Signal parameter prediction method and system based on deep learning, terminal and storage medium

Abstract

The invention relates to the technical field of radar and communication signal processing, and discloses a signal parameter prediction method and system based on deep learning, a terminal and a storage medium, and the method comprises the steps: obtaining a multi-component linear frequency modulation signal, inputting the multi-component linear frequency modulation signal into a parameterized symmetric instantaneous autocorrelation function for calculation, and obtaining a multi-component linear frequency modulation signal; obtaining a two-dimensional complex matrix; performing decoupling processing on the two-dimensional complex matrix according to a scaling transformation strategy to obtain a two-dimensional feature matrix; and inputting the two-dimensional feature matrix into a deep neural network model for prediction to obtain a signal parameter prediction value. According to the method, separation and parameter prediction are carried out on aliasing components in the multi-component linear frequency modulation signals according to the parameterized symmetric instantaneous autocorrelation function and scaling decoupling, and high-precision extraction of the center frequency and the frequency modulation slope is achieved.

Description

Technical Field

[0001] This invention relates to the field of radar and communication signal processing technology, and in particular to a signal parameter prediction method, system, terminal, and computer-readable storage medium based on deep learning. Background Technology

[0002] Linear Frequency Modulated (LFM) signals are widely used in radar, sonar, and communications, and their parameter estimation is a key technology in signal processing. However, existing CFCR (Chirp-Fourier Correlation Representation) domain parameter estimation techniques suffer from the following technical problems: Limited estimation accuracy: Traditional Lv's Distribution (LVD) methods employ 2D-FFT (Two-Dimensional Fast Fourier Transform), which, limited by the inherent resolution of FFT, struggles to achieve high-precision parameter separation in dense multi-target environments. Strong search dependence: Methods such as Fractional Fourier Transform (FRFT) and Radon-Wigner Transform (RWT) require angle or order searches, introducing quantization errors and failing to reach the theoretical accuracy limit. Poor noise robustness: Traditional FFT methods experience a sharp performance drop in low signal-to-noise ratio environments, significantly deteriorating estimation accuracy and failing to meet the high-precision application requirements in complex electromagnetic environments. Insufficient super-resolution capability: Limited by Rayleigh resolution, it cannot separate multi-component LFM signals with similar parameters, and its performance degrades in dense signal environments. Lack of adaptive capability: Traditional methods cannot adaptively adjust processing strategies based on signal characteristics and noise environment, resulting in a lack of flexibility in complex and ever-changing application scenarios.

[0003] Existing time-frequency analysis and parameter estimation methods include: CFCR domain methods based on the Lv distribution (LVD), search methods based on fractional Fourier transform (FRFT), and transform methods based on Radon-Wigner transform (RWT). These traditional methods are limited by the inherent resolution limitations of the Fourier transform, quantization errors introduced during the search process, and sensitivity to noise. They cannot achieve high-precision parameter separation and super-resolution prediction of sound signals in dense multi-target environments. Therefore, existing CFCR domain parameter estimation methods for linear frequency modulated signals still need improvement and optimization. Summary of the Invention

[0004] The main objective of this invention is to provide a signal parameter prediction method, system, terminal, and computer-readable storage medium based on deep learning, aiming to solve the problem that existing linear frequency modulated signal CFCR domain parameter estimation methods cannot achieve high-precision parameter separation and super-resolution prediction of sound signals in dense multi-target environments.

[0005] To achieve the above objectives, the present invention provides a signal parameter prediction method based on deep learning, the signal parameter prediction method based on deep learning comprising the following steps: A multi-component linear frequency modulated signal is acquired, and the multi-component linear frequency modulated signal is input into a parameterized symmetric instantaneous autocorrelation function for calculation to obtain a two-dimensional complex matrix. The two-dimensional complex matrix is ​​decoupled according to a scaling transformation strategy to obtain a two-dimensional feature matrix; The two-dimensional feature matrix is ​​input into a deep neural network model for prediction to obtain predicted values ​​of signal parameters.

[0006] Optionally, the deep learning-based signal parameter prediction method, wherein obtaining a multi-component linear frequency modulated signal and inputting the multi-component linear frequency modulated signal into a parameterized symmetric instantaneous autocorrelation function for calculation to obtain a two-dimensional complex matrix, specifically includes: A multi-component linear frequency modulated (LFM) signal is acquired, and then down-converted and ADC-processed to obtain an LFM signal. ; in, Indicates LFM signal, Indicates the number of LFM signals. Indicates the first LFM signal amplitude, Indicates a time index. Represents the imaginary unit. Indicates the first The center frequency of the LFM signal Indicates the first The frequency modulation slope of an LFM signal; The LFM signal is input into a parameterized symmetric instantaneous autocorrelation function for calculation, resulting in a two-dimensional complex matrix.

[0007] Optionally, the deep learning-based signal parameter prediction method, wherein obtaining a multi-component linear frequency modulated (LFM) signal, performing down-conversion and ADC processing on the LFM signal to obtain an LFM signal, specifically includes: A multi-component linear frequency modulated (LFM) signal is acquired, and the LFM signal is down-converted according to a preset down-conversion strategy to obtain a baseband signal. The baseband signal is processed by an ADC according to a preset ADC strategy to obtain an LFM signal.

[0008] Optionally, the deep learning-based signal parameter prediction method, wherein inputting the LFM signal into a parameterized symmetric instantaneous autocorrelation function for calculation to obtain a two-dimensional complex matrix, specifically includes: The LFM signal is input into a parameterized symmetric instantaneous autocorrelation function for calculation, resulting in a two-dimensional complex matrix: ; in, Indicates the time delay parameter. Represents the time delay variable. Represents a two-dimensional complex matrix. This indicates the conjugate operation. This represents the input value.

[0009] Optionally, in the deep learning-based signal parameter prediction method, the step of decoupling the two-dimensional complex matrix according to a scaling transformation strategy to obtain a two-dimensional feature matrix specifically includes: The two-dimensional complex matrix is ​​decoupled according to the scaling factor of the scaling transformation strategy to obtain the two-dimensional feature matrix: ; in, This represents the input signal of the neural network. Indicates the scaling operator. This indicates the scaled timeline. Indicates the scaling factor. This represents the time delay parameter.

[0010] Optionally, the deep learning-based signal parameter prediction method, wherein inputting the two-dimensional feature matrix into a deep neural network model for prediction to obtain predicted signal parameter values, specifically includes: The two-dimensional feature matrix is ​​input into a deep neural network model for super-resolution analysis to obtain the amplitude spectrum. The amplitude spectrum is estimated using the deep neural network model to obtain predicted signal parameters.

[0011] Optionally, in the deep learning-based signal parameter prediction method, the deep neural network model includes a width matrix transformation layer, a height matrix transformation layer, a super-resolution layer, a transposed 2D convolutional layer, and an output layer.

[0012] Optionally, in the deep learning-based signal parameter prediction method, the step of inputting the two-dimensional feature matrix into a deep neural network model for super-resolution analysis to obtain the amplitude spectrum specifically includes: The two-dimensional feature matrix is ​​input into the width matrix transformation layer and the height matrix transformation layer for complex linear transformation to obtain the target width matrix and the target height matrix: ; ; in, Represents the real part of the width target matrix. The imaginary part of the width target matrix is ​​represented. Denotes the real part of the two-dimensional characteristic matrix. Denotes the imaginary part of a two-dimensional characteristic matrix. Indicates the transpose of the imaginary part. This indicates the transpose of the real part. The real part of the height target matrix is ​​represented. The imaginary part of the height target matrix is ​​represented. Represents the real part of the coefficient matrix. Represents the imaginary part of the coefficient matrix; The amplitude spectrum is obtained by performing super-resolution analysis on the width target matrix and the height target matrix through the super-resolution layer.

[0013] Optionally, the deep learning-based signal parameter prediction method, wherein the step of estimating the amplitude spectrum parameters using the deep neural network model to obtain predicted signal parameter values ​​specifically includes: The amplitude spectrum is parameter-estimated using the sigmoid function of the deep neural network model and the transposed 2D convolutional layer to obtain predicted signal parameter values. ; in, Indicates amplitude spectrum, Represents a constant.

[0014] Optionally, in the deep learning-based signal parameter prediction method, the predicted signal parameter values ​​include a center frequency prediction value and a frequency modulation slope prediction value.

[0015] Furthermore, to achieve the above objectives, the present invention also provides a signal parameter prediction system based on deep learning, wherein the signal parameter prediction system based on deep learning: The two-dimensional complex matrix calculation module is used to acquire a multi-component linear frequency modulated signal, input the multi-component linear frequency modulated signal into a parameterized symmetric instantaneous autocorrelation function for calculation, and obtain a two-dimensional complex matrix. A two-dimensional feature matrix scaling module is used to decouple the two-dimensional complex matrix according to a scaling transformation strategy to obtain a two-dimensional feature matrix; The signal numerical prediction module is used to input the two-dimensional feature matrix into a deep neural network model for prediction to obtain predicted values ​​of signal parameters.

[0016] Optionally, in the deep learning-based signal parameter prediction system, the two-dimensional complex matrix calculation module includes: The signal processing unit is used to acquire a multi-component linear frequency modulated (LFM) signal, perform down-conversion and ADC processing on the multi-component LFM signal, and obtain an LFM signal. ; in, Indicates LFM signal, Indicates the number of LFM signals. Indicates the first LFM signal amplitude, Indicates a time index. Represents the imaginary unit. Indicates the first The center frequency of the LFM signal Indicates the first The frequency modulation slope of an LFM signal; The complex matrix calculation unit is used to input the LFM signal into a parameterized symmetric instantaneous autocorrelation function for calculation to obtain a two-dimensional complex matrix.

[0017] Optionally, in the deep learning-based signal parameter prediction system, the signal processing unit includes: The signal down-conversion processing subunit is used to acquire a multi-component linear frequency modulated signal and perform down-conversion processing on the multi-component linear frequency modulated signal according to a preset down-conversion strategy to obtain a baseband signal. The baseband signal processing subunit is used to perform ADC processing on the baseband signal according to a preset ADC strategy to obtain an LFM signal.

[0018] Optionally, in the deep learning-based signal parameter prediction system, the complex matrix calculation unit includes: The matrix calculation subunit is used to input the LFM signal into a parameterized symmetric instantaneous autocorrelation function for calculation, resulting in a two-dimensional complex matrix: ; in, Indicates the time delay parameter. Represents the time delay variable. Represents a two-dimensional complex matrix. This indicates the conjugate operation. This represents the input value.

[0019] Optionally, in the deep learning-based signal parameter prediction system, the two-dimensional feature matrix scaling module includes: The decoupling unit is used to decouple the two-dimensional complex matrix according to the scaling factor of the scaling transformation strategy to obtain a two-dimensional feature matrix: ; in, This represents the input signal of the neural network. Indicates the scaling operator. This indicates the scaled timeline. Indicates the scaling factor. This represents the time delay parameter.

[0020] Optionally, in the deep learning-based signal parameter prediction system, the signal numerical prediction module includes: The amplitude spectrum analysis unit is used to input the two-dimensional feature matrix into a deep neural network model for super-resolution analysis to obtain the amplitude spectrum. The parameter estimation unit is used to estimate the parameters of the amplitude spectrum through the deep neural network model to obtain the predicted values ​​of the signal parameters.

[0021] Optionally, in the deep learning-based signal parameter prediction system, the amplitude spectrum analysis unit includes: The matrix linear transformation subunit is used to input the two-dimensional feature matrix into the width matrix transformation layer and the height matrix transformation layer of the deep neural network model for complex linear transformation to obtain the target width matrix and the target height matrix. ; ; in, Represents the real part of the width target matrix. The imaginary part of the width target matrix is ​​represented. Denotes the real part of the two-dimensional characteristic matrix. Denotes the imaginary part of a two-dimensional characteristic matrix. Indicates the transpose of the imaginary part. This indicates the transpose of the real part. The real part of the height target matrix is ​​represented. The imaginary part of the height target matrix is ​​represented. Represents the real part of the coefficient matrix. Represents the imaginary part of the coefficient matrix; The amplitude spectrum super-resolution analysis subunit is used to perform super-resolution analysis on the width target matrix and the height target matrix through the super-resolution layer of the deep neural network model to obtain the amplitude spectrum.

[0022] Optionally, in the deep learning-based signal parameter prediction system, the parameter estimation unit includes: The amplitude spectrum parameter estimation subunit is used to estimate the parameters of the amplitude spectrum using the sigmoid function and the transposed 2D convolutional layer of the deep neural network model, to obtain predicted signal parameter values. ; in, Indicates amplitude spectrum, Represents a constant.

[0023] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a deep learning-based signal parameter prediction program, which, when executed by a processor, implements the steps of the deep learning-based signal parameter prediction method as described above.

[0024] In this invention, a multi-component linear frequency modulated (LFM) signal is acquired, and then input into a parameterized symmetric instantaneous autocorrelation function for calculation to obtain a two-dimensional complex matrix. The two-dimensional complex matrix is ​​then decoupled according to a scaling transformation strategy to obtain a two-dimensional feature matrix. This feature matrix is ​​then input into a deep neural network model for prediction to obtain predicted signal parameters. This invention separates and predicts the aliasing components in a multi-component LFM signal based on the parameterized symmetric instantaneous autocorrelation function and scaling decoupling, achieving high-precision extraction of the center frequency and modulation slope. Attached Figure Description

[0025] Figure 1 This is a flowchart of a preferred embodiment of the signal parameter prediction method based on deep learning of the present invention; Figure 2 This is a flowchart of signal CFCR domain parameter estimation, a preferred embodiment of the signal parameter prediction method based on deep learning of the present invention. Figure 3 This is a flowchart of the LFM estimation network of a preferred embodiment of the signal parameter prediction method based on deep learning of the present invention; Figure 4 This is a flowchart of the residual module of a preferred embodiment of the signal parameter prediction method based on deep learning of the present invention; Figure 5 This is a flowchart illustrating the specific implementation process of step S10 in a preferred embodiment of the signal parameter prediction method based on deep learning of the present invention. Figure 6 This is a flowchart illustrating the specific implementation process of step S11 in a preferred embodiment of the signal parameter prediction method based on deep learning of the present invention. Figure 7 This is a flowchart of the conjugate operation of a preferred embodiment of the signal parameter prediction method based on deep learning of the present invention; Figure 8 This is a flowchart of step S20 of a preferred embodiment of the signal parameter prediction method based on deep learning of the present invention; Figure 9 This is a flowchart of step S30 of a preferred embodiment of the signal parameter prediction method based on deep learning of the present invention; Figure 10 This is a flowchart of step S31 of a preferred embodiment of the signal parameter prediction method based on deep learning of the present invention; Figure 11 This is a flowchart of the complex linear transformation of a preferred embodiment of the signal parameter prediction method based on deep learning of the present invention; Figure 12 This is a flowchart of amplitude spectrum iterative optimization, which is a preferred embodiment of the signal parameter prediction method based on deep learning of the present invention. Figure 13 This is a schematic diagram of a signal parameter prediction system based on deep learning according to the present invention; Figure 14 This is another schematic diagram of the signal parameter prediction system based on deep learning according to the present invention; Figure 15 This is a structural diagram of a preferred embodiment of the terminal of the device of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0027] Existing time-frequency analysis and parameter estimation methods include: CFCR domain methods based on the Lv distribution (LVD), search methods based on fractional Fourier transform (FRFT), and transform methods based on Radon-Wigner transform (RWT). These traditional methods are limited by the inherent resolution constraints of the Fourier transform, quantization errors introduced during the search process, and sensitivity to noise. They cannot achieve high-precision parameter separation and super-resolution prediction of sound signals in dense, multi-target environments. Therefore, a deep learning-based signal parameter prediction method is needed to separate and predict the aliasing components in multi-component linear frequency modulated (LFM) signals based on a parameterized symmetric instantaneous autocorrelation function and scaling decoupling. This would enable high-precision extraction of the center frequency and modulation slope, avoiding the problem of not being able to achieve high-precision parameter separation and super-resolution prediction of sound signals in dense, multi-target environments.

[0028] The signal parameter prediction method based on deep learning described in the preferred embodiment of the present invention, such as... Figure 1 , Figure 2 , Figure 3 and Figure 4 As shown, the deep learning-based signal parameter prediction method includes the following steps: Step S10: Obtain a multi-component linear frequency modulated signal, input the multi-component linear frequency modulated signal into a parameterized symmetric instantaneous autocorrelation function for calculation, and obtain a two-dimensional complex matrix.

[0029] like Figure 5 As shown, step S10 includes: Step S11: Obtain a multi-component linear frequency modulated (LFM) signal, perform down-conversion and ADC processing on the multi-component LFM signal to obtain an LFM signal; Step S12: Input the LFM signal into the parameterized symmetric instantaneous autocorrelation function for calculation to obtain a two-dimensional complex matrix.

[0030] Specifically, a multi-component linear frequency modulated (LFM) signal is acquired, and down-conversion and ADC processing are performed on the LFM signal to obtain an LFM signal. The LFM signal is then input into a parameterized symmetric instantaneous autocorrelation function for calculation to obtain a two-dimensional complex matrix. (Preprocessing may include down-conversion (converting the RF signal to baseband) and ADC (converting the analog signal to a digital signal). If the acquired signal is already a baseband digital signal, no preprocessing step is required.) ; in, Indicates LFM signal, Indicates the number of LFM signals. Indicates the first LFM signal amplitude, Indicates a time index. Represents the imaginary unit. Indicates the first The center frequency of the LFM signal Indicates the first The frequency modulation slope of an LFM signal.

[0031] like Figure 6 As shown, step S11 includes: Step S111: Obtain a multi-component linear frequency modulation signal, and perform down-conversion processing on the multi-component linear frequency modulation signal according to a preset down-conversion strategy to obtain a baseband signal; Step S112: Perform ADC processing on the baseband signal according to the preset ADC strategy to obtain the LFM signal.

[0032] Specifically, a multi-component linear frequency modulated (LFM) signal is acquired, and the LFM signal is down-converted according to a preset down-conversion strategy to obtain a baseband signal (converting the radio frequency signal (multi-component LFM signal) to a baseband signal). The baseband signal is then processed by an ADC according to a preset ADC strategy (converting the analog signal (baseband signal) into a digital signal) to obtain an LFM signal.

[0033] As an example, a preset downconversion strategy is used to shift the entire RF signal spectrum down to the baseband (near zero frequency), facilitating subsequent digital sampling and processing. Implementation: A mixer (frequency converter) is used to mix the RF signal with the local oscillator (LO) signal, generating a difference frequency component, i.e., the baseband signal. Preset parameters: The LO frequency, filter bandwidth, and filtering characteristics of the downconversion are preset according to the system design and the target signal frequency band, ensuring that the baseband signal contains all target LFM components without aliasing. Filtering: After mixing, a low-pass filter removes high-frequency components, retaining the baseband signal and suppressing image frequency interference. Baseband signal characteristics: The baseband signal is an analog signal, containing the amplitude, phase, and frequency modulation characteristics of multiple LFM components. Its frequency range is significantly reduced compared to the RF signal, facilitating subsequent sampling. Maintaining signal integrity and time-frequency characteristics provides an accurate basis for parameter estimation. Preset ADC strategy, sampling rate selection: Based on the baseband signal bandwidth and the Nyquist sampling theorem, the sampling rate is preset to ensure no spectral aliasing and to meet signal reconstruction requirements. Quantization Accuracy: Based on the system's dynamic range and signal-to-noise ratio requirements, a suitable ADC resolution (bit depth) is selected to ensure effective preservation of signal details and noise characteristics. Sampling Synchronization: The sampling clock is synchronized with the system clock to ensure sampling time accuracy and avoid the impact of sampling jitter on parameter estimation. Anti-Aliasing Filtering: An anti-aliasing filter is configured at the ADC front end to further suppress out-of-band interference and ensure the quality of the sampled signal. The ADC processing converts the analog baseband signal into a digital signal, forming a discrete-time series digital LFM signal. The digital signal contains amplitude and phase information of the multi-component LFM signal, serving as input for subsequent parameterized symmetric instantaneous autocorrelation function calculations. The resulting digital LFM signal meets the requirements for subsequent deep learning network input, exhibiting high fidelity and complete time-frequency characteristics. This digital signal forms the basis of the CFCR domain parameter estimation method of this invention, achieving high-precision, super-resolution multi-component LFM signal parameter estimation.

[0034] Step S12 includes: Step S121: Input the LFM signal into the parameterized symmetric instantaneous autocorrelation function for calculation to obtain a two-dimensional complex matrix.

[0035] Specifically, the LFM signal is input into the parameterized symmetric instantaneous autocorrelation function (PSIAF calculation module) for calculation, resulting in a two-dimensional complex matrix: ; in, Indicates the time delay parameter. Represents the time delay variable. Represents a two-dimensional complex matrix. This indicates the conjugate operation. This represents the input value.

[0036] Furthermore, such as Figure 7 As shown, the present invention may further include: Step S41: Perform a conjugate operation on the LFM signal based on the time delay parameter of the symmetric instantaneous autocorrelation function to obtain the first parameter; Step S42: Multiply the first parameter according to the time variable and time delay variable of the symmetric instantaneous autocorrelation function to obtain the target two-dimensional complex matrix.

[0037] Specifically, the LFM signal is conjugate according to the time delay parameter (which is usually set to 1) of the symmetric instantaneous autocorrelation function to obtain the first parameter. The first parameter is then multiplied according to the time variable and the time delay variable of the symmetric instantaneous autocorrelation function to obtain the target two-dimensional complex matrix.

[0038] Step S20: Decouple the two-dimensional complex matrix according to the scaling transformation strategy to obtain the two-dimensional feature matrix.

[0039] like Figure 8 As shown, step S20 includes: Step S21: Decouple the two-dimensional complex matrix according to the scaling transformation strategy to obtain the two-dimensional feature matrix.

[0040] Specifically, the two-dimensional complex matrix is ​​decoupled according to the scaling factor (scaling operator) of the scaling transformation strategy to obtain the two-dimensional feature matrix: ; in, This represents the input signal of the neural network. Indicates the scaling operator. This indicates the scaled timeline. Indicates the scaling factor. This represents the time delay parameter.

[0041] As an example, the two-dimensional complex matrix obtained by calculating the parameterized symmetric instantaneous autocorrelation function contains the coupled information of the time and delay variables of the LFM signal, i.e., a time-frequency coupled structure. This coupled structure is characterized by signal components distributed along a diagonal line in the matrix. Directly inputting this into a deep learning network increases the estimation difficulty and affects parameter separation and estimation accuracy. The core purpose of the scaling transformation module is to decouple the time and delay variables by applying a scaling operator (scaling factor) to the two-dimensional complex matrix, transforming the diagonal structure into a horizontal line structure, thereby simplifying the time-frequency structure of the signal and improving the feature extraction capability and estimation accuracy of the subsequent deep learning network. The scaled two-dimensional complex matrix is ​​the decoupled two-dimensional feature matrix, possessing a clearer time-frequency structure. This matrix facilitates feature extraction and parameter estimation by the deep learning network, improving the network's ability to identify the center frequency and modulation slope. The decoupled matrix effectively reduces interference between signal components and improves the separation capability of multi-component LFM signals. The scaled two-dimensional complex matrix is ​​the decoupled two-dimensional feature matrix, possessing a clearer time-frequency structure. This matrix facilitates feature extraction and parameter estimation in deep learning networks, enhancing their ability to identify center frequency and frequency modulation slope. The decoupled matrix effectively reduces interference between signal components, improving the separation capability of multi-component LFM signals. The resulting two-dimensional feature matrix serves as input to the deep learning estimation module, fed into a 2D-ResFreq network for high-precision estimation of center frequency and frequency modulation slope parameters. The complex nature of this matrix allows the network to utilize amplitude and phase information to achieve finer parameter estimation and super-resolution separation.

[0042] Furthermore, step S22 of the present invention further includes: Step S22: Based on the scaling transformation strategy, the diagonal component of the LFM signal is transformed to obtain the horizontal structure component.

[0043] Specifically, based on the scaling transformation strategy, the sloping component of the LFM signal is transformed to obtain a horizontal structure component. The sloping component refers to the energy distribution of the LFM signal with a certain slope in the time-frequency domain, which reflects the frequency modulation characteristics of the signal. By performing coordinate transformation or filtering on the sloping component, it is mapped to a horizontal structure component, that is, an energy distribution parallel to the time axis in the time-frequency diagram, thereby simplifying the signal structure and highlighting its features.

[0044] As an example, in the time-frequency domain representation of a linear frequency modulated (LFM) signal, the signal energy is typically distributed along a line with a certain slope, which directly reflects the frequency modulation slope of the signal. This sloping structure embodies the characteristic that the signal frequency changes linearly with time, a typical feature of LFM signals. Specifically, in a two-dimensional time-frequency matrix, the energy distribution on the time and frequency axes presents a sloping shape, with the slope proportional to the frequency modulation slope. The scaling transformation strategy maps the sloping components to horizontal structural components by applying a specific coordinate transformation to the two-dimensional time-frequency matrix. The transformation is essentially a linear scaling and offset adjustment of the time axis, "straightening" the signal energy originally distributed along the sloping line into a horizontal line parallel to the time axis. After the transformation, the time variable and the frequency (or delay) variable are decoupled. In some implementations, the scaling transformation can be combined with filtering operations to further suppress noise and cross-term interference, highlighting the main energy distribution of the signal. Filter design can be optimized for the transformed horizontal structure to enhance the salience of signal features. Structural simplification: The signal energy distribution changes from a complex sloping structure to a simple horizontal line structure, reducing the complexity of the signal. **Feature Highlighting:** The horizontal structure facilitates the extraction of key features by deep learning networks, improving the accuracy of parameter estimation. **Decoupling Effect:** Decoupling of time and delay variables makes parameter estimation more stable, especially in environments with multiple superimposed components and low signal-to-noise ratios. **Super-Resolution Support:** The simplified signal structure provides a good input foundation for subsequent super-resolution modules, breaking through the resolution limitations of traditional FFTs.

[0045] Step S30: Input the two-dimensional feature matrix into the deep neural network model for prediction to obtain the predicted values ​​of the signal parameters.

[0046] like Figure 9 As shown, step S30 includes: Step S31: Input the two-dimensional feature matrix into a deep neural network model for super-resolution analysis to obtain the amplitude spectrum; Step S32: Estimate the parameters of the amplitude spectrum using the deep neural network model to obtain the predicted values ​​of the signal parameters.

[0047] Specifically, such as Figure 4 As shown, the two-dimensional feature matrix is ​​input into a deep neural network model (2D-ResFreq, a super-resolution module consisting of width / height matrix transformation layers and stacked residual blocks, which is composed of 32 stacked residual blocks, as shown). Figure 4 Super-resolution analysis is performed on the deep neural network model (which consists of an upsampling layer module (composed of three stacked two-dimensional deconvolutional layers), a transposed 2D convolutional layer, and an output layer) to obtain the amplitude spectrum. The predicted values ​​of the signal parameters include the predicted center frequency and the predicted frequency modulation slope.

[0048] As an example, the network includes the following key modules: Width / Height Matrix Transformation Layer: Using two two-dimensional discrete Fourier transform matrices based on learnable matrices, complex linear transformations are performed on the input matrix along the width and height directions, respectively, to achieve frequency domain feature extraction and transformation of the signal. Super-Resolution Module: Composed of 32 stacked residual modules, the residual modules effectively alleviate the gradient vanishing problem in deep network training through skip connections, enhancing the network's ability to express complex signal features. Upsampling Layer Module: Consists of three two-dimensional deconvolutional layers, which enlarge the spatial size of the feature map, improve the output resolution, and achieve a super-resolution effect.

[0049] Furthermore, in the super-resolution analysis process, the network maps the input two-dimensional complex matrix to a higher-resolution feature space through multi-layer nonlinear mapping and deep feature extraction. The width / height matrix transformation layer first performs a linear transformation on the input signal in the complex domain to extract frequency-related features. The super-resolution module further refines the detailed features of the signal, enhancing the ability to distinguish dense multi-component signals. The upsampling layer enlarges the feature map size through deconvolution operations, breaking through the Rayleigh resolution limitation of traditional FFT and achieving super-resolution amplitude spectrum output. Output amplitude spectrum: The network finally outputs an amplitude spectrum matrix in the CFCR domain. The amplitude spectrum reflects the energy distribution of the signal in the two parameter dimensions of center frequency and frequency modulation slope. The peak value of the amplitude spectrum corresponds to the parameter estimation target of the LFM signal, and the peak position is the estimated value of center frequency and frequency modulation slope.

[0050] In this embodiment, the parameter estimation principle is that the peak position in the amplitude spectrum directly corresponds to the center frequency and modulation slope parameters of the LFM signal. By learning the mapping relationship in a large number of training samples, the network can accurately locate the peak, overcoming the insufficient resolution and noise interference problems of traditional FFT methods. Peak detection and parameter extraction involve performing two-dimensional peak detection on the amplitude spectrum output by the network to identify local maxima. The two-dimensional coordinates of the peak are mapped to estimated values ​​of the center frequency and modulation slope, respectively. This process can be combined with interpolation algorithms to further improve the peak positioning accuracy, achieving sub-pixel-level parameter estimation. The accuracy enhancement mechanism utilizes the network's super-resolution capability to make the amplitude spectrum exhibit a more refined structure in the parameter space, with sharper peaks that are easier to locate accurately. The deep learning model possesses strong nonlinear fitting capabilities, enabling adaptive suppression of noise and cross-term interference, improving the robustness and accuracy of the estimation. Combined with residual modules and upsampling layers, the network can effectively separate dense multi-component signals, achieving high-precision multi-target parameter estimation. The final output is a high-precision estimate of the center frequency and modulation slope for each LFM component. This estimation result can be used for subsequent signal detection, tracking, and recognition applications.

[0051] In this embodiment, as Figure 3As shown, the input layer first receives a low-resolution image. Initial feature extraction is performed through a 2D convolutional layer, followed by a residual module to refine the features and ensure gradient flow. Subsequently, the network increases the feature map size through upsampling and 2D transposed convolution operations, and again uses a residual module for feature optimization. At the core of the process is a dedicated super-resolution module focused on learning the complex mapping from low-resolution to high-resolution features. Afterward, the network further performs upsampling and transposed convolution to continue amplifying the features, and finally, a residual module performs final feature refinement. Finally, all processed features are integrated through the output layer to generate the reconstructed high-resolution image. The deep neural network model includes a width matrix transformation layer, a height matrix transformation layer, a super-resolution layer, a transposed 2D convolutional layer, and an output layer.

[0052] As an example, the input layer and preprocessing: The input layer receives a scaled two-dimensional complex matrix. Typically, the complex matrix is ​​split into real and imaginary channels to form a multi-channel input, facilitating subsequent convolutional operations to process complex signal features. Width / Height Matrix Transformation Layer: The network first implements complex linear transformations in the width and height directions using two two-dimensional discrete Fourier transform matrices based on learnable matrices. The width transform layer uses learnable coefficient matrices to perform a complex linear transformation on the input signal along the width dimension, enhancing the signal's frequency resolution. Similarly, the height transform layer performs a complex linear transformation on the height dimension to further extract the signal's two-dimensional frequency features. This stage is equivalent to adaptive frequency domain filtering of the input features, improving the signal's feature representation ability. Super-resolution Module (Residual Block Stacking): The core super-resolution module is composed of 32 stacked residual modules (e.g., ...). Figure 2As shown in the diagram, each residual module contains multiple layers of convolutional, activation, and normalization layers, utilizing residual connections to alleviate the gradient vanishing problem in deep network training. This module learns the relationship between complex signal features and parameters through deep nonlinear mapping, achieving super-resolution separation of dense, multi-component LFM signals with similar parameters. Through residual learning, the network can capture minute frequency and slope differences, significantly improving the accuracy and robustness of parameter estimation. Upsampling layer module: To further improve the spatial resolution of the output, the network designs an upsampling module consisting of three stacked two-dimensional deconvolutional layers (transposed convolutional layers). The upsampling layer module progressively enlarges the feature map processed by the super-resolution module, restoring a higher resolution amplitude spectrum representation, enhancing the distinguishability of peaks, and facilitating subsequent parameter extraction. Output layer: The output layer generates the final center frequency-frequency modulation slope domain (CFCR domain) representation matrix through convolutional layers and a sigmoid activation function. The two-dimensional peak positions of this matrix correspond to the predicted values ​​of the center frequency and frequency modulation slope of each component LFM signal, achieving accurate parameter estimation. Amplitude Spectrum and Parameter Estimation: The amplitude spectrum calculated by the intermediate layer of the network reflects the energy distribution of the signal in the CFCR domain. Through processing by the super-resolution module and the upsampling layer, the signal peaks in the amplitude spectrum are sharper and have higher resolution. The peak coordinates are located in the output CFCR domain matrix using a peak detection algorithm, directly obtaining the predicted center frequency and frequency modulation slope of each LFM signal component. This method overcomes the Rayleigh resolution limitation of traditional FFT, achieving high-precision separation and parameter estimation of dense multi-component signals.

[0053] like Figure 10 As shown, step S31 includes: Step S311: Input the two-dimensional feature matrix into the width matrix transformation layer and the height matrix transformation layer of the deep neural network model to perform complex linear transformation, and obtain the width target matrix and the height target matrix; Step S312: Perform super-resolution analysis on the width target matrix and height target matrix through the super-resolution layer of the deep neural network model to obtain the amplitude spectrum.

[0054] Specifically, the two-dimensional feature matrix is ​​input into the width matrix transformation layer and the height matrix transformation layer for complex linear transformation to obtain the target width matrix and the target height matrix: ; ; in, Represents the real part of the width target matrix. The imaginary part of the width target matrix is ​​represented. Denotes the real part of the two-dimensional characteristic matrix. Denotes the imaginary part of a two-dimensional characteristic matrix. Indicates the transpose of the imaginary part. This indicates the transpose of the real part. The real part of the height target matrix is ​​represented. The imaginary part of the height target matrix is ​​represented. Represents the real part of the coefficient matrix. This represents the imaginary part of the coefficient matrix.

[0055] In this embodiment, firstly, a two-dimensional complex feature matrix, after scaling transformation, is used as input. This matrix contains key information about the linear frequency modulated (LFM) signal in the time-delay domain. This input matrix is ​​sequentially fed into a width matrix transformation layer and a height matrix transformation layer. In the width matrix transformation layer, the input matrix undergoes a complex linear transformation along the width direction. Specifically, this layer contains a learnable complex coefficient matrix, which extracts frequency features in the signal's width dimension through matrix multiplication with the input matrix. This transformation is similar to the one-dimensional transformation in a two-dimensional Fourier transform, but the difference lies in the fact that the elements of the transformation matrix are automatically optimized through deep learning training, rather than fixed Fourier basis functions. This learnable transformation allows the network to adaptively adjust the transformation kernel based on training data, more effectively capturing the frequency components and structural features of the signal, thereby improving the accuracy of parameter estimation. Subsequently, the width-transformed matrix is ​​input into the height matrix transformation layer, which also undergoes a complex linear transformation along the height direction. The height matrix transformation layer also contains a learnable complex coefficient matrix, extracting the signal's frequency features in the height dimension through matrix multiplication. Through continuous complex linear transformations in both width and height, the network achieves a two-dimensional frequency domain mapping of the input two-dimensional complex matrix, converting the signal's time-domain structure into a frequency-domain feature representation in the parameter space. This bidirectional complex linear transformation not only inherits the physical meaning and mathematical rigor of the traditional two-dimensional Fourier transform but also enhances the network's nonlinear expressive and adaptive capabilities through a learnable transformation matrix. This allows the network to overcome the resolution limitations of traditional FFTs and more accurately separate and estimate the parameters of densely multi-component linear frequency modulated signals. After the transformation, the network calculates the amplitude spectrum of the transformation result, which serves as the input to the subsequent super-resolution module. The amplitude spectrum reflects the energy distribution of the signal along the two parameter dimensions of center frequency and frequency modulation slope, providing rich feature information for high-precision parameter estimation.

[0056] Furthermore, further, such as Figure 11 As shown, the present invention may further include: Step S61: Obtain the first coefficient matrix, and perform a width complex linear transformation on the two-dimensional feature matrix based on the first coefficient matrix and the width matrix transformation layer to obtain the width target matrix; Step S62: Obtain the second coefficient matrix, and perform a height complex linear transformation on the two-dimensional feature matrix based on the second coefficient matrix and the height matrix transformation layer to obtain the height target matrix.

[0057] Specifically, a first coefficient matrix is ​​obtained, and a width complex linear transformation is performed on the two-dimensional feature matrix based on the first coefficient matrix and the width matrix transformation layer to obtain a width target matrix. A second coefficient matrix is ​​obtained, and a height complex linear transformation is performed on the two-dimensional feature matrix based on the second coefficient matrix and the height matrix transformation layer to obtain a height target matrix.

[0058] As an example, the system first obtains a first coefficient matrix for width transformation. This first coefficient matrix is ​​a learnable complex weight matrix, the size of which matches the width dimension of the input feature matrix, and is used to perform a complex linear transformation on the input feature matrix along the width direction. By performing matrix multiplication on the input two-dimensional feature matrix and the first coefficient matrix, a complex linear mapping along the width dimension is completed, resulting in the target width matrix. This width complex linear transformation not only effectively extracts feature information in the width direction but also achieves adaptive enhancement of signal features by learning and optimizing the parameters of the coefficient matrix. Subsequently, the system obtains a second coefficient matrix for height transformation, which is also a learnable complex weight matrix, the size of which matches the height dimension of the input feature matrix. The second coefficient matrix is ​​used to perform a complex linear transformation on the input two-dimensional feature matrix in the height direction, and a feature mapping along the height dimension is achieved through matrix multiplication, resulting in the target height matrix. This height complex linear transformation further enriches the feature representation capability and enhances the network's ability to identify signal parameters. Through the complex linear transformations in both width and height directions, the system can achieve efficient feature extraction and transformation in the two-dimensional feature space, providing more accurate and richer input features for the subsequent super-resolution module, thereby improving the estimation accuracy and robustness of the LFM signal center frequency and modulation slope.

[0059] In this embodiment, the matched filtering module consists of two two-dimensional discrete Fourier transform matrices based on learnable matrices. The width target matrix performs a width transform: a complex linear transform is achieved along the width dimension using a learnable coefficient matrix. The height target matrix performs a height transform: a complex linear transform is achieved along the width dimension using a learnable coefficient matrix. Super-resolution analysis is performed on the width and height target matrices through the super-resolution layer to obtain the amplitude spectrum.

[0060] Furthermore, the amplitude spectrum is parameter-estimated using the sigmoid function of the deep neural network model and the transposed 2D convolutional layer to obtain predicted signal parameters (generating a CFCR domain representation matrix, where the peak two-dimensional coordinates of the CFCR domain representation matrix correspond to the estimated values ​​of two parameters, including the predicted center frequency and the predicted frequency modulation slope). ; in, Indicates amplitude spectrum, Represents a constant.

[0061] Furthermore, such as Figure 12 As shown, the present invention may further include: Step S71: Reconstruct the amplitude spectrum by using the upsampling layer and transposed 2D convolutional layer of the deep neural network model to obtain the initial amplitude spectrum; Step S72: Iteratively process the initial amplitude spectrum using multiple residual modules of the deep neural network model to obtain the target amplitude spectrum.

[0062] Specifically, the amplitude spectrum is reconstructed using the upsampling layer and transposed 2D convolutional layer of the deep neural network model to obtain an initial amplitude spectrum. The initial amplitude spectrum is then iteratively processed using multiple residual modules of the deep neural network model to obtain the target amplitude spectrum.

[0063] In this embodiment, feature reconstruction is first performed using an upsampling layer in the deep neural network model. This upsampling layer consists of multiple stacked transposed 2D convolutional layers (also known as deconvolutional layers), which effectively improves the spatial resolution of the amplitude spectrum and recovers the detailed features of the signal in the CFCR domain. Specifically, the transposed 2D convolutional layers learn trainable convolutional kernel parameters to amplify the spatial dimension and enhance features of the input amplitude spectrum, generating a larger and more detailed initial amplitude spectrum. Subsequently, the initial amplitude spectrum is input into the super-resolution module of the deep neural network, which consists of multiple stacked residual modules. Each residual module effectively alleviates the gradient vanishing problem in deep network training by introducing residual connection structures, promoting efficient information transfer and fine-grained feature extraction. The residual module contains multiple convolutional layers, activation functions, and normalization layers, which can capture complex nonlinear features and detailed structures in the amplitude spectrum. Through iterative processing of the initial amplitude spectrum, the residual module gradually optimizes the expression of the amplitude spectrum, suppresses noise interference, and enhances the super-resolution capability of the signal. After deep iterative processing by multiple residual modules, the network finally outputs the target amplitude spectrum. The target amplitude spectrum significantly outperforms the initial amplitude spectrum in terms of both spatial resolution and signal-to-noise ratio, enabling high-precision separation and parameter estimation of dense multi-component LFM signals. The peak positions of the target amplitude spectrum correspond to the center frequency and frequency modulation slope parameters of each LFM signal, providing an accurate basis for subsequent peak detection and parameter extraction.

[0064] Furthermore, such as Figure 13 As shown, based on the above-described deep learning-based signal parameter prediction method, this invention also provides a deep learning-based signal parameter prediction system, wherein the deep learning-based signal parameter prediction system includes: The two-dimensional complex matrix calculation module 50 is used to acquire a multi-component linear frequency modulated signal, input the multi-component linear frequency modulated signal into a parameterized symmetric instantaneous autocorrelation function for calculation, and obtain a two-dimensional complex matrix. The two-dimensional feature matrix scaling module 60 is used to decouple the two-dimensional complex matrix according to the scaling transformation strategy to obtain the two-dimensional feature matrix. The signal numerical prediction module 70 is used to input the two-dimensional feature matrix into a deep neural network model for prediction to obtain the predicted values ​​of signal parameters.

[0065] like Figure 14 As shown in this embodiment of the signal parameter prediction system based on deep learning, in this embodiment, the two-dimensional complex matrix calculation module 50 includes: Signal processing unit 501 is used to acquire a multi-component linear frequency modulated (LFM) signal, perform down-conversion and ADC processing on the multi-component LFM signal, and obtain an LFM signal. ; in, Indicates LFM signal, Indicates the number of LFM signals. Indicates the first LFM signal amplitude, Indicates a time index. Represents the imaginary unit. Indicates the first The center frequency of the LFM signal Indicates the first The frequency modulation slope of an LFM signal; The complex matrix calculation unit 502 is used to input the LFM signal into a parameterized symmetric instantaneous autocorrelation function for calculation to obtain a two-dimensional complex matrix.

[0066] In this embodiment, the signal processing unit 501 includes: The signal downconversion processing subunit 5011 is used to acquire a multi-component linear frequency modulation signal and perform downconversion processing on the multi-component linear frequency modulation signal according to a preset downconversion strategy to obtain a baseband signal. The baseband signal processing subunit 5012 is used to perform ADC processing on the baseband signal according to a preset ADC strategy to obtain an LFM signal.

[0067] In this embodiment, the complex matrix calculation unit 502 includes: Matrix calculation subunit 5021 is used to input the LFM signal into a parameterized symmetric instantaneous autocorrelation function for calculation, to obtain a two-dimensional complex matrix: ; in, Indicates the time delay parameter. Represents the time delay variable. Represents a two-dimensional complex matrix. This indicates the conjugate operation. This represents the input value.

[0068] In this embodiment, the two-dimensional feature matrix scaling module 60 includes: Decoupling processing unit 601 is used to decouple the two-dimensional complex matrix according to the scaling factor of the scaling transformation strategy to obtain a two-dimensional feature matrix: ; in, This represents the input signal of the neural network. Indicates the scaling operator. This indicates the scaled timeline. Indicates the scaling factor. This represents the time delay parameter.

[0069] In this embodiment, the signal numerical prediction module 70 includes: The amplitude spectrum analysis unit 701 is used to input the two-dimensional feature matrix into a deep neural network model for super-resolution analysis to obtain the amplitude spectrum. The parameter estimation unit 702 is used to perform parameter estimation on the amplitude spectrum through the deep neural network model to obtain predicted values ​​of signal parameters.

[0070] In this embodiment, the amplitude spectrum analysis unit 701 includes: The matrix linear transformation subunit 7011 is used to input the two-dimensional feature matrix into the width matrix transformation layer and the height matrix transformation layer for complex linear transformation to obtain the target width matrix and the target height matrix. ; ; in, Represents the real part of the width target matrix. The imaginary part of the width target matrix is ​​represented. Denotes the real part of the two-dimensional characteristic matrix. Denotes the imaginary part of a two-dimensional characteristic matrix. Indicates the transpose of the imaginary part. This indicates the transpose of the real part. The real part of the height target matrix is ​​represented. The imaginary part of the height target matrix is ​​represented. Represents the real part of the coefficient matrix. Represents the imaginary part of the coefficient matrix; The amplitude spectrum super-resolution analysis subunit 7012 is used to perform super-resolution analysis on the width target matrix and the height target matrix through the super-resolution layer to obtain the amplitude spectrum.

[0071] In this embodiment, the parameter estimation unit 702 includes: The amplitude spectrum parameter estimation subunit 7021 is used to estimate the parameters of the amplitude spectrum using the sigmoid function of the deep neural network model and the transposed 2D convolutional layer to obtain predicted signal parameter values. ; in, Indicates amplitude spectrum, Represents a constant.

[0072] Furthermore, such as Figure 15 As shown, based on the above-mentioned deep learning-based signal parameter prediction method and system, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 15 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0073] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a deep learning-based signal parameter prediction program 40, which can be executed by the processor 10 to implement the deep learning-based signal parameter prediction method of this application.

[0074] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the deep learning-based signal parameter prediction method.

[0075] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The terminals communicate with each other via a system bus.

[0076] In one embodiment, when the processor 10 executes the deep learning-based signal parameter prediction program 40 in the memory 20, the following steps are performed: A multi-component linear frequency modulated signal is acquired, and the multi-component linear frequency modulated signal is input into a parameterized symmetric instantaneous autocorrelation function for calculation to obtain a two-dimensional complex matrix. The two-dimensional complex matrix is ​​decoupled according to a scaling transformation strategy to obtain a two-dimensional feature matrix; The two-dimensional feature matrix is ​​input into a deep neural network model for prediction to obtain predicted values ​​of signal parameters.

[0077] Specifically, the step of acquiring the multi-component linear frequency modulated signal involves inputting the multi-component linear frequency modulated signal into a parameterized symmetric instantaneous autocorrelation function for calculation to obtain a two-dimensional complex matrix, which includes: A multi-component linear frequency modulated (LFM) signal is acquired, and then down-converted and ADC-processed to obtain an LFM signal. ; in, Indicates LFM signal, Indicates the number of LFM signals. Indicates the first LFM signal amplitude, Indicates a time index. Represents the imaginary unit. Indicates the first The center frequency of the LFM signal Indicates the first The frequency modulation slope of an LFM signal; The LFM signal is input into a parameterized symmetric instantaneous autocorrelation function for calculation, resulting in a two-dimensional complex matrix.

[0078] Specifically, the step of acquiring a multi-component linear frequency modulated (LFM) signal, performing down-conversion and ADC processing on the LFM signal to obtain an LFM signal, includes: A multi-component linear frequency modulated (LFM) signal is acquired, and the LFM signal is down-converted according to a preset down-conversion strategy to obtain a baseband signal. The baseband signal is processed by an ADC according to a preset ADC strategy to obtain an LFM signal.

[0079] Specifically, the step of inputting the LFM signal into a parameterized symmetric instantaneous autocorrelation function for calculation to obtain a two-dimensional complex matrix includes: The LFM signal is input into a parameterized symmetric instantaneous autocorrelation function for calculation, resulting in a two-dimensional complex matrix: ; in, Indicates the time delay parameter. Represents the time delay variable. Represents a two-dimensional complex matrix. This indicates the conjugate operation. This represents the input value.

[0080] Specifically, the step of decoupling the two-dimensional complex matrix according to the scaling transformation strategy to obtain the two-dimensional feature matrix includes: The two-dimensional complex matrix is ​​decoupled according to the scaling factor of the scaling transformation strategy to obtain the two-dimensional feature matrix: ; in, This represents the input signal of the neural network. Indicates the scaling operator. This indicates the scaled timeline. Indicates the scaling factor. This represents the time delay parameter.

[0081] Specifically, the step of inputting the two-dimensional feature matrix into a deep neural network model for prediction to obtain predicted signal parameter values ​​includes: The two-dimensional feature matrix is ​​input into a deep neural network model for super-resolution analysis to obtain the amplitude spectrum. The amplitude spectrum is estimated using the deep neural network model to obtain predicted signal parameters.

[0082] The deep neural network model includes a width matrix transformation layer, a height matrix transformation layer, a super-resolution layer, a transposed 2D convolutional layer, and an output layer.

[0083] Specifically, the step of inputting the two-dimensional feature matrix into a deep neural network model for super-resolution analysis to obtain the amplitude spectrum includes: The two-dimensional feature matrix is ​​input into the width matrix transformation layer and the height matrix transformation layer for complex linear transformation to obtain the target width matrix and the target height matrix: ; ; in, Represents the real part of the width target matrix. The imaginary part of the width target matrix is ​​represented. Denotes the real part of the two-dimensional characteristic matrix. Denotes the imaginary part of a two-dimensional characteristic matrix. Indicates the transpose of the imaginary part. This indicates the transpose of the real part. The real part of the height target matrix is ​​represented. The imaginary part of the height target matrix is ​​represented. Represents the real part of the coefficient matrix. Represents the imaginary part of the coefficient matrix; The amplitude spectrum is obtained by performing super-resolution analysis on the width target matrix and the height target matrix through the super-resolution layer.

[0084] Specifically, the step of estimating the amplitude spectrum using the deep neural network model to obtain predicted signal parameters includes: The amplitude spectrum is parameter-estimated using the sigmoid function of the deep neural network model and the transposed 2D convolutional layer to obtain predicted signal parameter values. ; in, Indicates amplitude spectrum, Represents a constant.

[0085] The predicted signal parameters include the predicted center frequency and the predicted frequency modulation slope.

[0086] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a deep learning-based signal parameter prediction program, which, when executed by a processor, implements the steps of the deep learning-based signal parameter prediction method as described above.

[0087] In summary, this invention provides a signal parameter prediction method, system, terminal, and storage medium based on deep learning. The method includes: acquiring a multi-component linear frequency modulated (LFM) signal; inputting the multi-component LFM signal into a parameterized symmetric instantaneous autocorrelation function for calculation to obtain a two-dimensional complex matrix; decoupling the two-dimensional complex matrix according to a scaling transformation strategy to obtain a two-dimensional feature matrix; and inputting the two-dimensional feature matrix into a deep neural network model for prediction to obtain predicted signal parameter values. This invention separates and predicts the aliasing components in a multi-component LFM signal based on a parameterized symmetric instantaneous autocorrelation function and scaling decoupling, achieving high-precision extraction of the center frequency and modulation slope.

[0088] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal system that includes that element.

[0089] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0090] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A signal parameter prediction method based on deep learning, characterized in that, The deep learning-based signal parameter prediction method includes: A multi-component linear frequency modulated signal is acquired, and the multi-component linear frequency modulated signal is input into a parameterized symmetric instantaneous autocorrelation function for calculation to obtain a two-dimensional complex matrix. The two-dimensional complex matrix is ​​decoupled according to a scaling transformation strategy to obtain a two-dimensional feature matrix; The two-dimensional feature matrix is ​​input into a deep neural network model for prediction to obtain predicted values ​​of signal parameters.

2. The signal parameter prediction method based on deep learning according to claim 1, characterized in that, The process of acquiring a multi-component linear frequency modulated (LFM) signal, inputting the LFM signal into a parameterized symmetric instantaneous autocorrelation function for calculation to obtain a two-dimensional complex matrix, specifically includes: A multi-component linear frequency modulated (LFM) signal is acquired, and then down-converted and ADC-processed to obtain an LFM signal. ； in, Indicates LFM signal, Indicates the number of LFM signals. Indicates the first LFM signal amplitude, Indicates a time index. Represents the imaginary unit. Indicates the first The center frequency of the LFM signal Indicates the first The frequency modulation slope of an LFM signal; The LFM signal is input into a parameterized symmetric instantaneous autocorrelation function for calculation, resulting in a two-dimensional complex matrix.

3. The signal parameter prediction method based on deep learning according to claim 2, characterized in that, The process of acquiring a multi-component linear frequency modulated (LFM) signal, performing down-conversion and ADC processing on the LFM signal to obtain an LFM signal specifically includes: A multi-component linear frequency modulated (LFM) signal is acquired, and the LFM signal is down-converted according to a preset down-conversion strategy to obtain a baseband signal. The baseband signal is processed by an ADC according to a preset ADC strategy to obtain an LFM signal.

4. The signal parameter prediction method based on deep learning according to claim 2, characterized in that, The step of inputting the LFM signal into a parameterized symmetric instantaneous autocorrelation function for calculation to obtain a two-dimensional complex matrix specifically includes: The LFM signal is input into a parameterized symmetric instantaneous autocorrelation function for calculation, resulting in a two-dimensional complex matrix: ； in, Indicates the time delay parameter. Represents the time delay variable. Represents a two-dimensional complex matrix. This indicates the conjugate operation. This represents the input value.

5. The signal parameter prediction method based on deep learning according to claim 1, characterized in that, The step of decoupling the two-dimensional complex matrix according to the scaling transformation strategy to obtain the two-dimensional feature matrix specifically includes: The two-dimensional complex matrix is ​​decoupled according to the scaling factor of the scaling transformation strategy to obtain the two-dimensional feature matrix: ； in, This represents the input signal of the neural network. Indicates the scaling operator. This indicates the scaled timeline. Indicates the scaling factor. This represents the time delay parameter.

6. The signal parameter prediction method based on deep learning according to claim 1, characterized in that, The step of inputting the two-dimensional feature matrix into a deep neural network model for prediction to obtain predicted signal parameter values ​​specifically includes: The two-dimensional feature matrix is ​​input into a deep neural network model for super-resolution analysis to obtain the amplitude spectrum. The amplitude spectrum is estimated using the deep neural network model to obtain predicted signal parameters.

7. The signal parameter prediction method based on deep learning according to claim 6, characterized in that, The deep neural network model includes a width matrix transformation layer, a height matrix transformation layer, a super-resolution layer, a transposed 2D convolutional layer, and an output layer.

8. The signal parameter prediction method based on deep learning according to claim 7, characterized in that, The step of inputting the two-dimensional feature matrix into a deep neural network model for super-resolution analysis to obtain the amplitude spectrum specifically includes: The two-dimensional feature matrix is ​​input into the width matrix transformation layer and the height matrix transformation layer for complex linear transformation to obtain the target width matrix and the target height matrix: ； ； in, Represents the real part of the width target matrix. The imaginary part of the width target matrix is ​​represented. Denotes the real part of the two-dimensional characteristic matrix. Denotes the imaginary part of a two-dimensional characteristic matrix. Indicates the transpose of the imaginary part. This indicates the transpose of the real part. The real part of the height target matrix is ​​represented. The imaginary part of the height target matrix is ​​represented. Represents the real part of the coefficient matrix. Represents the imaginary part of the coefficient matrix; The amplitude spectrum is obtained by performing super-resolution analysis on the width target matrix and the height target matrix through the super-resolution layer.

9. The signal parameter prediction method based on deep learning according to claim 7, characterized in that, The step of estimating the amplitude spectrum using the deep neural network model to obtain predicted signal parameters specifically includes: The amplitude spectrum is parameter-estimated using the sigmoid function of the deep neural network model and the transposed 2D convolutional layer to obtain predicted signal parameter values. ； in, Indicates amplitude spectrum, Represents a constant.

10. The signal parameter prediction method based on deep learning according to claim 9, characterized in that, The predicted signal parameters include the predicted center frequency and the predicted frequency modulation slope.

11. A signal parameter prediction system based on deep learning, characterized in that, The deep learning-based signal parameter prediction system includes: The two-dimensional complex matrix calculation module is used to acquire a multi-component linear frequency modulated signal, input the multi-component linear frequency modulated signal into a parameterized symmetric instantaneous autocorrelation function for calculation, and obtain a two-dimensional complex matrix. A two-dimensional feature matrix scaling module is used to decouple the two-dimensional complex matrix according to a scaling transformation strategy to obtain a two-dimensional feature matrix; The signal numerical prediction module is used to input the two-dimensional feature matrix into a deep neural network model for prediction to obtain predicted values ​​of signal parameters.

12. The signal parameter prediction system based on deep learning according to claim 11, characterized in that, The two-dimensional complex matrix calculation module includes: The signal processing unit is used to acquire a multi-component linear frequency modulated (LFM) signal, perform down-conversion and ADC processing on the multi-component LFM signal, and obtain an LFM signal. ； in, Indicates LFM signal, Indicates the number of LFM signals. Indicates the first LFM signal amplitude, Indicates a time index. Represents the imaginary unit. Indicates the first The center frequency of the LFM signal Indicates the first The frequency modulation slope of an LFM signal; The complex matrix calculation unit is used to input the LFM signal into a parameterized symmetric instantaneous autocorrelation function for calculation to obtain a two-dimensional complex matrix.

13. The signal parameter prediction system based on deep learning according to claim 12, characterized in that, The signal processing unit includes: The signal down-conversion processing subunit is used to acquire a multi-component linear frequency modulated signal and perform down-conversion processing on the multi-component linear frequency modulated signal according to a preset down-conversion strategy to obtain a baseband signal. The baseband signal processing subunit is used to perform ADC processing on the baseband signal according to a preset ADC strategy to obtain an LFM signal.

14. The signal parameter prediction system based on deep learning according to claim 12, characterized in that, The complex matrix calculation unit includes: The matrix calculation subunit is used to input the LFM signal into a parameterized symmetric instantaneous autocorrelation function for calculation, resulting in a two-dimensional complex matrix: ； in, Indicates the time delay parameter. Represents the time delay variable. Represents a two-dimensional complex matrix. This indicates the conjugate operation. This represents the input value.

15. The signal parameter prediction system based on deep learning according to claim 11, characterized in that, The two-dimensional feature matrix scaling module includes: The decoupling unit is used to decouple the two-dimensional complex matrix according to the scaling factor of the scaling transformation strategy to obtain a two-dimensional feature matrix: ； in, This represents the input signal of the neural network. Indicates the scaling operator. This indicates the scaled timeline. Indicates the scaling factor. This represents the time delay parameter.

16. The signal parameter prediction system based on deep learning according to claim 11, characterized in that, The signal numerical prediction module includes: The amplitude spectrum analysis unit is used to input the two-dimensional feature matrix into a deep neural network model for super-resolution analysis to obtain the amplitude spectrum. The parameter estimation unit is used to estimate the parameters of the amplitude spectrum through the deep neural network model to obtain the predicted values ​​of the signal parameters.

17. The signal parameter prediction system based on deep learning according to claim 16, characterized in that, The amplitude spectrum analysis unit includes: The matrix linear transformation subunit is used to input the two-dimensional feature matrix into the width matrix transformation layer and the height matrix transformation layer of the deep neural network model for complex linear transformation to obtain the target width matrix and the target height matrix. ； ； in, Represents the real part of the width target matrix. The imaginary part of the width target matrix is ​​represented. Denotes the real part of the two-dimensional characteristic matrix. Denotes the imaginary part of a two-dimensional characteristic matrix. Indicates the transpose of the imaginary part. This indicates the transpose of the real part. The real part of the height target matrix is ​​represented. The imaginary part of the height target matrix is ​​represented. Represents the real part of the coefficient matrix. Represents the imaginary part of the coefficient matrix; The amplitude spectrum super-resolution analysis subunit is used to perform super-resolution analysis on the width target matrix and the height target matrix through the super-resolution layer of the deep neural network model to obtain the amplitude spectrum.

18. The signal parameter prediction system based on deep learning according to claim 16, characterized in that, The parameter estimation unit includes: The amplitude spectrum parameter estimation subunit is used to estimate the parameters of the amplitude spectrum using the sigmoid function and the transposed 2D convolutional layer of the deep neural network model, to obtain predicted signal parameter values. ； in, Indicates amplitude spectrum, Represents a constant.

19. A terminal, characterized in that, The terminal includes: a memory, a processor, and a deep learning-based signal parameter prediction program stored in the memory and executable on the processor. When the deep learning-based signal parameter prediction program is executed by the processor, it implements the steps of the deep learning-based signal parameter prediction method as described in any one of claims 1-10.

20. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a deep learning-based signal parameter prediction program, which, when executed by a processor, implements the steps of the deep learning-based signal parameter prediction method as described in any one of claims 1-10.

Images