Method and device for mixed modulation recognition of multi-type signals

Abstract

The application relates to a mixed modulation identification method and device for multiple types of signals and relates to the technical field of communication signal processing. The method comprises the following steps: processing original discrete signals to obtain denoised signals; classifying based on power mutations of the denoised signals to obtain burst type signals or non-burst type signals; performing frequency mutation detection on the burst type signals to obtain frequency hopping signals or burst signals; performing cyclic prefix detection on the non-burst type signals to obtain OFDM signals or continuous signals; extracting fuzzy high-order cyclic cumulative quantity features from the continuous signals, the burst signals and the frequency hopping signals; reconstructing the OFDM signals into three-dimensional signal frames; identifying the fuzzy high-order cyclic cumulative quantity features of the signals; and outputting the modulation formats of the corresponding signals. The application can realize integrated identification of multiple types of signals.

Description

Technical Field

[0001] This invention relates to the field of communication signal processing technology, and in particular to a hybrid modulation identification method and apparatus for multiple types of signals. Background Technology

[0002] Modulation format identification is a core component of blind signal processing in communication. Its function is to determine the modulation type through signal feature analysis when the modulation parameters of the signal are unknown, thus providing a foundation for subsequent demodulation, decoding, and signal analysis. With the development of technologies such as 5G and cognitive radio, signal types in actual communication scenarios exhibit diversified characteristics: traditional continuous wave signals, short-time burst signals, anti-interference frequency hopping signals, and wideband OFDM signals coexist, and various signals cover multiple modulation formats such as BPSK, QPSK, and 16QAM.

[0003] Existing modulation format recognition technologies suffer from three major pain points: First, they lack versatility. Most existing methods are designed for "single signal adaptation." For example, methods based on cyclic cumulants can only efficiently identify continuous signals, but they suffer from incomplete feature extraction when faced with the short duration of burst signals. Furthermore, recognition algorithms for frequency-hopping signals are incompatible with wideband signals such as OFDM. Second, preprocessing mechanisms are lacking or inefficient. The positioning deviation of the start / end time of burst signals and the ambiguity of the transition points of frequency-hopping signals lead to increased errors in subsequent feature extraction. Existing preprocessing methods often use single threshold detection, resulting in high rates of missed detections and false detections at low signal-to-noise ratios. Third, robustness and efficiency are difficult to balance. To improve applicability, some methods employ complex feature fusion strategies, leading to increased recognition latency. Meanwhile, simplified feature methods experience a sharp drop in recognition accuracy for complex signals such as OFDM at low signal-to-noise ratios, failing to meet practical engineering needs.

[0004] In addition, existing technologies generally lack a recognition framework for "unified scheduling of multiple types of signals". In complex electromagnetic environments, different recognition modules need to be switched, which increases the system response delay and makes it difficult to adapt to real-time monitoring scenarios. Summary of the Invention

[0005] Therefore, it is necessary to provide a hybrid modulation recognition method and apparatus for multiple types of signals that can combine high versatility, high-precision preprocessing, and strong robustness to address the above-mentioned technical problems.

[0006] A hybrid modulation identification method for multiple signal types, the method comprising:

[0007] Adaptive noise suppression is applied to the original discrete input signal to obtain a denoised signal;

[0008] Based on the power abrupt changes in the denoised signal, burst signals or non-burst signals are obtained;

[0009] Frequency abrupt change detection is performed on the burst-type signals to distinguish them from frequency hopping signals or burst signals; cyclic prefix detection is performed on the non-burst-type signals to distinguish them from OFDM signals or continuous signals.

[0010] The continuous signal is directly extracted with fuzzy high-order cyclic cumulative features; the burst signal is preprocessed with long and short window energy detection and then fuzzy high-order cyclic cumulative features are extracted; the frequency hopping signal is preprocessed with long and short window energy detection and time-frequency ridge analysis and then fuzzy high-order cyclic cumulative features are extracted; the OFDM signal is reconstructed into a three-dimensional signal frame.

[0011] The fuzzy high-order cyclic cumulant features of the continuous signal, the burst signal, and the frequency hopping signal are input into a trained fully connected neural network classifier for recognition. The three-dimensional signal frame of the OFDM signal is input into a trained dedicated deep learning network for recognition, and the modulation format of the corresponding signal is output.

[0012] On the other hand, a hybrid modulation identification device for multiple types of signals is also provided, comprising:

[0013] The noise reduction signal acquisition module is used to perform adaptive noise suppression on the input original discrete signal to obtain the noise reduction signal;

[0014] The noise reduction signal classification module is used to classify the noise reduction signal based on the power mutation of the noise reduction signal to obtain burst signals or non-burst signals;

[0015] The burst and non-burst signal classification module is used to perform frequency change detection on the burst signals to distinguish them as frequency hopping signals or burst signals; and to perform cyclic prefix detection on the non-burst signals to distinguish them as OFDM signals or continuous signals.

[0016] The classification signal processing module is used to directly extract fuzzy higher-order cyclic cumulative features from the continuous signal; extract fuzzy higher-order cyclic cumulative features from the burst signal after preprocessing with long and short window energy detection; extract fuzzy higher-order cyclic cumulative features from the frequency hopping signal after preprocessing with long and short window energy detection and time-frequency ridge analysis; and reconstruct the OFDM signal into a three-dimensional signal frame.

[0017] The classification signal modulation format recognition module is used to input the fuzzy high-order cyclic cumulant features of the continuous signal, the burst signal, and the frequency hopping signal into a trained fully connected neural network classifier for recognition, input the three-dimensional signal frame of the OFDM signal into a trained dedicated deep learning network for recognition, and output the modulation format of the corresponding signal.

[0018] Compared with existing technologies, the hybrid modulation identification method and apparatus for multiple signal types provided by this invention have the following advantages:

[0019] 1. Through hierarchical classification logic of power mutation classification, frequency mutation detection, and cyclic prefix detection, it uniformly adapts to four types of signals: continuous, burst, frequency hopping, and OFDM. It can achieve integrated recognition of multiple signal types without switching different recognition modules, significantly improving versatility. It can simplify the deployment and maintenance costs of the system, avoid the performance loss caused during the switching process, and better meet the needs of random signal type alternation in actual communication scenarios.

[0020] 2. Design of long and short window energy detection preprocessing for burst signals; for the analysis of time-frequency ridges superimposed on frequency hopping signals, it solves the problems of positioning deviation of the start and end times of burst signals and ambiguity of the jump points of frequency hopping signals in the existing technology. This differentiated preprocessing for different signals improves the processing accuracy and provides a high-quality signal foundation for subsequent feature extraction.

[0021] 3. Fuzzy high-order cyclic cumulant feature extraction is used for continuous, burst, and frequency-hopping signals, and noise interference is reduced by multi-value averaging near the cyclic frequency; the three-dimensional signal frame of OFDM signal is reconstructed and input into a dedicated deep learning network to enhance the robustness of OFDM signal recognition under low signal-to-noise ratio.

[0022] 4. The fully connected neural network classifier and dedicated deep learning network used in the preprocessing and recognition processes are both lightweight and do not require complex feature fusion, which reduces computational complexity. They can be deployed on edge devices such as drones and terminal receivers to meet the engineering requirements of real-time monitoring. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention, and those skilled in the art can obtain other related drawings based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating the hybrid modulation recognition method for multiple signal types provided in Example 1.

[0025] Figure 2 This is a flowchart of the hybrid modulation recognition method for multiple signal types provided in Example 1;

[0026] Figure 3 This is a schematic diagram of the structure of the fully connected neural network classifier provided in Example 1;

[0027] Figure 4 This is a schematic diagram of the structure of the dedicated deep learning network provided in Example 1;

[0028] Figure 5This is a structural block diagram of the hybrid modulation recognition device for multiple signal types provided in Example 2;

[0029] Figure 6 This is an internal structural diagram of the computer device provided in Example 3.

[0030] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0032] It should be noted that in this invention, the use of terms such as "first," "second," etc., is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0033] It is understood that the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0034] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0035] Example 1

[0036] like Figure 1 As shown, this embodiment provides a hybrid modulation identification method for multiple signal types, including the following steps:

[0037] Step 201: Perform adaptive noise suppression on the input original discrete signal to obtain a denoised signal.

[0038] Step 202: Classify the signal based on the power abrupt change of the denoised signal to obtain burst-type or non-burst-type signals.

[0039] Step 203: For burst-type signals, frequency change detection is performed to distinguish between frequency hopping signals and burst signals; for non-burst-type signals, cyclic prefix detection is performed to distinguish between OFDM signals and continuous signals.

[0040] Step 204: Extract fuzzy higher-order cyclic cumulative features directly from continuous signals; extract fuzzy higher-order cyclic cumulative features from burst signals after long and short window energy detection preprocessing; extract fuzzy higher-order cyclic cumulative features from frequency hopping signals after long and short window energy detection and time-frequency ridge analysis preprocessing; reconstruct OFDM signals into three-dimensional signal frames.

[0041] Step 205: Input the fuzzy high-order cyclic cumulant features of continuous signals, burst signals, and frequency hopping signals into the trained fully connected neural network classifier for recognition, and input the three-dimensional signal frame of OFDM signal into the trained dedicated deep learning network for recognition, and output the modulation format of the corresponding signal.

[0042] In the specific implementation of step 201, adaptive noise suppression is a process of dynamically adjusting processing parameters to selectively filter noise interference in the original signal and retain effective signal components.

[0043] In the processing, the original discrete signal is first decomposed into three levels using wavelet transform to obtain high-frequency detail coefficients and low-frequency approximation coefficients. Then, the adaptive threshold of each level is calculated, and the high-frequency detail coefficients are subjected to soft thresholding based on the adaptive threshold to obtain the processed high-frequency detail coefficients. Finally, the processed high-frequency detail coefficients and low-frequency approximation coefficients are reconstructed using inverse discrete wavelet transform to obtain the denoised signal.

[0044] Specifically, the db4 wavelet is selected for the original discrete signal. ( A three-level decomposition is performed using the sampling time sequence number to obtain high-frequency detail coefficients. , , Approximation coefficient with low frequency Then, the adaptive threshold for each layer is calculated using the adaptive threshold formula, and the calculation expression is:

[0045] ;

[0046] In the formula, Indicates the first Adaptive threshold of the layer; Indicates the number of decomposition layers, in this embodiment ; This represents an estimate of the noise standard deviation. For the first High-frequency detail factor () The length is , The length is , The length is ), Indicates the signal length.

[0047] The high-frequency detail coefficients are subjected to soft thresholding based on an adaptive threshold, expressed as follows:

[0048] ;

[0049] In the formula, For the processed first High frequency coefficient of layer, It is a symbolic function.

[0050] Finally, the processed high-frequency detail coefficients are... Approximation coefficient with low frequency The denoised signal is obtained by reconstructing the signal using inverse discrete wavelet transform. .

[0051] This step, through the multi-scale decomposition characteristics of wavelet transform combined with the dynamic adjustment of adaptive threshold, can effectively suppress noise while preserving key signal features, laying a high-quality foundation for subsequent signal classification and recognition.

[0052] In the specific implementation of step 202, power mutation detection is a processing method that determines whether the signal has sudden characteristics by analyzing the difference in instantaneous power changes.

[0053] During the processing, the instantaneous power of the noise-reduced signal is first calculated, and a power difference sequence is constructed based on the instantaneous power. Then, the coefficient of variation of the power difference sequence is calculated. Then, a first threshold is preset, and the noise-reduced signal is determined to be a burst signal or a non-burst signal by comparing the coefficient of variation with the first threshold.

[0054] Specifically, calculate the noise-reduced signal instantaneous power And construct the power difference sequence: Calculate the coefficient of variation of the difference series. ,in: for standard deviation for The mean.

[0055] Preset first threshold Determine the coefficient of variation Is it greater than the first threshold? .like If so, the corresponding noise reduction signal is determined to be a burst signal (including burst signals and frequency hopping signals); if If so, the corresponding noise reduction signal is determined to be a non-burst signal (including continuous signals and OFDM signals).

[0056] It is worth noting that, This is an empirical threshold, set to 0.8 for low signal-to-noise ratios.

[0057] This step utilizes the characteristics of power mutation to achieve preliminary signal classification. It is logically simple, computationally inexpensive, and can quickly identify the approximate affiliation of different types of signals, improving the targeting of subsequent processing.

[0058] In the specific implementation of step 203, frequency change detection is a detection method that determines whether a frequency jump exists by analyzing the time-frequency entropy change of burst-type signals.

[0059] During processing, the time-frequency matrix of burst signals is calculated using short-time Fourier transform, and the time-frequency entropy of each frame is calculated based on this time-frequency matrix. A second threshold is preset, and it is determined whether the difference between the time-frequency entropy of two consecutive frames is greater than the second threshold. If it is, the corresponding burst signal is determined to be a frequency hopping signal; otherwise, the corresponding burst signal is determined to be a burst signal.

[0060] Specifically, for burst-type signals, the time-frequency entropy is calculated using the Short-Time Fourier Transform (STFT) to verify whether a frequency jump exists. The time-frequency matrix is ​​as follows:

[0061] ;

[0062] In the formula, Represents the time-frequency matrix; It is a Hanning window; It's a long window; It is the frame shift step size; It is the FFT point count; It is the frame number; It is a frequency index.

[0063] Then, the time-frequency entropy of each frame is calculated based on the time-frequency matrix:

[0064] ;

[0065] in, Represents time-frequency entropy; This is the normalized power.

[0066] Preset second threshold Determine whether the time-frequency entropy difference between two consecutive frames is greater than a second threshold. If there is a time-frequency entropy difference between two consecutive frames... If the corresponding burst signal is determined to be a frequency hopping signal, then the time-frequency entropy difference between two consecutive frames is... If so, the corresponding burst signal is determined to be a burst signal. It is worth noting that... This is an empirical threshold.

[0067] Cyclic prefix detection is a method to distinguish OFDM signals from continuous signals by detecting the presence of a cyclic prefix structure unique to OFDM signals.

[0068] During the processing, first set the candidate FFT point set and the candidate CP length ratio set, calculate the corresponding candidate CP length based on the candidate FFT point set and the candidate CP length ratio set, and then form a candidate combination with the candidate FFT point set and the corresponding candidate CP length.

[0069] The non-burst signal is divided into several overlapping segments according to the preset overlap rate. For each overlapping segment and each candidate combination, the autocorrelation function of the first delay and the second delay is calculated.

[0070] Then, the first autocorrelation average is calculated based on the autocorrelation function of all first delays, and the second autocorrelation average is calculated based on the autocorrelation function of all second delays.

[0071] Select a delay window and calculate the average autocorrelation value within the window; calculate the peak-to-peak ratio based on the average first autocorrelation value, the average second autocorrelation value, and the average autocorrelation value within the window.

[0072] A third threshold is preset. It is determined whether the peak ratio is greater than the third threshold. If it is, and the peak ratio is the maximum value among all candidates, the corresponding non-burst signal is determined to be an OFDM signal; otherwise, the corresponding non-burst signal is determined to be a continuous signal.

[0073] Specifically, let the set of candidate FFT points be... The candidate CP length ratio set is For each Calculate the length of the corresponding candidate CP. ( , (to round to the nearest integer); then, the candidate FFT points in the candidate FFT point set... With the corresponding candidate CP length Form candidate combinations .

[0074] The overlap rate is set to 50% in this embodiment. Based on this preset overlap rate, non-burst signals are divided into... There are overlapping segments, each with a length of [number]. ( ), for each segment ( ) and each candidate combination Calculate the first delay respectively Second delay The autocorrelation function is expressed as:

[0075] ;

[0076] ;

[0077] In the formula, Indicates the first The autocorrelation function of the first delay in the segment; Indicates the first The autocorrelation function of the second delay in the segment; Indicates the first The signal after noise reduction at the sampling time The conjugate of the sampled values; Indicates the first The signal after noise reduction at the sampling time The conjugate of the sampled values; Indicates the first Discrete sampling sequence of non-burst type signals after segment noise reduction.

[0078] Then, the first autocorrelation average value is calculated based on the autocorrelation function of the first delay in segment K. ; Calculate the average value of the second autocorrelation based on the autocorrelation function of the second delay in segment K. It is worth noting that the CP portion of the OFDM symbol repeats the corresponding position of the following symbol, therefore... and Autocorrelation peaks will appear in all of them.

[0079] Therefore, a delay window is selected. (exclude (Itself), calculate the average autocorrelation value within the window. Then, the peak-to-peak ratio is calculated based on the average value of the first autocorrelation, the average value of the second autocorrelation, and the average autocorrelation value within the window. The calculation expression is as follows:

[0080] ;

[0081] In the formula, Indicates peak-to-peak ratio; Indicates the number of candidate FFT points; Indicates the length of the candidate CP; This represents the average value of the first autocorrelation. This represents the average value of the second autocorrelation. This represents the average autocorrelation value within the window.

[0082] Preset third threshold If there exists a candidate combination whose peak ratio is greater than the third threshold If the peak ratio of this combination is the largest among all candidates, then this candidate combination is determined to be the optimal screening result. The corresponding non-burst type signal is the OFDM signal, and the estimated number of FFT points is recorded. With CP length If the peak ratio of all candidate combinations is If so, the corresponding non-burst signal is determined to be a continuous signal. It is worth noting that... This is an empirical threshold for low signal-to-noise ratio.

[0083] This step accurately distinguishes between burst signals and frequency-hopping signals, avoiding confusion between the two types of signals that could lead to deviations in subsequent preprocessing and feature extraction. Simultaneously, it utilizes the cyclic prefix characteristic of OFDM signals to achieve accurate identification, solving the challenge of distinguishing between continuous signals and OFDM signals in non-burst signal categories.

[0084] In the specific implementation of step 204, long-short window energy detection is a preprocessing method that uses a long window to estimate noise power and a short window to capture local signal energy, thereby extracting effective segments of burst signals. The preprocessing steps for long-short window energy detection are the same for both burst signals and frequency-hopping signals.

[0085] During processing, for burst signals, the first instantaneous power of the burst signal is calculated; the burst signal is divided into several non-overlapping long windows, and the first average power of each long window is calculated based on the first instantaneous power; a preset proportional threshold is used to filter long windows whose first average power is lower than the proportional threshold, and the first noise power is estimated; a first dynamic threshold is calculated based on the first noise power and a preset false alarm probability; a first local average power is calculated using a sliding short window; it is determined whether the first local average power is greater than the first dynamic threshold. If so, the signal interval corresponding to the short window is determined to be the effective signal interval of the burst signal; if not, the signal interval corresponding to the short window is determined to be the noise or invalid signal interval of the burst signal; the short windows containing the effective burst signal are spliced ​​together to obtain a burst signal segment containing the effective burst signal.

[0086] For a frequency-hopping signal, calculate the second instantaneous power of the frequency-hopping signal; divide the frequency-hopping signal into several non-overlapping long windows, and calculate the second average power of each long window based on the second instantaneous power; preset a proportional threshold, filter long windows whose second average power is lower than the proportional threshold, and estimate the second noise power; calculate the second dynamic threshold based on the second noise power and a preset false alarm probability; calculate the second local average power using a sliding short window; determine whether the second local average power is greater than the second dynamic threshold. If so, determine that the signal interval corresponding to the short window is the effective signal interval of the frequency-hopping signal; if not, determine that the signal interval corresponding to the short window is the noise or invalid signal interval of the frequency-hopping signal; splice the short windows containing the effective frequency-hopping signal to obtain a frequency-hopping signal segment containing the effective frequency-hopping signal.

[0087] Specifically, taking burst signals as an example, the instantaneous power of burst signals is calculated. ; Divide the sudden signal into A length of Non-overlapping long windows are used to calculate the average power of each long window based on instantaneous power. .

[0088] Then preset the ratio threshold. Screening average power below the proportion threshold A long window is used to estimate the noise power. Based on the Neyman-Pearson criterion, according to noise power and the preset false alarm probability The dynamic threshold is calculated using the following expression:

[0089] ;

[0090] In the formula, It is the short window length. It is the inverse function of the standard Gaussian Q-function, satisfying ,in, t is the independent variable of the standard Gaussian Q-function; t is the integral dummy variable.

[0091] Then, using a length of Sliding short window calculation of local average power Where r represents the start time of the sliding short window; determine whether the local average power is greater than the dynamic threshold, if... If the signal interval corresponding to the short window is valid, it is determined to be a valid signal interval of the burst signal; otherwise, it is determined to be a noise or invalid signal interval of the burst signal. Finally, the short windows containing valid burst signals are spliced ​​together to obtain a burst signal segment containing valid burst signals.

[0092] After the frequency-hopping signal undergoes long and short window energy detection preprocessing, it still needs to be preprocessed by time-frequency ridge analysis. In the time-frequency ridge analysis preprocessing, a short-time Fourier transform is first performed on the effective frequency-hopping signal segments obtained from the long and short window energy detection to obtain the time-frequency matrix. The time-frequency ridge is then extracted by searching for the maximum energy frequency frame by frame. The index difference component of the time-frequency ridges in adjacent frames is calculated, a second dynamic threshold is set, the frequency jump time is located, and a set of frequency-hopping times is formed. Based on the set of frequency-hopping times, steady-state frequency segments are divided, the center frequency of each segment is estimated, and the frequency-hopping signal is reconstructed.

[0093] Specifically, for frequency hopping signal segments containing valid frequency hopping signals. Perform a short-time Fourier transform (STFT) to obtain the time-frequency matrix. Based on the time-frequency matrix, the maximum energy frequency is searched frame by frame to extract the time-frequency ridge. , Indicates frequency index; Indicates the frame number; This represents the signal length. The difference component of the ridge index between adjacent frames is defined as... Its statistical characteristics reflect the intensity of frequency jumps. Based on its statistical characteristics, a third dynamic threshold is set to... ,in: and They are the difference index and difference components respectively. The mean and standard deviation.

[0094] Determine the index difference component Is it less than the third dynamic threshold? ,when At that time, determine the corresponding frequency transition moment, i.e. Frequency jumps occur at any given moment, thus forming a set of frequency hopping moments. ;like If no frequency jump occurs, it is determined that there is no frequency jump and no record is made.

[0095] The frequency-hopping signal segment is divided into several continuous steady-state frequency segments based on the set of frequency-hopping times. For each steady-state frequency segment... Based on the corresponding time-frequency ridge, its center frequency is estimated to be... ,in This is an index-frequency mapping function.

[0096] Within each steady-state frequency band, based on the center frequency Frequency hopping signal segment and frequency hopping time set The signal is reconstructed to obtain the reconstructed frequency-hopping signal. The final expression of the reconstructed frequency-hopping signal is as follows: ,in, Indicates time The index of the frequency hopping band to which it belongs; Indicates the first The starting time of each steady-state frequency band (frequency hopping band), which is taken from the set of frequency hopping times. , is the The time boundary for the start of each frequency hopping band; Indicates the first The end time of each steady-state frequency band (frequency hopping band), which is taken from the set of frequency hopping times. , is the The time boundary for the termination of each frequency hopping band;

[0097] For all three types of signals, fuzzy higher-order cyclic cumulant features need to be extracted. During processing, the cyclic frequencies of the continuous signal, burst signal segments, and reconstructed frequency-hopping signal are estimated respectively; a preset number of frequency values ​​near the cyclic frequencies are selected to form a frequency set; the continuous signal and its corresponding frequency set, the burst signal segment and its corresponding frequency set, and the reconstructed frequency-hopping signal and its corresponding frequency set are substituted into the cyclic time-varying moment function for calculation, respectively, to obtain the fuzzy higher-order cyclic cumulant features of the continuous signal, burst signal, and frequency-hopping signal.

[0098] Specifically, the formula for estimating the cycle frequency is as follows: ,in: It is the cycle frequency; It is the order of the cumulative quantity; It is the conjugate order; frequency offset The estimate is obtained through the fourth power spectrum method, and the formula is: ,in, Indicates Fast Fourier Transform; It is the symbol rate, taking .

[0099] Select A frequency set is formed by a predetermined number of nearby frequency values. In this embodiment, the frequency set is selected. The set consists of 100 nearby frequency values. (Unit: Hz), substitute Time-varying moment function of the first cycle Specifically:

[0100] ;

[0101] in, , This indicates optional conjugation, with c factors being conjugated. Let be the delay vector, take , It indicates an average over time.

[0102] Calculation based on cyclic time-varying moment function The formula for the cumulative amount of a cycle is expressed as follows:

[0103] ;

[0104] in, It is an index set All possible partitions, each partition Include Subset . formula Indicates to The accumulation of variable torque during the first cycle. It is necessary to... All of the above satisfy of Perform a traversal. ,in, Indicates the first The cycle frequency corresponding to each sub-partition is determined by the local cumulative quantum order, local conjugate quantum order, frequency offset, and symbol rate of that sub-partition. Indicates the first The delay components corresponding to each sub-partition are delay vectors. The elements in the sub-partition that match it; Represents the set of indices The The number of elements contained in a subset is the size of the local index corresponding to that subpartition. Indicates the first The local conjugate order corresponding to each sub-partition is the global conjugate order. The amount allocated on this sub-partition.

[0105] Through the above The formula for the fourth-order cyclic cumulant is used to calculate the fourth-order cyclic cumulant. Then, the average value of all cyclic frequencies is taken to obtain the mean of the fuzzy cumulant. (4th-order 2-conjugate cumulant). Similarly, calculate the 6th-order cyclic cumulant, and then average all cyclic frequency values ​​to obtain the mean fuzzy cumulant. (6th-order 3-conjugate cumulants) ultimately form the fuzzy higher-order cyclic cumulant feature vector. .

[0106] Through the above calculation process, the fuzzy higher-order cyclic cumulant feature vectors of continuous signals are obtained respectively. Fuzzy high-order cyclic cumulant feature vector of burst signal fuzzy higher-order cyclic cumulant eigenvectors of frequency-hopping signals .

[0107] For OFDM signals, their IQ components are... Reconstructed into 3D signal frames ;in, This represents the complex baseband signal of an OFDM signal; The serial number of a single OFDM symbol (within a range of values) ), For the subcarrier index within the corresponding symbol; Indicates OFDM complex baseband signal The quadrature components (Q-path signals); ( ) represents the operation of extracting the real part of a complex signal, used to extract the in-phase component of a complex baseband signal; ( The ) represents the operation of extracting the imaginary part of a complex signal, used to extract the quadrature components of a complex baseband signal; Indicates OFDM complex baseband signal The in-phase component (I-channel signal); It is based on the single-symbol length of the blind detection results. This represents the estimated number of FFT points of the OFDM signal obtained through blind cyclic prefix detection. This represents the estimated length of the cyclic prefix (CP) of the OFDM signal obtained through blind cyclic prefix detection; It is the number of OFDM symbols. The complex baseband signal representing the OFDM signal The total number of sampling points.

[0108] This step requires no additional preprocessing for continuous signals; it directly extracts the fuzzy high-order cyclic cumulant features. This is because continuous signals are highly stable, and direct feature extraction reduces information loss caused by preprocessing, improving the efficiency and completeness of feature extraction.

[0109] The burst signal is first preprocessed by long and short window energy detection and then the fuzzy high-order cyclic cumulative feature is extracted. This can accurately locate the start and end time of the burst signal, eliminate pure noise, and solve the problem of incomplete feature extraction caused by the short duration of burst signals.

[0110] Frequency hopping signals need to be preprocessed by long and short window energy detection and time-frequency ridge analysis before fuzzy high-order cyclic cumulative features are extracted. This not only eliminates noise interference but also accurately captures the frequency jump characteristics of the frequency hopping signal, solving the problem of fuzzy jump points in frequency hopping signals and providing a clear signal basis for subsequent feature extraction.

[0111] For OFDM signals, the IQ components of the OFDM signal are reorganized according to the symbol length and the number of symbols, so that the reconstructed three-dimensional signal frame can completely retain the intra-symbol and inter-symbol correlation of the OFDM signal, adapt to the input requirements of dedicated deep learning networks, and improve the accuracy of OFDM signal recognition.

[0112] In the specific implementation of step 205, such as Figure 3As shown, the fully connected neural network classifier (F-FCNN) has a 3-layer structure, including an input layer, hidden layers, and an output layer. The input is a feature vector F = [x1, x2], and the output is y1, y2, y3. The training samples cover BPSK, QPSK, and 16QAM single-carrier signals with low signal-to-noise ratios (-10dB to 10dB), with 1000 samples for each modulation and each signal-to-noise ratio. The optimizer is Adam, the learning rate is 0.001, and the training is performed for 50 epochs. Step 204 yields the fuzzy high-order cyclic cumulant feature vector of the continuous signal. Fuzzy high-order cyclic cumulant feature vector of burst signal fuzzy higher-order cyclic cumulant eigenvectors of frequency-hopping signals The input is a pre-trained fully connected neural network classifier for recognition, and the output is a modulation format of BPSK, QPSK or 16QAM.

[0113] like Figure 4 As shown on the left, the dedicated deep learning network (OFDM-Net) includes a downsampling block consisting of two 1×7 convolutional kernels, two batch normalization layers, two ReLU activations, and two max pooling layers; an XYXY alternating module; and a classification layer consisting of a global average pooling layer, two fully connected layers, a dropout layer, and a softmax layer. In the XYXY alternating module, the X module uses a one-dimensional horizontal filter to extract intra-symbol correlations; the Y module uses a one-dimensional vertical filter to extract inter-symbol correlations. Each XYXY alternating module contains a summation layer (add) and a concatenation layer (concat) to avoid gradient vanishing. Specifically, as shown... Figure 4 As shown on the right, module X includes one 1×1 convolutional kernel, three batch normalization layers, three ReLU activations, one max pooling layer, two 1×3 horizontal convolutional kernels, two summation layers, and one concatenation layer. The structure of module Y is the same as module X, except that the two 1×3 horizontal convolutional kernels are replaced with two 3×1 vertical convolutional kernels.

[0114] The training samples used included frequency-selective multipath Rayleigh fading, noise (-10dB~10dB), frequency offset (0~1000Hz), and phase offset. OFDM sample set (covering) and All combinations (32,000 samples for each modulation) are used to obtain a trained dedicated deep learning network (OFDM-Net). The 3D signal frame obtained in step 204 is then processed. The input is processed by OFDM-Net, and the output is the modulation format BPSK, QPSK or 16QAM.

[0115] In one embodiment, in order to better describe the ability of the present invention to identify the modulation format of multiple types of signals such as continuous, burst, frequency hopping and OFDM, the signal shown in Table 1 is selected as the signal under test.

[0116] Table 1 Parameters of the measured signal

[0117]

[0118] The implementation steps of the hybrid modulation recognition method for multiple signal types provided by this invention are as follows:

[0119] 1) Perform adaptive noise suppression on the input signal to improve signal quality.

[0120] 2) By detecting power surges, the signal can be initially determined to be either bursty or non-burst.

[0121] 3) Perform refined classification and preprocessing for different types of signals:

[0122] a) For burst signals, perform frequency change detection to distinguish between frequency hopping signals and burst signals.

[0123] b) For non-burst signals, perform cyclic prefix detection to identify OFDM signals and continuous signals.

[0124] 4) Feature extraction and modulation recognition of various signals:

[0125] a) Extract fuzzy high-order cyclic cumulant features from continuous, burst, and frequency-hopping signals, and input them into a trained fully connected neural network classifier for recognition.

[0126] b) For OFDM signals, reconstruct their three-dimensional signal frames and input them into a dedicated deep learning network for recognition;

[0127] 5) Output the modulation format identification results for all signal types.

[0128] The above steps constitute the complete process for modulation format recognition according to an embodiment of the present invention. The modulation format recognition performance of the present invention can be tested through multiple experiments.

[0129] After 4096 tests, the accuracy of the hybrid modulation recognition method for multi-type signals provided by this invention compared with existing methods for signal modulation format recognition is shown in Table 2 (average).

[0130] Table 2 Accuracy Comparison

[0131]

[0132] As shown in the table above, the hybrid modulation recognition method for multiple signal types provided by this invention can effectively identify the modulation format of unknown signals. Although it does not demonstrate an absolute advantage over existing methods in terms of recognition accuracy, the advantage of this invention lies in its unified recognition framework, which is compatible with the detection process of multiple signal types. It eliminates the need to adjust processing strategies for different signal types, simplifying system deployment and maintenance costs, avoiding performance losses during switching, and better meeting the needs of random signal type alternation in real-world communication scenarios. Based on the data in the table, the recognition accuracy of this invention at various signal-to-noise ratios and for various signal types is on par with existing methods, sufficient to cover the recognition needs of common scenarios.

[0133] It should be understood that, although this embodiment Figure 1 The steps are shown sequentially as indicated by the arrows, but they are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order in which these steps are performed; they can be executed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0134] Example 2

[0135] Based on the hybrid modulation recognition method for multiple signal types in Embodiment 1, this embodiment discloses a hybrid modulation recognition device for multiple signal types, such as... Figure 5 As shown, the hybrid modulation recognition device for multiple signal types includes: a noise reduction signal acquisition module 401, a noise reduction signal classification module 402, a burst and non-burst signal classification module 403, a classification signal processing module 404, and a classification signal modulation format recognition module 405, wherein:

[0136] The noise reduction signal acquisition module 401 is used to perform adaptive noise suppression on the input original discrete signal to obtain a noise reduction signal.

[0137] The noise reduction signal classification module 402 is used to classify the noise reduction signal based on the power change, and obtain burst signal or non-burst signal.

[0138] The burst and non-burst signal classification module 403 is used to perform frequency change detection on burst signals to distinguish between frequency hopping signals and burst signals; and to perform cyclic prefix detection on non-burst signals to distinguish between OFDM signals and continuous signals.

[0139] The classification signal processing module 404 is used to directly extract fuzzy higher-order cyclic cumulative features from continuous signals; extract fuzzy higher-order cyclic cumulative features from burst signals after long and short window energy detection preprocessing; extract fuzzy higher-order cyclic cumulative features from frequency hopping signals after long and short window energy detection and time-frequency ridge analysis preprocessing; and reconstruct OFDM signals into three-dimensional signal frames.

[0140] The classification signal modulation format recognition module 405 is used to input the fuzzy high-order cyclic cumulant features of continuous signals, burst signals, and frequency hopping signals into a trained fully connected neural network classifier for recognition, input the three-dimensional signal frame of OFDM signals into a trained dedicated deep learning network for recognition, and output the modulation format of the corresponding signal.

[0141] In this embodiment, the specific working process and working principle of the noise reduction signal acquisition module 401, the noise reduction signal classification module 402, the burst and non-burst signal classification module 403, the classification signal processing module 404, and the classification signal modulation format recognition module 405 are the same as those in Embodiment 1, and therefore will not be described again in this embodiment. Each unit module can be implemented entirely or partially through software, hardware, or a combination thereof. Each unit module can be embedded in or independent of the processor in the computer device in hardware form, or it can be stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to the above unit modules.

[0142] Example 3

[0143] like Figure 6 The diagram illustrates a terminal device disclosed in this embodiment, comprising a transmitter, a receiver, a memory, and a processor. The transmitter transmits instructions and data, the receiver receives instructions and data, the memory stores computer-executed instructions, and the processor executes the computer-executed instructions stored in the memory to implement the method described in Embodiment 1 above.

[0144] It is important to note that the aforementioned memory can be either standalone or integrated with the processor. When the memory is set up independently, the terminal device also includes a bus for connecting the memory and the processor.

[0145] Example 4

[0146] This embodiment discloses a computer-readable storage medium storing computer-executable instructions. When a processor executes the computer-executable instructions, it implements the method in Embodiment 1 above.

[0147] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0148] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0149] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A hybrid modulation recognition method for multiple signal types, characterized in that, The method includes: Adaptive noise suppression is applied to the original discrete input signal to obtain a denoised signal; Based on the power abrupt changes in the denoised signal, burst signals or non-burst signals are obtained; Frequency abrupt change detection is performed on the burst-type signals to distinguish them from frequency hopping signals or burst signals; cyclic prefix detection is performed on the non-burst-type signals to distinguish them from OFDM signals or continuous signals. The continuous signal is directly extracted with fuzzy high-order cyclic cumulative features; the burst signal is preprocessed with long and short window energy detection and then fuzzy high-order cyclic cumulative features are extracted; the frequency hopping signal is preprocessed with long and short window energy detection and time-frequency ridge analysis and then fuzzy high-order cyclic cumulative features are extracted; the OFDM signal is reconstructed into a three-dimensional signal frame. The fuzzy high-order cyclic cumulant features of the continuous signal, the burst signal, and the frequency hopping signal are input into a trained fully connected neural network classifier for recognition. The three-dimensional signal frame of the OFDM signal is input into a trained dedicated deep learning network for recognition, and the modulation format of the corresponding signal is output.

2. The hybrid modulation recognition method for multiple signal types according to claim 1, characterized in that, Adaptive noise suppression is applied to the original discrete input signal to obtain a denoised signal, including: Wavelet transform is used to perform a three-level decomposition on the original discrete signal to obtain high-frequency detail coefficients and low-frequency approximation coefficients. Calculate the adaptive threshold for each layer, and then perform soft thresholding on the high-frequency detail coefficients based on the adaptive thresholds to obtain the processed high-frequency detail coefficients; The processed high-frequency detail coefficients and the low-frequency approximation coefficients are reconstructed using inverse discrete wavelet transform to obtain the denoised signal.

3. The hybrid modulation recognition method for multiple signal types according to claim 1, characterized in that, Based on the power abrupt changes in the denoised signal, burst signals or non-burst signals are obtained, including: Calculate the instantaneous power of the noise-reduced signal, construct a power difference sequence based on the instantaneous power, and then calculate the coefficient of variation of the power difference sequence; A first threshold is preset, and it is determined whether the coefficient of variation is greater than the first threshold; if yes, the corresponding noise reduction signal is determined to be a burst signal; if no, the corresponding noise reduction signal is determined to be a non-burst signal.

4. The hybrid modulation recognition method for multiple signal types according to any one of claims 1 to 3, characterized in that, Frequency change detection is performed on the burst-type signals to distinguish between frequency hopping signals and burst signals, including: The time-frequency matrix of the burst-type signal is calculated using short-time Fourier transform; the time-frequency entropy of each frame is calculated based on the time-frequency matrix. A second threshold is preset, and it is determined whether the time-frequency entropy difference between two consecutive frames is greater than the second threshold. If it is, the corresponding burst signal is determined to be a frequency hopping signal; if not, the corresponding burst signal is determined to be a burst signal.

5. The hybrid modulation recognition method for multiple signal types according to any one of claims 1 to 3, characterized in that, Cyclic prefix detection is performed on the non-burst type signal to distinguish between OFDM signal and continuous signal, including: Set a set of candidate FFT points and a set of candidate CP length ratios; calculate the corresponding candidate CP length based on the set of candidate FFT points and the set of candidate CP length ratios, and then form a candidate combination by combining the candidate FFT points in the set of candidate FFT points and the corresponding candidate CP lengths. The non-burst signal is divided into several overlapping segments according to a preset overlap rate. For each overlapping segment and each candidate combination, the autocorrelation function of the first delay and the second delay is calculated. Then, the first autocorrelation average is calculated based on the autocorrelation function of all first delays, and the second autocorrelation average is calculated based on the autocorrelation function of all second delays; Select a delay window and calculate the average autocorrelation value within the window; The peak ratio is calculated based on the first autocorrelation average value, the second autocorrelation average value, and the average autocorrelation value within the window. A third threshold is preset, and it is determined whether the peak ratio is greater than the third threshold. If it is, and the peak ratio is the maximum value among all candidates, then the corresponding non-burst signal is determined to be an OFDM signal; otherwise, the corresponding non-burst signal is determined to be a continuous signal.

6. The hybrid modulation recognition method for multiple signal types according to claim 5, characterized in that, The formula for calculating the peak value ratio is: ； In the formula, Indicates peak-to-peak ratio; Indicates the number of candidate FFT points; Indicates the length of the candidate CP; This represents the average value of the first autocorrelation. This represents the average value of the second autocorrelation; This represents the average autocorrelation value within the window.

7. The hybrid modulation recognition method for multiple signal types according to any one of claims 1 to 3, characterized in that, The burst signal undergoes long and short window energy detection preprocessing, including: Calculate the first instantaneous power of the burst signal; divide the burst signal into several non-overlapping long windows, and calculate the first average power of each long window based on the first instantaneous power; A preset ratio threshold is used to filter long windows where the first average power is lower than the ratio threshold, and the first noise power is estimated. A first dynamic threshold is calculated based on the first noise power and the preset false alarm probability; a first local average power is calculated using a sliding short window. Determine whether the first local average power is greater than the first dynamic threshold. If yes, determine that the signal interval corresponding to the short window is the effective signal interval of the burst signal; otherwise, determine that the signal interval corresponding to the short window is the noise or invalid signal interval of the burst signal. By splicing together the short windows containing valid burst signals, a burst signal segment containing valid burst signals is obtained.

8. The hybrid modulation recognition method for multiple signal types according to claim 7, characterized in that, The frequency-hopping signal undergoes preprocessing including long and short window energy detection and time-frequency ridge analysis, comprising: Calculate the second instantaneous power of the frequency hopping signal; divide the frequency hopping signal into several non-overlapping long windows, and calculate the second average power of each long window based on the second instantaneous power; A preset ratio threshold is used to filter long windows where the second average power is lower than the ratio threshold, and the second noise power is estimated. The second dynamic threshold is calculated based on the second noise power and the preset false alarm probability; the second local average power is calculated using a sliding short window. Determine whether the second local average power is greater than the second dynamic threshold. If yes, determine that the signal interval corresponding to the short window is the effective signal interval of the frequency hopping signal; otherwise, determine that the signal interval corresponding to the short window is the noise or invalid signal interval of the frequency hopping signal. By splicing together the short windows containing the effective frequency hopping signal, a frequency hopping signal segment containing the effective frequency hopping signal is obtained; A short-time Fourier transform is performed on the frequency-hopping signal segment to obtain a time-frequency matrix; based on the time-frequency matrix, the maximum energy frequency is searched frame by frame to extract the time-frequency ridge. Calculate the index difference component of the time-frequency ridge line between adjacent frames, and set a third dynamic threshold based on the statistical characteristics of the index difference component; Determine whether the index difference component is less than the third dynamic threshold. If not, determine that there is a frequency jump, locate the corresponding frequency jump time, and form a set of frequency jump times. If yes, determine that there is no frequency jump and do not record it. The frequency hopping signal segment is divided into several continuous steady-state frequency segments according to the set of frequency hopping times. The center frequency of each steady-state frequency segment is estimated based on the time-frequency ridge line corresponding to each steady-state frequency segment. Within each steady-state frequency band, the signal is reconstructed based on the center frequency, the frequency hopping signal segment, and the set of frequency hopping times to obtain the reconstructed frequency hopping signal.

9. The hybrid modulation recognition method for multiple signal types according to claim 8, characterized in that, Extracting fuzzy higher-order cyclic cumulant features, including: Estimate the cycle frequencies of the continuous signal, the burst signal segment, and the reconstructed frequency-hopping signal respectively; select a preset number of frequency values ​​near the cycle frequency to form a frequency set; Substituting the continuous signal and its corresponding frequency set, the burst signal segment and its corresponding frequency set, and the reconstructed frequency hopping signal and its corresponding frequency set into the cyclic time-varying moment function for calculation, the fuzzy higher-order cyclic cumulative features of the continuous signal, the burst signal, and the frequency hopping signal are obtained respectively.

10. A hybrid modulation recognition device for multiple signal types, characterized in that, The device includes: The noise reduction signal acquisition module is used to perform adaptive noise suppression on the input original discrete signal to obtain the noise reduction signal; The noise reduction signal classification module is used to classify the noise reduction signal based on the power mutation of the noise reduction signal to obtain burst signals or non-burst signals; The burst and non-burst signal classification module is used to perform frequency change detection on the burst signals to distinguish them as frequency hopping signals or burst signals; and to perform cyclic prefix detection on the non-burst signals to distinguish them as OFDM signals or continuous signals. The classification signal processing module is used to directly extract fuzzy higher-order cyclic cumulative features from the continuous signal; extract fuzzy higher-order cyclic cumulative features from the burst signal after preprocessing with long and short window energy detection; extract fuzzy higher-order cyclic cumulative features from the frequency hopping signal after preprocessing with long and short window energy detection and time-frequency ridge analysis; and reconstruct the OFDM signal into a three-dimensional signal frame. The classification signal modulation format recognition module is used to input the fuzzy high-order cyclic cumulant features of the continuous signal, the burst signal, and the frequency hopping signal into a trained fully connected neural network classifier for recognition, input the three-dimensional signal frame of the OFDM signal into a trained dedicated deep learning network for recognition, and output the modulation format of the corresponding signal.

Images