A method, device and equipment for identifying non-cooperative composite modulation signals in a clutter background

Abstract

The application relates to a non-cooperative composite modulation signal recognition method, device and equipment in a clutter background, and belongs to the technical field of modulation signal recognition. The method comprises the following steps: acquiring an airborne radar echo signal containing a non-cooperative composite modulation signal in a clutter background; acquiring a radar echo RD spectrum through RD transformation and inputting the radar echo RD spectrum into a pre-trained complex signal feature extraction network for clutter RD spectrum feature learning, and removing the clutter RD spectrum to obtain a composite modulation signal RD spectrum, and then sequentially performing maximum value-based pulse selection, inverse matching filtering and short-time Fourier transformation to acquire time-frequency spectra of each sub-modulation signal component in the non-cooperative composite modulation signal, and inputting the time-frequency spectra into a residual neural network for modulation signal type recognition, and finally completing non-cooperative composite modulation signal recognition based on the types of the sub-modulation signal components. The method can improve the accuracy and robustness of composite modulation signal recognition in a complex electromagnetic environment.

Description

Technical Field

[0001] This application relates to the field of modulation signal recognition technology, and in particular to a method, apparatus and equipment for recognizing non-cooperative composite modulation signals under clutter background. Background Technology

[0002] Low-power intra-pulse modulated signals and strong clutter pose significant challenges to airborne radar target detection. Identifying the type of modulated signal is a prerequisite and key to the accurate implementation of anti-modulation measures by airborne radar. During look-down detection, strong clutter severely impairs the airborne radar's ability to identify active modulated signals. Therefore, identifying composite intra-pulse modulated signals against a strong clutter background remains a pressing problem to be solved.

[0003] In recent years, methods based on manual feature extraction, such as intrinsic integral sequence features, Doppler compensation results, and fractional Fourier transform information, have been successfully applied to modulated signal detection and identification. However, these methods fail when clutter is present. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, and device for identifying non-cooperative composite modulation signals under clutter backgrounds to address the aforementioned technical problems.

[0005] A method for identifying non-cooperative composite modulation signals against a clutter background, the method comprising: Acquire airborne radar echo signals containing non-cooperative composite modulation signals against a clutter background; The radar echo RD spectrum of the airborne radar echo signal is obtained by RD transformation. The radar echo RD spectrum is input into a pre-trained complex-valued signal feature extraction network for clutter RD spectrum feature learning. Clutter RD spectrum in radar echo RD spectrum is removed and purified to obtain composite modulation signal RD spectrum. Based on the RD spectrum of the composite modulation signal, pulse selection based on the maximum value, inverse matched filtering, and short-time Fourier transform are performed sequentially to obtain the time spectrum of each sub-modulation signal component in the non-cooperative composite modulation signal. The spectrum is then input into the residual neural network for modulation signal type identification. Finally, the non-cooperative composite modulation signal identification is completed based on the type of each sub-modulation signal component.

[0006] In one embodiment, the airborne radar echo signal containing non-cooperative composite modulation signals against a clutter background is represented as: ; in, for The airborne radar echo signal after frequency conversion at that time. , , and These represent modulation signal 1, modulation signal 2, target echo, and the first... The amplitude of the radar echo corresponding to each clutter scattering point Indicates the pulse repetition interval. , , , These represent modulation signal 1, modulation signal 2, target echo, and the first... The echo delay corresponding to each clutter scattering point , , and These represent modulation signal 1, modulation signal 2, target echo, and the first... The Doppler frequency shift corresponding to each clutter scattering point , , , and These represent the radar signal, modulation signal 1, modulation signal 2, Gaussian white noise, and the number of pulses within a coherent processing interval, respectively. For pulse count index, The superscript indicates the number of clutter scattering points. The imaginary unit is used; the combination of modulation signal 1 and modulation signal 2 is a non-cooperative composite modulation signal, and the modulation methods of modulation signal 1 and modulation signal 2 are different.

[0007] In one embodiment, the complex-valued signal feature extraction network consists of an input layer, a bi-branch convolutional layer, a feature concatenation unit, a residual unit, a two-dimensional average pooling unit, a pointwise convolutional unit, an attention unit, and an output layer, which are connected in series. The two-dimensional average pooling unit, the pointwise convolutional unit, and the attention unit form a channel attention mechanism module, which is used to generate attention weights to perform channel attention weighting on the output feature map of the residual unit, thereby strengthening the channel features that are strongly correlated with clutter components and suppressing redundant information, thus achieving precise focusing and enhancement of clutter features. The input layer is a convolutional kernel with a size of [size missing]. A 2D complex-valued convolutional layer is used to process the single-channel radar echo RD spectrum of the input. Complex feature modeling is performed to provide a basic representation for subsequent complex feature extraction. The dimension of the output feature map of the input layer is [missing value]. ;in, For complex numbers; A dual-branch convolutional layer consists of two parallel feature extraction units, each containing two convolutional kernels with a size of [missing information]. The 2D complex-valued convolutional layers are used to extract the amplitude and phase features of clutter and target echo under different receptive fields, respectively, to achieve parallel modeling of clutter components of different scales in the radar echo RD spectrum. The feature concatenation unit is used to concatenate the output feature maps of two feature extraction units along the channel dimension, fusing complementary information from the real / imaginary parts of features extracted from different branches to enhance feature representation and form a high-dimensional complex feature representation containing multi-scale clutter information, expressed as: ;in, and These are the output feature maps of the first feature extraction unit and the second feature extraction unit, respectively. and All dimensions are , The real part is , The imaginary part is , , They are respectively and The real part quantity, , They are respectively and The imaginary part; The residual unit constructs a cross-layer information transmission path through the residual connection mechanism, further representing the high-dimensional complex features after splicing while preserving the basic features of the underlying layers. The process involves processing the data and obtaining the processed results to alleviate the gradient vanishing problem during network training and enhance the deep representation capability of clutter features. The specific process is as follows: First, a convolutional kernel with a size of... 2D complex value convolutional layer pairs Compression is performed along the channel dimension, resulting in a feature map with the following dimension: The compressed feature map is then summed with the output feature map of the input layer for the first time to obtain the first residual summed feature map. This first residual summed feature map is then input into two convolutional kernels with a size of [missing value]. A 2D complex-valued convolutional layer is used to obtain the output feature map, which is then added to the first residual. A second residual is then added to the feature map to obtain the processed result. ; Two-dimensional average pooling units are used for... Global downsampling is performed in the spatial dimension. While preserving the amplitude and phase information of complex features, the spatial dimension of the feature map is reduced, and abstract features containing the global clutter distribution pattern are extracted, reducing subsequent computational load. The pooling kernel size is [size missing]. Step size is Fill with Output feature map of a two-dimensional average pooling unit The spatial dimensions are calculated as follows: ; in, and These represent the height and width of the two-dimensional pooling window, respectively. and These represent the stride of the two-dimensional pooling window in each sliding motion along the height and width directions, respectively. and These represent the number of zero-padding elements at the height and width edges of the input feature map of the two-dimensional average pooling unit, respectively. and These represent the height and width of the input feature map for the two-dimensional average pooling unit, respectively. and These represent the height and width of the output feature map of the two-dimensional average pooling unit, respectively. For the number of channels, This is the floor function; The element-wise mathematical definition of pooling operations is as follows: ; in, for In the Channel, First line, number Complex output values ​​for column positions, and They are respectively The real and imaginary components and For local index variables within the pooling kernel, for In the Channel, First line, number Complex output values ​​for column positions, and They are respectively The real and imaginary components; combined into complex form, it is: ; After completing the pooling operation, the output feature map of the two-dimensional average pooling unit is... This is denoted as the channel attention weight map. ,in, For the real number field; Pointwise convolutional units are used to pass through two convolutional kernels with a size of 2D complex-valued convolutional layer channel attention weight map The channel attention weights obtained after channel adjustment through two pointwise convolutions are expressed as follows: ; in, These are the weight matrices for the first and second pointwise convolutions, respectively. These are the biases for the first and second pointwise convolutions, respectively; and For activation functions; Attention action unit is used to assign channel attention weights Broadcasting to spatial dimensions, that is, to the original Element-wise multiplication yields a multi-channel feature map, represented as follows: ; The output layer is used to pass through a convolutional kernel with a size of 2D complex-valued convolutional layers achieve feature fusion, combining multi-channel feature maps. Mapping to a single channel to achieve RD spectrum reconstruction is represented as: ; in, For single-channel reconstructed clutter RD spectrum, This is the equivalent weight matrix of the output layer. This is the equivalent bias vector for the output layer. A vectorization operator that flattens a 3D tensor into column vectors in sequence. For matrix reshaping operators.

[0008] In one embodiment, the loss function for training the complex-valued signal feature extraction network is expressed as: ; in, and Let the distance dimension and the Doppler dimension of the RD spectrum represent the total number of elements, respectively. and These represent the unit numbers in the distance dimension and Doppler dimension of the RD spectrum, respectively; and These represent the RD spectrum of the radar echo, respectively. The real and imaginary parts of each unit; This represents the clutter RD spectrum estimated by the feature extraction network for complex-valued signals. and These represent the first and second halves of the RD spectrum of the composite modulated signal, respectively. The real and imaginary parts of each unit.

[0009] In one embodiment, clutter RD spectra in the radar echo RD spectrum are removed to purify the composite modulation signal RD spectrum, including: The normalized clutter RD spectrum, estimated by the complex-valued signal feature extraction network, is removed from the normalized radar echo RD spectrum to obtain the composite modulated signal RD spectrum, which is represented as: ; in, , and These represent the radar echo RD spectrum, the maximum value of the radar echo RD spectrum, and the clutter RD spectrum output by the network estimation, respectively.

[0010] In one embodiment, based on the RD spectrum of the composite modulation signal, pulse selection based on the maximum value, inverse matched filtering, and short-time Fourier transform are performed sequentially to obtain the time spectrum of each sub-modulation signal component within the non-cooperative composite modulation signal, including: The maximum value of each pulse sequence is extracted from the RD spectrum of the composite modulated signal to form the Doppler vector. , and select The two pulse sequences corresponding to the two peaks with the largest amplitude are subjected to matched filtering inverse transform to recover and obtain each sub-modulation signal component in the non-cooperative composite modulation signal. The time spectrum of each sub-modulation signal component is obtained by performing short-time Fourier transform on each sub-modulation signal component.

[0011] A device for identifying non-cooperative composite modulation signals under clutter background, the device comprising: The signal acquisition module is used to acquire airborne radar echo signals containing non-cooperative composite modulation signals against a clutter background. The clutter removal module is used to obtain the radar echo RD spectrum of the airborne radar echo signal through RD transformation, input the radar echo RD spectrum into a pre-trained complex value signal feature extraction network for clutter RD spectrum feature learning, and remove the clutter RD spectrum in the radar echo RD spectrum to obtain the composite modulation signal RD spectrum. The composite modulation signal identification module is used to perform pulse selection based on the maximum value, inverse matched filtering, and short-time Fourier transform sequentially based on the RD spectrum of the composite modulation signal to obtain the time spectrum of each sub-modulation signal component in the non-cooperative composite modulation signal. The spectrum is then input into the residual neural network for modulation signal type identification. Finally, the non-cooperative composite modulation signal identification is completed based on the type of each sub-modulation signal component.

[0012] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps: Acquire airborne radar echo signals containing non-cooperative composite modulation signals against a clutter background; The radar echo RD spectrum of the airborne radar echo signal is obtained by RD transformation. The radar echo RD spectrum is input into a pre-trained complex-valued signal feature extraction network for clutter RD spectrum feature learning. Clutter RD spectrum in radar echo RD spectrum is removed and purified to obtain composite modulation signal RD spectrum. Based on the RD spectrum of the composite modulation signal, pulse selection based on the maximum value, inverse matched filtering, and short-time Fourier transform are performed sequentially to obtain the time spectrum of each sub-modulation signal component in the non-cooperative composite modulation signal. The spectrum is then input into the residual neural network for modulation signal type identification. Finally, the non-cooperative composite modulation signal identification is completed based on the type of each sub-modulation signal component.

[0013] The aforementioned method, apparatus, and equipment for identifying non-cooperative composite modulation signals under clutter backgrounds fully consider the distribution characteristics of airborne clutter in the RD domain and construct an end-to-end composite modulation signal identification process. Its beneficial effects are as follows: First, by learning the RD domain distribution characteristics of clutter through a complex-valued signal feature extraction network, non-stationary airborne clutter components in the radar echo RD spectrum under complex backgrounds can be accurately identified and eliminated, significantly improving the purification accuracy of non-cooperative composite modulation signals. Then, based on the purified composite modulation signal RD spectrum, fine decomposition of sub-modulation signal components is achieved through pulse selection based on maximum values, inverse matched filtering, and short-time Fourier transform. This solves the problem of traditional methods treating composite modulation signals as a whole and failing to distinguish different modulation signal types, providing a reliable time-frequency feature basis for subsequent identification. Finally, by using a residual neural network to discriminate the modulation signal type of each sub-modulation signal component, the differentiated patterns of different modulation signal types can be effectively captured, significantly improving the accuracy and robustness of composite modulation signal identification in complex electromagnetic environments. The entire process forms a positive collaborative closed loop of "clutter suppression - modulation signal purification - sub-component decomposition - type discrimination", which systematically improves the overall performance of airborne radar composite modulation signal recognition from the source to the end, and effectively supports the airborne radar target detection performance in complex scenarios. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating a method for identifying non-cooperative composite modulation signals against a clutter background in one embodiment. Figure 2 This is a schematic diagram of airborne radar looking down detection in one embodiment; Figure 3 This is a schematic diagram of the complex-valued signal feature extraction network (CV-SFEN) in one embodiment; Figure 4 This is a schematic diagram illustrating the clutter RD spectrum extraction results of a composite modulated signal containing ISRJ (Intermittent Sample-and-Forward Modulated Signal) + MDFJ (Partial Pulse Dense Forward Modulated Signal) by CV-SFEN in one embodiment when the modulation signal clutter-to-noise ratio (JCNR) is -20 dB; wherein, Figure 4 (a) is the clutter RD spectrum containing ISRJ+MDFJ. Figure 4 (b) The clutter RD spectrum extracted by the network; Figure 5This is a schematic diagram illustrating the clutter RD spectrum extraction results of a composite modulated signal containing ISRJ+SMSPJ (spectral dispersion modulated signal) using CV-SFEN at JCNR = -20 dB in one embodiment; where, Figure 5 (a) is the clutter RD spectrum containing ISRJ+SMSPJ. Figure 5 (b) The clutter RD spectrum extracted by the network; Figure 6 This is a schematic diagram illustrating the clutter RD spectrum extraction results of a composite modulated signal containing SMSPJ+MDFJ using CV-SFEN at JCNR = -20 dB in one embodiment; wherein, Figure 6 (a) is the clutter RD spectrum containing SMSPJ+MDFJ. Figure 6 (b) The clutter RD spectrum extracted by the network; Figure 7 This is a schematic diagram illustrating pulse selection for different composite modulation signals in one embodiment when JCNR = -20 dB; wherein, Figure 7 (a) is a schematic diagram of pulse selection including ISRJ+MDFJ. Figure 7 (b) is a schematic diagram of pulse selection including ISRJ+SMSPJ. Figure 7 (c) is a schematic diagram of pulse selection including SMSPJ+MDFJ; Figure 8 This is a schematic diagram showing the comparison of modulation signal recognition accuracy of different algorithms under different JCNR in one embodiment; Figure 9 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0016] In one embodiment, such as Figure 1As shown, a method for identifying non-cooperative composite modulation signals under clutter backgrounds is provided, comprising the following steps: acquiring airborne radar echo signals containing non-cooperative composite modulation signals under clutter backgrounds; obtaining the radar echo RD spectrum of the airborne radar echo signal through RD (range-Doppler) transform, inputting the radar echo RD spectrum into a pre-trained complex-valued signal feature extraction network for clutter RD spectrum feature learning, and removing the clutter RD spectrum from the radar echo RD spectrum to purify and obtain the composite modulation signal RD spectrum; based on the composite modulation signal RD spectrum, sequentially performing pulse selection based on maximum value, inverse matched filtering, and short-time Fourier transform to obtain the time spectrum of each sub-modulation signal component within the non-cooperative composite modulation signal, and inputting it into a residual neural network for modulation signal type identification; finally, completing the identification of non-cooperative composite modulation signals based on the type of each sub-modulation signal component.

[0017] The aforementioned method for identifying non-cooperative composite modulation signals under clutter backgrounds firstly learns the RD domain distribution characteristics of clutter through a complex-valued signal feature extraction network. This accurately identifies and removes non-stationary airborne clutter components from the RD spectrum of radar echoes under complex backgrounds, significantly improving the purification accuracy of non-cooperative composite modulation signals. Then, based on the purified composite modulation signal RD spectrum, refined decomposition of sub-modulation signal components is achieved through pulse selection based on maximum values, inverse matched filtering, and short-time Fourier transform. This solves the problem of traditional methods treating composite modulation signals as a whole and failing to distinguish different modulation signal types, providing a reliable time-frequency feature basis for subsequent identification. Finally, the residual neural network is used to discriminate the modulation signal type of each sub-modulation signal component, effectively capturing the differentiated patterns of different modulation signal types and significantly improving the accuracy and robustness of composite modulation signal identification in complex electromagnetic environments.

[0018] In one embodiment, when the airborne radar is in look-down detection mode, such as Figure 2 As shown, assume that a target aircraft carrying a modulation signal transmitter transmits modulation signal 1 within its main beam, and another modulation signal transmitter transmits modulation signal 2 (i.e., ...) near the target aircraft. Figure 2 (The two red dashed arrows in the image) indicate that the airborne radar echo signal containing non-cooperative composite modulation signals against the clutter background is represented as follows: ; in, for The airborne radar echo signal after frequency conversion at that time. , , and These represent modulation signal 1, modulation signal 2, target echo, and the first... The amplitude of the radar echo corresponding to each clutter scattering point Indicates the pulse repetition interval. , , , These represent modulation signal 1, modulation signal 2, target echo, and the first... The echo delay corresponding to each clutter scattering point , , and These represent modulation signal 1, modulation signal 2, target echo, and the first... The Doppler frequency shift corresponding to each clutter scattering point , , , and These represent the radar signal, modulation signal 1, modulation signal 2, Gaussian white noise, and the number of pulses within a coherent processing interval, respectively. For pulse count index, The superscript indicates the number of clutter scattering points. The imaginary unit is used; the combination of modulation signal 1 and modulation signal 2 is a non-cooperative composite modulation signal, and the modulation methods of modulation signal 1 and modulation signal 2 are different.

[0019] In one embodiment, such as Figure 3 As shown, the complex-valued signal feature extraction network consists of an input layer, a dual-branch convolutional layer, a feature concatenation unit, a residual unit, a two-dimensional average pooling unit, a pointwise convolutional unit, an attention unit, and an output layer, all connected in series. The two-dimensional average pooling unit, the pointwise convolutional unit, and the attention unit form a channel attention mechanism module, used to generate attention weights to apply channel attention weights to the output feature map of the residual unit. This strengthens channel features strongly correlated with clutter components and suppresses redundant information, achieving precise focusing and enhancement of clutter features. The network aims to learn the mapping from the (clutter-containing) radar echo RD spectrum to the estimated clutter RD spectrum, and then extract the composite modulation signal RD spectrum by subtracting the estimated clutter RD spectrum from the original radar echo RD spectrum. This avoids the influence of clutter on subsequent modulation signal type identification and improves the accuracy of non-cooperative composite modulation signal identification and classification. The specific components of this network structure are described in detail below: 1. Input layer: A convolutional kernel with a size of [size missing]. A 2D complex-valued convolutional layer is used to process the single-channel radar echo RD spectrum of the input. Complex feature modeling is performed to provide a basic representation for subsequent complex feature extraction. The dimension of the output feature map of the input layer is [missing value]. ;in, It is a complex field. The real and imaginary components are respectively .

[0020] The output of the complex convolutional layer follows the complex algebraic multiplication rule, defined as: ; In the formula, and These represent the real and imaginary parts of the convolution kernel, respectively, and the cross-fusion of real / imaginary features is processed separately. and These represent the real and imaginary parts of the input feature map, respectively. This is a 2D convolution operation over the real number field.

[0021] 2. Dual-branch convolutional layer: Consists of two parallel feature extraction units, each containing two convolutional kernels with a size of [missing information]. The 2D complex-valued convolutional layers are used to extract the amplitude and phase features of clutter and target echoes under different receptive fields, thereby realizing parallel modeling of clutter components of different scales in the radar echo RD spectrum.

[0022] 3. Feature Concatenation Unit (C): This unit concatenates the output feature maps of two feature extraction units along the channel dimension, fusing complementary information from the real / imaginary parts of features extracted from different branches. This enhances feature representation and forms a high-dimensional complex feature representation containing multi-scale clutter information, represented as: ;in, and These are the output feature maps of the first feature extraction unit and the second feature extraction unit, respectively. and All dimensions are , The real part is , The imaginary part is , , They are respectively and The real part quantity, , They are respectively and The imaginary part of the quantity.

[0023] 4. Residual Unit: A cross-layer information transmission path is constructed through a residual connection mechanism. While preserving the basic features of the lower layers, the high-dimensional complex features of the spliced ​​unit are further represented. The process involves processing the data and obtaining the processed results to alleviate the gradient vanishing problem during network training and enhance the deep representation capability of clutter features. The specific process is as follows: First, a convolutional kernel with a size of... 2D complex value convolutional layer pairs Compression is performed along the channel dimension, resulting in a feature map with the following dimension: The compressed feature map is then summed with the output feature map of the input layer for the first time to obtain the first residual summed feature map. This first residual summed feature map is then input into two convolutional kernels with a size of [missing value]. A 2D complex-valued convolutional layer is used to obtain the output feature map, which is then added to the first residual. A second residual is then added to the feature map to obtain the processed result. .

[0024] 5. Two-dimensional average pooling unit: used for... Global downsampling is performed in the spatial dimension. While preserving the amplitude and phase information of complex features, the spatial dimension of the feature map is reduced, and abstract features containing the global clutter distribution pattern are extracted, reducing subsequent computational load. The pooling kernel size is [size missing]. Step size is Fill with Output feature map of a two-dimensional average pooling unit The spatial dimensions are calculated as follows: ; in, and These represent the height and width of the two-dimensional pooling window, respectively. and These represent the stride of the two-dimensional pooling window in each sliding motion along the height and width directions, respectively. and These represent the number of zero-padding elements at the height and width edges of the input feature map of the two-dimensional average pooling unit, respectively. and These represent the height and width of the input feature map for the two-dimensional average pooling unit, respectively. and These represent the height and width of the output feature map of the two-dimensional average pooling unit, respectively. For the number of channels, This is the floor function; The element-wise mathematical definition of pooling is as follows (the real and imaginary parts are independently averaged locally): ; in, for In the Channel, First line, number Complex output values ​​for column positions, and They are respectively The real and imaginary components and For local index variables within the pooling kernel, for In the Channel, First line, number Complex output values ​​for column positions, and They are respectively The real and imaginary components; combined into complex form, it is: ; After completing the pooling operation, the output feature map of the two-dimensional average pooling unit is... This is denoted as the channel attention weight map. ,in, It is the field of real numbers.

[0025] 6. Pointwise convolution unit: used to pass through two convolution kernels with a size of... 2D complex-valued convolutional layer channel attention weight map The channel attention weights obtained after channel adjustment through two pointwise convolutions are expressed as follows: ; in, , are the weight matrices for the first and second pointwise convolutions, respectively; , are the biases for the first and second pointwise convolutions, respectively. and This is the activation function.

[0026] 7. Attention Activation Unit: Used to assign channel attention weights Broadcasting to spatial dimensions, that is, to the original Element-wise multiplication yields a multi-channel feature map, represented as follows: .

[0027] 8. Output layer: Used to pass through a convolutional kernel with a size of 2D complex-valued convolutional layers achieve feature fusion, combining multi-channel feature maps. Mapping to a single channel to achieve RD spectrum reconstruction is represented as: ; in, For single-channel reconstructed clutter RD spectrum, This is the equivalent weight matrix of the output layer. This is the equivalent bias vector for the output layer. A vectorization operator that flattens a 3D tensor into column vectors in sequence. For matrix reshaping operators.

[0028] To quantify the estimation error during training, in one embodiment, the loss function for training the complex-valued signal feature extraction network is the mean squared error (MSE), expressed as: ; in, and Let the distance dimension and the Doppler dimension of the RD spectrum represent the total number of elements, respectively. and These represent the unit numbers in the distance dimension and Doppler dimension of the RD spectrum, respectively; and These represent the RD spectrum of the radar echo, respectively. The real and imaginary parts of each unit; This represents the clutter RD spectrum estimated by the feature extraction network for complex-valued signals. and These represent the first and second halves of the RD spectrum of the composite modulated signal, respectively. The real and imaginary parts of each unit.

[0029] In one embodiment, clutter RD spectra in the radar echo RD spectrum are removed to purify the composite modulation signal RD spectrum, including: The normalized clutter RD spectrum, estimated by the complex-valued signal feature extraction network, is removed from the normalized radar echo RD spectrum to obtain the composite modulated signal RD spectrum, which is represented as: ; in, , and These represent the radar echo RD spectrum, the maximum value of the radar echo RD spectrum, and the clutter RD spectrum output by the network estimation, respectively.

[0030] In one embodiment, based on the RD spectrum of the composite modulation signal, pulse selection based on the maximum value, inverse matched filtering, and short-time Fourier transform are performed sequentially to obtain the time spectrum of each sub-modulation signal component within the non-cooperative composite modulation signal, including: The maximum value of each pulse sequence is extracted from the RD spectrum of the composite modulated signal to form the Doppler vector. , and select The two pulse sequences corresponding to the two peaks with the largest amplitude are subjected to matched filtering inverse transform to recover and obtain each sub-modulation signal component in the non-cooperative composite modulation signal. The time spectrum of each sub-modulation signal component is obtained by performing short-time Fourier transform on each sub-modulation signal component.

[0031] To demonstrate the performance of the CV-SFEN network proposed in this application, the following experiments were also conducted: Experimental dataset: Considering the airborne radar operates in look-down detection mode, the bandwidth of the transmitted linear frequency modulated signal is... MHz, pulse width The pulse repetition interval is 50µs, and 256 pulses are transmitted within one coherent processing interval (CPI). Clutter data is simulated based on an equidistant loop model, considering a clutter distribution model based on the Weibull distribution. The clutter simulation parameters are shown in Table 1. The experiment considers three types of composite pulse in-transmission modulated signals: (1) ISRJ + MDFJ; (2) ISRJ + SMSPJ; (3) SMSPJ + MDFJ. In each case, the sub-modulated signals have equal energy, and their basic parameters are listed in Table 2. Radar echo data is generated by a semi-physical platform.

[0032] Table 1 Clutter Simulation Parameters

[0033] Table 2 Modulation signal parameter settings

[0034] The experiment was set up to benchmark CV-SFEN against two state-of-the-art methods: the Non-Uniform Pulse Repetition Interval (NUPRI) method and the Transformer-based Hybrid Attention Module (HAM) method.

[0035] Figures 4 to 6 The results of clutter RD spectrum extraction for three types of composite modulated signals using CV-SFEN at JCNR = -20 dB are presented. Figures 4 to 6 As can be seen, at JCNR=-20 dB, the CV-SFEN of this application can extract the RD spectrum of clutter relatively completely, and the modulation signal component in the RD spectrum is effectively suppressed. Subtracting the normalized extracted clutter RD spectrum from the radar echo RD spectrum can purify the modulation signal. Figure 7 The pulse selection for different composite modulation signals is shown at JCNR = -20 dB. From Figure 7 It can be concluded that after the modulation signal is purified, the energy of the modulation signal is significantly greater than the energy of the clutter, which is beneficial for correctly selecting pulses containing the modulation signal.

[0036] Figure 8 The results of three comparative algorithms for modulated signal recognition under different JCNRs are presented. For the ResNet-based modulated signal recognition model, there are 500 samples for each single modulated signal type (JNR=10 dB) in the training set; and 1000 samples for each composite modulated signal type at each JCNR in the test set. Figure 8It can be seen that, compared with the other two modulation signal recognition algorithms, the CV-SFEN proposed in this application can adapt to modulation signal recognition scenarios with stronger clutter energy. Specifically, when JCNR = -25 dB, the modulation signal recognition accuracy of CV-SFEN is about 12 percentage points higher than the NUPRI-based method, not to mention the HAM-based algorithm. Experimental results show that when JCNR ≥ -25 dB, the CV-SFEN proposed in this application achieves a recognition accuracy of over 95% for three types of composite interrupted forwarding modulation signals.

[0037] In one embodiment, a non-cooperative composite modulation signal identification device under clutter background is provided, comprising: The signal acquisition module is used to acquire airborne radar echo signals containing non-cooperative composite modulation signals against a clutter background. The clutter removal module is used to obtain the radar echo RD spectrum of the airborne radar echo signal through RD transformation, input the radar echo RD spectrum into a pre-trained complex value signal feature extraction network for clutter RD spectrum feature learning, and remove the clutter RD spectrum in the radar echo RD spectrum to obtain the composite modulation signal RD spectrum. The composite modulation signal identification module is used to perform pulse selection based on the maximum value, inverse matched filtering, and short-time Fourier transform sequentially based on the RD spectrum of the composite modulation signal to obtain the time spectrum of each sub-modulation signal component in the non-cooperative composite modulation signal. The spectrum is then input into the residual neural network for modulation signal type identification. Finally, the non-cooperative composite modulation signal identification is completed based on the type of each sub-modulation signal component.

[0038] Specific limitations regarding the non-cooperative composite modulation signal identification device under clutter backgrounds can be found in the limitations of the non-cooperative composite modulation signal identification method under clutter backgrounds described above, and will not be repeated here. Each module in the aforementioned non-cooperative composite modulation signal identification device under clutter backgrounds can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independently of the processor in a computer device, or stored in software in the memory of a computer device, so that the processor can call and execute the corresponding operations of each module.

[0039] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 9As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a method for identifying non-cooperative composite modulation signals against a clutter background. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0040] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0041] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to perform the following steps: Acquire airborne radar echo signals containing non-cooperative composite modulation signals against a clutter background; The radar echo RD spectrum of the airborne radar echo signal is obtained by RD transformation. The radar echo RD spectrum is input into a pre-trained complex-valued signal feature extraction network for clutter RD spectrum feature learning. Clutter RD spectrum in radar echo RD spectrum is removed and purified to obtain composite modulation signal RD spectrum. Based on the RD spectrum of the composite modulation signal, pulse selection based on the maximum value, inverse matched filtering, and short-time Fourier transform are performed sequentially to obtain the time spectrum of each sub-modulation signal component in the non-cooperative composite modulation signal. The spectrum is then input into the residual neural network for modulation signal type identification. Finally, the non-cooperative composite modulation signal identification is completed based on the type of each sub-modulation signal component.

[0042] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0043] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0044] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application.

Claims

1. A method for identifying non-cooperative composite modulation signals against a clutter background, characterized in that, The method includes: Acquire airborne radar echo signals containing non-cooperative composite modulation signals against a clutter background; The radar echo RD spectrum of the airborne radar echo signal is obtained by RD transformation. The radar echo RD spectrum is input into a pre-trained complex value signal feature extraction network for clutter RD spectrum feature learning. The clutter RD spectrum in the radar echo RD spectrum is removed and purified to obtain the composite modulation signal RD spectrum. Based on the RD spectrum of the composite modulation signal, pulse selection based on the maximum value, inverse matched filtering, and short-time Fourier transform are performed sequentially to obtain the time spectrum of each sub-modulation signal component in the non-cooperative composite modulation signal. The spectrum is then input into the residual neural network for modulation signal type identification. Finally, the non-cooperative composite modulation signal identification is completed based on the type of each sub-modulation signal component.

2. The method for identifying non-cooperative composite modulation signals under clutter background according to claim 1, characterized in that, Airborne radar echo signals containing non-cooperative composite modulation signals against a clutter background are represented as follows: ； in, for The airborne radar echo signal after frequency conversion at that time. , , and These represent modulation signal 1, modulation signal 2, target echo, and the first... The amplitude of the radar echo corresponding to each clutter scattering point Indicates the pulse repetition interval. , , , These represent modulation signal 1, modulation signal 2, target echo, and the first... The echo delay corresponding to each clutter scattering point , , and These represent modulation signal 1, modulation signal 2, target echo, and the first... The Doppler frequency shift corresponding to each clutter scattering point , , , and These represent the radar signal, modulation signal 1, modulation signal 2, Gaussian white noise, and the number of pulses within a coherent processing interval, respectively. For pulse count index, The superscript indicates the number of clutter scattering points. The imaginary unit is used; the combination of modulation signal 1 and modulation signal 2 is a non-cooperative composite modulation signal, and the modulation methods of modulation signal 1 and modulation signal 2 are different.

3. The method for identifying non-cooperative composite modulation signals under clutter background according to claim 1, characterized in that, The complex-valued signal feature extraction network consists of an input layer, a bi-branch convolutional layer, a feature concatenation unit, a residual unit, a two-dimensional average pooling unit, a pointwise convolutional unit, an attention unit, and an output layer, all connected in series. The two-dimensional average pooling unit, the pointwise convolutional unit, and the attention unit form a channel attention mechanism module, which generates attention weights to apply channel attention weights to the output feature map of the residual unit. This strengthens channel features strongly correlated with clutter components and suppresses redundant information, achieving precise focusing and enhancement of clutter features. The input layer is a convolutional kernel with a size of [missing value]. A 2D complex-valued convolutional layer is used to process the single-channel radar echo RD spectrum of the input. Complex feature modeling is performed to provide a basic representation for subsequent complex feature extraction. The dimension of the output feature map of the input layer is [missing value]. ;in, For complex numbers; The dual-branch convolutional layer consists of two parallel feature extraction units, each containing two convolutional kernels with a size of [missing information]. The 2D complex-valued convolutional layers are used to extract the amplitude and phase features of clutter and target echo under different receptive fields, respectively, to achieve parallel modeling of clutter components of different scales in the radar echo RD spectrum. The feature concatenation unit is used to concatenate the output feature maps of the two feature extraction units along the channel dimension, fusing complementary information from the real / imaginary parts of the features extracted from different branches to enhance feature representation and form a high-dimensional complex feature representation containing multi-scale clutter information, expressed as: ;in, and These are the output feature maps of the first feature extraction unit and the second feature extraction unit, respectively. and All dimensions are , The real part is , The imaginary part is , , They are respectively and The real part quantity, , They are respectively and The imaginary part; The residual unit constructs a cross-layer information transmission path through a residual connection mechanism, further enhancing the high-dimensional complex feature representation after splicing while preserving the basic features of the underlying layers. The process involves processing the data and obtaining the processed results to alleviate the gradient vanishing problem during network training and enhance the deep representation capability of clutter features. The specific process is as follows: First, a convolutional kernel with a size of... 2D complex value convolutional layer pairs Compression is performed along the channel dimension, resulting in a feature map with the following dimension: The compressed feature map is then summed with the output feature map of the input layer for the first time to obtain the first residual summed feature map. This first residual summed feature map is then input into two convolutional kernels with a size of [missing value]. A 2D complex-valued convolutional layer is used to obtain the output feature map, which is then added to the first residual. A second residual is then added to the feature map to obtain the processed result. ; The two-dimensional average pooling unit is used for... Global downsampling is performed in the spatial dimension. While preserving the amplitude and phase information of complex features, the spatial dimension of the feature map is reduced, and abstract features containing the global clutter distribution pattern are extracted, reducing subsequent computational load. The pooling kernel size is [size missing]. Step size is Fill with Output feature map of a two-dimensional average pooling unit The spatial dimensions are calculated as follows: ； in, and These represent the height and width of the two-dimensional pooling window, respectively. and These represent the stride of the two-dimensional pooling window in each sliding motion along the height and width directions, respectively. and These represent the number of zero-padding elements at the height and width edges of the input feature map of the two-dimensional average pooling unit, respectively. and These represent the height and width of the input feature map for the two-dimensional average pooling unit, respectively. and These represent the height and width of the output feature map of the two-dimensional average pooling unit, respectively. For the number of channels, This is the floor function; The element-wise mathematical definition of pooling operations is as follows: ； in, for In the Channel, First line, number Complex output values ​​for column positions, and They are respectively The real and imaginary components and For local index variables within the pooling kernel, for In the Channel, First line, number Complex output values ​​for column positions, and They are respectively The real and imaginary components; combined into complex form, it is: ； After completing the pooling operation, the output feature map of the two-dimensional average pooling unit is... This is denoted as the channel attention weight map. ,in, For the real number field; The pointwise convolutional unit is used to pass two convolutional kernels with a size of [missing information]. 2D complex-valued convolutional layer channel attention weight map The channel attention weights obtained after channel adjustment through two pointwise convolutions are expressed as follows: ； in, These are the weight matrices for the first and second pointwise convolutions, respectively. These are the biases for the first and second pointwise convolutions, respectively; and For activation functions; The attention unit is used to assign channel attention weights. Broadcasting to spatial dimensions, that is, to the original Element-wise multiplication yields a multi-channel feature map, represented as follows: ； The output layer is used to pass a convolutional kernel with a size of [missing information]. 2D complex-valued convolutional layers achieve feature fusion, combining multi-channel feature maps. Mapping to a single channel to achieve RD spectrum reconstruction is represented as: ； in, For single-channel reconstructed clutter RD spectrum, This is the equivalent weight matrix of the output layer. This is the equivalent bias vector for the output layer. A vectorization operator that flattens a 3D tensor into column vectors in sequence. For matrix reshaping operators.

4. The method for identifying non-cooperative composite modulation signals under clutter background according to claim 3, characterized in that, The loss function for training the complex-valued signal feature extraction network is expressed as: ； in, and Let the distance dimension and the Doppler dimension of the RD spectrum represent the total number of elements, respectively. and These represent the unit numbers in the distance dimension and Doppler dimension of the RD spectrum, respectively; and These represent the RD spectrum of the radar echo, respectively. The real and imaginary parts of each unit; This represents the clutter RD spectrum estimated by the feature extraction network for complex-valued signals. and These represent the first and second halves of the RD spectrum of the composite modulated signal, respectively. The real and imaginary parts of each unit.

5. The method for identifying non-cooperative composite modulation signals under clutter background according to claim 1, characterized in that, After removing clutter RD spectra from the radar echo RD spectrum, the composite modulation signal RD spectrum is purified, including: The normalized clutter RD spectrum, estimated by the complex-valued signal feature extraction network, is removed from the normalized radar echo RD spectrum to obtain the composite modulated signal RD spectrum, which is represented as: ； in, , and These represent the radar echo RD spectrum, the maximum value of the radar echo RD spectrum, and the clutter RD spectrum output by the network estimation, respectively.

6. The method for identifying non-cooperative composite modulation signals under clutter background according to claim 5, characterized in that, Based on the RD spectrum of the composite modulation signal, pulse selection based on the maximum value, inverse matched filtering, and short-time Fourier transform are performed sequentially to obtain the time spectrum of each sub-modulation signal component within the non-cooperative composite modulation signal, including: The maximum value of each pulse sequence is extracted from the RD spectrum of the composite modulated signal to form the Doppler vector. , and select The two pulse sequences corresponding to the two peaks with the largest amplitude are subjected to matched filtering inverse transform to recover and obtain each sub-modulation signal component in the non-cooperative composite modulation signal. The time spectrum of each sub-modulation signal component is obtained by performing short-time Fourier transform on each sub-modulation signal component.

7. A device for identifying non-cooperative composite modulation signals under clutter background, characterized in that, The device includes: The signal acquisition module is used to acquire airborne radar echo signals containing non-cooperative composite modulation signals against a clutter background. The clutter removal module is used to obtain the radar echo RD spectrum of the airborne radar echo signal through RD transformation, input the radar echo RD spectrum into a pre-trained complex value signal feature extraction network for clutter RD spectrum feature learning, and remove the clutter RD spectrum in the radar echo RD spectrum to obtain the composite modulation signal RD spectrum. The composite modulation signal identification module is used to perform pulse selection based on the maximum value, inverse matched filtering and short-time Fourier transform sequentially based on the RD spectrum of the composite modulation signal to obtain the time spectrum of each sub-modulation signal component in the non-cooperative composite modulation signal, and input it into the residual neural network for modulation signal type identification. Finally, the non-cooperative composite modulation signal identification is completed based on the type of each sub-modulation signal component.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

Images