to be processed) is divided into tiles of a fixed size through a linear layer, each tile is converted into a vector to be processed with a fixed dimension (i.e., a token) through a linear normalization layer, and a position encoding representing the relative position of each tile in the input image is added. It should be noted that Figure 5 the circle containing "C" represents feature splicing of the first deep feature and the second deep feature.

[0130] Step 303: Determine the attention weights of the vector to be processed, and perform attention calculation on the vector to be processed using the attention weights to obtain the first encoded feature.

[0131] Then, the encoder is used to learn the features of the vector to be processed, and this process will be described below.

[0132] Step 304: Perform linear transformation and non-linear activation on the first encoded feature to learn the first encoded feature using a feed-forward neural network, and obtain the second encoded feature.

[0133] Among them, the specific method of using the feed-forward neural network to learn the first encoded feature is determined by those skilled in the art with reference to the prior art according to the specific usage scenario, and will not be limited here.

[0134] Step 305: Perform an addition operation on the first encoded feature and the second encoded feature to obtain the second fused feature.

[0135] After further feature extraction by the asymmetric convolution module, the embodiments of the present invention use the multi-head attention mechanism to fuse the features in the time-frequency domain and the bispectrum domain, enhance and converge the key features of the feature space distribution and between channels, promote the interaction and information complementarity between different features, and achieve the adaptive weight allocation and fusion of the time-frequency domain and the bispectrum domain.

[0136] To illustrate the process of obtaining the first encoded feature, as Figure 9 shown, the step 303 includes:

[0137] Step 3031: 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.

[0138] Among them, the query weight parameter, the key weight parameter, and the value weight parameter are all weight matrices.

[0139] Step 3032: Perform a convolution operation on the original key vector in the spatial dimension according to the reduction ratio to obtain the first reduced feature; perform a linear projection operation on the first reduced feature using the reduction ratio to compress the channel dimension of the first reduced feature to obtain the target key vector.

[0140] Step 3033: Perform a convolution operation on the original value vector in the spatial dimension according to the reduction ratio to obtain the second reduced feature; perform a linear projection operation on the second reduced feature using the reduction ratio to compress the channel dimension of the second reduced feature to obtain the target value vector.

[0141] Step 3034: Calculate the dot product of the query vector and the target key vector to obtain the attention weight.

[0142] Step 3035: Use the attention weight to perform weighted summation on the target value vector to generate the first encoded feature.

[0143] In the embodiment of the present invention, by calculating the weighted sum of the query vector, key vector, and value vector, the multi-head attention mechanism is utilized to capture the global dependencies between different positions in the sequence of the vector to be processed. Furthermore, for the time-frequency domain and bispectrum domain, the key information in the global context dimension can be obtained respectively, so that when the first fusion feature and the second fusion feature are stitched together subsequently, the global context information is used to guide the attention to the local detailed features, realizing better extraction of the effective features in the composite signal under the condition of low signal-to-noise ratio, and greatly improving the recognition ability of the composite signal.

[0144] When calculating the weighted sum of the query vector, key vector, and value vector to obtain the first encoded feature, by first calculating the dot product of the query vector and the target key vector to obtain the attention weight, and then calculating the weighted sum of the attention weight and the value vector, the spatial dimension of the calculation is reduced, effectively alleviating the problems of computational complexity and memory consumption caused by high-dimensional data, and can greatly reduce the computational memory overhead. This enables the neural network model in the embodiment of the present invention to process large-scale DRFM forwarding signal data more efficiently, greatly improving the efficiency of extracting global features during the training and inference processes without sacrificing the representational ability of the model.

[0145] Embodiment 2:

[0146] As Figure 10 shown, it is a schematic architecture diagram of a feature extraction device for a composite signal according to an embodiment of the present invention. The feature extraction device for a composite signal in this embodiment includes one or more processors 21 and a memory 22. Among them, Figure 10 Taking one processor 21 as an example.

[0147] The processor 21 and the memory 22 can be connected through a bus or other means, Figure 10 Taking the connection through a bus as an example.

[0148] The memory 22, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs and non-volatile computer-executable programs, such as the feature extraction method for a composite signal in this embodiment. The processor 21 executes the feature extraction method for a composite signal by running the non-volatile software programs and instructions stored in the memory 22.

[0149] The memory 22 may include high-speed random access memory, and may also include non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other non-volatile solid-state storage devices. In some embodiments, the memory 22 optionally includes a memory remotely disposed relative to the processor 21, and these remote memories may be connected to the processor 21 through a network. Examples of the above network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0150] The program instructions / modules are stored in the memory 22 and, when executed by the one or more processors 21, perform the feature extraction method of the composite signal in the above embodiments. For example, perform each step of the feature extraction method of the composite signal of the embodiments of the present invention described above.

[0151] Embodiments of the present invention also provide a non-volatile computer storage medium storing computer-executable instructions that, when executed by one or more processors, such as Figure 10 a processor 21, enable the above one or more processors to execute the feature extraction method of the composite signal in the specific embodiments of the present invention. For example, perform each step of the feature extraction method of the composite signal of the embodiments of the present invention described above; it can also implement Figure 10 the various modules and units described above; or execute the feature extraction method of the composite signal in the specific embodiments of the present invention. For example, perform each step of the feature extraction method of the composite signal of the embodiments of the present invention described above; it can also implement Figure 10 the various modules and units described above.

[0152] It should be noted that the information interaction, execution process, etc. between the modules and units in the above device and system, due to being based on the same concept as the method embodiments of the present invention, the specific content can be referred to the description in the method embodiments of the present invention and will not be elaborated here.

[0153] Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the embodiments can be completed by instructing relevant hardware through a program, and the program can be stored in a computer-readable storage medium. The storage medium may include: a read-only memory (ROM, Read Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk, or an optical disc, etc.

[0154] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for extracting features of a composite signal, characterized in that, Including: Obtaining the time-frequency diagram to be extracted and the bispectrum diagram to be extracted of the composite signal; extracting the local detail features and global context information of the time-frequency diagram to be extracted to obtain time-frequency diagram features; Extracting the local detail features and global context information of the bispectrum diagram to be extracted to obtain bispectrum diagram features; Performing channel swapping on the time-frequency diagram features and the bispectrum diagram features, and determining the effective channel features corresponding to each channel in the time-frequency diagram features and the bispectrum diagram features; obtaining a first fusion feature based on all the effective channel features; Extracting the local detail features at each position in the time-frequency diagram to be extracted to obtain a first discrimination feature; extracting the local detail features at each position in the bispectrum diagram to be extracted to obtain a second discrimination feature; Performing feature fusion on the first discrimination feature and the second discrimination feature based on attention weights to obtain a second fusion feature; Performing feature splicing on the first fusion feature and the second fusion feature to obtain the target feature of the composite signal.

2. The feature extraction method of the composite signal according to claim 1, wherein The performing channel swapping on the time-frequency diagram features and the bispectrum diagram features, and determining the effective channel features corresponding to each channel in the time-frequency diagram features and the bispectrum diagram features; The obtaining a first fusion feature based on all the effective channel features includes: Performing batch normalization on the time-frequency diagram features to obtain a first feature map to be evaluated; performing batch normalization on the bispectrum diagram features to obtain a second feature map to be evaluated; If the channel features in the first feature map to be evaluated overlap with the channel features at the corresponding positions in the second feature map to be evaluated, then the overlapping channel features are determined as the effective channel features in the first feature map to be evaluated and the effective channel features in the second feature map to be evaluated; If the channel features in the first feature map to be evaluated do not overlap with the channel features at the corresponding positions in the second feature map to be evaluated, then the non-overlapping channel features in the first feature map to be evaluated are determined as first original channel features, and the non-overlapping channel features in the second feature map to be evaluated are determined as second original channel features; determining whether the first original channel features are effective channel features, determining whether the second original channel features are effective channel features, and performing channel swapping on the original channel features that are not effective channel features to obtain corresponding effective channel features; Determining all the effective channel features in the first feature map to be evaluated as time-frequency intermediate features; determining all the effective channel features in the second feature map to be evaluated as bispectrum intermediate features; performing feature splicing on the time-frequency intermediate features and the bispectrum intermediate features to obtain a first fusion feature.

3. The method for extracting the characteristics of the composite signal according to claim 2, wherein, The performing channel swapping on the original channel features that are not effective channel features to obtain corresponding effective channel features includes: When the first original channel features are not effective channel features, calculating the mean of the channel features of all channels in the current second feature map to be evaluated to obtain a first feature mean, and determining the first feature mean as the effective channel feature of the channel in the current first feature map to be evaluated to perform channel swapping on the first original channel features; When the second original channel feature is not a valid channel feature, calculate the mean of the channel features of all channels in the current first feature map to be evaluated to obtain a second feature mean, and determine the second feature mean as the valid channel feature of the channel in the current second feature map to be evaluated, so as to perform channel swapping between the first feature map to be evaluated and the second feature map to perform channel swapping on the second original channel feature.

4. The method for extracting the characteristics of the composite signal according to claim 2, wherein, The determination of whether the first original channel feature is a valid channel feature includes: When performing batch normalization on the time-frequency map feature, set a scaling factor parameter for each channel in advance; Add a penalty function to the total loss function to use the penalty function to constrain the sparsity of the scaling factor parameter; wherein, the total loss function is used to extract the target feature; When the scaling factor parameter of the channel is lower than the threshold, the first original channel feature is not a valid channel feature; When the scaling factor parameter of the channel is not lower than the threshold, the first original channel feature is a valid channel feature.

5. The method for extracting the characteristics of a composite signal according to claim 4, characterized in that, The expression of the valid channel feature is: Among them, f c represents the first original channel feature of the c-th channel of the time-frequency diagram feature, f c ' represents the effective channel feature corresponding to the c-th channel, μ c represents the mean of the feature values of all channels in the current second feature map to be evaluated, σ c represents the standard deviation of the feature values of all channels in the current second feature map to be evaluated, γ c represents the scaling factor parameter, β c represents the bias parameter, and ε is a constant.

6. The method for extracting the characteristics of a composite signal according to claim 1, wherein The feature fusion of the first discrimination feature and the second discrimination feature based on the attention weight to obtain a second fusion feature includes: Calculate the dot product of each position in the first discrimination feature using multiple asymmetric convolution kernels to extract the spatial feature of the first discrimination feature to obtain a first deep feature; calculate the dot product of each position in the second discrimination feature using multiple asymmetric convolution kernels to extract the spatial feature of the second discrimination feature to obtain a second deep feature; Perform feature concatenation on the first deep feature and the second deep feature to obtain a feature to be processed; divide the feature to be processed into tiles of a fixed size; convert each tile into a vector to be processed with a fixed dimension; Determine the attention weight of the vector to be processed, and use the attention weight to perform attention calculation on the vector to be processed to obtain a first encoded feature; Perform linear transformation and non-linear activation on the first encoded feature to use a feed-forward neural network to learn the first encoded feature to obtain a second encoded feature; Perform an addition operation on the first encoded feature and the second encoded feature to obtain a second fusion feature.

7. The method for extracting the characteristics of the composite signal according to claim 6, wherein The determination of the attention weight of the vector to be processed, and the use of the attention weight to perform attention calculation on the vector to be processed to obtain a first encoded feature includes: Calculate the product of the vector to be processed and the query weight parameter to obtain a query vector; calculate the product of the vector to be processed and the key weight parameter to obtain an original key vector; calculate the product of the vector to be processed and the value weight parameter to obtain an original value vector; Perform a convolution operation on the original key vector in the spatial dimension according to a reduction ratio to obtain a first reduced feature; perform a 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; Perform a convolution operation on the original value vector in the spatial dimension according to the reduction ratio to obtain a second reduced feature; use the reduction ratio to perform a linear projection operation on the second reduced feature to compress the channel dimension of the second reduced feature and obtain a target value vector; Calculate the dot product of the query vector and the target key vector to obtain attention weights; Use the attention weights to perform a weighted sum on the target value vector to generate a first encoded feature.

8. The method for extracting the characteristics of a composite signal according to any one of claims 1-7, wherein, Obtain the time-frequency map to be extracted and the bispectrum map to be extracted of the composite signal; extract the local detailed features and global context information of the time-frequency map to be extracted to obtain time-frequency map features; Extract the local detailed features and global context information of the bispectrum map to be extracted, and the obtained bispectrum map features include: Convert the original time-frequency map of the composite signal into a grayscale map to obtain the time-frequency map to be extracted; perform convolution on the time-frequency map to be extracted at multiple feature extraction scales to determine the first discriminative feature; model the global context information in the time-frequency map to be extracted to obtain the first global feature; fuse the first discriminative feature and the first global feature to obtain time-frequency map features; Convert the original bispectrum map of the composite signal into a grayscale map to obtain the bispectrum map to be extracted; perform convolution on the bispectrum map to be extracted at multiple feature extraction scales to determine the second discriminative feature; model the global context information in the bispectrum map to be extracted to obtain the second global feature; fuse the second discriminative feature and the second global feature to obtain bispectrum map features.

9. A feature extraction device for a composite signal, characterized in that, The feature extraction device for the composite signal includes at least one processor and a memory, the at least one processor and the memory are connected through a data bus, the memory stores instructions executable by the at least one processor, and after being executed by the processor, the instructions are used to implement the feature extraction method for the composite signal according to any one of claims 1-8.

10. A non-volatile computer storage medium, characterized in that, The computer storage medium stores computer-executable instructions, and the computer-executable instructions are executed by one or more processors to complete the feature extraction method for the composite signal according to any one of claims 1-8.

Images