An intermediate frequency signal conversion system based on S-band frequency conversion processing

Abstract

The present invention relates to the field of digital signal processing technology, and in particular to an intermediate frequency signal conversion system based on S-band frequency conversion processing, which includes the following steps: S1, denoising the original signal, S2, dynamically adjusting the local frequency characteristics of the signal through an adaptive time-frequency window, S3, extracting the time-frequency characteristics of different frequency bands through wavelet packet transform, and the wavelet packet transform decomposes the signal into multiple frequency bands through multi-level decomposition, S4, dynamically adjusting the resolution of the time-frequency graph through a reconstruction algorithm, and optimizing the reconstruction accuracy of different frequency components, S5, extracting signal features from the processed signal time-frequency graph and wavelet packet coefficients, and extracting features related to the S-band from the signal, and S6, learning the potential representation of the signal through a variational autoencoder to extract a low-dimensional feature space. The present invention can efficiently remove high-frequency noise in the S-band signal, improve the recognition ability of the signal pattern, process the S-band signal in real time, and dynamically optimize according to the signal characteristics.

Description

Technical Field

[0001] The present invention relates to the technical field of digital signal processing, and in particular to an intermediate frequency signal conversion system based on S-band frequency conversion processing. Background Art

[0002] With the rapid development of modern communication technology, S-band signals are widely used in wireless communications, radar, satellite navigation and other fields due to their superior transmission characteristics. However, S-band signals are easily interfered by complex electromagnetic environments in practical applications, especially noise and spectrum interference, which can significantly affect the accuracy and stability of signals. At the same time, due to the high bandwidth and frequency distribution of S-band signals, traditional signal processing methods are insufficient in time-frequency resolution and feature extraction capabilities, making it difficult to accurately capture the key patterns and features in the signal.

[0003] At present, mainstream signal processing technologies mostly use single analysis methods based on time domain or frequency domain. Although they can remove noise to a certain extent, they still have limitations in frequency resolution and signal reconstruction accuracy. In addition, in complex signal environments, how to efficiently extract signal features and optimize signal classification capabilities is still an urgent problem to be solved. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides an intermediate frequency signal conversion system based on S-band frequency conversion processing.

[0005] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] An intermediate frequency signal conversion system based on S-band frequency conversion processing comprises the following steps:

[0007] S1: De-noising the original signal. For the high-frequency noise in the S-band signal, the wavelet denoising algorithm is used to remove the noise components in the wavelet packet coefficients by decomposing the signal through wavelet packets. The time-frequency characteristics of different frequency bands are extracted through wavelet packet transform. The wavelet packet transform decomposes the signal into multiple frequency bands through multi-level decomposition.

[0008] S2: Dynamically adjust the local frequency characteristics of the signal through an adaptive time-frequency window to improve the frequency resolution. Based on wavelet packet decomposition, dynamically adjust the resolution of the time-frequency diagram through a reconstruction algorithm to optimize the reconstruction accuracy of different frequency components and fully capture the signal characteristics.

[0009] S3: Extract signal features from the processed signal time-frequency diagram and wavelet packet coefficients, and extract features related to the S frequency band from the signal;

[0010] S4: The latent representation of the signal is learned through a variational autoencoder to extract a low-dimensional feature space, and is adaptively optimized in combination with reinforcement learning to identify typical patterns in the S-band signal.

[0011] Furthermore, the step S2 specifically includes the following sub-steps:

[0012] S201: After the input signal is transformed by wavelet packet, a multi-level frequency band decomposition is generated, time-frequency analysis is performed on each frequency band, and a time-frequency diagram is preliminarily constructed according to the characteristics of the signal;

[0013] S202: dynamically adjusting the time-frequency resolution based on the local characteristics of the signal, and using a dynamic window function of time-frequency reconstruction to capture changes of different frequency components in different time periods;

[0014] S203: Perform local frequency reconstruction on each frequency band, use spectrum smoothing and transition band adjustment technology, combine spectrum filtering method to remove artifacts or noise on each frequency band of wavelet packet decomposition, and then use inverse wavelet packet transform to re-combine the time-frequency information of multiple frequency bands to obtain a reconstructed version of the signal.

[0015] Furthermore, in step S202, the dynamic window function The width is dynamically adjusted according to the local characteristics of the signal. The process is as follows:

[0016] ;

[0017] in, is a dynamic window function, is the window width, is the center position of the window, Represents a time variable.

[0018] Furthermore, in step S203, the inverse transform process performs signal recovery by reconstructing the wavelet coefficients of all frequency bands, and the process is as follows:

[0019] ;

[0020] in, is the input signal, is the mother wavelet function Shifted into time The translation form of position, represents the index of the frequency band, is the restored wavelet coefficient, represents the time variable, is the sampling period.

[0021] Furthermore, the step S3 specifically includes the following sub-steps:

[0022] S301: extracting time-frequency graph features of the reconstructed signal, obtaining a description of the signal change by analyzing the energy distribution, frequency offset and instantaneous frequency change characteristics in different time and frequency segments, clustering the time-frequency graph of the signal using a spectrum clustering algorithm, and distinguishing different frequency bands or signal modes in the signal;

[0023] S302: extracting the instantaneous frequency characteristics of the signal from the time-frequency diagram using Hilbert transform, and further capturing the frequency fluctuations of the signal at different time points through instantaneous frequency estimation;

[0024] S303: Perform resonance peak analysis on the signal to identify the resonance frequency points in the signal, and use a feature fusion method to fuse multiple extracted signal features into a unified feature vector. The signal features include frequency offset, instantaneous frequency and resonance peak.

[0025] Furthermore, in step S302, the Hilbert transform process is:

[0026] The Hilbert transform process is:

[0027] ;

[0028] ;

[0029] in, is a real signal, is the analytical signal, , is the imaginary unit, are the orthogonal components of the input signal, t is the time variable, is the integral variable. The Hilbert transform is a real signal Convert to analytical signal process.

[0030] The rate of change of instantaneous frequency is as follows:

[0031] ;

[0032] The expression for instantaneous frequency is:

[0033] ;

[0034] in, yes The instantaneous phase of is the imaginary part of the analytical signal, is the real part of the analytical signal, is the time variable, is the instantaneous phase versus time The first derivative of .

[0035] Furthermore, in step S303, the resonance peak recognition is based on the spectrum peak detection of the signal, and firstly a fast Fourier transform is performed to obtain the spectrum, and the process is as follows:

[0036] ;

[0037] in, represents the spectrum of the signal, represents the time domain signal, represents the time variable, represents the complex exponential function.

[0038] Furthermore, the step S4 specifically includes the following sub-steps:

[0039] S401: Input the extracted signal features into the encoder part of the variational autoencoder, and perform latent space representation of the signal features through a convolutional neural network or a fully connected layer. The variational autoencoder captures the complex patterns and noise components in the S-band signal by learning the latent space of the signal;

[0040] S402: Combine the reinforcement learning strategy for adaptive optimization and strengthen the model's exploration of signal features by designing a reward function;

[0041] S403: Input the latent space representation generated by the variational autoencoder and the policy output optimized by reinforcement learning into a deep neural network classifier for training.

[0042] Furthermore, in step S401, the encoder part maps the input features into latent variables The distribution process is:

[0043] ;

[0044] in, and is the input signal The decoder generates signal reconstruction through the inferred mean and standard deviation, and the process is:

[0045] ;

[0046] in, is the conditional probability distribution of the latent variables obtained by the encoder network, Represents latent variables The Gaussian distribution with mean is , the standard deviation is , by input data control, is the input data obtained through the decoder network The conditional probability distribution of Represents a given latent variable In the case of reconstructing data The Gaussian distribution of is given by the mean and standard deviation Decide.

[0047] Furthermore, in step S403, the output layer process input into the deep neural network for training is: ;

[0048] in, is the activation function, is the weight matrix, is the bias term, It is the latent space representation obtained by the variational autoencoder.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] 1. The present invention uses wavelet packet transform and adaptive time-frequency window design to enable the system to efficiently remove high-frequency noise in the S-band signal, dynamically adjust the time-frequency resolution, optimize the accuracy of different frequency components during signal reconstruction, and ensure the high fidelity and comprehensiveness of signal reconstruction;

[0051] 2. The present invention uses a variational autoencoder to model the potential space of the signal, and combines it with a reinforcement learning strategy to achieve adaptive optimization of signal features, thereby improving the ability to recognize signal patterns, especially being able to accurately capture key features in complex electromagnetic environments, thereby improving the accuracy of the system's analysis of S-band signals;

[0052] 3. Through dynamic window function adjustment, time-frequency reconstruction and deep learning classifier, the present invention can process S-band signals in real time and perform dynamic optimization according to signal characteristics, thereby reducing the limitations of traditional fixed parameter models in non-stationary signal processing and greatly improving the efficiency and robustness of signal processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0054] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0055] Embodiment 1, as Figure 1 As shown, the present invention discloses an intermediate frequency signal conversion system based on S-band frequency conversion processing, comprising the following steps:

[0056] S1: De-noise the original signal. For the high-frequency noise in the S-band signal, the wavelet denoising algorithm is used. The signal is decomposed into wavelet packets to remove the noise components in the wavelet packet coefficients. The time-frequency characteristics of different frequency bands are extracted through wavelet packet transform. The wavelet packet transform decomposes the signal into multiple frequency bands through multi-level decomposition. Among them, wavelet packet decomposition is a multi-level frequency band decomposition of the signal, and its process is as follows: ;

[0057] in, is the input signal, is the mother wavelet function shifted to time The translation form of position, is the wavelet coefficient, which is used to describe the representation of the signal under the wavelet basis function. is the sampling period.

[0058] S2: In order to cope with the frequency offset and rapid changes of S-band signals, an adaptive time-frequency window is used to replace the traditional fixed window function. The size of the adaptive window is dynamically adjusted according to the local frequency characteristics of the signal to make the time-frequency distribution more accurate.

[0059] Based on wavelet packet decomposition, the resolution of the time-frequency diagram is dynamically adjusted through the reconstruction algorithm, the reconstruction accuracy of different frequency components is optimized, and the signal characteristics are fully captured.

[0060] S201: After the input signal is transformed by wavelet packet, a multi-level frequency band decomposition is generated, and each frequency band represents the time-frequency information of the signal in the frequency band. Time-frequency analysis is performed on each frequency band, and a time-frequency diagram is preliminarily constructed according to the characteristics of the signal. The time-frequency diagram is used to show the frequency component distribution of the signal at different time points.

[0061] S202: In the process of reconstructing the time-frequency graph, adaptive parameters are used to control the size, frequency resolution and time resolution of each time window. Traditional methods often use fixed window functions, which will result in insufficient resolution when the frequency of the signal changes rapidly. Therefore, it is necessary to dynamically adjust the time-frequency resolution based on the local characteristics of the signal. The dynamic window function used in time-frequency reconstruction is used to capture the changes of different frequency components in different time periods. These dynamic window functions will adaptively adjust the width according to the instantaneous frequency of the signal to avoid reconstruction errors when the frequency changes rapidly.

[0062] For time-frequency analysis, when the instantaneous frequency of the signal changes quickly, a narrower window is required to improve the time resolution, and when the instantaneous frequency changes slowly, a wider window is required to improve the frequency resolution. Therefore, the window width can be designed as a function of the instantaneous frequency. The specific formula is:

[0063] ;

[0064] in, is the window width at time t, is the scaling factor used to adjust the overall size of the window, is the instantaneous frequency of the signal, is a small constant used to prevent the window width from being too large when the instantaneous frequency is close to zero.

[0065] Adapt window width Substitute the original dynamic window function , the process is:

[0066] ;

[0067] in, is a dynamic window function, is the window width, is the center position of the window, Represents a time variable.

[0068] S203: Perform local frequency reconstruction for each frequency band. For example, for areas where the frequency changes suddenly or jumps, use a smaller time window for frequency reconstruction, thereby improving the reconstruction accuracy of this part of the signal. Use spectral smoothing and transition band adjustment techniques to reduce boundary effects, ensure smooth connection of signals in transition frequency bands, and avoid artifacts between frequency bands. In each frequency band of wavelet packet decomposition, combine spectral filtering methods to remove possible artifacts or noise. Then use the inverse wavelet packet transform to re-merge the time-frequency information of multiple frequency bands to obtain a reconstructed version of the signal. The final signal reconstruction result can fully retain the frequency components of the signal while eliminating artifacts and distortion that may occur during the reconstruction process. The inverse transform process recovers the signal by reconstructing the wavelet coefficients of all frequency bands. The process is as follows:

[0069] ;

[0070] in, is the input signal, is the mother wavelet function Shifted into time The translation form of position, represents the index of the frequency band, is the restored wavelet coefficient, represents the time variable, is the sampling period.

[0071] S3: Extract signal features from the processed signal time-frequency diagram and wavelet packet coefficients, and extract features related to the S frequency band from the signal.

[0072] S301: Extract time-frequency diagram features of the reconstructed signal, obtain a description of the signal change by analyzing the energy distribution, frequency offset and instantaneous frequency change characteristics in different time and frequency segments, cluster the time-frequency diagram of the signal using a spectrum clustering algorithm, and distinguish different frequency bands or signal modes in the signal.

[0073] S302: Use Hilbert transform to extract the instantaneous frequency characteristics of the signal from the time-frequency diagram. Through instantaneous frequency estimation, the frequency fluctuations of the signal at different time points are further captured. This facilitates the identification of frequency jumps, rapid frequency changes, and other phenomena, and is particularly suitable for frequency hopping and frequency modulation in S-band signals. In order to capture the frequency of the signal that changes rapidly in time, the Hilbert transform is used to calculate the instantaneous frequency. The Hilbert transform process is:

[0074] The Hilbert transform process is:

[0075] ;

[0076] ;

[0077] in, is a real signal, is the analytical signal, , is the imaginary unit, are the orthogonal components of the input signal, t is the time variable, is the integral variable. The Hilbert transform is a real signal Convert to analytical signal process.

[0078] The rate of change of instantaneous frequency is as follows:

[0079] ;

[0080] The expression for instantaneous frequency is:

[0081] ;

[0082] in, yes The instantaneous phase of is the imaginary part of the analytical signal, is the real part of the analytical signal, is the time variable, is the instantaneous phase versus time The first derivative of .

[0083] S303: Perform resonance peak analysis on the signal, identify the resonance frequency points in the signal, and use feature fusion method to fuse multiple extracted signal features into a unified feature vector. Signal features include frequency offset, instantaneous frequency and resonance peak. Reduce the dimension of feature space through dimensionality reduction technology to optimize subsequent classification and pattern recognition processes. Resonance peak recognition is usually based on the peak detection of the signal spectrum. First, fast Fourier transform is performed to obtain the spectrum. The process is as follows:

[0084] ;

[0085] in, represents the spectrum of the signal, represents the time domain signal, represents the time variable, represents the complex exponential function.

[0086] Then, based on the amplitude response of the spectrum , identify the resonant frequency , which satisfies the peak condition ; This means that the signal has an obvious resonance peak at a certain frequency in the spectrum. Multiple extracted features (such as instantaneous frequency, resonance peak) are combined to generate a feature vector , In order to reduce the dimension of the feature space, the principal component analysis method is applied as follows: ,in, is the principal component matrix, is the feature vector after dimensionality reduction.

[0087] S401: The extracted signal features are input into the encoder part of the variational autoencoder, and the signal features are represented in the latent space through a convolutional neural network or a fully connected layer. The variational autoencoder captures the complex patterns and noise components in the S-band signal by learning the latent space of the signal, thereby providing a more accurate feature representation for subsequent analysis of the signal. The encoder part maps the input features to latent variables The distribution process is:

[0088] ;

[0089] in, and is the input signal The decoder generates signal reconstruction through the inferred mean and standard deviation, and the process is:

[0090] ;

[0091] in, is the latent variable obtained by the encoder network The conditional probability distribution of Represents latent variables The Gaussian distribution with mean is , the standard deviation is , by input data control, is the input data obtained through the decoder network The conditional probability distribution of Represents a given latent variable In the case of reconstructing data The Gaussian distribution of is given by the mean and standard deviation Decide.

[0092] S402: Adaptive optimization is performed in combination with reinforcement learning strategies. By designing a reward function, the model's exploration of signal features is strengthened. The reinforcement learning model adjusts the latent space representation so that the generated signal can be closer to the distribution of the real signal. At the same time, it can help the model adaptively optimize its feature extraction strategy in different signal environments.

[0093] S403: Input the latent space representation generated by the variational autoencoder and the policy output optimized by reinforcement learning into the deep neural network classifier for training. During the classification process, the model divides the signal into multiple different categories (such as frequency hopping signals, frequency modulation signals, abnormal signals, etc.). The feedback mechanism of reinforcement learning helps the classifier better adapt to the special fluctuations and interference of S-band signals. The output layer process input into the deep neural network for training is:

[0094] ;

[0095] in, is the activation function, is the weight matrix, is the bias term, It is the latent space representation obtained by the variational autoencoder.

[0096] Embodiment 2: Based on embodiment 1, this embodiment proposes a method of using S transform to effectively capture the local characteristics of the spectrum.

[0097] In step S2, the S-band signal is located in the frequency band of 2.0 GHz to 4.0 GHz, and usually has a high frequency modulation or frequency hopping characteristic. Therefore, in the time-frequency analysis, the S transform is used to effectively capture the local characteristics of the spectrum. The S transform process is:

[0098] ;

[0099] in, is the input signal, Represents the transformation result of the signal in the time-frequency domain, where is the scale parameter, is the time position, is the scale parameter of the time window, is the frequency of the signal, It is an imaginary unit, and the S transform is used to provide the local time-frequency information of the signal.

[0100] For frequency hopping signals, the frequency of the signal will jump frequently within a specific time interval. The process of obtaining the instantaneous frequency through Hilbert transform is as follows: calculate the Hilbert transform through convolution and construct the analytical signal , get the instantaneous phase , calculate the instantaneous frequency , the formula is:

[0101] ;

[0102] ;

[0103] ;

[0104] ;

[0105] In the time-frequency reconstruction process, the instantaneous frequency With estimated frequency By comparison, the process of adjusting the band weights is:

[0106] ;

[0107] in is the original time-frequency coefficient, is the instantaneous frequency, is the estimated frequency of reconstruction, is the adjustment factor. By weighting different frequency bands of the time-frequency graph, the reconstruction quality of the frequency hopping signal is enhanced and noise interference is reduced.

[0108] It should be noted that, in the first embodiment, is the standard definition of Hilbert transform, which reflects the global characteristics of the signal through integral calculation. In this embodiment, It is a simplified form of Hilbert transform, ignoring the integral operation, and significantly reducing the computational complexity by directly multiplying with the attenuation function. It is suitable for scenarios with limited signal frequency range and relatively gentle local change characteristics, and can improve real-time performance.

[0109] Of course, the present invention may have many other embodiments. Without departing from the spirit and essence of the present invention, technicians familiar with the field may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. An intermediate frequency signal conversion system based on S-band frequency conversion processing, characterized in that: The following steps are involved: S1: De-noising the original signal. For the high-frequency noise in the S-band signal, the wavelet denoising algorithm is used to remove the noise components in the wavelet packet coefficients by decomposing the signal through wavelet packets. The time-frequency characteristics of different frequency bands are extracted through wavelet packet transform. The wavelet packet transform decomposes the signal into multiple frequency bands through multi-level decomposition. S2: Dynamically adjust the local frequency characteristics of the signal through an adaptive time-frequency window to improve the frequency resolution. Based on wavelet packet decomposition, dynamically adjust the resolution of the time-frequency diagram through a reconstruction algorithm to optimize the reconstruction accuracy of different frequency components and fully capture the signal characteristics. S3: Extract signal features from the processed signal time-frequency diagram and wavelet packet coefficients, and extract features related to the S frequency band from the signal; S4: The latent representation of the signal is learned through a variational autoencoder to extract a low-dimensional feature space, and is adaptively optimized in combination with reinforcement learning to identify typical patterns in the S-band signal.

2. The intermediate frequency signal conversion system based on S-band frequency conversion processing according to claim 1, characterized in that: The step S2 specifically includes the following sub-steps: S201: After the input signal is transformed by wavelet packet, a multi-level frequency band decomposition is generated, time-frequency analysis is performed on each frequency band, and a time-frequency diagram is preliminarily constructed according to the characteristics of the signal; S202: dynamically adjusting the time-frequency resolution based on the local characteristics of the signal, and using a dynamic window function of time-frequency reconstruction to capture changes of different frequency components in different time periods; S203: Perform local frequency reconstruction on each frequency band, use spectrum smoothing and transition band adjustment technology, combine spectrum filtering method to remove artifacts or noise on each frequency band of wavelet packet decomposition, and then use inverse wavelet packet transform to re-combine the time-frequency information of multiple frequency bands to obtain a reconstructed version of the signal.

3. The intermediate frequency signal conversion system based on S-band frequency conversion processing according to claim 2 is characterized in that: In step S202, the dynamic window function The width is dynamically adjusted according to the local characteristics of the signal. The process is as follows: ； in, is a dynamic window function, is the window width, is the center position of the window, Represents a time variable.

4. The intermediate frequency signal conversion system based on S-band frequency conversion processing according to claim 2 is characterized in that: In step S203, the inverse transformation process performs signal recovery by reconstructing the wavelet coefficients of all frequency bands, and the process is as follows: ； in, is the input signal, is the mother wavelet function Shifted into time The translation form of position, represents the index of the frequency band, is the restored wavelet coefficient, represents the time variable, is the sampling period.

5. The intermediate frequency signal conversion system based on S-band frequency conversion processing according to claim 1, characterized in that: The step S3 specifically includes the following sub-steps: S301: extracting time-frequency graph features of the reconstructed signal, obtaining a description of the signal change by analyzing the energy distribution, frequency offset and instantaneous frequency change characteristics in different time and frequency segments, clustering the time-frequency graph of the signal using a spectrum clustering algorithm, and distinguishing different frequency bands or signal modes in the signal; S302: extracting the instantaneous frequency characteristics of the signal from the time-frequency diagram using Hilbert transform, and further capturing the frequency fluctuations of the signal at different time points through instantaneous frequency estimation; S303: Perform resonance peak analysis on the signal to identify the resonance frequency points in the signal, and use a feature fusion method to fuse multiple extracted signal features into a unified feature vector. The signal features include frequency offset, instantaneous frequency and resonance peak.

6. The intermediate frequency signal conversion system based on S-band frequency conversion processing according to claim 5, characterized in that: In step S302, the Hilbert transform process is: ； ； in, is a real signal, is the analytical signal, , is the imaginary unit, are the orthogonal components of the input signal, t is the time variable, is the integration variable, The rate of change of instantaneous frequency is as follows: ； The expression for instantaneous frequency is: ； in, yes The instantaneous phase of is the imaginary part of the analytical signal, is the real part of the analytical signal, is the time variable, is the instantaneous phase versus time The first derivative of .

7. The intermediate frequency signal conversion system based on S-band frequency conversion processing according to claim 5, characterized in that: In step S303, the resonance peak recognition is based on the spectrum peak detection of the signal. First, a fast Fourier transform is performed to obtain the spectrum. The process is as follows: ； in, represents the spectrum of the signal, represents the time domain signal, represents the time variable, represents the complex exponential function.

8. The intermediate frequency signal conversion system based on S-band frequency conversion processing according to claim 1, characterized in that: The step S4 specifically includes the following sub-steps: S401: Input the extracted signal features into the encoder part of the variational autoencoder, and perform latent space representation of the signal features through a convolutional neural network or a fully connected layer. The variational autoencoder captures the complex patterns and noise components in the S-band signal by learning the latent space of the signal; S402: Combine the reinforcement learning strategy for adaptive optimization and strengthen the model's exploration of signal features by designing a reward function; S403: Input the latent space representation generated by the variational autoencoder and the policy output optimized by reinforcement learning into a deep neural network classifier for training.

9. The intermediate frequency signal conversion system based on S-band frequency conversion processing according to claim 8, characterized in that: In step S401, the encoder part maps the input features into latent variables The distribution process is: ； in, and is the input signal The decoder generates signal reconstruction through the inferred mean and standard deviation, and the process is: ； in, is the latent variable obtained by the encoder network The conditional probability distribution of Represents latent variables The Gaussian distribution with mean is , the standard deviation is , by input data control, is the input data obtained through the decoder network The conditional probability distribution of Represents a given latent variable In the case of reconstructing data The Gaussian distribution of is given by the mean and standard deviation Decide.

10. The intermediate frequency signal conversion system based on S-band frequency conversion processing according to claim 8, characterized in that: In step S403, the output layer process of inputting into the deep neural network for training is as follows: ; in, is the activation function, is the weight matrix, is the bias term, It is the latent space representation obtained by the variational autoencoder.

Images