A method and apparatus for feature extraction of a composite signal

Abstract

The present application relates to the field of radar signal processing, and provides a feature extraction method and device for a composite signal. Local detail features and global context information of a to-be-extracted time-frequency graph and a to-be-extracted bispectrum graph of the composite signal are extracted respectively to obtain time-frequency graph features and bispectrum graph features, and channel exchange is performed on the two features to determine effective channel features corresponding to each channel; a first fusion feature is obtained according to all the effective channel features; local detail features of each position in the to-be-extracted time-frequency graph and the to-be-extracted bispectrum graph are extracted respectively to obtain first distinguishing features and second distinguishing features, and feature fusion is performed on the first distinguishing features and the second distinguishing features based on attention weights to obtain a second fusion feature; and feature splicing is performed on the first fusion feature and the second fusion feature to obtain a target feature, thereby solving the problem of low recognition rate of the composite signal in the case of low signal-to-noise ratio.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing technology, and in particular to a method and apparatus for feature extraction of composite signals. Background Technology

[0002] The relayed signals of Digital Radio Frequency Memory (DRFM) are used to reduce the radar's ability to detect target signals, or even render the radar completely ineffective. In modern radar systems, the characteristics of DRFM relayed signals are often highly overlapping and complex; that is, the radar often detects composite signals. In such scenarios, the radar may be unable to detect the target signal.

[0003] Traditional time-domain or frequency-domain analysis methods often fail to effectively address this problem, especially when the DRFM relay signal has multi-scale structural features. The difficulty of feature extraction increases with increasing noise; when the noise power increases to a certain level, the noise will completely overwhelm the DRFM relay signal. In such cases, it is difficult to extract effective features from the composite signal detected by radar. Existing technologies, under low signal-to-noise ratio and complex signal conditions, result in poor signal quality for composite signals detected by radar, leading to low recognition capabilities and severely impacting the normal operation of the radar.

[0004] Therefore, overcoming the shortcomings of the existing technology is an urgent problem to be solved in this technical field. Summary of the Invention

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

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

[0007] In a first aspect, the present invention provides a feature extraction method for composite signals, comprising:

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

[0009] Channel swapping is performed on the time-frequency plot features and the bispectral plot features to determine the effective channel features corresponding to each channel in the time-frequency plot features and the bispectral plot features; a first fusion feature is obtained based on all the effective channel features.

[0010] Local detail features are extracted from each position in the time-frequency image to be extracted to obtain a first distinguishing feature; local detail features are extracted from each position in the bispectral image to be extracted to obtain a second distinguishing feature; the first distinguishing feature and the second distinguishing feature are fused based on attention weights to obtain a second fused feature;

[0011] The first fusion feature and the second fusion feature are spliced ​​together to obtain the target feature of the composite signal.

[0012] Further, the process of channel swapping between the time-frequency plot feature and the bispectral plot feature to determine the effective channel features corresponding to each channel in the time-frequency plot feature and the bispectral plot feature; and obtaining the first fused feature based on all the effective channel features, includes:

[0013] The time-frequency graph features are batch normalized to obtain a first feature map to be evaluated; the bispectral graph features are batch normalized to obtain a second feature map to be evaluated.

[0014] If a channel feature in the first feature map to be evaluated overlaps with a channel feature at the corresponding position in the second feature map to be evaluated, then the overlapping channel feature is determined as a valid channel feature in the first feature map to be evaluated and a valid channel feature 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 corresponding channel features 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 the first original channel features, and the non-overlapping channel features in the second feature map to be evaluated are determined as the second original channel features; it is determined whether the first original channel feature is a valid channel feature, and whether the second original channel feature is a valid channel feature, and channel swapping is performed on the original channel features that are not valid channel features to obtain the corresponding valid channel features;

[0016] All effective channel features in the first feature map to be evaluated are identified as time-frequency intermediate features; all effective channel features in the second feature map to be evaluated are identified as bispectral intermediate features; the time-frequency intermediate features and the bispectral intermediate features are spliced ​​together to obtain the first fused feature.

[0017] Furthermore, the process of channel swapping the original channel features that are not valid channel features to obtain the corresponding valid channel features includes:

[0018] When the first original channel feature is not a valid channel feature, the mean of the channel features of all channels in the current second feature map to be evaluated is calculated to obtain the first feature mean. The first feature mean is determined 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, the mean value of the channel features of all channels in the current first feature map to be evaluated is calculated to obtain the second feature mean value. The second feature mean value is determined as the valid channel feature of the channel in the current second feature map to be evaluated, so as to perform channel exchange between the first feature map to be evaluated and the second feature map to be evaluated, and to perform channel exchange on the second original channel feature.

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

[0021] When performing batch normalization on the time-frequency graph features, a scaling factor parameter is pre-set for each channel;

[0022] A penalty function is added to the total loss function to constrain the sparsity of the scaling factor parameters; wherein the total loss function is used to extract the target features;

[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] Furthermore, the expression for the effective channel feature is:

[0026]

[0027] Among them, f c f represents the first original channel feature of the c-th channel of the time-frequency plot. c ′ represents the effective channel feature corresponding to the c-th channel, μ c σ represents the mean of the eigenvalues ​​of all channels in the current second feature map to be evaluated. c γ represents the standard deviation of the eigenvalues ​​of all channels in the current second feature map to be evaluated. c This represents the scaling factor parameter, β. c This represents the bias parameter, where ε is a constant.

[0028] Furthermore, the feature fusion of the first distinguishing feature and the second distinguishing feature based on attention weights to obtain the second fused feature includes:

[0029] Multiple asymmetric convolution kernels are used to calculate the dot product at each position in the first distinguishing feature to extract the spatial features of the first distinguishing feature and obtain the first deep feature; multiple asymmetric convolution kernels are used to calculate the dot product at each position in the second distinguishing feature to extract the spatial features of the second distinguishing feature and obtain the second deep feature.

[0030] The first deep feature and the second deep feature are concatenated to obtain the feature to be processed; the feature to be processed is segmented into fixed-size patches; each patch is converted into a fixed-dimensional vector to be processed.

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

[0032] The first encoded feature is subjected to linear transformation and nonlinear activation to learn the first encoded feature using a feedforward neural network, thereby obtaining the second encoded feature;

[0033] The first encoded feature and the second encoded feature are added together to obtain the second fused feature.

[0034] Further, 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:

[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] 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.

[0037] 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.

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

[0039] The target value vector is weighted and summed using the attention weights to generate the first encoded feature.

[0040] Further, the steps of acquiring the time-frequency plot and bispectral plot of the composite signal to be extracted; extracting local detail features and global context information of the time-frequency plot to be extracted to obtain time-frequency plot features; and extracting local detail features and global context information of the bispectral plot to be extracted to obtain bispectral plot features include:

[0041] The original time-frequency map of the composite signal is converted into a grayscale image to obtain the time-frequency map to be extracted; the time-frequency map to be extracted is convolved at multiple feature extraction scales to determine the first distinguishing feature; the global context information in the time-frequency map to be extracted is modeled to obtain the first global feature; the first distinguishing feature and the first global feature are fused to obtain the time-frequency map feature;

[0042] The original bispectral image of the composite signal is converted into a grayscale image to obtain the bispectral image to be extracted; the bispectral image to be extracted is convolved at multiple feature extraction scales to determine the second distinguishing feature; the global context information in the bispectral image to be extracted is modeled to obtain the second global feature; the second distinguishing feature and the second global feature are fused to obtain the bispectral image feature.

[0043] Secondly, the present invention also provides a feature extraction device for composite signals, comprising:

[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 for performing the feature extraction method for the composite signal described in the first aspect.

[0045] Thirdly, the present invention also provides a non-volatile computer storage medium storing computer-executable instructions that are executed by one or more processors to perform the feature extraction method for composite signals described in the first aspect.

[0046] Fourthly, a computer program product containing instructions is provided, which, when executed on a computer or processor, causes the computer or processor to perform a feature extraction method for composite signals as described in the first aspect.

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

[0048] Unlike existing technologies, the present invention has at least the following beneficial effects:

[0049] This invention extracts local detail features and global context information from the time-frequency map and bispectral map to be extracted, respectively. Since different types of DRFM forwarding signals have low distinguishability in the time-frequency domain, information from the composite signal in the bispectral domain is used to capture nonlinear features, supplementing the information in the time-frequency domain to obtain corresponding time-frequency map features and bispectral map features. By performing channel swapping on the time-frequency map features and bispectral map features, the corresponding feature maps in both the time-frequency domain and bispectral domain become effective channel features after the swap, achieving deep inter-domain fusion of the time-frequency map features and bispectral map features, obtaining key information in the local dimension of the time-frequency domain and bispectral domain. Attention weights are used to fuse the local detail features in the time-frequency map features and bispectral map features, obtaining key information in the global context dimension of the time-frequency domain and bispectral domain. Finally, by concatenating the first and second fused features, the key information in the local and global context dimensions is integrated. The global context information guides the attention to local detail features, thereby enabling better extraction of effective features from composite signals even with low signal-to-noise ratios, thus improving the ability to identify composite signals. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments of the present invention will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0051] Figure 1 This is a flowchart illustrating a feature extraction method for composite signals provided in an embodiment of the present invention;

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

[0053] Figure 3 This is a schematic diagram illustrating a specific example of channel switching provided in an embodiment of the present invention;

[0054] Figure 4 This is a flowchart illustrating step 10 provided in an embodiment of the present invention;

[0055] Figure 5 This is a schematic diagram of the network structure of a neural network model according to an embodiment of the present invention.

[0056] Figure 6 This is a flowchart illustrating step 20 provided in an embodiment of the present invention;

[0057] Figure 7 This is a flowchart illustrating step 203 provided in an embodiment of the present invention;

[0058] Figure 8 This is a flowchart illustrating step 30 provided in an embodiment of the present invention;

[0059] Figure 9 This is a flowchart illustrating step 303 provided in an embodiment of the present invention;

[0060] Figure 10 This is a schematic diagram of the architecture of a feature extraction device for composite signals provided in an embodiment of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this invention clearer, the 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 merely illustrative and not intended to limit the invention.

[0062] 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.

[0063] 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.

[0064] 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.

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

[0066] In the description of this invention, the expression “A and / or B” (where A and B are used to formally represent specific features) will be used. The corresponding expression includes the following three combinations: only A, only B, and a combination of A and B.

[0067] As used in this invention, “about,” “approximately,” or “approximately” includes the stated value and the average value within an acceptable range of deviation from a particular value, wherein the acceptable range of deviation is determined by a person skilled in the art taking into account the measurement under discussion and the error associated with the measurement of the particular quantity (i.e., the limitations of the measurement system).

[0068] Furthermore, 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] Example 1:

[0070] To solve the above problems, such as Figure 1 As shown, this embodiment of the invention provides a feature extraction method for composite signals, including:

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

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

[0073] It should be noted that the steps for processing the extracted time-frequency image to obtain its features are the same as those for processing the extracted bispectral image to obtain its features. This process will be explained below.

[0074] Due to the diverse and complex characteristics of DRFM forwarding signals in complex electromagnetic environments, single-dimensional features are insufficient to distinguish the differences between composite DRFM forwarding signals. Therefore, it is necessary to extract multi-dimensional and multi-domain features to describe DRFM forwarding signals, which will help in the subsequent identification of DRFM forwarding signal types. Among them, composite DRFM forwarding signals refer to signals composed of multiple superimposed DRFM forwarding signals, i.e., composite signals in the embodiments of the present invention.

[0075] Since the frequency distribution of different types of DRFM forwarding signals is different at different time points, the time-frequency domain can better display the different DRFM forwarding signals, and the differences between DRFM forwarding signals are intuitively visible. Therefore, the embodiments of the present invention extract the features of composite signals based on time-frequency domain images.

[0076] However, for several types of DRFM forwarding signals, such as range gate pull-off (RGPO), range-velocity gate pull-off (RGPO-VGPO), velocity gate pull-off (VGPO), and angle gate pull-off (AGPO), the frequency components of the corresponding DRFM forwarding signals evolve similarly over time, making them difficult to distinguish in the time-frequency domain.

[0077] Therefore, based on the time-frequency domain image, this embodiment of the invention also combines the bispectral domain image of the composite signal to obtain the features of the DRFM relay signal. The bispectral image of the DRFM relay signal projected onto the two frequency components forms a plane. This plane can reflect the phase coupling relationship between different frequency components in the composite signal. Brightness or color represents the degree of phase coupling between frequency pairs, revealing the nonlinear interaction between different frequency components through the phase relationship. Since the bispectral image can effectively suppress Gaussian noise while preserving the amplitude and phase information of the signal, this embodiment of the invention introduces a corresponding bispectral image to capture the nonlinear features of the DRFM relay signal, which has low discriminative power in the time-frequency domain. Through multi-level feature learning and enhancement, effective complementarity and deep fusion of cross-domain (i.e., time-frequency domain and bispectral domain) features are achieved.

[0078] To address situations where composite signal quality is poor and signal-to-noise ratio is low, this embodiment of the invention first extracts local detail features and global context information from the time-frequency graph to be extracted, respectively, to obtain time-frequency graph features; and then extracts local detail features and global context information from the bispectral graph to be extracted, to obtain bispectral graph features; the global context information is used to enhance local detail features that help identify DRFM forwarded signals, thereby extracting effective features from the time-frequency graph and the bispectral graph to be extracted.

[0079] In one embodiment, such as Figure 2 As shown, after the time-frequency image to be extracted is input into the multi-scale module, the multi-scale module performs convolution on the time-frequency image at multiple feature extraction scales to extract local detail features at various locations in the time-frequency image, thus obtaining the first discriminative feature. After the bispectral image to be extracted is input into the multi-scale module, the global feature module models the global information in the time-frequency image to be extracted, thus obtaining the first global feature. The first discriminative feature and the first global feature are fused to obtain the time-frequency image feature. The processing flow for obtaining the time-frequency image feature is the same as that for obtaining the bispectral image feature, and will not be described again 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 swapping on the time-frequency plot feature and the bispectral plot feature to determine the effective channel features corresponding to each channel in the time-frequency plot feature and the bispectral plot feature; obtain the first fusion feature based on all the effective channel features.

[0082] Here, the channel is one dimension of the feature map (i.e., time-frequency map feature or bispectral map feature); such as Figure 3As shown, the feature map in this embodiment of the 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] To achieve effective complementarity and deep fusion of time-frequency domain features and bispectral domain features, this invention implements a phased fusion of cross-domain features. First, channel features in the time-frequency map features and bispectral features are exchanged between different domains (i.e., the time-frequency domain and the bispectral domain) to perform cross-domain feature fusion. In one embodiment, since the time-frequency map features and bispectral features of a composite signal may have overlapping regions with identical features, it is necessary to compare the time-frequency map features and bispectral features to determine the non-overlapping regions with different features, and then determine effective features helpful for identifying the DRFM forwarding signal from these non-overlapping regions. Figure 3 As shown, in the process of determining effective features from the non-overlapping regions of the two, it is first determined whether the features in the time-frequency plot features and the bispectral plot features are effective features. In one embodiment, when a feature is not effective, the mean value of the features in another domain is used as the effective feature and replaced with that feature. This channel exchange ensures that the non-overlapping regions of the two are all effective features, achieving the goal of deep fusion of the time-frequency plot features and the bispectral plot features. The channel exchange process will be further explained below.

[0084] Step 30: Extract local detail features from each position in the time-frequency image to be extracted to obtain the first distinguishing feature; extract local detail features from each position in the bispectral image to be extracted to obtain the second distinguishing feature; perform feature fusion on the first distinguishing feature and the second distinguishing feature based on attention weight to obtain the second fused feature.

[0085] The process of obtaining the first distinguishing feature will be described below. In one embodiment, a convolution kernel is used to calculate the dot product at each position in the time-frequency map to be extracted, so as to extract the DRFM forwarding category information of the time-frequency map to be extracted at various spatial resolutions, thereby obtaining the first distinguishing feature.

[0086] The second fusion feature is a feature map with channel attention.

[0087] Compared to the fusion process through channel exchange described above, this embodiment of the invention uses an attention mechanism to enhance and aggregate key information on the cross-spatial distribution and channels in time-frequency plot features and bispectral plot features based on the first and second distinguishing features. By effectively capturing the complex correlation between time-frequency plot features and bispectral plot features, it promotes their interaction and information complementarity.

[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] This invention extracts features from the time-frequency domain and the bispectral domain (i.e., the first fusion feature and the second fusion feature) using two fusion methods, and then concatenates the two to obtain the target feature fused from the time-frequency domain and the bispectral domain. It should be noted that the following... Figure 5 The circle containing "C" indicates that the first fusion feature and the second fusion feature are spliced ​​together.

[0090] This invention extracts local detail features and global context information from the time-frequency map and bispectral map to be extracted, respectively. Since different types of DRFM forwarding signals have low distinguishability in the time-frequency domain, information from the composite signal in the bispectral domain is used to capture nonlinear features, supplementing the information in the time-frequency domain to obtain corresponding time-frequency map features and bispectral map features. By performing channel swapping on the time-frequency map features and bispectral map features, the corresponding feature maps in both the time-frequency domain and bispectral domain become effective channel features after the swap, achieving deep inter-domain fusion of the time-frequency map features and bispectral map features, obtaining key information in the local dimension of the time-frequency domain and bispectral domain. Attention weights are used to fuse the local detail features in the time-frequency map features and bispectral map features, obtaining key information in the global context dimension of the time-frequency domain and bispectral domain. Finally, by concatenating the first and second fused features, the key information in the local and global context dimensions is integrated. The global context information guides the attention to local detail features, thereby enabling better extraction of effective features from composite signals even with low signal-to-noise ratios, thus improving the ability to identify composite signals.

[0091] To illustrate the process of obtaining time-frequency map features by processing the time-frequency map to be extracted, and obtaining bispectral map features by processing the bispectral map to be extracted, as follows: Figure 4 As shown, step 10 includes:

[0092] Step 101: 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 distinguishing feature; model the global context information in the time-frequency map to be extracted to obtain the first global feature; fuse the first distinguishing feature and the first global feature to obtain the time-frequency map feature.

[0093] On the one hand, although DRFM forwarding signals are difficult to distinguish in the time-frequency domain, their subtle features and energy distributions in the time-frequency domain are fundamentally different. Therefore, embodiments of this invention employ multi-scale convolution to extract local detail features of the time-frequency map or bispectral map to be extracted at multiple feature extraction scales (i.e., using convolution kernels of different sizes), thereby enhancing the sensitivity of the neural network of this invention to the subtle local features of the DRFM forwarding signal. In one embodiment, such as... Figure 5As shown, the multi-scale module convolves the time-frequency image to be extracted at multiple feature extraction scales to obtain the first distinguishing feature, and convolves the bispectral image to be extracted at multiple feature extraction scales to obtain the second distinguishing feature.

[0094] On the other hand, existing technologies cannot determine the location of features important for DRFM forwarding signal identification within the signal. Furthermore, the contextual information characterizing the DRFM forwarding signal type occupies a larger area than the DRFM forwarding signal itself. Therefore, extracting only artificial features from the DRFM forwarding signal is insufficient to effectively extract features helpful for DRFM forwarding signal identification. In contrast, this invention models the global information in the time-frequency diagram to be extracted, thereby extracting key features helpful for DRFM forwarding identification, i.e., the first global feature. This key feature is then used as context, combined with local detail features (i.e., the first distinguishing feature or the second distinguishing feature hereinafter), to obtain features helpful for DRFM forwarding signal type identification (i.e., time-frequency domain features or bispectral domain features hereinafter) from the DRFM forwarding signal in the time-frequency diagram or bispectral diagram to be extracted. In one embodiment, such as... Figure 5 As shown, the global feature module models the global context information in the time-frequency image to be extracted, obtaining the first global feature; the global feature module also models the global context information in the bispectral image to be extracted, obtaining the second global feature.

[0095] Finally, the fusion module integrates local detailed features and global information to capture the deeper features of the composite signal, thereby improving the robustness and accuracy of identifying the type of DRFM forwarding signal for composite signals composed of multiple types of DRFM forwarding signals.

[0096] Step 102: Convert the original bispectral image of the composite signal into a grayscale image to obtain the bispectral image to be extracted; perform convolution on the bispectral image to be extracted at multiple feature extraction scales to determine the second distinguishing feature; model the global context information in the bispectral image to be extracted to obtain the second global feature; fuse the second distinguishing feature and the second global feature to obtain the bispectral image feature.

[0097] The processing flow for the original time-frequency diagram and the original bispectral diagram is the same in the embodiments of the present invention, and will not be repeated here.

[0098] In this embodiment, it is necessary to first obtain the effective channel features, and then obtain the first fusion feature based on all the effective channel features. The determination process of the effective channel features is divided into two stages: (1) For overlapping channel features, the overlapping channel features can be directly used as effective channel features; (2) For non-overlapping channel features, it is necessary to determine whether they are effective channel features based on the channel scaling factor parameter. If they are not effective channel features, further channel swapping is required to obtain 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 graph features to obtain a first feature map to be evaluated; perform batch normalization on the bispectral graph features to obtain a second feature map to be evaluated.

[0100] In one embodiment, such as Figure 5 As shown, the time-frequency map features and dual-spectral map features output by the fusion module are input into the channel exchange module to perform channel exchange on the non-overlapping regions of the two. Before comparing the overlapping and non-overlapping regions, batch normalization is first performed to ensure that the feature dimensions are consistent.

[0101] Step 202: If the channel feature in the first feature map to be evaluated overlaps with the channel feature at the corresponding position in the second feature map to be evaluated, then the overlapping channel feature is determined as the valid channel feature in the first feature map to be evaluated and the valid channel feature in the second feature map to be evaluated.

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

[0103] In one embodiment, such as Figure 2 As 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 bispectral 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 corresponding channel features 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 the first original channel features, and the non-overlapping channel features in the second feature map to be evaluated are determined as the second original channel features; determine whether the first original channel feature is a valid channel feature, determine whether the second original channel feature is a valid channel feature, and perform channel swapping on the original channel features that are not valid channel features to obtain the corresponding valid channel features.

[0105] The criteria for determining the characteristics of an effective channel will be explained below.

[0106] In one embodiment, when the first original channel feature is not a valid channel feature, the mean value of the channel features of all channels in the current second feature map to be evaluated is calculated to obtain a first feature mean value. The first feature mean value is determined as a 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. When the second original channel feature is not a valid channel feature, the mean value of the channel features of all channels in the current first feature map to be evaluated is calculated to obtain a second feature mean value. The second feature mean value is determined as a 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.

[0107] For example, such as Figure 2 As shown, when the first original channel feature is the feature of the channel shown by the light-colored square in the time-frequency domain, during the channel exchange process 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 exchanged to the corresponding mean. In this embodiment of the invention, it is determined one by one whether the first original channel feature in the first feature map to be evaluated is a valid channel feature. If it is not a valid channel feature, the mean of all channel feature values ​​in the current second feature map to be evaluated during the channel exchange process is calculated as the first feature mean and determined as the valid channel feature of the channel shown by the light-colored square.

[0108] For example, such as Figure 2 As shown, when the second original channel feature is the feature of the channel shown by the dark square in the dual-spectral domain, it is determined one by one whether the second original channel feature in the second feature map to be evaluated is a valid channel feature. If it is not a valid channel feature, the mean value of all channel feature values ​​in the current first feature map to be evaluated during the channel exchange process is calculated as the mean value of the second feature and determined as the valid channel feature of the channel shown by the dark square.

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

[0110] This invention compares the features of each channel in the time-frequency plot with the features of each channel in the bispectral plot, and through channel swapping, replaces the channel features that are not effective channel features in the non-overlapping regions with the feature mean of the other domain. On the one hand, by selectively swapping channels, feature fusion between the time-frequency domain and the bispectral domain is achieved; on the other hand, during the channel swapping process, the key information that helps identify the composite signal in both the first and second feature maps to be evaluated continuously increases. Compared to the channel features that are not effective channel features, the feature mean of the feature map to be evaluated is 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 criteria for determining effective channel characteristics, in one embodiment, such as... Figure 7 As shown, in step 203, determining whether the first original channel feature is a valid channel feature includes:

[0112] Step 2031: When performing batch normalization on the time-frequency graph features, a scaling factor parameter is pre-set for each channel.

[0113] The specific values ​​of the scaling factor parameters pre-set for each channel are determined by those skilled in the art based on the specific application scenario.

[0114] A penalty function is added to the total loss 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 switching module, a scaling factor parameter is used in the normalization layer to evaluate the importance of each channel; in an optional embodiment, such as Figure 2 As shown, a 1-norm penalty is applied to the scaling factor parameter, and filters that satisfy the sparsity criterion are explicitly pruned, thereby enabling direct fusion of the switching channels between the time-frequency domain and the bispectral domain.

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

[0117]

[0118] Among them, f c f represents the first original channel feature of the c-th channel of the time-frequency plot. c ′ represents the effective channel feature corresponding to the c-th channel, μ c σ represents the mean of the eigenvalues ​​of all channels in the current second feature map to be evaluated. cγ represents the standard deviation of the eigenvalues ​​of all channels in the current second feature map to be evaluated. c This represents the scaling factor parameter, β. c This represents the bias parameter, where ε is a constant.

[0119] The bias parameters and constants are selected by those skilled in the art based on the specific application scenario, and are not limited here.

[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] The threshold is selected by those skilled in the art based on the specific use case, and the threshold should be 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, the sparsity constraint of the scaling factor parameter is applied to the non-overlapping regions of the time-frequency domain and the bispectral domain. If the scaling factor parameter of a certain channel of the feature map of a domain is lower than a certain threshold, it means that it has little impact on the final result and is redundant. Therefore, in this embodiment of the invention, the feature mean of another domain is used instead.

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

[0125] Step 301: 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.

[0126] Because the shallow features extracted initially often have a limited receptive field, the neural network model in this embodiment of the invention is insufficient to fully capture the complex and potential DRFM forwarding signal features in the image. Therefore, this embodiment of the invention utilizes asymmetric convolution (i.e., Figure 5The asymmetric convolution module independently further deepens the feature extraction process in the time-frequency domain and the bispectral domain, enhancing the expressive power and diversity of the features. Specifically, after the first distinguishing feature is input into the asymmetric convolution module for the time-frequency domain, the asymmetric convolution module performs convolution on the first distinguishing feature at multiple feature extraction scales to extract local detail features at various positions in the first distinguishing feature, thus obtaining the first deep feature. After the second distinguishing feature is input into the asymmetric convolution module for the bispectral domain, the asymmetric convolution module performs convolution on the second distinguishing feature at multiple feature extraction scales to extract local detail features at various positions in the second distinguishing feature, thus obtaining the second deep feature. In this embodiment of the invention, a traditional single large-size (e.g., n×n) convolutional kernel is decomposed into two smaller and asymmetric convolutional kernels (e.g., 1×n and n×1) to significantly reduce the number of parameters in the neural network model. When the asymmetric convolution module performs convolution at multiple feature extraction scales, it uses multiple asymmetric convolutional kernels to calculate the dot product at each position in the first or second discriminative feature to extract spatial features, thereby improving the efficiency of training and inference and effectively mitigating the risk of overfitting. More importantly, due to the addition of an extra nonlinear transformation layer (i.e., Figure 5 The asymmetric convolution module enables the neural network model in this embodiment of the invention to capture richer and more detailed spatial features, thereby improving the ability to represent complex DRFM forwarding signals.

[0127] It should be noted that both asymmetric convolution modules and multi-scale convolution modules use multiple convolution kernels of different sizes to extract local detail features. The core idea of ​​feature extraction in asymmetric convolution modules and multi-scale convolution modules is the same: both use convolution kernels of different sizes to extract features in parallel. The only difference is the size and distribution of the convolution kernels used by asymmetric convolution modules and multi-scale convolution modules.

[0128] Step 302: Perform feature concatenation on the first deep feature and the second deep feature to obtain the feature to be processed; divide the feature to be processed into fixed-size patches; convert each patch into a fixed-dimensional vector to be processed.

[0129] In one embodiment, after the input image (i.e., the features to be processed) is segmented into fixed-size patches through a linear layer, each patch is converted into a fixed-dimensional vector to be processed (i.e., a token) through a line normalization layer, and a positional encoding representing the relative position of each patch in the input image is added. It should be noted that... Figure 5 The circle containing "C" indicates that the first deep feature and the second deep feature are spliced ​​together.

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

[0131] The encoder then learns the features of the vector to be processed, a process that will be explained below.

[0132] Step 304: Perform linear transformation and nonlinear activation on the first encoded feature to learn the first encoded feature using a feedforward neural network and obtain the second encoded feature.

[0133] The specific method of using a feedforward neural network to learn the first encoded feature shall be determined by those skilled in the art based on the specific application scenario and with reference to existing technologies, and is not 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, this embodiment of the invention uses a multi-head attention mechanism to fuse features in the time-frequency domain and the bispectral domain, enhance and converge key features in the feature spatial distribution and channels, promote interaction and information complementarity between different features, and achieve adaptive weight allocation and fusion in the time-frequency domain and the bispectral domain.

[0136] To illustrate the process of obtaining the first encoded feature, as follows: Figure 9 As shown, 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, key weight parameter, and 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 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.

[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; 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 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 weights to perform a weighted summation on the target value vector to generate the first encoded feature.

[0143] This invention utilizes a multi-head attention mechanism by calculating the weighted sum of the query vector, key vector, and value vector to capture the global dependencies between different positions in the sequence of vectors to be processed. This allows for the acquisition of key information in the global context dimension in both the time-frequency domain and the bispectral domain. This global context information guides the focus on local details when concatenating the first and second fused features, enabling better extraction of effective features from composite signals even with low signal-to-noise ratios, thus significantly improving the ability to recognize composite signals.

[0144] When calculating the weighted sum of the query vector, key vector, and value vector to obtain the first encoded feature, the attention weight is obtained by first calculating the dot product of the query vector and the target key vector, and then the weighted sum of the attention weight and the value vector is calculated. This reduces the spatial dimension of the computation and effectively alleviates the computational complexity and memory consumption problems caused by high-dimensional data, significantly reducing computational memory overhead. This enables the neural network model of this invention to process large-scale DRFM forwarding signal data more efficiently, greatly improving the efficiency of global feature extraction during training and inference without sacrificing the model's representational ability.

[0145] Example 2:

[0146] like Figure 10 The diagram shown is an architectural schematic of a feature extraction device for composite signals according to an embodiment of the present invention. The feature extraction device for composite signals in this embodiment includes one or more processors 21 and a memory 22. Figure 10 Take a processor 21 as an example.

[0147] Processor 21 and memory 22 can be connected via a bus or other means. Figure 10 Taking the example of a connection between China and Israel via a bus.

[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 composite signals in this embodiment. The processor 21 executes the feature extraction method for composite signals by running the non-volatile software program and instructions stored in the memory 22.

[0149] Memory 22 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, memory 22 may optionally include memory remotely located relative to processor 21, which can be connected to processor 21 via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

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

[0151] This invention also provides a non-volatile computer storage medium storing computer-executable instructions that are executed by one or more processors, for example... Figure 10 A processor 21 can enable one or more processors to execute the feature extraction method for composite signals in the specific embodiments of the present invention, for example, to execute the various steps of the feature extraction method for composite signals in the embodiments of the present invention described above; it can also implement Figure 10 The various modules and units described above; or the feature extraction method for composite signals in the specific embodiments of the present invention, for example, executing the various steps of the feature extraction method for composite signals in the embodiments of the present invention described above; can also achieve Figure 10 The aforementioned modules and units.

[0152] It is worth noting that the information interaction and execution process between the modules and units in the above-mentioned device and system are based on the same concept as the processing method embodiment of the present invention. For details, please refer to the description in the method embodiment of the present invention, and will not be repeated here.

[0153] Those skilled in the art will understand that all or part of the steps in the various methods of the embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, etc.

[0154] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A feature extraction method for composite signals, characterized in that, include: Obtain the time-frequency plot and bispectral plot of the composite signal to be extracted; extract the local detail features and global context information of the time-frequency plot to be extracted to obtain the time-frequency plot features; Local detail features and global context information of the bispectral image to be extracted are extracted to obtain bispectral image features; wherein, the composite signal is the forwarding signal of the digital radio frequency memory; The time-frequency graph features are batch normalized to obtain a first feature map to be evaluated; the bispectral graph features are batch normalized to obtain a second feature map to be evaluated. If a channel feature in the first feature map to be evaluated overlaps with a channel feature at the corresponding position in the second feature map to be evaluated, then the overlapping channel feature is determined as a valid channel feature in the first feature map to be evaluated and a valid channel feature 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 corresponding channel features 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 the first original channel features, and the non-overlapping channel features in the second feature map to be evaluated are determined as the second original channel features; it is determined whether the first original channel feature is a valid channel feature, and whether the second original channel feature is a valid channel feature, and channel swapping is performed on the original channel features that are not valid channel features to obtain the corresponding valid channel features; All effective channel features in the first feature map to be evaluated are identified as time-frequency intermediate features; all effective channel features in the second feature map to be evaluated are identified as bispectral intermediate features; the time-frequency intermediate features and the bispectral intermediate features are concatenated to obtain the first fused feature; Local detail features are extracted from each position in the time-frequency image to be extracted to obtain a first distinguishing feature; local detail features are extracted from each position in the bispectral image to be extracted to obtain a second distinguishing feature; the first distinguishing feature and the second distinguishing feature are fused based on attention weights to obtain a second fused feature; The first fusion feature and the second fusion feature are concatenated to obtain the target feature of the composite signal; The step of determining whether the first original channel feature is a valid channel feature includes: When performing batch normalization on the time-frequency graph features, a scaling factor parameter is pre-set for each channel; A penalty function is added to the total loss function to constrain the sparsity of the scaling factor parameters; wherein the total loss function is used to extract the target features; 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.

2. The feature extraction method for composite signals according to claim 1, characterized in that, The process of exchanging original channel features that are not valid channel features to obtain corresponding valid channel features includes: When the first original channel feature is not a valid channel feature, the mean of the channel features of all channels in the current second feature map to be evaluated is calculated to obtain the first feature mean. The first feature mean is determined 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. When the second original channel feature is not a valid channel feature, the mean value of the channel features of all channels in the current first feature map to be evaluated is calculated to obtain the second feature mean value. The second feature mean value is determined as the valid channel feature of the channel in the current second feature map to be evaluated, so as to perform channel exchange between the first feature map to be evaluated and the second feature map to be evaluated, and to perform channel exchange on the second original channel feature.

3. The feature extraction method for composite signals according to claim 1, characterized in that, The expression for the effective channel feature is: in, The first characteristic representing the time-frequency graph The first original channel feature of each channel Indicates the first The effective channel characteristics corresponding to each channel This represents the mean of the eigenvalues ​​of all channels in the current second feature map to be evaluated. This represents the standard deviation of the eigenvalues ​​of all channels in the current second feature map to be evaluated. This represents the scaling factor parameter. Indicates the bias parameter. It is a constant.

4. The feature extraction method for composite signals according to claim 1, characterized in that, The feature fusion based on attention weights of the first distinguishing feature and the second distinguishing feature to obtain the second fused feature includes: Multiple asymmetric convolution kernels are used to calculate the dot product at each position in the first distinguishing feature to extract the spatial features of the first distinguishing feature and obtain the first deep feature; multiple asymmetric convolution kernels are used to calculate the dot product at each position in the second distinguishing feature to extract the spatial features of the second distinguishing feature and obtain the second deep feature. The first deep feature and the second deep feature are concatenated to obtain the feature to be processed; the feature to be processed is segmented into fixed-size patches; each patch is converted into a fixed-dimensional vector to be processed. Determine the attention weights of the vector to be processed, and use the attention weights to perform attention calculation on the vector to be processed to obtain the 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, thereby obtaining the second encoded feature; The first encoded feature and the second encoded feature are added together to obtain the second fused feature.

5. The feature extraction method for composite signals according to claim 4, 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 target value vector is weighted and summed using the attention weights to generate the first encoded feature.

6. The feature extraction method for composite signals according to any one of claims 1-5, characterized in that, The process involves acquiring the time-frequency plot and bispectral plot of the composite signal to be extracted; extracting the local detail features and global context information of the time-frequency plot to be extracted to obtain the time-frequency plot features; Extracting the local detail features and global context information of the bispectral image to be extracted yields bispectral features including: The original time-frequency map of the composite signal is converted into a grayscale image to obtain the time-frequency map to be extracted; the time-frequency map to be extracted is convolved at multiple feature extraction scales to determine the first distinguishing feature; the global context information in the time-frequency map to be extracted is modeled to obtain the first global feature; the first distinguishing feature and the first global feature are fused to obtain the time-frequency map feature; The original bispectral image of the composite signal is converted into a grayscale image to obtain the bispectral image to be extracted; the bispectral image to be extracted is convolved at multiple feature extraction scales to determine the second distinguishing feature; the global context information in the bispectral image to be extracted is modeled to obtain the second global feature; the second distinguishing feature and the second global feature are fused to obtain the bispectral image feature.

7. A feature extraction device for composite signals, characterized in that, The feature extraction device for the composite signal 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 feature extraction method for the composite signal according to any one of claims 1-6.

8. 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 feature extraction method for composite signals according to any one of claims 1-6.

Images