Satellite communication filtering method based on handheld terminal

Abstract

The application discloses a satellite communication filtering method based on a handheld terminal and belongs to the technical field of satellite communication. The application firstly performs EMD decomposition on a satellite communication signal collected by the handheld terminal, obtains a plurality of intrinsic mode functions and a residual signal, and extracts energy signals of the intrinsic mode functions; then, first and second time-domain attenuation components are respectively calculated through neighborhood ratio detection and range impact detection, time-domain attenuation weight signals are obtained through fusion, and frequency-domain attenuation weight signals are calculated according to the energy signal in-band and out-of-band total energy difference value; each intrinsic mode function is processed by using a trained lightweight neural network, adaptive attention weighting is realized in combination with the time-domain and frequency-domain attenuation weight signals, a reconstructed signal is obtained, the reconstructed signal is superimposed with the residual signal, and finally, a filtered signal is output. The application effectively improves the satellite communication signal filtering precision through the time-domain and frequency-domain two-dimensional weight constraint combined with the lightweight neural network feature focusing.

Description

Technical Field

[0001] This invention relates to the field of satellite communication technology, and more specifically to a satellite communication filtering method based on a handheld terminal. Background Technology

[0002] Satellite communication plays a vital role in modern communications, especially in global mobile communication services, where handheld terminal devices (such as satellite phones and handheld data terminals) are widely used. The L-band (1-2 GHz) is a commonly used operating frequency band for satellite communication because of its strong penetration capabilities, enabling stable signal transmission even in adverse weather conditions. However, in complex electromagnetic environments, the satellite signals received by handheld terminal devices are often weak and subject to significant interference, leading to a substantial deterioration in signal quality and affecting the stability and accuracy of communication.

[0003] In existing technologies, filtering methods for satellite communication signals typically rely on traditional filters (such as bandpass filters and Kalman filters). These methods are mostly fixed in structure and lack adaptability. Fixed-structure filtering methods often cannot provide sufficient filtering accuracy when faced with dynamically changing satellite signals.

[0004] Especially in high-noise environments, existing filtering technologies cannot be adjusted in real time according to the specific characteristics of satellite signals, resulting in poor filtering effects and failing to effectively improve signal quality, thereby affecting the effectiveness and reliability of communication. Summary of the Invention

[0005] In view of the above-mentioned shortcomings in the prior art, the present invention provides a satellite communication filtering method based on a handheld terminal, which solves the problem of low filtering accuracy in the prior art.

[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a satellite communication filtering method based on a handheld terminal, comprising the following steps:

[0007] S1. Perform EMD decomposition on the satellite communication signal collected by the handheld terminal to obtain multiple intrinsic mode functions and residual signals, and obtain the energy signal for each intrinsic mode function;

[0008] S2. Detect the neighborhood ratio of the energy signal and calculate the first time-domain attenuation component;

[0009] S3. Perform range impact detection on the energy signal, calculate the second time-domain attenuation component, and combine it with the first time-domain attenuation component to obtain the time-domain attenuation weight signal.

[0010] S4. Based on the difference between the total out-of-band energy and the total in-band energy of the energy signal, calculate the frequency domain attenuation weight to obtain the frequency domain attenuation weight signal.

[0011] S5. Multiple intrinsic mode functions are processed using a trained lightweight neural network. Attention is applied based on the time-domain attenuation weight signal and the frequency-domain attenuation weight signal respectively to obtain the reconstructed signal.

[0012] S6. Add the reconstructed signal to the residual signal to obtain the filtered signal.

[0013] Furthermore, S2 includes the following sub-steps:

[0014] S21. Using the energy at each moment in the energy signal as the reference energy, obtain the energy within the neighborhood of the reference energy;

[0015] S22. Take the average value of each energy within the neighborhood of the reference energy as the neighborhood mean.

[0016] S23. Calculate the first time-domain decay component based on the ratio of the reference energy to the mean of the corresponding neighborhood.

[0017] Further, the process of S23 includes: taking the ratio of the reference energy to the mean of the corresponding neighborhood; when the ratio is less than or equal to the ratio threshold, setting the corresponding first time-domain decay component to 1; when the ratio is greater than the ratio threshold, subtracting the ratio from the ratio threshold to obtain the difference; using the natural constant e as the base and the negative of the square of the difference as the exponent, obtaining the first time-domain decay component in the range of 0 to 1.

[0018] Furthermore, S3 includes the following sub-steps:

[0019] S31. Using the energy at each moment in the energy signal as the reference energy, obtain the energy within the neighborhood of the reference energy;

[0020] S32. Take the average value of each energy in the neighborhood of the reference energy as the neighborhood mean, and extract the maximum and minimum energies in the neighborhood of the reference energy.

[0021] S33. Subtract the maximum and minimum energies in the neighborhood to obtain the energy range. Use the ratio of the energy range to the neighborhood mean as the range impact value.

[0022] S34. Calculate the second time-domain attenuation component based on the extreme impact value;

[0023] S35. Take the average of the first time-domain attenuation component and the second time-domain attenuation component to obtain the time-domain attenuation weight signal.

[0024] Furthermore, the process of S34 includes: when the range impact value is less than or equal to the range impact threshold, setting the corresponding second time-domain attenuation component to 1; when the range impact value is greater than the range impact threshold, subtracting the range impact value from the range impact threshold to obtain the difference; using the natural constant e as the base and the negative of the square of the difference as the exponent, obtaining the second time-domain attenuation component in the range of 0 to 1.

[0025] Furthermore, S4 includes the following sub-steps:

[0026] S41. Perform a short-time Fourier transform on each energy signal to obtain the global spectrum at each time step;

[0027] S42. Extract the spectrum of the effective signal frequency band from the global spectrum at each time moment. The remaining spectrum is the out-of-band spectrum.

[0028] S43. Add up all the amplitudes in the out-of-band spectrum to get the total out-of-band energy;

[0029] S44. Sum all the amplitude values ​​in the spectrum of the effective signal band to obtain the total energy within the band;

[0030] S45. Based on the difference between the total energy in the band and the total energy out of the band, calculate the frequency domain attenuation weight to obtain the frequency domain attenuation weight signal.

[0031] Furthermore, the S45 process includes: when the total energy in the band is less than the noise floor energy, the corresponding frequency domain attenuation weight is reset to 0; when the total energy in the band is greater than or equal to the noise floor energy, it is determined whether the total energy in the band is greater than or equal to the total energy outside the band. If so, the corresponding frequency domain attenuation weight is reset to 1; if not, the total energy outside the band is subtracted from the total energy in the band to obtain the difference. Using the natural constant e as the base and the negative of the square of the difference as the exponent, the frequency domain attenuation weight in the range of 0 to 1 is obtained. The frequency domain attenuation weights at each time point belonging to the same intrinsic mode function are used to form the frequency domain attenuation weight signal.

[0032] Furthermore, the lightweight neural network in S5 includes: multiple dual-domain weight adaptive modulation modules, adder A1, adder A2, a first temporal convolutional layer, a second temporal convolutional layer, and a weighting module;

[0033] Each dual-domain weighted adaptive modulation module is used to input an intrinsic mode function, as well as the corresponding time-domain attenuation weight signal and frequency-domain attenuation weight signal;

[0034] The input of adder A1 is connected to the first output of multiple dual-domain weight adaptive modulation modules; the input of adder A2 is connected to the second output of multiple dual-domain weight adaptive modulation modules; the input of the first temporal convolutional layer is connected to the output of adder A1; the input of the second temporal convolutional layer is connected to the output of adder A2; the first input of the weighting module is connected to the output of the first temporal convolutional layer, its second input is connected to the output of the second temporal convolutional layer, and its output serves as the output of the lightweight neural network.

[0035] Furthermore, S5 includes the following sub-steps:

[0036] S51. Each dual-domain weighted adaptive modulation module generates attention to the time-domain attenuation weight signal and the frequency-domain attenuation weight signal respectively. The two attentions are applied to the features of the corresponding intrinsic mode function to obtain the time-domain weighted correction feature and the frequency-domain weighted correction feature.

[0037] S52. Add all time-domain weighted correction features element-wise using adder A1 to obtain time-domain weighted correction fusion features. Add all frequency-domain weighted correction features element-wise using adder A2 to obtain frequency-domain weighted correction fusion features.

[0038] S53. The first temporal convolutional layer is used to extract features from the temporal weight correction and fusion features to obtain the temporal weight correction and fusion deep features. The second temporal convolutional layer is used to extract features from the frequency domain weight correction and fusion features to obtain the frequency domain weight correction and fusion deep features.

[0039] S54. The deep features of time-domain weighted correction fusion and frequency-domain weighted correction fusion are weighted using a weighting module to obtain the reconstructed signal.

[0040] Furthermore, the dual-domain weight adaptive modulation module includes: a first pointwise convolutional layer, a second pointwise convolutional layer, a third pointwise convolutional layer, a first sigmoid layer, a second sigmoid layer, a multiplier M1, and a multiplier M2;

[0041] The input of the first pointwise convolutional layer is used to input the temporal decay weight signal, and its output is connected to the input of the first sigmoid layer.

[0042] The input of the second pointwise convolutional layer is used to input the intrinsic mode function, and its output is connected to the first input of multiplier M1 and the first input of multiplier M2, respectively.

[0043] The input of the third pointwise convolutional layer is used to input the frequency domain attenuation weight signal, and its output is connected to the input of the second sigmoid layer.

[0044] The second input of multiplier M1 is connected to the output of the first sigmoid layer, and its output serves as the first output of the dual-domain weight adaptive modulation module, used to output the temporal weight correction feature.

[0045] The second input of multiplier M2 is connected to the output of the second sigmoid layer, and its output serves as the second output of the dual-domain weighted adaptive modulation module, used to output the frequency domain weighted correction feature.

[0046] The beneficial effects of this invention are as follows:

[0047] 1. This invention decomposes non-stationary and nonlinear satellite communication signals into multiple intrinsic mode functions (IMFs) through EMD decomposition, avoiding the problem that traditional fixed filter structures cannot adapt to dynamic signal changes. Combined with subsequent time-domain and frequency-domain dual-weight analysis, it can adaptively refine the noise characteristics of different mode components, significantly improving filtering accuracy.

[0048] 2. In the time domain, the first and second time-domain attenuation components are calculated through neighborhood ratio detection and range impulse detection, respectively, to construct a time-domain attenuation weight signal, which can effectively identify and suppress sudden impulse noise and environmental noise. In the frequency domain, the frequency-domain attenuation weight signal is calculated using the difference between the total energy outside the band and inside the band, enhancing the ability to suppress outside-band interference, which is especially suitable for satellite communication scenarios in high-noise and complex electromagnetic environments.

[0049] 3. This invention employs a trained lightweight neural network to process multiple intrinsic mode functions (EMFs), and applies attention based on both time-domain and frequency-domain attenuation weight signals. This enables the network to effectively attenuate noise-rich mode components while preserving the dominant signal mode components. Compared to traditional filtering methods, this dual-domain attention mechanism better preserves signal details and improves the quality of the reconstructed signal. Attached Figure Description

[0050] Figure 1 This is a flowchart of a satellite communication filtering method based on a handheld terminal;

[0051] Figure 2 This is a schematic diagram of the structure of a lightweight neural network;

[0052] Figure 3 This is a schematic diagram of the dual-domain weighted adaptive modulation module. Detailed Implementation

[0053] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0054] like Figure 1 As shown, a satellite communication filtering method based on a handheld terminal includes the following steps:

[0055] S1. Perform EMD decomposition on the satellite communication signal collected by the handheld terminal to obtain multiple intrinsic mode functions and residual signals, and obtain the energy signal for each intrinsic mode function;

[0056] S2. Detect the neighborhood ratio of the energy signal and calculate the first time-domain attenuation component;

[0057] S3. Perform range impact detection on the energy signal, calculate the second time-domain attenuation component, and combine it with the first time-domain attenuation component to obtain the time-domain attenuation weight signal.

[0058] S4. Based on the difference between the total out-of-band energy and the total in-band energy of the energy signal, calculate the frequency domain attenuation weight to obtain the frequency domain attenuation weight signal.

[0059] S5. Multiple intrinsic mode functions are processed using a trained lightweight neural network. Attention is applied based on the time-domain attenuation weight signal and the frequency-domain attenuation weight signal respectively to obtain the reconstructed signal.

[0060] S6. Add the reconstructed signal to the residual signal to obtain the filtered signal.

[0061] The expression for obtaining the energy signal for each intrinsic mode function is: ,in, For the first The energy signal of the eigenmode function of the th intrinsic mode function Energy at all times For the first The eigenmode function of the eigenmode function Moment amplitude, Used as time number.

[0062] In this embodiment, S2 includes the following sub-steps:

[0063] S21. Using the energy at each moment in the energy signal as the reference energy, obtain the energy within the neighborhood of the reference energy;

[0064] S22. Take the average value of each energy within the neighborhood of the reference energy as the neighborhood mean.

[0065] S23. Calculate the first time-domain decay component based on the ratio of the reference energy to the mean of the corresponding neighborhood.

[0066] In this embodiment, process S23 includes: taking the ratio of the reference energy to the mean of the corresponding neighborhood. When the ratio is less than or equal to the ratio threshold ( ), for the corresponding first time-domain attenuation component Set to 1 when the ratio is greater than the ratio threshold ( Subtracting the ratio from the ratio threshold yields the difference. Using the natural constant e as the base and the negative of the square of this difference as the exponent, the first time-domain decay component in the range of 0 to 1 is obtained. ,in, For the first The eigenmode function of the eigenmode function The first time-domain decay component at time step 1. For the first The energy signal of the eigenmode function of the th intrinsic mode function The ratio corresponding to time, For the first The energy signal of the eigenmode function of the th intrinsic mode function Energy at all times For the first The energy signal of the eigenmode function of the th intrinsic mode function Constantly refer to the neighborhood mean of the energy. To avoid parameters with a denominator of 0, It is a natural constant. The ratio threshold, This is the numbering of the intrinsic mode functions. This is the time number.

[0067] This invention, by dynamically comparing the instantaneous energy of each intrinsic mode function with its local neighborhood mean, can accurately identify anomalous points in the signal where instantaneous energy changes abruptly—that is, when the reference energy is significantly higher than the mean of the surrounding neighborhood (the ratio is greater than a threshold), an exponential decay function is used. A first time-domain attenuation component between 0 and 1 is generated, which effectively suppresses this moment in subsequent reconstruction; while when the ratio is normal, the attenuation component is directly set to 1 to retain the original components of the signal.

[0068] In this embodiment, the ratio threshold is usually set between 2.5 and 3, and the specific value can be adjusted according to the actual noise situation. When the background noise is large, the threshold can be increased to 3 to 5 to avoid excessive attenuation of normal signal fluctuations.

[0069] In this embodiment, S3 includes the following sub-steps:

[0070] S31. Using the energy at each moment in the energy signal as the reference energy, obtain the energy within the neighborhood of the reference energy;

[0071] S32. Take the average value of each energy in the neighborhood of the reference energy as the neighborhood mean, and extract the maximum and minimum energies in the neighborhood of the reference energy.

[0072] S33. Subtract the maximum and minimum energies in the neighborhood to obtain the energy range. Use the ratio of the energy range to the neighborhood mean as the range impact value.

[0073] S34. Calculate the second time-domain attenuation component based on the extreme impact value;

[0074] S35. Take the average of the first time-domain attenuation component and the second time-domain attenuation component to obtain the time-domain attenuation weight signal.

[0075] The expression for the energy range obtained in S33 is: ,in, For the first The eigenmode function of the eigenmode function At times, energy levels are extremely low. For the first The energy signal of the eigenmode function. The maximum energy within the reference energy neighborhood at time [time]. For the first The energy signal of the eigenmode function. The minimum energy within the neighborhood of the reference energy at a given time.

[0076] The expression for the extreme impact value obtained in S33 is: ,in, For the first The energy signal of the eigenmode function of the th intrinsic mode function The extreme impact corresponding to each moment, For the first The energy signal of the eigenmode function. The average value of the neighborhood energy is always referenced.

[0077] In this embodiment, process S34 includes: when the range shock value is less than or equal to the range shock threshold ( ), for the corresponding second time-domain attenuation component Set to 1 when the range shock value is greater than the range shock threshold ( Subtracting the extreme impact value from the extreme impact threshold yields the difference. Using the natural constant e as the base and the negative of the square of this difference as the exponent, the second time-domain attenuation component in the range of 0 to 1 is obtained. ,in, For the first The eigenmode function of the eigenmode function The second time-domain decay component at time 10:00. This is the extreme impact threshold.

[0078] The expression for averaging the first and second time-domain attenuation components in S35 is as follows: ,in, For the first The time-domain decaying weighted signal corresponding to the eigenmode function. The time-domain decay weight at time step.

[0079] The extreme impact threshold is adaptively set based on experiments and experience.

[0080] This invention calculates the energy range by extracting the maximum and minimum values ​​within the neighborhood of the reference energy, and then divides this value by the neighborhood mean to obtain the range impact value. This effectively measures the instantaneous fluctuation amplitude of energy within a local window—a larger range impact value indicates a more severe energy jump or spike interference within the neighborhood. When the range impact value exceeds a threshold, an exponential decay function is used. A second time-domain decay component, ranging from 0 to 1, is generated to adaptively suppress anomalous moments with drastic fluctuations; otherwise, it is set to 1 to preserve the signal. This method complements the first time-domain decay component (based on the ratio of reference energy to the neighborhood mean, focusing on detecting single-point energy spikes): the former is sensitive to single-point anomalies, while the latter is sensitive to the overall fluctuation amplitude within a local window. The time-domain decay weight signal obtained by averaging the two components can more comprehensively and robustly characterize the time-domain noise features of different forms, avoiding missed or false detections by a single detection mechanism.

[0081] In this embodiment, the length of the neighborhood range is set to 7 to 10 time intervals, and the length of the neighborhood range is the same in S2 and S3.

[0082] In this embodiment, S4 includes the following sub-steps:

[0083] S41. Perform a short-time Fourier transform on each energy signal to obtain the global spectrum at each time step;

[0084] S42. Extract the spectrum of the effective signal frequency band from the global spectrum at each time moment. The remaining spectrum is the out-of-band spectrum.

[0085] S43. Add up all the amplitudes in the out-of-band spectrum to get the total out-of-band energy;

[0086] S44. Sum all the amplitude values ​​in the spectrum of the effective signal band to obtain the total energy within the band;

[0087] S45. Based on the difference between the total energy in the band and the total energy out of the band, calculate the frequency domain attenuation weight to obtain the frequency domain attenuation weight signal.

[0088] In this embodiment, the effective signal frequency band is set as follows: ,in, B is the center frequency of the satellite communication signal (determined by the current communication frequency of the handheld terminal), and B is the bandwidth, B=25kHz (narrowband voice / data scenario) or =200kHz (broadband data scenario).

[0089] In this embodiment, the process S45 includes: when the total in-band energy is less than the noise floor energy ( ), corresponding frequency domain attenuation weights Set to 0 when the total energy within the band is greater than or equal to the noise floor energy ( ), determine whether the total in-band energy is greater than or equal to the total out-of-band energy, and if so, attenuate the corresponding frequency domain weights. If set to 1, then subtract the total out-of-band energy from the total in-band energy to obtain the difference. Using the natural constant e as the base and the negative of the square of this difference as the exponent, the frequency domain attenuation weight in the range of 0 to 1 is obtained. The frequency domain attenuation weights at each time step belonging to the same eigenmode function are used to construct a frequency domain attenuation weight signal, where, For the first The frequency domain attenuation weight signal corresponding to the eigenmode function. Time-domain attenuation weights, For the first The energy signal of the eigenmode function of the th intrinsic mode function Always carry external total energy. For the first The energy signal of the eigenmode function of the th intrinsic mode function Always carry internal total energy. This represents the noise floor energy.

[0090] The noise floor energy is preset to a fixed value. It can be obtained by adding all amplitudes within the effective signal frequency band when the satellite signal is not locked and is in an idle state. Its specific value can be adjusted based on experience or implementation.

[0091] When the total energy within the band is lower than the noise floor energy, it indicates that there is likely no effective signal at that moment, and the frequency domain attenuation weight is directly reset to 0 to suppress that moment. When the total energy within the band is greater than or equal to the noise floor and greater than or equal to the total energy outside the band, it indicates that the signal is dominant, and the weight is reset to 1 to retain it. When the total energy within the band is greater than or equal to the noise floor but less than the total energy outside the band, it indicates that the out-of-band interference is strong, and an exponential attenuation function is used. Weights in the range of 0 to 1 are generated, and adaptive decay is performed when the out-of-band energy is relatively too high.

[0092] For each energy signal, a Short-Time Fourier Transform (STFT) is performed with a window length of 64 sampling points and an overlap rate of 50% between adjacent windows. This yields a series of global spectra corresponding to time frames. Then, the frequency domain attenuation weights at each time point are interpolated (e.g., linear interpolation) or assigned piecewise to pad the frequency domain attenuation weight signals to match the length of the intrinsic mode function. For example, performing an STFT on a signal of length 1000 results in approximately 30 time frames, each corresponding to a global spectrum and thus 30 frequency domain attenuation weight values. For instance, if frame 1 covers sampling points 1-64 with a weight of 0.85, and frame 2 covers sampling points 33-96 with a weight of 0.42, then the sampling points in the overlapping region take the average of the two weights, 0.635, while the weights of other sampling points remain unchanged.

[0093] like Figure 2 As shown, the lightweight neural network in S5 includes: multiple dual-domain weight adaptive modulation modules, adder A1, adder A2, a first temporal convolutional layer, a second temporal convolutional layer, and a weighting module;

[0094] Each dual-domain weighted adaptive modulation module is used to input an intrinsic mode function, as well as the corresponding time-domain attenuation weight signal and frequency-domain attenuation weight signal;

[0095] The input of adder A1 is connected to the first output of multiple dual-domain weight adaptive modulation modules; the input of adder A2 is connected to the second output of multiple dual-domain weight adaptive modulation modules; the input of the first temporal convolutional layer is connected to the output of adder A1; the input of the second temporal convolutional layer is connected to the output of adder A2; the first input of the weighting module is connected to the output of the first temporal convolutional layer, its second input is connected to the output of the second temporal convolutional layer, and its output serves as the output of the lightweight neural network.

[0096] S5 includes the following steps:

[0097] S51. Each dual-domain weighted adaptive modulation module generates attention to the time-domain attenuation weight signal and the frequency-domain attenuation weight signal respectively. The two attentions are applied to the features of the corresponding intrinsic mode function to obtain the time-domain weighted correction feature and the frequency-domain weighted correction feature.

[0098] S52. Add all time-domain weighted correction features element-wise using adder A1 to obtain time-domain weighted correction fusion features. Add all frequency-domain weighted correction features element-wise using adder A2 to obtain frequency-domain weighted correction fusion features.

[0099] S53. The first temporal convolutional layer is used to extract features from the temporal weight correction and fusion features to obtain the temporal weight correction and fusion deep features. The second temporal convolutional layer is used to extract features from the frequency domain weight correction and fusion features to obtain the frequency domain weight correction and fusion deep features.

[0100] S54. The deep features of time-domain weighted correction fusion and frequency-domain weighted correction fusion are weighted using a weighting module to obtain the reconstructed signal.

[0101] This invention achieves independent adaptive modulation of each intrinsic mode function in both the time and frequency domains. Two adders fuse the frequency and time domain correction features respectively, enabling the final reconstructed signal to comprehensively utilize noise suppression information in both domains. Temporal convolutional layers extract temporal dimension features from the fused features in both domains, enhancing the temporal continuity of the satellite communication signal. A weighting module further balances the contributions of time and frequency domain attention, ensuring that the reconstructed signal effectively suppresses time-domain impulse interference and out-of-band frequency noise while retaining the key features of the original signal. The entire network adopts a lightweight design (without complex fully connected or convolutional stacking), resulting in low computational overhead, making it suitable for real-time deployment on handheld terminals and significantly improving filtering accuracy and communication reliability in high-noise dynamic environments.

[0102] There are M dual-domain weighted adaptive modulation modules. Each dual-domain weighted adaptive modulation module is input with an intrinsic mode function and its corresponding time-domain attenuation weight signal and frequency-domain attenuation weight signal.

[0103] In this embodiment, the kernel of the first temporal convolutional layer is 1×3 and the kernel of the second temporal convolutional layer is 1×3. The stride is set to 1, and the boundary padding is 1, so that the length of the feature sequence before and after the temporal convolution remains unchanged and the scale of the input feature is not changed.

[0104] like Figure 3 As shown, the dual-domain weight adaptive modulation module includes: a first pointwise convolutional layer, a second pointwise convolutional layer, a third pointwise convolutional layer, a first sigmoid layer, a second sigmoid layer, a multiplier M1, and a multiplier M2;

[0105] The input of the first pointwise convolutional layer is used to input the temporal decay weight signal, and its output is connected to the input of the first sigmoid layer.

[0106] The input of the second pointwise convolutional layer is used to input the intrinsic mode function, and its output is connected to the first input of multiplier M1 and the first input of multiplier M2, respectively.

[0107] The input of the third pointwise convolutional layer is used to input the frequency domain attenuation weight signal, and its output is connected to the input of the second sigmoid layer.

[0108] The second input of multiplier M1 is connected to the output of the first sigmoid layer, and its output serves as the first output of the dual-domain weight adaptive modulation module, used to output the temporal weight correction feature.

[0109] The second input of multiplier M2 is connected to the output of the second sigmoid layer, and its output serves as the second output of the dual-domain weighted adaptive modulation module, used to output the frequency domain weighted correction feature.

[0110] The first and third pointwise convolutional layers of this invention perform learnable feature transformations on the time-domain and frequency-domain weight signals, respectively, and then the attention is mapped to the range of 0 to 1 by the sigmoid layer, realizing adaptive nonlinear modulation of the dual-domain weights; the intrinsic mode function features extracted by the second pointwise convolutional layer are multiplied by the two attentions at the same time, realizing independent application of the time-domain and frequency-domain attentions on the same feature; each intrinsic mode function can obtain targeted suppression strength according to its own time-domain noise distribution (such as impulse interference) and frequency-domain noise distribution (such as out-of-band spurious emissions), which significantly improves the adaptive filtering capability of the neural network in complex noise environments.

[0111] In this embodiment, the kernel of the first pointwise convolutional layer is 1×1, the kernel of the second pointwise convolutional layer is 1×1, and the kernel of the third pointwise convolutional layer is 1×1.

[0112] The intrinsic mode function has a length of L. The input and output lengths of the second pointwise convolutional layer are the same. The lengths of the frequency domain weight correction feature and the temporal domain weight correction feature are both L, containing feature values ​​at L time points. The number of intrinsic mode functions is M. That is, the M frequency domain weight correction features are added together according to the feature values ​​at the same time point to obtain the feature values ​​at the corresponding time point in the frequency domain weight correction fusion feature. The M temporal weight correction features are added together according to the feature values ​​at the same time point to obtain the feature values ​​at the corresponding time point in the temporal domain weight correction fusion feature.

[0113] The expression for the weighted module is: ,in, For the reconstructed signal, the first Amplitude at time, To correct the temporal weights in the fusion of deep features The eigenvalues ​​at time t, For frequency domain weight correction of the fused deep features, the first The eigenvalues ​​at time t, For the weighted module Time-domain weights For the weighted module Time-domain weights.

[0114] In this embodiment, when training the lightweight neural network, a large amount of labeled sample data needs to be generated (input: noisy signal, label: clean signal). For L-band satellite communication signals, a modulation method (BPSK / QPSK / GMSK) is selected to generate multiple clean signals. Various noise parameters (SNR, impulse probability, interference type, etc.) are randomly selected and added sequentially to obtain noisy signals. These noisy signals are then used to generate training samples (multiple intrinsic mode functions, and corresponding time-domain and frequency-domain attenuation weight signals) according to steps S1-S4. Each training sample is input into the lightweight neural network. The mean square error between the filtered signal and the target signal (clean signal) is used as the loss function. The backpropagation algorithm and gradient descent optimizer (such as Adam) are used to iteratively update the network parameters until the loss function converges. After training, the network parameters are fixed. The loss function is: ,in, The loss value. The first of the filtered signals Moment amplitude, The first of the target signal Moment amplitude.

[0115] This invention first utilizes Empirical Mode Decomposition (EMD) to adaptively decompose the non-stationary, nonlinear, and extremely low signal-to-noise ratio original satellite signal into multiple Intrinsic Mode Functions (IMFs) and residual signals. For each IMF, the time-domain attenuation weight and frequency-domain attenuation weight are obtained at each time step. A lightweight neural network then applies attention to the features at the corresponding time step based on the time-domain and frequency-domain attenuation weights, thereby forming an adaptive filtering scheme based on the anomalies at each time step. This solves the problem that existing filtering techniques cannot adjust in real time according to the specific characteristics of satellite signals, resulting in poor filtering effects and failing to effectively improve signal quality, thus affecting the effectiveness and reliability of communication.

[0116] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A satellite communication filtering method based on a handheld terminal, characterized in that, Includes the following steps: S1. Perform EMD decomposition on the satellite communication signal collected by the handheld terminal to obtain multiple intrinsic mode functions and residual signals, and obtain the energy signal for each intrinsic mode function; S2. Detect the neighborhood ratio of the energy signal and calculate the first time-domain attenuation component; S3. Perform range impact detection on the energy signal, calculate the second time-domain attenuation component, and combine it with the first time-domain attenuation component to obtain the time-domain attenuation weight signal. S4. Based on the difference between the total out-of-band energy and the total in-band energy of the energy signal, calculate the frequency domain attenuation weight to obtain the frequency domain attenuation weight signal. S5. Multiple intrinsic mode functions are processed using a trained lightweight neural network. Attention is applied based on the time-domain attenuation weight signal and the frequency-domain attenuation weight signal respectively to obtain the reconstructed signal. S6. Add the reconstructed signal to the residual signal to obtain the filtered signal.

2. The satellite communication filtering method based on a handheld terminal according to claim 1, characterized in that, S2 includes the following steps: S21. Using the energy at each moment in the energy signal as the reference energy, obtain the energy within the neighborhood of the reference energy; S22. Take the average value of each energy within the neighborhood of the reference energy as the neighborhood mean. S23. Calculate the first time-domain decay component based on the ratio of the reference energy to the mean of the corresponding neighborhood.

3. The satellite communication filtering method based on a handheld terminal according to claim 2, characterized in that, The process of S23 includes: taking the ratio of the reference energy to the mean of the corresponding neighborhood; when the ratio is less than or equal to the ratio threshold, setting the corresponding first time-domain decay component to 1; when the ratio is greater than the ratio threshold, subtracting the ratio threshold from the ratio threshold to obtain the difference; using the natural constant e as the base and the negative of the square of the difference as the exponent, obtaining the first time-domain decay component in the range of 0 to 1.

4. The satellite communication filtering method based on a handheld terminal according to claim 1, characterized in that, S3 includes the following steps: S31. Using the energy at each moment in the energy signal as the reference energy, obtain the energy within the neighborhood of the reference energy; S32. Take the average value of each energy in the neighborhood of the reference energy as the neighborhood mean, and extract the maximum and minimum energies in the neighborhood of the reference energy. S33. Subtract the maximum and minimum energies in the neighborhood to obtain the energy range. Use the ratio of the energy range to the neighborhood mean as the range impact value. S34. Calculate the second time-domain attenuation component based on the extreme impact value; S35. Take the average of the first time-domain attenuation component and the second time-domain attenuation component to obtain the time-domain attenuation weight signal.

5. The satellite communication filtering method based on a handheld terminal according to claim 4, characterized in that, The process of S34 includes: when the range impact value is less than or equal to the range impact threshold, setting the corresponding second time-domain attenuation component to 1; when the range impact value is greater than the range impact threshold, subtracting the range impact value from the range impact threshold to obtain the difference; using the natural constant e as the base and the negative of the square of the difference as the exponent, obtaining the second time-domain attenuation component in the range of 0 to 1.

6. The satellite communication filtering method based on a handheld terminal according to claim 1, characterized in that, S4 includes the following steps: S41. Perform a short-time Fourier transform on each energy signal to obtain the global spectrum at each time step; S42. Extract the spectrum of the effective signal frequency band from the global spectrum at each time moment. The remaining spectrum is the out-of-band spectrum. S43. Add up all the amplitudes in the out-of-band spectrum to get the total out-of-band energy; S44. Sum all the amplitude values ​​in the spectrum of the effective signal band to obtain the total energy within the band; S45. Based on the difference between the total energy in the band and the total energy out of the band, calculate the frequency domain attenuation weight to obtain the frequency domain attenuation weight signal.

7. The satellite communication filtering method based on a handheld terminal according to claim 6, characterized in that, The process of S45 includes: when the total energy in the band is less than the noise floor energy, the corresponding frequency domain attenuation weight is reset to 0; when the total energy in the band is greater than or equal to the noise floor energy, it is determined whether the total energy in the band is greater than or equal to the total energy outside the band. If so, the corresponding frequency domain attenuation weight is reset to 1; if not, the total energy outside the band is subtracted from the total energy in the band to obtain the difference. Using the natural constant e as the base and the negative of the square of the difference as the exponent, the frequency domain attenuation weight in the range of 0 to 1 is obtained. The frequency domain attenuation weights at each time point belonging to the same intrinsic mode function are used to form the frequency domain attenuation weight signal.

8. The satellite communication filtering method based on a handheld terminal according to claim 1, characterized in that, The lightweight neural network in S5 includes: multiple dual-domain weight adaptive modulation modules, adder A1, adder A2, a first temporal convolutional layer, a second temporal convolutional layer, and a weighting module; Each dual-domain weighted adaptive modulation module is used to input an intrinsic mode function, as well as the corresponding time-domain attenuation weight signal and frequency-domain attenuation weight signal; The input of adder A1 is connected to the first output of multiple dual-domain weight adaptive modulation modules; the input of adder A2 is connected to the second output of multiple dual-domain weight adaptive modulation modules; the input of the first temporal convolutional layer is connected to the output of adder A1; the input of the second temporal convolutional layer is connected to the output of adder A2; the first input of the weighting module is connected to the output of the first temporal convolutional layer, its second input is connected to the output of the second temporal convolutional layer, and its output serves as the output of the lightweight neural network.

9. The satellite communication filtering method based on a handheld terminal according to claim 8, characterized in that, S5 includes the following steps: S51. Each dual-domain weighted adaptive modulation module generates attention to the time-domain attenuation weight signal and the frequency-domain attenuation weight signal respectively. The two attentions are applied to the features of the corresponding intrinsic mode function to obtain the time-domain weighted correction feature and the frequency-domain weighted correction feature. S52. Add all time-domain weighted correction features element-wise using adder A1 to obtain time-domain weighted correction fusion features. Add all frequency-domain weighted correction features element-wise using adder A2 to obtain frequency-domain weighted correction fusion features. S53. The first temporal convolutional layer is used to extract features from the temporal weight correction and fusion features to obtain the temporal weight correction and fusion deep features. The second temporal convolutional layer is used to extract features from the frequency domain weight correction and fusion features to obtain the frequency domain weight correction and fusion deep features. S54. The deep features of time-domain weighted correction fusion and frequency-domain weighted correction fusion are weighted using a weighting module to obtain the reconstructed signal.

10. The satellite communication filtering method based on a handheld terminal according to claim 8 or 9, characterized in that, The dual-domain weight adaptive modulation module includes: a first pointwise convolutional layer, a second pointwise convolutional layer, a third pointwise convolutional layer, a first sigmoid layer, a second sigmoid layer, a multiplier M1, and a multiplier M2; The input of the first pointwise convolutional layer is used to input the temporal decay weight signal, and its output is connected to the input of the first sigmoid layer. The input of the second pointwise convolutional layer is used to input the intrinsic mode function, and its output is connected to the first input of multiplier M1 and the first input of multiplier M2, respectively. The input of the third pointwise convolutional layer is used to input the frequency domain attenuation weight signal, and its output is connected to the input of the second sigmoid layer. The second input of multiplier M1 is connected to the output of the first sigmoid layer, and its output serves as the first output of the dual-domain weight adaptive modulation module, used to output the temporal weight correction feature. The second input of multiplier M2 is connected to the output of the second sigmoid layer, and its output serves as the second output of the dual-domain weighted adaptive modulation module, used to output the frequency domain weighted correction feature.

Images