Mesoscale time shortwave signal feature analysis and processing method based on multi-modal fusion

Abstract

The invention discloses a mesoscale time shortwave signal feature analysis and processing method based on multi-modal fusion, and belongs to the technical field of wireless communication signal processing. The mesoscale time short-wave signal feature analysis and processing method based on the multi-modal fusion solves the problems that in an existing mesoscale time short-wave signal feature extraction method, dynamic property exists, signal and time coupling is enhanced, and feature instability is caused by the dynamic property. Comprising the steps of extracting time-domain and frequency-domain statistical features, extracting time-frequency distribution diagram features and extracting one-dimensional time sequence dynamic features to form a multi-level feature analysis framework, and fusing the extracted multiple modal features by utilizing complementarity among the time-domain and frequency-domain statistical features, the time-frequency distribution diagram features and the one-dimensional time sequence dynamic features to form a multi-level feature analysis framework. The description capability of signal features is effectively improved, and the accuracy of modulation identification of the mesoscale time shortwave signal is improved at the same time.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless communication signal processing, and specifically to a method for analyzing and processing medium-scale time short-wave signal characteristics based on multimodal fusion. Background Art

[0002] Signal modulation recognition (MR) is a key link in wireless communication systems. It identifies the modulation mode of received signals with limited or completely no prior knowledge. Whether as the primary task in electronic reconnaissance or to achieve link adaptation in civilian communication, modulation recognition provides fundamental support for subsequent signal processing.

[0003] Short-wave communication has received extensive attention due to its unique propagation characteristics. However, due to the complex channel environment brought about by ionospheric reflection, short-wave signals exhibit characteristics such as multipath, fading, and Doppler frequency shift, making it more difficult to extract their features. Traditional short-wave signal feature extraction methods mainly rely on signal processing and statistical analysis techniques. These methods extract signal features from different angles such as the time domain and frequency domain, providing a basis for signal analysis and recognition. However, these methods still have the following problems for the feature extraction of medium-scale time short-wave signals:

[0004] (1) The channel characteristics of medium-scale time short-wave signals introduce time-varying characteristics during dynamic changes, randomly changing the channel conditions. The dynamic characteristics of drastic amplitude changes and spectrum broadening affect their stability, resulting in an increase in signal uncertainty.

[0005] (2) The time-varying characteristics of medium-scale time short-wave channels cause signal characteristics to be interdependent between time segments, resulting in an enhanced coupling between the signal and time.

[0006] (3) Due to the weakening of the coupling between modalities of medium-scale time short-wave signals, the differences in the performance of different modal characteristics under channel changes increase, making it difficult to fuse information between modalities; and due to randomly changing the signal amplitude, phase, etc., resulting in the superposition of channel fading and noise, it is difficult for a single modal characteristic to stably represent the signal. Summary of the Invention

[0007] Aiming at the shortcomings and deficiencies in the prior art, the present invention provides a method for analyzing and processing medium-scale time short-wave signal characteristics based on multimodal fusion, which can effectively capture the dynamic changes of signals in the time series, accurately describe and compensate for the dynamic characteristics of signals, significantly enhance the stability and robustness of modal characteristics, and significantly improve the accuracy of signal modulation recognition.

[0008] To achieve the above objectives, the present invention is realized through the following technical solutions: The method for analyzing and processing medium-scale time shortwave signal features based on multimodal fusion provided by the present invention comprehensively represents the static attributes and dynamic change characteristics of the signal, and mines the multimodal features of the signal at multiple levels. Specifically, it includes extracting time-domain and frequency-domain statistical features, deep learning features of time-frequency distribution maps, and one-dimensional time-series dynamic features to form a multi-level feature analysis framework;

[0009] And by utilizing the complementarity of the above different modal features, the extracted multimodal features are fused to enhance the ability to describe signal features, improve the accuracy of modulation recognition of medium-scale time shortwave signals, and improve the performance of the recognition system under complex channel conditions.

[0010] Preferably, it specifically includes the following steps:

[0011] S1. Generation of the dataset: First, simulate modulation signals, including amplitude modulation signals, frequency modulation signals, phase modulation signals, and quadrature amplitude modulation signals, and obtain shortwave signals therefrom; use the lognormal distribution to describe the channel changes in the medium and long term, introduce the channel change models in the medium and long term, and generate a medium-scale time shortwave signal dataset;

[0012] S2. Multimodal feature extraction: Extract time-domain and frequency-domain statistical features, extract time-frequency distribution map features, and extract one-dimensional time-series dynamic features, specifically as follows:

[0013] The time-domain and frequency-domain statistical features are descriptions of the overall distribution of the signal, used to distinguish the average amplitude and spectral distribution of the above different modulation methods; the time-frequency distribution map features effectively capture modulation signals with sudden changes or local features through joint representation in time and frequency; the one-dimensional time-series dynamic features are used to capture the evolution patterns and dependencies of the signal in time, effectively distinguish the above different phase modulation methods, and learn the interaction between amplitude and phase changes in quadrature modulation signals;

[0014] Using the medium-scale time shortwave signal dataset generated in step S1, the multi-dimensional characteristics of the medium-scale time shortwave signal are extracted by the above multiple modalities to form a multi-level feature analysis framework, comprehensively representing the static attributes and dynamic change characteristics of the shortwave signal;

[0015] S3. Multimodal feature fusion: Utilize the complementarity between the time-domain and frequency-domain statistical features, time-frequency distribution map features, and one-dimensional time-series dynamic features to fuse the multimodal features extracted in step S2, and enhance the ability to describe signal features and the accuracy of modulation recognition of medium-scale time shortwave signals; the specific operation steps are as follows:

[0016] The three modal features of time-domain frequency-domain statistical features, time-frequency distribution map features, and one-dimensional time-series dynamic features complement each other according to their respective characteristics, jointly constituting a complete representation of the modulation methods of amplitude modulation signals, frequency modulation signals, phase modulation signals, and quadrature amplitude modulation signals, effectively improving the accuracy of signal modulation recognition.

[0017] Preferably, in step S1, the amplitude modulation signals include 2ASK, 4ASK, 8ASK; the frequency modulation signals include 2FSK, 4FSK; the phase modulation signals include BPSK, QPSK; the quadrature amplitude modulation signals include 16QAM;

[0018] The medium-term and long-term channel change models are introduced. Among them, the mesoscale effect is realized by means of the LTV and ITV models. According to the measured data, its time constant and standard deviation are obtained, and are achieved through the combination of an Alpha filter and a high-pass filter to obtain the signal-to-noise ratio SNR sequence. Select N data points from the generated SNR sequence, sample the N SNR points at intervals of 1 second, and combine with the Watterson model to make each SNR value data point correspond to a specific time period, ensuring that the signal is affected by different SNRs at different time points, so as to simulate the real channel environment, and finally generate a complete data set containing N-second mesoscale time short-wave signals.

[0019] Preferably, it includes step S2.1 of extracting time-domain frequency-domain statistical features, including extracting time-domain statistical features and extracting frequency-domain statistical features;

[0020] The time-domain statistical features include the mean value, skewness, and kurtosis factor; the frequency-domain statistical features include the mean value of the frequency-domain amplitude, the center frequency, and the frequency standard deviation;

[0021] Normalize the above six time-domain frequency-domain statistical features extracted to make the multi-modal features have better compatibility and consistency during the fusion process. The specific calculation formula is:

[0022]

[0023] Among them, x norm is the normalized standard feature; x is the original feature; μ is the mean value of each feature; α is the standard deviation of each feature.

[0024] Preferably, (1) The mean value is an index of the overall DC offset or the signal base value of the signal, which can distinguish whether there is a significant offset in the DC component of the signal and reflect the overall DC offset degree of the modulation signal:

[0025] By using different modulation depths for different signals, resulting in the changing trend of the average value, 2ASK, 4ASK, 8ASK, 2FSK, and 16QAM can be clearly distinguished. Specifically, for the amplitude modulation signals of 2ASK, 4ASK, and 8ASK, the amplitude size determines the change in their average value. In the 2FSK frequency modulation signal, different frequency components cause the change in the average value. The average value of the 16QAM quadrature amplitude modulation signal reflects the distribution characteristics of the constellation diagram points;

[0026] The above average value is used to distinguish different signal modulation methods. The specific calculation formula is as follows:

[0027]

[0028] In the formula, S i is the amplitude of the signal sampling data points; N is the number of signal sampling data points for each sample;

[0029] (2) Skewness is used to measure the asymmetry degree of the signal waveform. For frequency modulation signals, skewness is used to describe whether there is an asymmetric amplitude distribution feature deviating from the mean value in the signal. Because during the switching process between different frequencies, the spectrum shows slight asymmetry, the 2FSK frequency modulation signal can be clearly distinguished. Therefore, skewness is used to distinguish frequency modulation signals in shortwave signals. The specific calculation formula is as follows:

[0030]

[0031] In the formula, S i is the amplitude of the signal sampling data points; N is the number of signal sampling data points for each sample; ρ t is the standard deviation, and the formula is as follows:

[0032]

[0033] ρ t 3 is the cube of the standard deviation;

[0034] (3) Kurtosis factor describes the sharpness of the signal amplitude and is an important feature in time-domain analysis. It is very effective especially for detecting impact signals, pulse signals, and high-peak value modulation signals;

[0035] When there are sudden amplitude changes in the amplitude modulation signal and rapid frequency jumps between different frequencies in the frequency modulation signal, the kurtosis factor will increase significantly. Signals with high kurtosis contain more spikes or high-energy instantaneous components. It can clearly distinguish 2ASK, 4ASK, 2FSK, and 4FSK as follows: Sudden amplitude changes in 2ASK and 4ASK amplitude modulation signals cause an increase in kurtosis; Rapid frequency jump signals between different frequencies in 2FSK and 4FSK will increase kurtosis;

[0036] The specific calculation formula is as follows:

[0037]

[0038] Wherein, S i is the amplitude of the sampled data points of the signal; N is the number of sampled data points of each sample.

[0039] Preferably, (4) The average frequency-domain amplitude reflects the central tendency of the signal spectrum amplitude distribution, describes the degree of energy concentration in the spectrum, and is used to distinguish different modulation methods;

[0040] The average frequency-domain amplitude can clearly distinguish 2ASK, 4ASK, 8ASK, 2FSK, 4FSK, BPSK, QPSK, 16QAM, specifically as follows: The spectrum amplitude distributions of 2ASK, 4ASK, 8ASK amplitude modulation signals have significant characteristics; The spectrum energy of 2FSK, 4FSK keying signals is concentrated at specific frequency positions; The energy distributions of BPSK, QPSK, 16QAM phase modulation or composite modulation signals are different in the frequency domain. Therefore, the average frequency-domain amplitude can be used as an identification basis;

[0041] The specific calculation formula is:

[0042]

[0043] Wherein, k is the spatial frequency of the wave; s(k) represents the frequency-domain amplitude at the k-th frequency point;

[0044] (5) The centroid frequency describes the dominant frequency position of the signal spectrum and is an important indicator for measuring the spectrum distribution; The centroid frequency positions of different signals have significant differences and can clearly distinguish 2FSK, 4FSK, BPSK, QPSK, specifically as follows: The different frequency components of 2FSK, 4FSK frequency keying signals determine the position of their centroid frequency; The spectrum distribution center of BPSK, QPSK phase modulation signals is closely related to their modulation methods. The specific calculation formula is:

[0045]

[0046] Wherein, k is the spatial frequency; s(k) represents the frequency-domain amplitude at the k-th frequency point; f k is the frequency value of the k-th frequency point;

[0047] (6) The frequency standard deviation describes the degree of dispersion or concentration of the spectrum and is one of the key features for distinguishing signals of different modulation methods. It can clearly distinguish 2FSK, 4FSK, BPSK, QPSK, specifically as follows: The jump signals of different frequencies of 2FSK, 4FSK will cause changes in the frequency standard deviation; The bandwidth of BPSK, QPSK modulation signals will directly affect the frequency standard deviation. The specific calculation formula is:

[0048]

[0049] where k is the spatial frequency; S i is the amplitude of the sampled data points of the signal; N is the number of sampled data points per sample; s(k) represents the frequency-domain amplitude at the k-th frequency point; f k is the frequency value at the k-th frequency point; S 2 is the center frequency at the k-th point.

[0050] Preferably, it further includes step S2.2 of extracting time-frequency distribution map features, specifically using the ResNet152 network to extract the features of two time-frequency distribution maps, namely the time-frequency distribution features of SPWVD (Smoothing Pseudo Wigner-Ville Distribution) and the time-frequency distribution features of BJD (Born-Jordan Distribution);

[0051] Among them, the SPWVD time-frequency distribution features capture local features through high resolution, which are used to observe the changes of the signal within a specific time range and improve the recognition ability of signal features;

[0052] The specific calculation formula is:

[0053] SPWVD x (t,f) = ∫∫x(t - v + τ / 2)x * (t - v - τ / 2)·h(τ)g(v)e -j2πfτ dvdτ

[0054] In the formula, SPWVD x (t,f) is the time-frequency result in the SPWVD time-frequency distribution features, τ is the frequency-domain window length; v is the time-domain window length; h(τ) is the frequency-domain smoothing window; g(v) is the time-domain smoothing window; x * represents the complex conjugate of x; x(t) is the analytic signal of r(t), which is expressed as:

[0055] x(t) = r(t) + jH[r(t)]

[0056] In the formula, H[·] represents the Hilbert transform; r(t) is the received signal;

[0057] The BJD time-frequency distribution features are used to provide a global time-frequency description of the signal, which can accurately concentrate the signal energy around its true time and frequency positions, making the spectral characteristics of the signal more obvious; the specific calculation formula is:

[0058]

[0059] In the formula, BJD x(t, f) is the time-frequency result in the time-frequency distribution characteristics of BJD; τ is the frequency-domain window length, and v is the time-domain window length; x * represents the complex conjugate of x;

[0060] where φ(t, τ) is:

[0061] Preferably, in step S2.2, it further includes fusing the 2048-dimensional features of the BJD time-frequency distribution characteristics and the SPWVD time-frequency distribution characteristics by using the Jensen–Shannon divergence (JS divergence), specifically as follows:

[0062] First, convert the features obtained from SPWVD and BJD into probability distributions p and q using the softmax function, and then calculate the JS divergence between them, which can capture details while not losing the overall signal; the calculation formula is as follows:

[0063]

[0064] In the formula, KL(P||Q) is the Kullback-Leibler divergence. Assuming two probability distributions are P and Q, and on the premise that they are set as continuous random variables, their corresponding probability density functions are p(x) and q(x) respectively. p is the probability distribution obtained by converting the features from SPWVD using the softmax function; q is the probability distribution obtained by converting the features from BJD using the softmax function;

[0065] The JS divergence represents the similarity between the two time-frequency distribution characteristics. The larger this divergence value, the less similar the two time-frequency distribution characteristics are, and the smaller the weight assigned; the smaller the divergence value, the more similar the two time-frequency distribution characteristics are, and the greater the weight assigned; each short-wave signal sample is assigned different weights according to the magnitude of the divergence between the multi-modal features, and the above multi-modal features are weighted and averaged according to the corresponding weights to obtain a fused high-dimensional feature vector.

[0066] Preferably, it includes step S2.3 to extract one-dimensional time-series dynamic features, specifically including using a Gated Recurrent Unit (GRU) to capture the dynamic changes and trends in the one-dimensional time-series signal, and obtain the features of the signal, enabling it to better selectively remember which information should be retained and which information should be forgotten. This helps to avoid the common problems of gradient disappearance or gradient explosion in the Recurrent Neural Network (RNN) and is used to process dynamic signals with long-range dependencies. The Gated Recurrent Unit GRU controls the flow of information by introducing an update gate and a reset gate, and the specific calculation formula is as follows:

[0067] z t = σ(W z ·[h t-1 , x t )

[0068] r t = σ(W r ·[h t-1 , x t )

[0069]

[0070] In the formula, x t is the input information at the current moment; h t-1 is the hidden state at the previous moment; h t is the hidden state passed to the next moment; is the candidate hidden state; r t is the reset gate, z t is the update gate, σ is the sigmoid function, and through this function, the data is changed to a value in the range of 0 - 1. The tanh function can change the data to a value in the range of [-1, 1].

[0071] Preferably, in step S3, the feature concatenation and fully - connected layer processing is to concatenate the time - domain frequency - domain statistical features, time - frequency distribution map features, and one - dimensional time - series features by dimension to form a high - dimensional feature vector. During the concatenation process, all features can be processed in a unified feature space;

[0072] The concatenated feature vector is processed by the fully - connected layer. The fully - connected layer uses a weight matrix to perform linear transformation and non - linear activation on the input features, and uses ReLU as the activation function. By calculating, each dimension of the input features is combined and re - encoded to learn the non - linear relationship between different features, thereby improving the performance of the recognition system under complex channel conditions.

[0073] The method for mesoscale short - wave signal feature analysis and processing based on multi - modal fusion provided by the present invention has the following beneficial effects:

[0074] (1) The method for analyzing and processing mesoscale shortwave signal features based on multimodal fusion of the present invention extracts time-frequency distribution features, time-domain and frequency-domain statistical features, and one-dimensional time-series dynamic features, mines the multi-dimensional characteristics of the signal from multiple levels, comprehensively characterizes the static attributes and dynamic change characteristics of the signal, and solves the difficult problems of dynamics, enhanced signal-time coupling, and feature instability existing in mesoscale time shortwave signals. And by making full use of the complementarity between time-frequency distribution features, time-domain and frequency-domain statistical features, and one-dimensional time-series dynamic features, multimodal fusion is carried out, so that it simultaneously has the high resolution of the time-frequency distribution diagram, the time-series modeling ability of the gated recurrent unit, and the robustness of statistical features, forming a multi-dimensional and multi-perspective signal analysis model, improving the overall performance of the signal, enhancing feature stability and robustness, and effectively improving the recognition accuracy of signal modulation while improving the performance of the recognition system under complex channel conditions.

[0075] (2) By extracting the mean value, skewness, kurtosis factor, average frequency-domain amplitude, and center frequency in the time-domain and frequency-domain statistical features, a comprehensive characterization of the mesoscale shortwave signal in the time domain and frequency domain is provided, which can better understand the changes of the signal in a dynamic environment; and they are normalized to make different modal features have better compatibility and consistency in the fusion process, improving the effect of feature fusion.

[0076] (3) Extract time-frequency distribution diagram features. The BJD time-frequency distribution feature provides a global time-frequency description of the signal, and the SPWVD time-frequency distribution feature can capture local features through high resolution. The two are fused through the JS divergence to further optimize the clarity of the time-frequency representation. The two cooperate with each other and work together to more comprehensively analyze the dynamic behavior of the signal, capturing details while not losing the overall signal. And a neural network is used and the time-domain and frequency-domain statistical features are fused to dynamically adjust the signal processing strategy to adapt to changing channel conditions, effectively solving the instability problem caused by dynamic characteristics.

[0077] (4) When extracting one-dimensional time-series features, the gated recurrent unit GRU is used to capture the dynamic changes and trends in the one-dimensional time-series signal. GRU is a recurrent neural network suitable for time-series data, which can capture the dynamic changes and trends in the one-dimensional time-series signal, improving the accuracy and reliability of feature extraction to process dynamic signals with long-range dependencies. By introducing an update gate and a reset gate to control the flow of information, it can efficiently and selectively retain or forget information, helping to avoid the common problems of gradient disappearance or gradient explosion in the recurrent neural network RNN, and thus better processing long-sequence data. Description of the Drawings

[0078] Figure 1 is the flowchart of the method in Embodiment 1 of the present invention;

[0079] Figure 2 It is the flowchart of the signal-to-noise ratio (SNR) sequence within the mesoscale time in Embodiment 2;

[0080] Figure 3 It is the simulation diagram of the signal-to-noise ratio (SNR) sequence within the mesoscale time in Embodiment 2;

[0081] Figure 4 It is the comparative simulation diagram of the short-time short-wave signal and the mesoscale short-wave signal in Embodiment 2

[0082] Figure 5 It is the simulation diagram of the average value in Embodiment 2;

[0083] Figure 6 It is the simulation diagram of skewness in Embodiment 2;

[0084] Figure 7 It is the simulation diagram of kurtosis factor in Embodiment 2;

[0085] Figure 8 It is the simulation diagram of the average value of the frequency domain amplitude in Embodiment 2;

[0086] Figure 9 It is the simulation diagram of the center frequency in Embodiment 2;

[0087] Figure 10 It is the simulation diagram of the frequency standard deviation in Embodiment 2;

[0088] Figure 11 It is the structural diagram of the multi-modal fusion network model in Embodiment 2;

[0089] Figure 12 It is the comparison diagram of the accuracy curve and the loss curve of the time-frequency distribution map features and the statistical features after multi-modal fusion in Embodiment 3;

[0090] Figure 13 It is the comparison diagram of the accuracy curve and the loss curve of the one-dimensional time series features and the statistical features after multi-modal fusion in Embodiment 3;

[0091] Figure 14 It is the schematic diagram of the confusion matrix after the fusion of three modalities in Embodiment 3. Detailed implementation manners

[0092] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

[0093] Embodiment 1

[0094] The method for analyzing and processing medium-scale time short-wave signal features based on multimodal fusion provided by the present invention includes extracting time-domain and frequency-domain statistical features, extracting time-frequency distribution map features, and extracting one-dimensional time-series dynamic features, forming a multi-level feature analysis framework, and fusing the above-mentioned extracted multiple modal features by using the complementarity between different modal features to improve the ability to describe signal features and the accuracy of modulating and identifying medium-scale time short-wave signals.

[0095] As Figure 1 shown, the present invention conducts feature analysis on medium-scale time short-wave signals and finds that the following difficult problems mainly exist in the current feature extraction of medium-scale time short-wave signals: First, the dynamic change of channel characteristics in medium-scale time short-wave signals introduces time-varying characteristics, randomly changes the channel conditions, and causes the dynamic characteristics of drastic amplitude change and spectrum broadening, resulting in the influence on its stability and increasing the uncertainty of the signal. Second, the time-varying characteristics of medium-scale time short-wave channels make the signal characteristics interdependent between time segments, leading to an enhanced coupling between the signal and time, and the traditional assumption that time segments are independent is no longer applicable, which brings challenges to signal processing and feature extraction. Third, due to the weakening of the coupling between modes in medium-scale time short-wave signals, the differences in the performance of different modal features under channel changes increase, making it difficult to fuse the information between modes, and the channel fading and noise are superimposed due to randomly changing the signal amplitude, phase, etc., making it difficult for a single modal feature to stably represent the signal.

[0096] The present invention aims at the above difficult problems. First, the statistical features of the time domain and frequency domain are extracted to provide a comprehensive representation of the signal in the time domain and frequency domain, so as to better understand the changes of the signal in a dynamic environment. The time-frequency distribution features are extracted to accurately describe and compensate for the dynamic characteristics of the signal. At the same time, the signal processing strategy is dynamically adjusted by combining deep learning technology with the attention mechanism and integrating the extracted statistical features of the time domain and frequency domain to adapt to the changing channel conditions. The dynamic behavior of the signal can be analyzed more comprehensively, while capturing the details without losing the overall signal characteristics. It is used to solve the problem of instability caused by dynamic characteristics. Secondly, by extracting one-dimensional time series features and using time series modeling technology, the dynamic changes and trends in one-dimensional time series signals can be captured, and the characteristics of the signal can be obtained to process dynamic signals with long-range dependencies, thereby improving the accuracy and reliability of feature extraction, and effectively solving the problem of enhanced coupling between signals and time due to time-varying characteristics. Finally, multimodal fusion technology is used to perform multimodal feature fusion by utilizing the complementarity between time-frequency distribution diagram features, time-domain and frequency-domain statistical features, and one-dimensional time series dynamic features. The multimodal model not only combines the high-resolution characteristics of the time-frequency distribution diagram and the time series modeling capabilities of the gated loop, but also utilizes the robustness of statistical features to construct a multi-dimensional, multi-perspective signal analysis to improve the overall signal performance and enhance feature stability and robustness, which is used to solve the problem of feature instability caused by dynamic channels.

[0097] Example 2

[0098] like Figures 2 - 11 As shown, the method for analyzing and processing characteristics of mesoscale time shortwave signals based on multimodal fusion of the present invention specifically includes the following steps:

[0099] S1. Dataset generation: Perform modulation signal simulation, including amplitude modulation signals 2ASK, 4ASK, 8ASK, frequency modulation signals 2FSK, 4FSK, phase modulation signals BPSK, QPSK and orthogonal amplitude modulation signals 16QAM, etc., to obtain shortwave signals;

[0100] like Figure 2 As shown, on the basis of the above shortwave signals, the dynamic change characteristics of the ionospheric channel are fully considered, the log-normal distribution is used to describe the medium-term and long-term channel changes, and the medium-term and long-term channel change models are introduced. Among them, the mesoscale effect is achieved with the help of LTV and ITV models. According to the measured data: the time constants are 160 seconds and 6.5 seconds, the standard deviation of LTV is 3.493dB, and the standard deviation of ITV is 3.889dB. It is realized through a combination of Alpha filter and high-pass filter, and then the signal-to-noise ratio SNR sequence is obtained.

[0101] like Figure 3As shown, 160 data points are selected from the generated SNR sequence, and 160 SNR points are sampled at 1-second intervals. Combining with the Watterson model, each data point value in the signal-to-noise ratio SNR corresponds to a specific time period, ensuring that the signal is affected by different data point values at different time points, so as to simulate the real channel environment. Finally, a mesoscale time shortwave signal dataset with a duration of 160 seconds is generated.

[0102] As Figure 4 shown, compared with the output signal spectrum of the Watterson model, the output signal spectrum of the mesoscale channel model generated in the present invention shows a more obvious spectrum broadening phenomenon, resulting in energy dispersion. Moreover, in terms of amplitude change, the amplitude change of the mesoscale time shortwave signal is more severe than that of the short-time shortwave signal. It truly reflects the change of channel quality caused by various fading within a long time period.

[0103] S2. Multi-modal feature extraction: including extracting time-domain and frequency-domain statistical features, extracting time-frequency distribution map features, and extracting one-dimensional time-series dynamic features; using the above-mentioned multiple modalities to extract the multi-dimensional characteristics of the mesoscale time shortwave signal from the mesoscale time shortwave signal dataset generated in step S1, forming a multi-level feature analysis framework to comprehensively characterize the static attributes and dynamic change characteristics of the shortwave signal.

[0104] S2.1 Extraction of time-domain and frequency-domain statistical features. A total of six statistical features are selected. The time-domain statistical features include the mean value, skewness, and kurtosis factor; the frequency-domain statistical features include the mean value of the frequency-domain amplitude, the center frequency, and the frequency standard deviation.

[0105] As Figure 5 shown, the mean value is an index of the overall DC offset or the signal base value of the signal, which can reflect the overall DC offset degree of the modulation signal. For some baseband signals or amplitude modulation signals, different modulation depths may lead to changes in the mean value. It can be seen from the figure that the amplitude sizes of the amplitude modulation signals such as 2ASK, 4ASK, and 8ASK determine the change of the mean value; the different frequency components of the frequency modulation signals represented by 2FSK and 4FSK will cause changes in the mean value; the mean value of the quadrature amplitude modulation represented by 16QAM reflects the distribution characteristics of the constellation diagram points. Therefore, the mean value can clearly distinguish 2ASK, 4ASK, 8ASK, 2FSK, and 16QAM. The specific calculation formula is:

[0106]

[0107] In the formula, S i is the amplitude of the sampled data points of the signal; N is the number of sampled data points of each sample.

[0108] As Figure 6As shown, skewness can measure the degree of asymmetry of the signal waveform. Due to the switching between different frequencies in the frequency modulation signal, the spectrum shows a slight asymmetry. Skewness is used to describe the characteristic of asymmetric amplitude distribution deviating from the mean in the signal, and to distinguish the frequency modulation signal 2FSK. The specific calculation formula is as follows:

[0109]

[0110] In the formula, S i is the amplitude of the sampled data points of the signal; N is the number of sampled data points for each sample; ρ t is the standard deviation, and the formula is as follows:

[0111]

[0112] ρ t 3 is the cube of the standard deviation;

[0113] As Figure 7 shown, the kurtosis factor can describe the sharpness of the signal amplitude and is an important feature in time-domain analysis. It is especially effective for detecting impact signals, pulse signals, and high-peak modulation signals. Signals with high kurtosis usually contain more spikes or high-energy instantaneous components. When there are sudden amplitude changes in the amplitude modulation signal and rapid frequency jumps between different frequencies in the frequency modulation signal, the kurtosis factor will increase significantly. It can be seen from the figure that the sudden amplitude changes in the amplitude modulation signals such as 2ASK and 4ASK lead to an increase in kurtosis. The rapid frequency jump signals between different frequencies in the frequency modulation signals such as 2FSK and 4FSK will increase the kurtosis. Therefore, the kurtosis factor can clearly distinguish 2ASK, 4ASK, 2FSK, and 4FSK, and the specific calculation formula is as follows:

[0114]

[0115] In the formula, S i is the amplitude of the sampled data points of the signal; N is the number of sampled data points for each sample.

[0116] As Figure 8 shown, the average frequency-domain amplitude reflects the central tendency of the signal spectrum amplitude distribution and can describe the degree of energy concentration in the spectrum. It can be seen from the figure that for 2ASK, 4ASK, and 8ASK: the spectrum amplitude distributions of the amplitude modulation signals have significant characteristics, with a peak of the same amplitude on each side of the carrier peak; the spectrum energy of the frequency modulation signals such as 2FSK and 4FSK is concentrated at specific frequency positions; the energy distributions of the phase modulation or composite modulation signals such as BPSK, QPSK, and 16QAM are different in the frequency domain. Therefore, the average frequency-domain amplitude is used to identify the above different signal modulation methods. The specific calculation formula is as follows:

[0117]

[0118] In the formula, k is the spatial frequency of the wave; s(k) represents the frequency-domain amplitude at the k-th frequency point.

[0119] As Figure 9 shown, the center frequency can describe the dominant frequency position of the signal spectrum and is an important indicator for measuring the spectrum distribution. It can be seen from the figure that the different frequency components in the 2FSK and 4FSK frequency modulation signals determine their center frequency positions; the spectral center distributions of the BPSK and QPSK phase modulation signals are the same as their modulation methods, and the center frequency is used to clearly distinguish the above 2FSK, 4FSK, BPSK, and QPSK signal modulation methods. The specific calculation formula is:

[0120]

[0121] In the formula, k is the spatial frequency; s(k) represents the frequency-domain amplitude at the k-th frequency point; f k is the frequency value of the k-th frequency point;

[0122] As Figure 10 shown, the frequency standard deviation can describe the distribution width state of the spectrum and is one of the key features for distinguishing signals of different modulation methods. It can be seen from the figure that the frequency standard deviations in the 2FSK and 4FSK frequency modulation signals change with the change of the jump signals of different frequencies. The frequency standard deviations of the BPSK and QPSK phase modulation signals change with the change of the frequency band width, and the frequency standard deviation is used to clearly distinguish the above 2FSK, 4FSK, BPSK, and QPSK signal modulation methods. The specific calculation formula is:

[0123]

[0124] In the formula, k is the spatial frequency; S i is the amplitude of the sampled data points of the signal; N is the number of sampled data points of each sample; s(k) represents the frequency-domain amplitude at the k-th frequency point; f k is the frequency value of the k-th frequency point; S 2 is the center frequency of the k-th point.

[0125] Normalize the above six time-domain and frequency-domain statistical features extracted to make different modal features have better compatibility and consistency during the fusion process and improve the effect of feature fusion.

[0126] The specific calculation formula is:

[0127]

[0128] Among them, x normis the normalized standard feature; x is the original feature; μ is the average value of each feature; α is the standard deviation of each feature.

[0129] S2.2 Extract time-frequency distribution features, including SPWVD time-frequency distribution features and BJD time-frequency distribution features; SPWVD time-frequency distribution features can provide good time and frequency resolutions simultaneously, making it excellent in capturing rapid frequency changes of signals. This is particularly important for frequency modulation signals 2FSK and 4FSK because these signals have frequent frequency jumps on the time axis. Moreover, SPWVD is especially suitable for observing signal changes within a specific time range. For amplitude modulation signals 2ASK and 4ASK, there are sudden amplitude changes at certain moments, and SPWVD can clearly display these changes, improving the ability to identify signal features; the formula for improving the ability to identify signal features is:

[0130] SPWVD x (t,f) = ∫∫x(t - v + τ / 2)x * (t - v - τ / 2)·h(τ)g(v)e -j2 π fτ dvdτ

[0131] In the formula, SPWVD x (t,f) is the time-frequency result in the SPWVD time-frequency distribution feature, τ is the frequency domain window length; v is the time domain window length; h(τ) is the frequency domain smoothing window; g(v) is the time domain smoothing window; x * represents the complex conjugate of x; x(t) is the analytic signal of r(t), expressed as:

[0132] x(t) = r(t) + jH[r(t)]

[0133] In the formula, H[·] represents the Hilbert transform; r(t) is the received signal;

[0134] BJD time-frequency distribution features can better concentrate the energy of the signal around its true time and frequency positions, making the spectral characteristics of the signal more obvious. This is particularly important for phase modulation signals BPSK and QPSK because the spectral characteristics of these signals are relatively complex. Moreover, BJD can more accurately reflect the spectral characteristics and phase changes of phase modulation signals, helping to better understand the modulation method of the signal. For example, in a QPSK signal, BJD can effectively display the distribution of its four phase states in the time-frequency domain, which is beneficial for signal discrimination in the actual demodulation process. For complex modulation methods such as 16QAM, BJD can provide a clearer representation of signal features, used to provide a global time-frequency description of the signal, and can accurately concentrate the signal energy around its true time and frequency positions, making the spectral characteristics of the signal more obvious. The specific calculation formula is:

[0135] BJD x (t,f) = ∫∫x(v + τ / 2)x * (v - τ / 2)·φ(t - v,τ)e -j2πfτ dvdτ

[0136] In the formula, BJD x (t,f) is the time-frequency result in the BJD time-frequency distribution feature; τ is the frequency-domain window length, v is the time-domain window length; x * represents the complex conjugate of x;

[0137] where φ(t,τ) is:

[0138] By combining the two time-frequency distribution features of SPWVD and BJD, the features of different modulation methods can be analyzed more comprehensively and accurately, providing strong support for signal recognition. The fusion of the BJD time-frequency distribution feature and the SPWVD time-frequency distribution feature through the JS divergence can further optimize the clarity of the time-frequency representation, making the extraction of signal features more refined. BJD provides a global time-frequency description of the signal, while SPWVD captures local features with high resolution. The combination of the two can analyze the dynamic behavior of the signal more comprehensively, capturing details while not losing the overall signal. The calculation formula is as follows:

[0139]

[0140] KL(P||Q) is the KL divergence (Kullback-Leibler divergence). We set two probability distributions as P and Q respectively. On the premise of setting them as continuous random variables, their corresponding probability density functions are p(x) and q(x) respectively. p is the probability distribution obtained by converting the features obtained by SPWVD using the softmax function; q is the probability distribution obtained by converting the features obtained by BJD using the softmax function;

[0141] The JS divergence represents the similarity of the two time-frequency distribution features. The larger the divergence value, the less similar the two time-frequency distribution features are, so the smaller the weight assigned; the smaller the divergence value, the more similar the two time-frequency distribution features are, so the larger the weight assigned. Each short-wave signal sample is assigned a different weight according to the size of the divergence between its different modal features, and the above-mentioned modal features are weighted and averaged according to the corresponding weights to obtain a fused 2048-dimensional high-dimensional feature vector.

[0142] S2.3 Extract one-dimensional time-series dynamic features, specifically including using a gated recurrent unit (GRU) with 128 hidden units in the GRU layer to capture the dynamic changes and trends in one-dimensional time-series signals, obtain the features of the signals, and process dynamic signals with long-range dependencies. The gated recurrent unit GRU controls the flow of information by introducing an update gate and a reset gate, enabling it to better selectively remember which information should be retained and which should be forgotten, helping to avoid the common problems of gradient vanishing or gradient explosion in recurrent neural networks (RNNs), and thus better process long-sequence data. The specific calculation formula is as follows:

[0143] z t =σ(W z ·[h t-1 ,x t )

[0144] r t =σ(W r ·[h t-1 ,x t )

[0145]

[0146] In the formula, x t is the input information at the current moment; h t-1 is the hidden state at the previous moment; h t is the hidden state passed to the next moment; is the candidate hidden state; r t is the reset gate, z t is the update gate, and σ is the sigmoid function. Through this function, the data can be transformed into values in the range of 0 - 1; the tanh function can transform the data into values in the range of [-1, 1].

[0147] For amplitude modulation signals such as 2ASK, 4ASK, and 8ASK, the amplitude change is their feature, and these amplitude changes occur over time. The GRU learns the patterns and rules of these amplitude changes, captures the dynamic changes in the signal amplitude, especially when the signal amplitude is affected by channel fading and other factors, the GRU can better identify these dynamic changes. Moreover, through the gating mechanism, the GRU remembers the start and end times of the signal amplitude changes, as well as the duration of different amplitude states, which are all key features for identifying amplitude modulation signals.

[0148] For frequency modulation signals such as 2FSK and 4FSK, the frequency hopping is their feature. The GRU learns the moments of frequency hopping, the relationships between frequencies, and the duration of each frequency state. And the GRU can identify the changes in the signal frequency at different time points, thereby determining the frequency switching mode of the modulated signal and helping to distinguish the modulation methods of different frequency modulation signals;

[0149] For phase modulation signals BPSK and QPSK, phase jumps are their characteristics. The GRU learns the moments of phase jumps, the relationships between phases, and the durations of phase states. Moreover, the GRU captures the patterns of signal phase changes, such as the amplitudes of phase mutations, the intervals between jumps, etc., which helps to distinguish the modulation methods of different phase modulation signals.

[0150] For the quadrature modulation signal 16QAM, both the amplitude and phase of the signal change, and there are certain relationships between these changes. The GRU learns the interactions and dynamic relationships between these amplitude and phase changes. Therefore, the GRU can capture the complex amplitude and phase change patterns in the 16QAM signal, thereby distinguishing different modulation states.

[0151] S3. Multimodal feature fusion: Utilize the complementarity between time-domain and frequency-domain statistical features, time-frequency distribution map features, and one-dimensional time-series dynamic features to fuse the multimodal features extracted in step S2. Time-domain and frequency-domain statistical features are descriptions of the overall distribution of the signal, used to distinguish the average amplitude and spectral distribution of different modulation methods; time-frequency distribution map features effectively capture modulation signals with sudden changes or local features through joint representation in time and frequency; one-dimensional time-series dynamic features are used to capture the evolution patterns and dependencies of the signal in time, effectively distinguishing different phase modulation methods, and learning the interactions between amplitude and phase changes in quadrature modulation signals.

[0152] The three modal features of time-domain and frequency-domain statistical features, time-frequency distribution map features, and one-dimensional time-series dynamic features complement each other according to their respective characteristics, jointly constituting a complete representation of the modulation methods of amplitude modulation signals 2ASK, 4ASK, 8ASK, frequency modulation signals 2FSK, 4FSK, phase modulation signals BPSK, QPSK, and quadrature amplitude modulation signal 16QAM, effectively improving the accuracy of signal modulation recognition.

[0153] As Figure 11 shown, feature fusion splicing and fully connected layer processing fuse and splice time-domain and frequency-domain statistical features, time-frequency distribution map features, and one-dimensional time-series features according to dimensions to form a high-dimensional feature vector of 2182 dimensions. During the fusion splicing process, all features can be processed in a unified feature space. The fully connected layer processes the spliced feature vector. The fully connected layer uses a weight matrix to perform linear transformation and non-linear activation on the input features, uses ReLU as the activation function, and recombines and encodes each dimension of the input features through calculation to learn the non-linear relationships between different features, thereby improving the performance of the recognition system under complex channel conditions.

[0154] Embodiment 3

[0155] To verify the complementarity between the statistical features in the time domain and frequency domain, the time-frequency distribution map features, and the one-dimensional time series features, the present invention conducted a comparative experiment on the statistical features after fusing the various modal features extracted above with the multi-modal features.

[0156] As Figure 12 shown, in this embodiment, the contrast curves of the accuracy and loss between the time-frequency distribution map features and the statistical features after extracting and fusing multi-modal features;

[0157] As Figure 13 shown, in this embodiment, the contrast curves of the accuracy and loss between the one-dimensional time series dynamic features and the statistical features after extracting and fusing multi-modal features.

[0158] It can be Figures 12 - 13 seen that after fusing the statistical features in the time domain and frequency domain, the time-frequency distribution map features, and the one-dimensional time series features into multi-modal features, dynamic features can be further extracted. Especially for signal modulation methods such as 2ASK, 4ASK, and 8ASK, by combining the statistical features in the time domain and frequency domain with the features of the time-frequency distribution map and the resolution characteristics of the one-dimensional time series features, significant advantages are demonstrated in the dynamic recognition of signals, enabling more effective differentiation of these modulation methods. This not only improves the model's ability to analyze complex signals but also can capture the dynamic features of signals even when the signal amplitude changes drastically.

[0159] In particular, Figure 12 after fusing the extracted multi-modal features, the accuracy is increased by 22.37% compared with only using the features of the time-frequency distribution map extracted by ResNet152, and the loss is reduced by 0.5. Figure 13 After fusing the extracted multi-modal features, the accuracy is increased by 10.5% compared with only using the one-dimensional time series dynamic features extracted by GRU, and the loss is reduced by 0.2.

[0160] Figure 14 The schematic diagram of the confusion matrix after further fusing the statistical features in the time domain and frequency domain, the time-frequency distribution map features, and the one-dimensional time series features in this embodiment is provided. It can be obtained that after fusing the extracted multi-modal features, the dynamic features of medium-scale signals can be effectively extracted, and the recognition rate of medium-scale short-wave signals can reach more than 99.88%. Fusing the statistical features in the time domain and frequency domain not only improves the model's ability to analyze complex signals but also demonstrates significant advantages in the dynamic recognition of signals.

[0161] In summary, the present invention extracts key signal features: time-frequency distribution features, time-domain and frequency-domain statistical features, and one-dimensional time-series dynamic features, and uses the complementarity between the above-mentioned modal features for multi-modal fusion, solving the difficult problems of medium-scale short-wave signals, including strong dynamics due to drastic changes in amplitude and spectral broadening, enhanced coupling due to the mutual dependence of signal characteristics in each time segment of the signal, and the difficulty of stably characterizing signal features with a single modal feature. While improving the ability to describe signal features, the recognition accuracy of signal modulation is effectively improved.

[0162] As described above, the above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A method for analyzing and processing characteristics of mesoscale time-shortwave signals based on multimodal fusion, characterized in that: By comprehensively characterizing the static properties and dynamic change characteristics of the signal, the multimodal features of the signal are mined at multiple levels, including extracting statistical features in the time and frequency domains, deep learning features of the time-frequency distribution diagram, and one-dimensional time series dynamic features, forming a multi-level feature analysis framework; The complementarity of the above-mentioned different modal features is used to fuse the extracted multimodal features, thereby improving the ability to describe signal characteristics, improving the accuracy of modulation recognition of mesoscale time shortwave signals, and improving the performance of the recognition system under complex channel conditions.

2. The method for analyzing and processing characteristics of mesoscale time shortwave signals based on multimodal fusion according to claim 1 is characterized in that: The specific steps include: S1. Dataset generation: First, modulated signal simulation is performed, including amplitude modulation signal, frequency modulation signal, phase modulation signal and orthogonal amplitude modulation signal, from which shortwave signals are obtained; log-normal distribution is used to describe medium-term and long-term channel changes, medium-term and long-term channel change models are introduced, and a medium-scale time shortwave signal dataset is generated; S2. Multimodal feature extraction: extracting statistical features in the time and frequency domains, extracting time-frequency distribution features, and extracting one-dimensional time series dynamic features, as follows: The time-domain and frequency-domain statistical features describe the overall distribution of the signal and are used to distinguish the average amplitude and spectrum distribution of the above-mentioned different modulation methods. The time-frequency distribution graph features effectively capture the modulation signals with sudden changes or local characteristics through the joint representation of time and frequency. The one-dimensional time series dynamic features are used to capture the evolution pattern and dependency of the signal in time, effectively distinguish the above-mentioned different phase modulation methods, and learn the interaction between amplitude and phase changes in orthogonal modulation signals. The mesoscale time shortwave signal data set generated in step S1 is used to extract the multi-dimensional characteristics of the mesoscale time shortwave signal using the above-mentioned multiple modes, forming a multi-level feature analysis framework to comprehensively characterize the static properties and dynamic change characteristics of the shortwave signal; S3, multimodal feature fusion: the multimodal features extracted in step S2 are fused by utilizing the complementarity between the statistical features in the time domain and frequency domain, the features of the time-frequency distribution diagram, and the one-dimensional time series dynamic features, so as to improve the description capability of the signal features and the accuracy of the modulation recognition of the mesoscale time shortwave signal; the specific operation steps are as follows: The three modal features of time domain and frequency domain statistical features, time-frequency distribution diagram features and one-dimensional time series dynamic features complement each other according to their respective characteristics, and together constitute a complete representation of the modulation modes of amplitude modulated signals, frequency modulated signals, phase modulated signals and orthogonal amplitude modulated signals, effectively improving the accuracy of signal modulation recognition.

3. The method for analyzing and processing characteristics of mesoscale time shortwave signals based on multimodal fusion according to claim 2 is characterized in that: In step S1, the amplitude modulation signal includes 2ASK, 4ASK, and 8ASK; the frequency modulation signal includes 2FSK and 4FSK; the phase modulation signal includes BPSK and QPSK; and the quadrature amplitude modulation signal includes 16QAM; The medium-term and long-term channel change models are introduced, wherein the mesoscale effect is obtained by means of LTV and ITV models, and its time constant and standard deviation are obtained according to the measured data, and are implemented through a combination of an Alpha filter and a high-pass filter to obtain a signal-to-noise ratio (SNR) sequence; N data points are selected from the generated signal-to-noise ratio (SNR) sequence, and the N SNR points are sampled at intervals of 1 second, and combined with the Watterson model, so that each signal-to-noise ratio (SNR) value data point corresponds to a specific time period, ensuring that the signal is affected by different SNRs at different time points, thereby simulating the real channel environment, and finally generating a complete data set containing N seconds of mesoscale time shortwave signals.

4. The method for analyzing and processing characteristics of mesoscale time shortwave signals based on multimodal fusion according to claim 2 is characterized in that: The method comprises the following steps: step S2.1 extracting statistical features in the time domain and the frequency domain, including extracting statistical features in the time domain and extracting statistical features in the frequency domain; The time domain statistical features include mean value, skewness, and kurtosis factor; the frequency domain statistical features include frequency domain amplitude mean value, centroid frequency, and frequency standard deviation; The six time-domain and frequency-domain statistical features extracted above are normalized to make the multimodal features more compatible and consistent during the fusion process. The specific calculation formula is: Among them, x norm is the normalized standard feature; x is the original feature; μ is the average value of each feature; α is the standard deviation of each feature.

5. The method for analyzing and processing characteristics of mesoscale time shortwave signals based on multimodal fusion according to claim 4 is characterized in that: (1) The average value is an indicator of the overall DC offset or base value of the signal, which can distinguish whether the DC component of the signal has a significant offset and reflect the overall DC offset degree of the modulated signal: Through the change trend of the average value caused by different modulation depths of different signals, 2ASK, 4ASK, 8ASK, 2FSK, and 16QAM can be clearly distinguished. Specifically, the amplitude of 2ASK, 4ASK, and 8ASK amplitude modulation signals determines the change of their average values. Different frequency components in 2FSK frequency modulation signals cause the change of the average value. The average value of 16QAM orthogonal amplitude modulation signal reflects the distribution characteristics of constellation points. The average value is used to distinguish the above different signal modulation methods, and the specific calculation formula is: In the formula, S i is the amplitude of the sampling data point of the signal; N is the number of sampling data points for each sample; (2) The skewness is used to measure the asymmetry of the signal waveform. For frequency modulation signals, the skewness is used to describe whether there is an asymmetric amplitude distribution characteristic that deviates from the mean in the signal. Because the spectrum shows a slight asymmetry during the switching process between different frequencies, the 2FSK frequency modulation signal can be clearly distinguished. Therefore, the skewness is used to distinguish the frequency modulation signal in the shortwave signal. The specific calculation formula is: In the formula, S i is the amplitude of the sampling data point of the signal; N is the number of sampling data points for each sample; ρ t is the standard deviation, and the formula is as follows: ρ t 3 is the cube of the standard deviation; (3) The kurtosis factor describes the sharpness of the signal amplitude and is an important feature in time domain analysis, especially for detecting impulse signals, pulse signals and high peak modulation signals. When there are sudden amplitude changes in the AM signal and fast jumps between different frequencies in the FM signal, the kurtosis factor will increase significantly. The signal with high kurtosis contains more peaks or high-energy instantaneous components, which can clearly distinguish 2ASK, 4ASK, 2FSK, and 4FSK. Specifically, the sudden amplitude changes in 2ASK and 4ASK AM signals lead to increased kurtosis; the fast jumps between different frequencies in 2FSK and 4FSK signals will increase the kurtosis. The specific calculation formula is: In the formula, S i is the amplitude of the sampling data point of the signal; N is the number of sampling data points for each sample.

6. The method for analyzing and processing characteristics of mesoscale time shortwave signals based on multimodal fusion according to claim 4 is characterized in that: (4) The frequency domain amplitude average value reflects the central trend of the signal spectrum amplitude distribution, describes the energy concentration in the spectrum, and is used to distinguish different modulation modes; The frequency domain amplitude average value can clearly distinguish 2ASK, 4ASK, 8ASK, 2FSK, 4FSK, BPSK, QPSK, and 16QAM, as follows: the spectrum amplitude distribution of 2ASK, 4ASK, and 8ASK amplitude modulation signals has significant characteristics; the spectrum energy of 2FSK and 4FSK keying signals is concentrated at a specific frequency position; BPSK, QPSK, and 16QAM phase modulation or composite modulation signals have different energy distributions in the frequency domain, so the frequency domain amplitude average value can be used as a basis for identification; The specific calculation formula is: In the formula, k is the spatial frequency of the wave; s(k) represents the frequency domain amplitude at the kth frequency point; (5) The center of gravity frequency describes the dominant frequency position of the signal spectrum and is an important indicator for measuring the spectrum distribution. The center of gravity frequency positions of different signals are significantly different, and 2FSK, 4FSK, BPSK, and QPSK can be clearly distinguished as follows: The different frequency components of 2FSK and 4FSK frequency keying signals determine the position of their center of gravity frequencies; the spectrum distribution center of BPSK and QPSK phase modulated signals is closely related to their modulation methods, and the specific calculation formula is: Where k is the spatial frequency; s(k) represents the frequency domain amplitude at the kth frequency point; f k is the frequency value of the kth frequency point; (6) The frequency standard deviation describes the dispersion or concentration of the spectrum and is one of the key features for distinguishing signals of different modulation modes. It can clearly distinguish 2FSK, 4FSK, BPSK, and QPSK. Specifically, the frequency jump signals of 2FSK and 4FSK will cause the frequency standard deviation to change; the bandwidth of BPSK and QPSK modulation signals will directly affect the frequency standard deviation. The specific calculation formula is: Where k is the spatial frequency; S i is the amplitude of the sampling data point of the signal; N is the number of sampling data points for each sample; s(k) represents the frequency domain amplitude at the kth frequency point; f k is the frequency value of the kth frequency point; S2 is the centroid frequency of the kth point.

7. The method for analyzing and processing characteristics of mesoscale time shortwave signals based on multimodal fusion according to claim 4 is characterized in that: The step S2.2 is also included to extract the time-frequency distribution features, specifically, to use the ResNet152 network to respectively extract the features of the two time-frequency distribution features, SPWVD time-frequency distribution features and BJD time-frequency distribution features; The SPWVD time-frequency distribution feature is used to capture local features through high resolution, observe the changes of signals within a specific time range, and improve the recognition ability of signal features; The specific calculation formula is: SPWVD x (t,f)=∫∫x(t-v+τ / 2)x * (t-v-τ / 2)·h(τ)g(v)e -j2πfτ dvdτ In the formula, SPWVD x (t,f) is the time-frequency result in the SPWVD time-frequency distribution feature, τ is the frequency domain window length; v is the time domain window length; h(τ) is the frequency domain smoothing window; g(v) is the time domain smoothing window; x * represents the complex conjugate of x; x(t) is the analytical signal of r(t), expressed as: x(t)=r(t)+jH[r(t)] Where, H[·] represents the Hilbert transform; r(t) is the received signal; The BJD time-frequency distribution feature is used to provide a global time-frequency description of the signal, which can accurately concentrate the signal energy around its real time and frequency position, making the spectral characteristics of the signal more obvious; the specific calculation formula is: BJD x (t,f)=∫∫x(v+τ / 2)x * (v-τ / 2)·φ(tv,τ)e -j2 p fτ DVD Where, BJD x (t,f) is the time-frequency result in the BJD time-frequency distribution feature; τ is the frequency domain window length, v is the time domain window length; x * represents the complex conjugate of x; where φ(t,τ) is:

8. The method for analyzing and processing characteristics of mesoscale time shortwave signals based on multimodal fusion according to claim 7 is characterized in that: Step S2.2 also includes fusing the BJD time-frequency distribution features and the SPWVD time-frequency distribution features, a total of 2048-dimensional features, by using JS divergence, as follows: First, the features obtained by SPWVD and BJD are converted into probability distributions p and q using the softmax function, and then the JS divergence (Jensen–Shannon divergence) between them is calculated to capture details while not losing the overall signal; the calculation formula is as follows: Where KL(P||Q) is the KL divergence (Kullback-Leibler divergence). The two probability distributions are set to P and Q. Under the premise of setting them as continuous random variables, their corresponding probability density functions are p(x) and q(x), respectively. p is the feature obtained by SPWVD converted to probability distribution using the softmax function; q is the feature obtained by BJD converted to probability distribution using the softmax function. JS divergence represents the similarity of two time-frequency distribution features. The larger the divergence value, the less similar the two time-frequency distribution features are, and the smaller the weight assigned; the smaller the divergence value, the more similar the two time-frequency distribution features are, and the larger the weight assigned; each shortwave signal sample is assigned different weights according to the size of the divergence between the multimodal features, and the above multimodal features are weighted averaged according to the corresponding weights to obtain a fused high-dimensional feature vector.

9. The method for analyzing and processing characteristics of mesoscale time shortwave signals based on multimodal fusion according to claim 2 is characterized in that: The method includes step S2.3 of extracting the one-dimensional time series dynamic features, specifically using a gated recurrent unit GRU to capture the dynamic changes and trends in the one-dimensional time series signal, and obtaining the features of the signal to process the dynamic signal with long-range dependencies; the gated recurrent unit GRU controls the flow of information by introducing an update gate and a reset gate, and the specific calculation formula is as follows: z t =σ(W z ·[h t-1 ,x t ]) r t =σ(W r ·[h t-1 ,x t ]) In the formula, x t Enter information for the current moment; h t-1 is the hidden state of the previous moment; h t is the hidden state passed to the next moment; is the candidate hidden state; r t is the reset gate, z t is the update gate, and σ is the sigmoid function.

10. The method for analyzing and processing characteristics of mesoscale time shortwave signals based on multimodal fusion according to claim 2 is characterized in that: In step S3, feature concatenation and full connection layer processing is to concatenate the time-domain and frequency-domain statistical features, the time-frequency distribution graph features, and the one-dimensional time series features by dimension to form a high-dimensional feature vector. In the concatenation process, all features can be processed in a unified feature space; The concatenated feature vector is processed by the fully connected layer. The fully connected layer uses the weight matrix to perform linear transformation and nonlinear activation on the input features, and uses ReLU as the activation function. Each dimension of the input features is combined and re-encoded by calculation to learn the nonlinear relationship between different features, thereby improving the performance of the recognition system under complex channel conditions.

Images