Radar active interference identification method and device based on SSR module depth feature mining

Abstract

The invention discloses a radar active interference identification method and device based on SSR module depth feature mining. The method comprises the following steps: extracting time-frequency data of an interference signal; inputting the time-frequency data into an interference identification model based on SSR to obtain an identification result of the interference signal; wherein an SSR feature deep mining module in the interference recognition model is composed of a self-attention network, an extrusion excitation network and a residual network in sequence. The SSR feature deep mining module in the model used in the recognition process is composed of the self-attention network, the extrusion excitation network and the residual network in sequence, the deep mining module can give full play to the advantages that the three networks can relieve gradient disappearance and enhance feature expression, and the recognition efficiency and the recognition robustness are improved.

Description

[0001] This application claims priority to the Chinese patent application filed with the State Intellectual Property Office on February 27, 2025, with application number 202510226657.8 and application name “Radar active interference identification method based on deep feature mining of SSR module”, the entire contents of which are incorporated by reference into this application. Technical Field

[0002] The present invention belongs to the field of radar technology, and in particular relates to a radar active interference identification method and device based on deep feature mining of an SSR module. Background Art

[0003] With the rapid development of Digital Radio Frequency Memory (DRFM) technology, various new and effective jamming methods have emerged, providing strong technical support for the innovation and advancement of electronic jamming technology. This has made electronic jamming methods increasingly flexible, complex, and intelligent. During operation, DRFM jammers can intelligently select appropriate jamming methods, set reasonable jamming parameters, and adjust the combination and strategy of multiple jamming modes based on the changing battlefield situation. This allows them to exert high-intensity and high-complexity jamming in the time, frequency, and spatial domains, and even in the joint domain, seriously affecting radar detection performance.

[0004] Current technical solutions are closely integrated with image processing techniques. They utilize information from the original signal's time, frequency, time-frequency, or wavelet domains, and after preprocessing, input it into a neural network for feature mining and classification mapping. Neural network selection has evolved from early convolutional neural networks to fine-tuned models of mature networks like Google Net and Alex Net. With the increasing maturity of deep learning, residual networks (ResNet), self-attention mechanisms, semantic segmentation models (UNet), transfer learning techniques, and multimodal technologies are increasingly being applied to interference recognition.

[0005] However, most current deep learning methods rely on a large number of samples, but obtaining sufficient samples in real-world environments is extremely difficult, resulting in poor practical applicability. There are relatively few interference recognition methods for the low-sample domain, and current low-sample recognition methods are insufficiently researched for complex electromagnetic environments, including complex interference. Current interference methods based on residual networks, self-attention mechanisms, and squeeze-excitation networks (SENs) still have many shortcomings. For example, while residual networks (ResNets) can mitigate gradient vanishing through residual connections, they lack the ability to dynamically model channel and spatial relationships. While self-attention mechanisms (Self-Attention) and squeeze-excitation networks (SEs) can enhance feature expression, when used alone, they are difficult to combine with the stable training characteristics of residual networks. This results in insufficient feature mining of input data in the low-sample domain, which in turn affects the accuracy and robustness of recognition of various interference signals in complex environments.

[0006] Therefore, the current interference recognition methods based on deep learning have poor recognition accuracy and robustness. Summary of the Invention

[0007] The embodiments of the present invention provide a radar active interference identification method and device based on deep feature mining of the SSR module, which can solve the above technical problems.

[0008] In a first aspect, an embodiment of the present invention provides a radar active jammer identification method based on deep feature mining of an SSR module, the method comprising:

[0009] Extracting time-frequency data of interference signals;

[0010] The time-frequency data is input into an SSR-based interference recognition model to obtain the recognition result of the interference signal; wherein, the SSR feature deep mining module in the interference recognition model is composed of a self-attention network, a squeeze excitation network and a residual network in sequence.

[0011] In a second aspect, an embodiment of the present invention provides a radar active jammer identification device based on deep feature mining of an SSR module, comprising:

[0012] A time-frequency extraction unit, configured to extract time-frequency data of an interference signal;

[0013] An identification unit is used to input the time-frequency data into an SSR-based interference identification model to obtain an identification result of the interference signal; wherein the SSR feature deep mining module in the interference identification model is composed of a self-attention network, a squeeze excitation network and a residual network.

[0014] In a third aspect, an embodiment of the present invention provides an electronic device comprising a processor and a memory, wherein the memory is used to store computer programs; the processor can be used to execute the computer program (instructions) stored in the memory to implement the method of the first aspect above.

[0015] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed, the method of the first aspect described above can be implemented.

[0016] It can be understood that the beneficial effects of the second to fourth aspects mentioned above can be found in the relevant description of the first aspect mentioned above, and will not be repeated here.

[0017] The beneficial effect of the embodiments of the present invention compared with the prior art is that since the SSR feature deep mining module in the model used in the recognition of the present invention is composed of a self-attention network, a squeeze excitation network and a residual network in sequence, this deep mining module can give full play to the advantages of the three networks in alleviating gradient disappearance and enhancing feature expression, thereby improving recognition efficiency and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A schematic diagram of the structure of an SSR feature deep mining module provided by an embodiment of the present invention;

[0019] Figure 2 A schematic structural diagram of an SSR-based interference recognition model provided in an embodiment of the present invention;

[0020] Figure 3 A flowchart of a radar active interference identification method based on deep feature mining of the SSR module provided in an embodiment of the present invention;

[0021] Figure 4 A schematic diagram of the structure of a radar active interference identification device based on deep feature mining of SSR modules provided by an embodiment of the present invention;

[0022] Figure 5a-5e The figure shows a comparison of recognition effects under different evaluation indicators provided by the embodiment of the present invention;

[0023] Figure 6 A schematic structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0024] In the following description, specific details such as particular system structures and techniques are provided for purposes of illustration, not limitation, to facilitate a thorough understanding of the embodiments of the present invention. However, it will be apparent to those skilled in the art that the present invention may be practiced in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present invention with unnecessary detail.

[0025] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0026] It will also be understood that the term "and / or" used in the present description and appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0027] As used in the present specification and the appended claims, the term "if" may be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" may be interpreted as meaning "upon determination" or "in response to determining" or "upon detection of [described condition or event]" or "in response to detecting [described condition or event]," depending on the context.

[0028] In addition, in the description of the present specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0029] References to "one embodiment" or "some embodiments" in the present specification mean that a particular feature, structure, or characteristic described in conjunction with that embodiment is included in one or more embodiments of the present invention. Thus, phrases such as "in one embodiment," "in some embodiments," "in other embodiments," and "in yet other embodiments" appearing in various places in this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "including," "comprising," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0030] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0031] Example 1

[0032] Figure 1 Shown is a structural diagram of an SSR feature deep mining module provided by an embodiment of the present invention.

[0033] In some embodiments, see Figure 1 , the SSR feature deep mining module can include a cascaded pre-feature extraction network 110, a self-attention network 120, a squeeze-excitation network 130 and a residual network 140.

[0034] For example, the pre-feature extraction network 110 can extract local features of the input features and perform normalization and linear transformation on them to obtain pre-features. The self-attention network 120 can perform self-attention operations on the pre-features to obtain self-attention features. The squeeze-excitation network 130 can determine weighted features based on the pre-features and the self-attention features. The residual network 140 can output deep features of the input features based on the input features and the weighted features, thereby completing a round of deep mining of the input features.

[0035] In one possible implementation, see Figure 1 , the front feature extraction network 110 may include a cascaded first convolution layer, a first batch normalization layer (BN1), a first activation layer, a second convolution layer, and a second batch normalization layer (BN2).

[0036] For example, the input features can be extracted by the first convolution layer to obtain their local features, then normalized by the first normalization layer, and then nonlinearly transformed by the first activation layer. The results of the linear transformation can then be convolved and normalized again by the second convolution layer and the second normalization layer.

[0037] Normalizing features through batch normalization layers BN1 and BN2 can accelerate network training and improve the generalization ability of the model.

[0038] In an example, the convolution kernel sizes of the first convolution layer and the second convolution layer may be 3×3, and the activation function used by the first activation layer may be a ReLU activation function.

[0039] In one possible implementation, the self-attention network 120 can specifically use the self-attention mechanism to multiply the input pre-features by different coefficient matrices to calculate the query matrix Q, key matrix K, and value matrix V, and then perform the self-attention operation through the following formula (1.1) to obtain the self-attention features.

[0040] The interaction between feature channels can be enhanced through self-attention operation.

[0041] In one example, the self-attention feature can satisfy the following formula:

[0042]

[0043] Among them, Z is the self-attention feature, d k is the dimension of the front feature.

[0044] For example, the softmax in formula (1.1) is an activation function that normalizes numerical values into probabilities. The softmax is used to calculate the i-th element x of the input vector x. i The calculation process can be expressed as:

[0045]

[0046] In one possible implementation, the squeeze excitation network 130 may include a cascaded first fully connected layer (FC1), a second activation layer, a second fully connected layer (FC2), a third activation layer, and a weighted layer.

[0047] For example, the self-attention feature can be reduced in dimension by the first fully connected layer, compressing the C channels of the self-attention feature into C / r channels to reduce the amount of computation; then it is transformed nonlinearly by the second activation layer, and then increased in dimension by the second fully connected layer to restore the number of channels to C; then it is normalized by the third activation layer to obtain a squeeze excitation weight of dimension 1×1×C; finally, the squeeze excitation weight is multiplied element-by-element by the previous feature through the weighted layer to weight the squeeze excitation weight to the previous feature to obtain a weighted feature.

[0048] Exemplarily, the second activation layer may use a Relu activation function, and the third activation layer may use a sigmoid activation function.

[0049] In one example, the squeeze activation weights may satisfy the following formula:

[0050] F ex (z,W)=σ(g(z,W))=σ(W2δ(W1z))(1.3)

[0051] Among them, F ex (z,W) is the squeeze activation weight, z is the self-attention feature, W is the feature width, δ(·) represents the ReLU activation function, and W1 and W2 are the weight coefficients of the two fully connected layers.

[0052] Accordingly, the weighted features can satisfy the following formula:

[0053]

[0054] in, is the weighted feature, F scale(·) represents the process of element-by-element multiplication, s c is the weight of the channel, u c is a two-dimensional matrix.

[0055] In one possible implementation, see Figure 1 , the residual network 140 may include a cascaded residual block and a fourth activation layer.

[0056] In one example, the residual block can add the weighted features to the input features to complete the residual connection, and perform a nonlinear transformation through the fourth activation layer to obtain the deep features.

[0057] For example, the activation functions used in each activation layer of the SSR feature deep mining module are not exactly the same, and the fourth activation layer may use the ReLU activation function.

[0058] Exemplarily, the processing of the residual block can be expressed as:

[0059] y=F(x,{W i})+W s x

[0060] Among them, y is the output result of the residual block, F(·) is the residual function, and W s is a weight matrix used to match the dimension of the input feature, and x is the input feature.

[0061] In one possible implementation, to avoid increasing the processing difficulty when the weighted features are not consistent with the input feature size, see Figure 1 , the SSR feature deep mining module can also include a downsampling network 150.

[0062] For example, before inputting the input feature into the downsampling network 150, it can be determined whether the size of the input feature meets the output size. If so, it is directly input into the residual network; if not, it is first downsampled by the downsampling network and then input into the residual network.

[0063] For example, the downsampling network 150 may use a 1x1 convolution kernel to adjust the shape of the input features.

[0064] In a possible implementation, an average pooling layer may be provided between the front feature extraction network 110 and the self-attention network 120 in the SSR feature deep mining module.

[0065] For example, the average pooling layer can perform adaptive global average pooling on the pre-processed features, compressing the spatial dimensions of the feature map to 1×1 to obtain a channel-level global feature description. The pooled features are then flattened into a one-dimensional vector and an unsqueeze operation is performed to add a dimension for input into the self-attention network 120 for inter-channel interaction calculations. The output of the self-attention network 120 can be restored to its original dimension through a squeeze operation.

[0066] Residual networks (ResNet), squeeze-excitation networks (SENet, SE), and self-attention mechanisms (Self-Attention) each have certain advantages, but these three networks are difficult to combine into a single module. Specifically, the self-attention mechanism has a high time complexity, which causes a surge in computational complexity when combined with the other two networks, exceeding the hardware's capacity. When combining the squeeze-excitation network with the self-attention mechanism, the SE network may suppress the expression of certain feature channels, while the self-attention mechanism requires the preservation of complete spatial information. Furthermore, the self-attention mechanism may introduce noise that interferes with residual learning.

[0067] Therefore, the present invention sets an average pooling layer before the self-attention network to reduce the amount of computation, and uses the first fully connected layer in the squeeze-excitation network to reduce the dimension of the output of the self-attention network, further reducing the amount of computation. By placing the self-attention network before the other two networks, the problem of poor self-attention operation caused by the SE network suppressing certain feature channels can be avoided. The SE network is placed between the self-attention network and the residual network. By squeezing the excitation weights, the weights of unimportant channels can be reduced, reducing the impact of noise introduced in the self-attention operation on the residual network.

[0068] Moreover, compared with some traditional methods that integrate squeeze-excitation networks and residual networks within self-attention networks, the present invention cascades the self-attention network, squeeze-excitation network and residual network. Through the local-channel-global feature mining process, different modules can process features in stages, forming a multi-level modeling from local to global, from channel to space, and enhancing the feature extraction effect; moreover, the cascade structure allows users to dynamically adjust the module ratio according to task requirements, enhancing the flexibility of feature extraction; in addition, the cascade structure also makes it easier for the model to be trained in stages, reducing the difficulty of multi-module joint optimization.

[0069] Example 2

[0070] Figure 2 The diagram below shows a schematic diagram of the structure of an SSR-based interference recognition model provided by an embodiment of the present invention. As an example and not a limitation, the model may include a feature extraction module 210 , multiple cascaded SSR feature deep mining modules 100 , and a category prediction module 220 .

[0071] In one possible implementation, see Figure 2 , the feature extraction module 210 can extract the first level features of the interference signal; see Figure 1 The x-th cascaded SSR feature deep mining module can use the x-1-th deep feature as the input feature, perform deep mining on the features of the interference signal, and obtain the x-th deep feature; the category prediction module 220 can predict the category of the interference signal based on the N-th deep feature to obtain the recognition result.

[0072] Exemplarily, x is a positive integer less than or equal to N, N is the total number of SSR feature deep mining modules, and the 0th deep feature is the first-level feature output by the feature extraction module 210.

[0073] Alternatively, see Figure 2 , an adaptive average pooling layer (AvgPool) and a fully connected layer (FC) can be set between the Nth SSR feature deep mining module and the category prediction module 220.

[0074] In one example, the model can be trained based on a back-propagation algorithm under the constraints of a cross-entropy loss.

[0075] For example, the loss function for calculating the cross entropy loss can be expressed as:

[0076]

[0077] Among them, N is the total number of samples in a batch, represents the true label of the i-th sample, Represents the result of exponentially scaling the original prediction score for a category.

[0078] Example 3

[0079] The radar active interference identification method based on deep feature mining of the SSR module provided in the embodiment of the present invention can be applied to electronic devices such as mobile terminals, personal laptops, supercomputers, etc. The embodiment of the present invention does not impose any restrictions on the specific type of electronic equipment.

[0080] Figure 3 The following is a flowchart illustrating an implementation of a radar active interference identification method based on deep feature mining of an SSR module, provided by an embodiment of the present invention. As an example and not a limitation, the method can be applied to the electronic device described above. The method may include steps S301 and S302, each of which is described below.

[0081] S301: Extract time-frequency data of the interference signal.

[0082] Exemplarily, the time-frequency data of the interference signal may include the real part, the imaginary part and the modulus value of the time-frequency graph of the interference signal.

[0083] In a possible implementation, the baseband data of the interference signal may be subjected to time-frequency analysis by short-time Fourier transform to obtain its time-frequency joint distribution function; and then the time-frequency data may be extracted according to the time-frequency joint distribution function.

[0084] In one example, the time-frequency signal x(t) may be segmented using a window function, and Fourier transform may be performed on each segmented signal to obtain a time-frequency joint distribution function of the signal x(t).

[0085] For example, the time-frequency joint distribution function may satisfy the following formula:

[0086]

[0087] Among them, STFT x (t,f) is the joint time-frequency distribution function, ω is the window function, f is the frequency, x(τ) is the value of x(t) at the center, and I(t,f) and R(t,f) are the real and imaginary parts of the time-frequency plot, respectively. The STFT cannot achieve high resolution in both the time and frequency domains. A shorter sliding window yields higher time resolution, while a longer sliding window yields higher frequency resolution.

[0088] Exemplarily, the modulus and phase of the interference signal time-frequency diagram may satisfy the following formulas:

[0089]

[0090] P(t,f)=arctan(I(t,f) / R(t,f))(1.8).

[0091] S302: Input the time-frequency data into an SSR-based interference recognition model to obtain a recognition result of the interference signal.

[0092] Exemplarily, the time-frequency data of the interference signal may be input into the model provided in the above-mentioned embodiment 2 to obtain the identification result of the interference signal.

[0093] Since the SSR feature deep mining module in the model used in the recognition of the present invention is composed of a self-attention network, a squeeze excitation network and a residual network in sequence, this deep mining module can give full play to the advantages of the three networks and improve the recognition efficiency and robustness.

[0094] Example 4

[0095] Figure 4The diagram shows a schematic diagram of a radar active jammer identification device based on deep feature mining of an SSR module according to an embodiment of the present invention. As an example and not a limitation, the device 400 may include a time-frequency extraction unit 410 and an identification unit 420.

[0096] Exemplarily, the time-frequency extraction unit 410 is used to extract the time-frequency data of the interference signal; the identification unit 420 is used to input the time-frequency data into the SSR-based interference recognition model to obtain the recognition result of the interference signal; wherein, the SSR feature deep mining module in the interference recognition model is composed of a self-attention network, a squeeze excitation network and a residual network.

[0097] In order to better illustrate the beneficial effects of the present invention, the following simulation experiments were conducted:

[0098] Exemplarily, in the hardware platform of the simulation experiment of the present invention, the CPU is Intel 8362 and the GPU is RTX 3090; the software platform of the simulation experiment of the present invention is MATLAB R2023b simulation software and PyCharm 2024.1 under the Windows 11 Home Chinese version 64-bit operating system.

[0099] For example, the simulation experiment includes 12 types of signals, namely, LFM signal of real target echo (Non-jamming), intermittent sampling forwarding interference (ISRJ), distance false target interference (DDJ), slice reconstruction interference (CI), smart noise interference (SNJ), comb spectrum interference (Comb), distance false target interference plus intermittent sampling interference (DDJ+ISRJ), distance false target interference plus noise convolution interference (DDJ+SNJ), distance false target interference plus comb spectrum interference (DDJ+Comb), comb spectrum interference plus intermittent sampling interference (Comb+ISRJ), blocking interference (BJ), and dense false target interference (DFTJ).

[0100] Exemplarily, the model provided in Example 2 of the present invention can be trained with 10 training times, a batch size of 16, a learning rate of 0.01, an optimizer of Adam, and a training data set ranging from 5% to 20% (step size 5%). Five independent experiments were conducted under each training ratio. At the same time, the recognition effects of the trained model and convolutional neural network (CNN2d), enhanced convolutional neural network (ECNN), weighted enhanced convolutional neural network (WECNN), and weighted enhanced convolutional neural network with integrated transfer learning (WECNN-TL) were compared; accuracy, precision, recall, F1 score, and visual confusion matrix were used as evaluation criteria for various methods.

[0101] Specifically, the number of predicted positive examples that are actually positive examples can be recorded as TP, the number of predicted positive examples that are actually negative examples can be recorded as FP, the number of predicted negative examples that are actually positive examples can be recorded as FN, and the number of predicted negative examples that are actually positive examples can be recorded as TN. Then the accuracy rate = (Tp+TN) / (TP+FP+FN+TN), the precision rate = TP / (TP+FP), the recall rate = TP / (TP+FN), and the F1 score = 2PR / (P+R).

[0102] Table 1 below shows the interference recognition accuracy (mean ± standard deviation) of 5 independent experiments with a training ratio of 5%. It can be seen from Table 1 that the recognition results of the method of the present invention are optimal when both single interference and composite interference exist. The number of individual evaluation indicators of each category with an average value of more than 95% and a standard deviation of less than 0.5% accounts for the largest number.

[0103] Figure 5a-5e Shown is a comparison chart of recognition effects under different evaluation indicators provided by an embodiment of the present invention.

[0104] Specifically, Figure 5a Shown are the average recognition accuracy under each training ratio and the standard deviation under 5% training ratio. Figure 5b Shown are the average accuracy at each training percentage and the standard deviation at 5% training percentage. Figure 5c Shown are the average recognition recall rate under each training ratio and the standard deviation under 5% training ratio. Figure 5d Shown are the average F1 scores for each training percentage and the standard deviation for the 5% training percentage. Figure 5e Shown is the confusion matrix of an experimental result.

[0105] from Figure 5a-5e It can be seen that the method provided by the present invention has the best recognition effect for the situations where both single interference and compound interference exist when the training ratio is low. The average recognition accuracy is as high as 95.09% when the training ratio is 5%.

[0106] Therefore, the present invention can give full play to the advantages of the three networks and improve recognition efficiency and recognition robustness.

[0107] Example 5

[0108] Figure 6 FIG. 1 is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. Figure 6 The electronic device 600 shown may include: at least one processor 610 ( Figure 6Only one processor is shown in the figure), a memory 620, and a computer program 630 stored in the memory 620 and executable on the at least one processor 610, wherein the processor 610 implements the steps of any of the above-mentioned method embodiments when executing the computer program 630.

[0109] The electronic device 600 may be a processing device such as a robot that can implement the above method. The embodiment of the present invention does not impose any limitation on the specific type of the electronic device.

[0110] Those skilled in the art will understand that Figure 6 The electronic device 600 is merely an example and does not limit the electronic device. The electronic device 600 may include more or fewer components than shown in the figure, or may combine certain components or different components. For example, the electronic device 600 may also include an input and output interface.

[0111] The processor 610 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASTC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gates, or transistor logic devices, or discrete hardware components. A general-purpose processor may be a microprocessor or any conventional processor.

[0112] In some embodiments, the memory 620 may be an internal storage unit, such as a hard disk or a memory. In other embodiments, the memory 620 may also be an external storage device, such as a plug-in hard disk, a smart memory card (SMC), a secure digital (SD) card, a flash card, etc. Furthermore, the memory 620 may include both an internal storage unit and an external storage device. The memory 620 is used to store an operating system, an application program, a boot loader, data, and other programs, such as the program code of the computer program. The memory 620 may also be used to temporarily store data that has been output or is about to be output.

[0113] It should be understood that the order of execution of the steps in the above embodiments does not necessarily mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0114] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of the present invention. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0115] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0116] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

Claims

1. A radar active interference identification method based on deep feature mining of SSR module, characterized in that: include: Extracting time-frequency data of interference signals; The time-frequency data is input into an SSR-based interference recognition model to obtain the recognition result of the interference signal; wherein, the SSR feature deep mining module in the interference recognition model is composed of a self-attention network, a squeeze excitation network and a residual network in sequence.

2. The method according to claim 1, characterized in that The time-frequency data includes the real part, imaginary part and modulus of the time-frequency diagram of the interference signal; The step of extracting the time-frequency data of the interference signal includes: Performing time-frequency analysis on the baseband data of the interference signal by short-time Fourier transform to obtain a time-frequency joint distribution function of the interference signal; The time-frequency data is determined according to the time-frequency joint distribution function.

3. The method according to claim 1, characterized in that The interference identification model includes: A feature extraction module, configured to extract first-level features of the interference signal; Multiple cascaded SSR feature deep mining modules, the xth SSR feature deep mining module is used to perform deep mining on the features of the interference signal according to the x-1th depth feature to obtain the xth depth feature, wherein x is a positive integer less than or equal to N, N is the total number of the SSR feature deep mining modules, and the 0th depth feature is the first-level feature; A category prediction module is used to predict the category of the interference signal according to the Nth depth feature to obtain the recognition result.

4. The method according to claim 3, characterized in that The x-th SSR feature deep mining module includes a cascaded front feature extraction network, a self-attention network, a squeeze excitation network and a residual network; The front feature extraction network is used to extract the local features of the x-1th depth feature and perform normalization and linear transformation on the extracted local features to obtain the xth front feature; The self-attention network is used to perform a self-attention operation on the x-th pre-feature to obtain an x-th self-attention feature; The squeeze-excitation network is used to determine an xth weighted feature based on the xth pre-feature and the xth self-attention feature; The residual network is used to determine the xth depth feature according to the xth depth feature and the xth weighted feature.

5. The method according to claim 4, characterized in that The front feature extraction network includes a cascaded first convolution layer, a first batch normalization layer, a first activation layer, a second convolution layer and a second batch normalization layer.

6. The method according to claim 4, characterized in that The squeeze excitation network includes a cascaded first fully connected layer, a second activation layer, a second fully connected layer, a third activation layer, and a weighted layer.

7. The method according to claim 6, characterized in that Between the front feature extraction network and the self-attention network, the x-th SSR feature deep mining module is further provided with an average pooling layer.

8. The method according to claim 6, characterized in that The residual network includes cascaded residual blocks and a fourth activation layer, wherein the activation functions used in each activation layer in the SSR feature deep mining module are not exactly the same.

9. A radar active interference identification device based on deep feature mining of SSR module, characterized in that: include: A time-frequency extraction unit, configured to extract time-frequency data of an interference signal; An identification unit is used to input the time-frequency data into an SSR-based interference identification model to obtain an identification result of the interference signal; wherein the SSR feature deep mining module in the interference identification model is composed of a self-attention network, a squeeze excitation network and a residual network.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

Images