A signal modulation recognition method based on instantaneous characteristics
The characteristics of communication signals are extracted through instantaneous power spectrum transformation and Hilbert transformation, and the CNN-LSTM model is used for identification, which solves the problem of low recognition accuracy under low signal-to-noise ratio, achieving higher recognition accuracy and robustness sex.
Patent Information
- Application Number
- CN202111506378.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-10
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2041-12-10
AI Technical Summary
The existing communication signal modulation recognition method based on instantaneous features has a low recognition accuracy under low signal-to-noise ratio conditions, and is greatly affected by noise interference, making it difficult to effectively identify the modulation mode of the communication signal.
The time-frequency domain information of the signal is extracted by instantaneous power spectrum transformation, feature extraction is performed through instantaneous frequency and instantaneous amplitude sequence, and phase characteristics are obtained by combining Hilbert transform to construct a three-branch parallel convolutional neural network-long and short-term memory Network (CNN-LSTM) model is used for classification identification.
Under low signal-to-noise ratio conditions, the recognition accuracy of communication signal modulation mode is improved, the dependence on classic instantaneous statistical parameters is reduced, and the recognition robustness and accuracy are improved.
Smart Images

Figure CN114520758B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of communication technology, and in particular relates to a signal modulation identification method based on instantaneous characteristics. Background Art
[0002] Signal modulation identification technology is an important problem to be solved in both the field of modern cooperative communications and non-cooperative communications. With the rapid development of communication technology in recent years, the new generation of information and communication technology represented by 5G has been deeply integrated with many fields, and people have higher and higher requirements for communication quality and technology. In the military field, modulation identification is a prerequisite for interference, reconnaissance and monitoring of the target's communication, and plays a vital role in the war situation. In the civilian field, the identification and monitoring of communication signals, the detection and management of spectrum all require the identification of signal modulation mode and other parameters to distinguish different users, so as to detect the existence of illegal and malicious users. Therefore, modulation identification technology is an indispensable technical foundation in both military and civilian fields.
[0003] In the process of modulation recognition system, the pattern recognition method based on feature extraction has been widely used in practical engineering due to its simple structure, less required prior information, less required data, strong adaptability and other advantages. As early as the 1990s, the multiple instantaneous statistical feature parameters proposed by Nandi and Azzouz et al. were called classic feature parameters due to their simple principles, easy extraction, stable performance and other advantages, and have been cited or improved by later generations. Most of the existing methods based on instantaneous features are based on five instantaneous statistical feature parameters: the maximum value of the spectral density of the zero-centered normalized instantaneous amplitude γmax; the standard deviation of the absolute value of the instantaneous phase nonlinear component of the zero-centered non-weak signal segment σap; the standard deviation of the instantaneous phase nonlinear component of the zero-centered non-weak signal segment σdp; the standard deviation of the absolute value of the zero-centered normalized instantaneous amplitude σaa and the standard deviation of the absolute value of the instantaneous frequency of the zero-centered normalized non-weak signal segment σaf. However, the instantaneous statistical feature parameters are greatly affected by noise, and the recognition accuracy is very low under low signal-to-noise ratio. Therefore, under low signal-to-noise ratio conditions, how to improve the anti-noise performance of instantaneous features and improve the accuracy of final recognition is an important problem to be solved.
[0004] Communication signal modulation recognition mainly refers to the information of the baseband signal being modulated on the amplitude, phase or frequency of the carrier. The differences in amplitude, phase and frequency of the modulated signal are the key to distinguishing different modulation methods. Therefore, the premise and basis for using instantaneous characteristics to identify the modulation method of the communication signal is the accurate extraction of the instantaneous information of the modulated signal. Summary of the invention
[0005] The technical problem to be solved by the present invention is to provide a communication signal modulation pattern recognition method based on instantaneous characteristics, so as to solve the problem that the noise interference of the communication signal under low signal-to-noise ratio conditions affects the recognition accuracy.
[0006] A signal modulation identification method based on instantaneous characteristics comprises the following steps, and the following steps are performed in sequence:
[0007] Step 1: Acquisition of communication signals
[0008] A baseband signal is processed in different modulation modes to obtain a modulated multi-level amplitude keying signal MASK, a multi-level frequency shift keying signal MFSK or a multi-level phase shift keying signal MPSK. Matlab software is used to simulate the above three types of modulation signals, and the expression of the communication modulation signal S(t) is obtained as follows:
[0009]
[0010] Where: S(t) is the communication signal; t is the time variable; f c is the carrier frequency; A(t) is the instantaneous amplitude; is the nonlinear phase;
[0011] Step 2: Make the instantaneous power spectrum of the signal
[0012] Perform instantaneous power spectrum transformation IPS on the communication signal S(t), and obtain the instantaneous power spectrum expression of the signal as follows:
[0013]
[0014] Where: IPS(t,f) is the instantaneous power spectrum (IPS) transform of signal S(t); t is the time variable; f is the frequency variable; τ is the time delay; S * is the conjugate of S; h(τ) is the window function, satisfying h(0)=1, where h(0) represents the value of the window function h(τ) at 0;
[0015] Step 3: Extract instantaneous frequency and instantaneous amplitude
[0016] According to the instantaneous power spectrum expression of the signal, the expressions in the time domain and frequency domain obtained by the properties of the instantaneous power spectrum are derived:
[0017]
[0018]
[0019] Wherein, S(t) represents the communication signal, S(f) represents the Fourier transform of S(t);
[0020] Instantaneous frequency fi (t) The expression derived from the above formula is:
[0021]
[0022] Instantaneous frequency f i The energy density at (t) represents the amplitude of the communication signal, and the instantaneous amplitude is obtained by the intensity of the energy value at the instantaneous frequency;
[0023] Step 4: Instantaneous phase extraction
[0024] Select the nonlinear phase as the phase characteristic sequence, and transform the communication signal S(t) in the time domain through Hilbert transform to obtain the analytical signal Z(t)=H[S(t)]=x(t)+j*y(t). And perform the unwinding operation to obtain the unwinding phase: Then remove the linear phase component caused by the carrier frequency to obtain the nonlinear phase sequence:
[0025]
[0026] In the formula, Z(t) is the analytical signal after Hilbert transform; x(t) is the real part of the analytical signal; y(t) is the imaginary part of the analytical signal; φ(i) is the nonlinear phase; f c represents the carrier frequency; t is the time variable; f s represents the sampling frequency; i represents the sampling point;
[0027] Step 5: Sampling and reducing the dimension of the feature sequence
[0028] The instantaneous amplitude sequence and instantaneous frequency sequence of the signal with a length of N obtained in step 3 and the instantaneous phase sequence with a length of N obtained in step 4 are subjected to multiple dimensionality reduction operations of sampling every other 1, so that the long sequence with a length of N is reduced to within 2000 sampling points, where N is a natural number;
[0029] Step 6: Extract the instantaneous feature sequence as a data set
[0030] Select a multi-level amplitude keying signal MASK, a multi-level frequency shift keying signal MFSK, and a multi-level phase shift keying signal MPSK, and randomly select M of each modulation signal, where M is a natural number, repeat steps one to five for each communication signal, and respectively obtain M corresponding instantaneous amplitude, instantaneous frequency, and instantaneous phase sequences after dimensionality reduction, save the M instantaneous amplitude sequences in a .csv file format to form an instantaneous amplitude sequence set, save the M instantaneous frequency sequences in a .csv file format to form an instantaneous frequency sequence set, save the M instantaneous phase sequences in a .csv file format to form an instantaneous phase sequence set, and the sequence data corresponding to the same sequence number in each data set comes from the same signal;
[0031] Step 7: Dataset Division
[0032] The three instantaneous feature sequence sets, namely, the instantaneous amplitude sequence set, the instantaneous frequency sequence set and the instantaneous phase sequence set, are divided into a training set and a validation set according to a set ratio;
[0033] Step 8. Convolutional Neural Network-Long Short-Term Memory Network CNN-LSTM Model Construction
[0034] Construct a three-branch parallel fusion convolutional neural network-long short-term memory network CNN-LSTM model. Using the data in the training set and the validation set, the three branches input the amplitude feature sequence, frequency feature sequence and phase feature sequence corresponding to the same sequence number one by one for synchronous training and validation, and obtain the optimal convolutional neural network-long short-term memory network CNN-LSTM model;
[0035] Step 9: Apply the convolutional neural network-long short-term memory network CNN-LSTM model to classify and identify modulation categories
[0036] Perform steps 1 to 5 on the multi-level amplitude keying signal MASK, multi-level frequency shift keying signal MFSK or multi-level phase shift keying signal MPSK that has been pre-classified and identified, and obtain the corresponding instantaneous amplitude sequence, instantaneous frequency sequence and instantaneous phase sequence after dimensionality reduction, all of which are saved in the reserved corresponding sequence set in .csv file format as a test set;
[0037] The .CSV file of the instantaneous amplitude sequence, instantaneous frequency sequence and instantaneous phase sequence of the communication signal that has been pre-classified and identified is used as input and input into the optimal convolutional neural network-long short-term memory network CNN-LSTM model obtained in step eight to identify the modulation category of the communication signal.
[0038] The method for constructing the three-branch parallel fusion convolutional neural network-long short-term memory network CNN-LSTM model in step eight is specifically as follows:
[0039] ① Each of the three branches first selects a one-dimensional convolutional neural network for processing one-dimensional data and a long short-term memory network specifically for processing time series, that is, a multi-layer convolutional neural network-long short-term memory network CNN-LSTM serial network structure is selected;
[0040] ② The three branches use the same network structure and are set up in sequence.
[0041] Each branch first customizes the number of one-dimensional convolutional layers, the number of pooling layers, and the number of long short-term memory networks (LSTMs), and then sets the number and size of convolution kernels in each convolutional layer, the parameters of the pooling layer, and the parameters of the long short-term memory layer;
[0042] ③ Connect the three set branches using the concatenate function, and splice the features extracted from the three network branches;
[0043] ④ Set the number of fully connected layers and the parameters of each layer in the convolutional neural network-long short-term memory network CNN-LSTM and input the result after splicing features in step ③;
[0044] ⑤ In the output layer of the convolutional neural network-long short-term memory network CNN-LSTM, the softmax function is used to output the multi-classification task neural network, and the modulation type of the output communication signal is the classification recognition result.
[0045] In step eight, the specific steps of using the data in the training set and the validation set to input the amplitude feature sequence, frequency feature sequence and phase feature sequence corresponding to the same sequence number in three branches one by one for synchronous training and validation are as follows:
[0046] ① Setting the initial parameters of the three-branch parallel fusion convolutional neural network-long short-term memory network CNN-LSTM and using the training set and the validation set to simultaneously train the neural network, the initial parameters include the number of iterations epoch, the learning rate, the gradient descent function and the loss function;
[0047] ② After the training is completed, the convolutional neural network-long short-term memory network CNN-LSTM generates two loss value curves corresponding to the training set and the validation set, and fits the two loss value curves. If underfitting or overfitting occurs, jump to step ③, and if there is no underfitting or overfitting, jump to step ⑤;
[0048] ③ Reset the number and parameters of convolutional layers, pooling layers, and long short-term memory layers of each branch, and reset the number of epochs, learning rate, gradient descent function, and loss function of the convolutional neural network-long short-term memory network CNN-LSTM;
[0049] ④ Repeat steps ① and ② for the convolutional neural network-long short-term memory network CNN-LSTM with new parameters set in step ③, using the training set and the validation set to train the convolutional neural network-long short-term memory network CNN-LSTM until the two loss value curves corresponding to the training set and the validation set are fitted and no underfitting or overfitting occurs;
[0050] ⑤Save the convolutional neural network-long short-term memory network CNN-LSTM as the optimal convolutional neural network-long short-term memory network CNN-LSTM model.
[0051] Through the above design scheme, the present invention can bring the following beneficial effects:
[0052] 1. The present invention discloses a communication signal modulation pattern recognition method based on instantaneous characteristics. By transforming the instantaneous power spectrum of the signal, the time-frequency domain information of the signal can be obtained; the instantaneous frequency sequence and the instantaneous amplitude sequence can be obtained through the properties of the instantaneous power spectrum. The instantaneous characteristic sequence is selected as the preprocessing feature instead of the classic instantaneous statistical parameter, which reduces the amount of calculation required to calculate the instantaneous statistical parameter;
[0053] 2. The communication signal modulation pattern recognition method based on instantaneous characteristics described in the present invention is an instantaneous feature sequence obtained through time-frequency domain transformation, which has information and characteristics in both time domain and frequency domain. It can improve the sensitivity of the classic Hilbert transform method for obtaining instantaneous statistical feature parameters to noise under low signal-to-noise ratio conditions, and improve the recognition accuracy under low signal-to-noise ratio conditions;
[0054] 3. The classification and recognition process using a deep learning classifier described in the present invention can extract the features of the data in a neural network and perform classification and recognition, without the need for the steps of extracting feature parameters after preprocessing, such as the machine learning classifiers using classic instantaneous statistical feature parameters, such as decision trees and support vector machines. By designing a network structure with three branches in parallel, the classification accuracy of identification by a single feature can be improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] The present invention is further described below with reference to the accompanying drawings and specific embodiments:
[0056] Figure 1 A flowchart of a signal modulation identification method based on instantaneous characteristics of the present invention;
[0057] Figure 2 A time domain waveform diagram of a 4ASK signal under noise-free conditions in an embodiment of a signal modulation identification method based on instantaneous characteristics of the present invention;
[0058] Figure 3 A time domain waveform diagram of a 4FSK signal under noise-free conditions in an embodiment of a signal modulation identification method based on instantaneous characteristics of the present invention;
[0059] Figure 4 A time domain waveform diagram of a 4PSK signal under noise-free conditions in an embodiment of a signal modulation identification method based on instantaneous characteristics of the present invention;
[0060] Figure 5 It is an instantaneous amplitude diagram of a 4ASK signal under noise-free conditions in an embodiment of a signal modulation identification method based on instantaneous characteristics of the present invention;
[0061] Figure 6It is a comparison diagram of instantaneous amplitudes of a classic method and an instantaneous power spectrum method of a 4ASK signal when the signal-to-noise ratio is 5dB in an embodiment of a signal modulation identification method based on instantaneous characteristics of the present invention;
[0062] Figure 7 It is an instantaneous frequency diagram of a 4FSK signal under noise-free conditions in an embodiment of a signal modulation identification method based on instantaneous characteristics of the present invention;
[0063] Figure 8 It is a comparison diagram of instantaneous frequency of a classic method and an instantaneous power spectrum method of a 4FSK signal when the signal-to-noise ratio is 5dB in an embodiment of a signal modulation identification method based on instantaneous characteristics of the present invention;
[0064] Fig. 9 A comparison diagram of the instantaneous phase of a 4PSK signal under noise-free conditions and the instantaneous phase of a 4PSK signal when the signal-to-noise ratio is 5 dB in an embodiment of a signal modulation identification method based on instantaneous characteristics of the present invention;
[0065] Fig.10 It is an instantaneous amplitude diagram after sampling and dimensionality reduction in an embodiment of a signal modulation identification method based on instantaneous features of the present invention;
[0066] Fig.11 It is an instantaneous frequency diagram after sampling and dimension reduction in an embodiment of a signal modulation identification method based on instantaneous features of the present invention;
[0067] Fig.12 It is an instantaneous phase image after sampling and dimension reduction in an embodiment of a signal modulation identification method based on instantaneous characteristics of the present invention;
[0068] Fig.13 This is a neural network structure diagram designed in an embodiment of a signal modulation recognition method based on instantaneous characteristics of the present invention. DETAILED DESCRIPTION
[0069] The following is a clear and complete description of the technical solutions of the embodiments of the present invention in conjunction with the drawings in the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.
[0070] For ease of understanding and explanation, see the attached Figures 1 to 13 As shown, the present invention provides a signal modulation identification method based on instantaneous characteristics, comprising the following steps:
[0071] Step 1, communication signal acquisition: A baseband signal is processed in different modulation modes to obtain a modulated multi-level amplitude keying signal MASK, a multi-level frequency shift keying signal MFSK or a multi-level phase shift keying signal MPSK. Each type of modulation signal can be generated by loading the baseband signal onto a carrier signal. The generation method has become mature. Matlab software is used to simulate the above three types of modulation signals. The expression of the communication modulation signal S(t) can be expressed as:
[0072]
[0073] Where: S(t) is the communication signal; t is the time variable; f c is the carrier frequency; A(t) is the instantaneous amplitude; is the nonlinear phase. If the signal center frequency f c and bandwidth Δf satisfies the condition f c ≥Δf, the signal can be regarded as a narrowband signal.
[0074] Step 2: Make the instantaneous power spectrum of the signal
[0075] The communication signal is subjected to instantaneous power spectrum transformation (IPS), which is a time-frequency analysis method proposed by Page and is quite insensitive to window parameters. Even at the analysis endpoint, it maintains good time-frequency localization characteristics, and its expression is:
[0076]
[0077] Where: IPS(t,f) is the instantaneous power spectrum (IPS) transform of signal S(t); t is the time variable; f is the frequency variable; τ is the time delay; S * is the conjugate of S; h(τ) is the window function, satisfying h(0)=1, where h(0) represents the value of the window function h(τ) at 0.
[0078] Step 3: Extract instantaneous frequency and instantaneous amplitude
[0079] For the instantaneous power spectrum transform (IPS) expression of the communication signal S(t), the expressions in the time domain and the frequency domain obtained from the instantaneous power spectrum properties can be derived:
[0080]
[0081]
[0082] Wherein, S(t) represents the communication signal, S(f) represents the Fourier transform of S(t);
[0083] Instantaneous frequency f i(t) The expression derived from the above formula is:
[0084]
[0085] Instantaneous frequency f i The energy density at (t) represents the amplitude of the communication signal, and the instantaneous amplitude is obtained by the intensity of the energy value at the instantaneous frequency;
[0086] Step 4: Instantaneous phase extraction
[0087] For the extraction of instantaneous phase sequence, nonlinear phase is selected as the phase characteristic sequence. The communication signal S(t) is transformed in the time domain through Hilbert transform to obtain the analytical signal Z(t) = H[S(t)] = x(t) + j*y(t). And perform the unwinding operation to obtain the unwinding phase: Then remove the linear phase component caused by the carrier frequency to obtain the nonlinear phase sequence:
[0088]
[0089] In the formula, Z(t) is the analytical signal after Hilbert transform; x(t) is the real part of the analytical signal; y(t) is the imaginary part of the analytical signal; φ(i) is the nonlinear phase; f c represents the carrier frequency; t is the time variable; f s represents the sampling frequency; i represents the sampling point;
[0090] Step 5: Sampling and reducing the dimension of feature sequences
[0091] The instantaneous amplitude sequence and instantaneous frequency sequence of the signal with a length of N obtained in step three and the instantaneous phase sequence with a length of N obtained in step four are subjected to multiple dimensionality reduction operations of sampling 1 every other time, so that the long sequence with a length of N is reduced to 2000 sampling points, where N is a natural number. This not only retains the instantaneous characteristics of the communication signal without loss, but also reduces the data length, which can reduce the time for subsequent neural network modulation recognition.
[0092] Step 6: Extract the instantaneous feature sequence as a data set
[0093] Three communication signals with different modulation formats, namely, multi-level amplitude keying signal MASK, multi-level frequency shift keying signal MFSK and multi-level phase shift keying signal MPSK, are selected. M signals are randomly selected from each communication signal, where M is a natural number and M≥500. Matlab software is used to simulate and generate the instantaneous amplitude, instantaneous frequency and phase sequences obtained in the above manner. M signals are randomly generated for each signal, and the obtained instantaneous amplitude, instantaneous frequency and instantaneous phase sequences are saved in .csv file format respectively. The three data sets are required to store the instantaneous amplitude sequence, instantaneous frequency sequence and instantaneous phase sequence respectively, and the sequence data corresponding to each data set comes from the same signal.
[0094] Step 7: Dataset Division
[0095] The instantaneous amplitude sequence set, the instantaneous frequency sequence set and the instantaneous phase sequence set are divided into a training set and a validation set according to a certain ratio, wherein the data in the training set and the validation set are sequence sets of known modulation types;
[0096] Step 8. Convolutional Neural Network-Long Short-Term Memory Network CNN-LSTM Model Construction
[0097] Advanced technologies such as deep learning can improve the accuracy of automatic modulation classification as classifiers because they can efficiently represent features, so features are extracted and classified at the same stage. This reduces the part of extracting feature parameters, and after data preprocessing, it is directly input into the neural network for feature extraction and classification and recognition process. Use python software to build a neural network classifier. Since the frequency of the MASK signal is constant, the amplitude reaches the maximum value at the carrier frequency, and the amplitude of the maximum value changes with the change of the signal amplitude; the amplitude of the maximum value of the MFSK signal is constant, and the maximum value appears at the instantaneous frequency of the signal; MPSK signals are distinguished by the phase of the modulated signal. Therefore, the six communication signals (2ASK, 4ASK, 2FSK, 4FSK, 2PSK, 4PSK) cannot be distinguished only by amplitude, frequency or phase;
[0098] In order to overcome the above reasons, a three-branch parallel fusion neural network is constructed. The three branches correspond to the amplitude feature sequence, frequency feature sequence and phase feature sequence corresponding to the same sequence number in the input training set or verification set for synchronous training and verification.
[0099] The training set is used to train the set three-branch parallel fusion convolutional neural network-long short-term memory network CNN-LSTM. Using the data in the training set and the validation set, the .CSV files of three different instantaneous feature sequences of known modulation types as training data are used as input, the training parameters are set, and the set convolutional neural network is trained to obtain the optimal convolutional neural network-long short-term memory network CNN-LSTM model and save it;
[0100] Step 9: Apply the convolutional neural network-long short-term memory network CNN-LSTM model to classify and identify the modulation category of communication signals
[0101] The trained neural network is saved, and the neural network model with the best training and parameter adjustment is saved. Finally, the multi-level amplitude keying signal MASK, multi-level frequency shift keying signal MFSK or multi-level phase shift keying signal MPSK that are pre-classified and identified are subjected to steps one to five, and the corresponding instantaneous amplitude sequence, instantaneous frequency sequence and instantaneous phase sequence after dimensionality reduction are obtained, and all are saved in the reserved corresponding sequence set as a test set in the .csv file format; the .CSV files of the three instantaneous feature sequences of the instantaneous amplitude sequence, instantaneous frequency sequence and instantaneous phase sequence of the communication signal pre-classified and identified in the test set are used as input, and are input into the optimal convolutional neural network-long short-term memory network CNN-LSTM model saved in step eight for classification and identification, so as to identify the modulation category of the communication signal.
[0102] The method for constructing the three-branch parallel fusion convolutional neural network-long short-term memory network CNN-LSTM model is specifically as follows:
[0103] ① Each of the three branches first selects a one-dimensional convolutional neural network for processing one-dimensional data and a long short-term memory network specifically for processing time series, that is, a multi-layer convolutional neural network-long short-term memory network (CNN-LSTM) serial network structure is selected;
[0104] ② The three branches use the same network structure and are set up in sequence.
[0105] Each branch first customizes the number of convolutional layers to I, the number of pooling layers to J, and the number of long short-term memory networks to K. Then, the number and size of convolution kernels in each convolutional layer, the parameters of the pooling layer, and the parameters of the long short-term memory layer are set.
[0106] ③ Connect the three set branches using the concatenate function and splice the features extracted from the three network branches;
[0107] ④ Set the number of fully connected layers and the parameters of each layer in the convolutional neural network-long short-term memory network CNN-LSTM and input the result after the splicing features in step ③;
[0108] ⑤ In the output layer of the convolutional neural network-long short-term memory network CNN-LSTM, the softmax function is used to output the multi-classification task neural network, and the modulation type of the output communication signal is the classification recognition result.
[0109] In step eight, the specific steps of using the data in the training set and the validation set to input the amplitude feature sequence, frequency feature sequence and phase feature sequence corresponding to the same sequence number in three branches one by one for synchronous training and validation are as follows:
[0110] ① Setting the initial parameters of the three-branch parallel fusion convolutional neural network-long short-term memory network CNN-LSTM and using the training set and the validation set to simultaneously train the neural network, the initial parameters include the number of iterations epoch, the learning rate, the gradient descent function and the loss function;
[0111] ② After the training is completed, the convolutional neural network - long short-term memory network CNN-LSTM will generate two accuracy curves corresponding to the training set and the validation set, as well as two loss value curves corresponding to the training set and the validation set. As the accuracy value increases, the loss value decreases accordingly. Due to the corresponding relationship between the two, only the loss value curve is used to judge the training situation. As the number of iterations increases, the loss value curve continues to decrease until it gradually flattens out. The flattening of the loss value curve means that the training of the convolutional neural network - long short-term memory network CNN-LSTM is basically completed, but it is not necessarily optimal. It depends on the degree of fit of the two loss value curves. Fit the two loss value curves corresponding to the training set and the validation set. If underfitting or overfitting occurs, jump to step ③. If there is no underfitting or overfitting, jump to step ⑤.
[0112] ③ Reset the number and parameters of convolutional layers, pooling layers, and long short-term memory layers of each branch, and reset the parameters such as the number of epochs, learning rate, gradient descent function, and loss function of the convolutional neural network-long short-term memory network CNN-LSTM;
[0113] ④ Repeat steps ① and ② for the convolutional neural network-long short-term memory network CNN-LSTM with new parameters set in step ③, using the training set and the validation set to train the convolutional neural network-long short-term memory network CNN-LSTM until the two loss value curves corresponding to the training set and the validation set are fitted and no underfitting or overfitting occurs;
[0114] ⑤Save the convolutional neural network-long short-term memory network CNN-LSTM as the optimal convolutional neural network-long short-term memory network CNN-LSTM model.
[0115] Example:
[0116] In order to enable those skilled in the art to better understand the method described in this embodiment, a specific example is given below to illustrate the method.
[0117] Simulation conditions: carrier frequency is 1000Hz, sampling frequency is 40000Hz, symbol rate is 50B, number of code elements is 16, number of sampling points is 12800, signal sequence S(n) is obtained, n=1, 2, ... 12800, window function selects Gaussian window, window length is set to 255, under Gaussian white noise condition, signal-to-noise ratio is set to 5dB;
[0118] 1. Communication signal acquisition
[0119] Set the modulation parameters and simulate the communication signals MASK, MFSK, and MPSK. Each type of modulation signal can be generated by loading the baseband signal onto the carrier signal. The generation method has become mature.
[0120] The communication modulation signal S(t) can be expressed as
[0121]
[0122] Where: S(t) is the communication signal; t is the time variable; f c is the carrier frequency; A(t) is the instantaneous amplitude; is the nonlinear phase. If the signal center frequency f c and bandwidth Δf satisfies the condition f c ≥Δf, the signal can be regarded as a narrowband signal.
[0123] Using matlab software simulation, the communication signal modulation results can be obtained under the above simulation conditions. Figure 2 The time domain waveform of the 4ASK signal under noise-free conditions is shown in FIG. Figure 3 The figure shows the time domain waveform of the 4FSK signal under noise-free conditions. Figure 4 Shown is the time domain waveform of the 4PSK signal under noise-free conditions.
[0124] 2. Calculate the instantaneous power spectrum of the signal
[0125] The communication signal S(t) is subjected to instantaneous power spectrum transformation (IPS). Instantaneous power spectrum transformation is a time-frequency analysis method proposed by Page, which is quite insensitive to window parameters. Even at the analysis endpoint, it maintains good time-frequency localization characteristics, and its expression is
[0126]
[0127] Where: IPS(t,f) is the instantaneous power spectrum (IPS) transform of signal S(t); t is the time variable; f is the frequency variable; τ is the time delay; S * is the conjugate of S; h(τ) is the window function, satisfying h(0)=1, where h(0) represents the value of the window function h(τ) at 0;
[0128] 3. Extraction of instantaneous frequency and instantaneous amplitude
[0129] For the instantaneous power spectrum transform (IPS) expression of the communication signal S(t), the expressions in the time domain and the frequency domain obtained from the instantaneous power spectrum properties can be derived:
[0130]
[0131]
[0132] Wherein, S(t) represents the communication signal, S(f) represents the Fourier transform of S(t);
[0133] Instantaneous frequency f i (t) The expression derived from the above formula is:
[0134]
[0135] Instantaneous frequency f i The energy density at (t) represents the amplitude of the communication signal, and the instantaneous amplitude is obtained by the intensity of the energy value at the instantaneous frequency;
[0136] Since the frequency of the MASK signal is constant, the amplitude reaches the maximum value at the carrier frequency, and the amplitude of the maximum value changes with the change of the signal amplitude; take the 4ASK signal as an example to observe the instantaneous amplitude. Figure 5 The figure shows the instantaneous amplitude of the 4ASK signal under noise-free conditions. Figure 6 This is a comparison chart of the instantaneous amplitude of the classic method and the instantaneous power spectrum method for 4ASK signal when the signal-to-noise ratio is 5dB;
[0137] Since the amplitude of the maximum value of the MFSK signal is constant, the maximum value appears at the instantaneous frequency of the signal; take the 4FSK signal as an example to observe the instantaneous frequency. Figure 7 It is the instantaneous frequency diagram of 4FSK signal under noise-free conditions. Figure 8 This is a comparison chart of the instantaneous frequency of the classic method and the instantaneous power spectrum method for 4FSK signals when the signal-to-noise ratio is 5dB;
[0138] 4. Instantaneous phase extraction
[0139] For the extraction of instantaneous phase sequence, nonlinear phase is selected as the phase characteristic sequence. The communication signal S(t) is transformed in the time domain through Hilbert transform to obtain the analytical signal Z(t) = H[S(t)] = x(t) + j*y(t). And perform the unwinding operation to obtain the unwinding phase: Then remove the linear phase component caused by the carrier frequency to obtain the nonlinear phase sequence:
[0140]
[0141] In the formula, Z(t) is the analytical signal after Hilbert transform; x(t) is the real part of the analytical signal; y(t) is the imaginary part of the analytical signal; φ(i) is the nonlinear phase; f c represents the carrier frequency; t is the time variable; f s represents the sampling frequency; i represents the sampling point;
[0142] Fig. 9 The instantaneous phase of the 4PSK signal under noise-free conditions is compared with the instantaneous phase of the 4PSK signal when the signal-to-noise ratio is 5dB;
[0143] 5. Sampling and dimensionality reduction of feature sequences
[0144] The instantaneous amplitude sequence, instantaneous frequency sequence and instantaneous phase sequence of the signal with a length of 12800 obtained above are subjected to multiple sampling operations every other 1, so that the long sequence with a length of 12800 is reduced to a sequence with a length of 1600 within 2000 sampling points. This not only retains the instantaneous characteristics of the communication signal without loss, but also reduces the data length, which can reduce the time for subsequent neural network modulation recognition.
[0145] like Fig.10 Shown is the instantaneous amplitude diagram after sampling and dimensionality reduction; Fig.11 It is the instantaneous frequency diagram after sampling and dimensionality reduction; Fig.12 is the instantaneous phase image after sampling and dimensionality reduction;
[0146] 6. Extract instantaneous feature sequences as data sets
[0147] Use matlab software to simulate and generate the instantaneous amplitude, instantaneous frequency and phase sequences of M multi-level amplitude keying signals MASK, M multi-level frequency shift keying signals MFSK and M multi-level phase shift keying signals MPSK in the above method. Generate 1000 signals of each type randomly, and save the instantaneous amplitude, instantaneous frequency and instantaneous phase sequences of each signal in .csv file format. When generating the data set, ensure that: 1) the three data sets store the instantaneous amplitude sequence, instantaneous frequency sequence and instantaneous phase sequence respectively, 2) the sequence data corresponding to each data set comes from the same signal sample, 3) label the signal samples, and when labeling, the three feature sequences of each signal sample are labeled with the same label, and 4) each type of signal is labeled with the same label.
[0148] 7. Convolutional Neural Network-Long Short-Term Memory Network CNN-LSTM Model Construction
[0149] Advanced technologies such as deep learning can improve the accuracy of automatic modulation classification as classifiers because they can efficiently represent features, so features are extracted and classified at the same stage. This reduces the part of extracting feature parameters and directly inputs them into the neural network for feature extraction and classification recognition after data preprocessing.
[0150] Use python software to build a neural network classifier and build a three-branch parallel fusion neural network. Each branch inputs the amplitude feature sequence, frequency feature sequence and phase feature sequence respectively. The input sequence length of each branch is 1600.
[0151] Set the structure and parameters of the convolutional neural network-long short-term memory network CNN-LSTM model. The neural network selected for each branch is a multi-layer convolutional neural network-long short-term memory network CNN-LSTM serial network structure. First, set 10 one-dimensional convolutional layers and 4 maximum pooling layers. Set 10 convolutional layers to use one-dimensional convolution. The first and second convolutional layers use 64 convolutional kernels of size 1*15, and then connect the pooling layer; the third and fourth convolutional layers use 128 convolutional kernels of size 1*13, and then connect the pooling layer; the fifth to seventh convolutional layers use Use 256 convolution kernels of size 1*11, and then connect the pooling layer; the eighth to tenth convolution layers all use 512 convolution kernels of size 1*9, and then connect the pooling layer; the four pooling layers all use the maximum pooling of size 1*2, and then set the parameter of the long short-term memory layer to 10, and then connect the three branches together through the concatenate function, and then set two fully connected layers, the number of neurons is 256 and 128 respectively, the activation function uses the relu activation function, and finally the softmax activation function is selected as the output layer, and the number of neurons in the output layer is 6;
[0152] 8. Input the extracted instantaneous feature sequence into the convolutional neural network-long short-term memory network CNN-LSTM model for modulation recognition
[0153] The original data set of 1000 signals of each type is divided into 70% training set and 30% validation set. The set CNN-LSTM network is trained with the training set and validation set at the same time. The .CSV files of the three different instantaneous feature sequences of the instantaneous amplitude sequence, instantaneous frequency sequence and instantaneous phase sequence of known modulation types as training data are used as the input of the three network branches. The training parameters are set to train the set convolutional neural network.
[0154] According to whether the loss function value curves drawn by the training set and the validation set are overfitting or underfitting, the hyperparameters of the neural network such as learning rate, number of iterations, loss function and optimizer are adjusted, and the internal weight parameters of the model are updated through the back propagation algorithm until the optimal network parameters are reached; the training is saved and the optimal convolutional neural network-long short-term memory network CNN-LSTM model is adjusted;
[0155] 9. Repeat steps 1-5 to regenerate 300 signal samples, and obtain 300 instantaneous amplitude sequences, instantaneous frequency sequences, and instantaneous phase sequences respectively. Make them into .CSV files as test sets, and input the test sets of the three instantaneous feature sequences into the above-trained optimal convolutional neural network-long short-term memory network CNN-LSTM model for classification and recognition.
[0156] The classic instantaneous statistical feature parameter extraction method is sensitive to noise under low signal-to-noise ratio. At the same time, the decision tree, support vector machine and BP neural network classifier of machine learning used in most classic methods for classification and identification need to extract parameters after signal preprocessing. The present invention provides a communication signal modulation mode recognition method based on instantaneous features. According to the characteristics of MASK, MFSK, MPSK (specifically six communication signals of 2ASK, 4ASK, 2FSK, 4FSK, 2PSK and 4PSK), the instantaneous power spectrum is used to obtain the instantaneous amplitude and instantaneous frequency sequence of the signal, the classical Hilbert transform is used to obtain the phase sequence, and a three-branch parallel convolutional neural network-long short-term memory network is used to classify and identify the one-dimensional instantaneous feature sequence, thereby reducing the steps of extracting feature parameters. The present invention is a communication signal modulation mode recognition method based on instantaneous feature sequence under Gaussian white noise background.
Claims
1. A signal modulation identification method based on instantaneous characteristics, characterized in that: The method comprises the following steps, and the following steps are performed in sequence: Step 1: Acquisition of communication signals A baseband signal is processed in different modulation modes to obtain a modulated multi-level amplitude keying signal MASK, a multi-level frequency shift keying signal MFSK or a multi-level phase shift keying signal MPSK. Matlab software is used to simulate the above three types of modulation signals, and the expression of the modulated communication signal S(t) is obtained as follows: Where: S(t) is the communication signal; t is the time variable; f c is the carrier frequency; A(t) is the instantaneous amplitude; is the nonlinear phase; Step 2: Make the instantaneous power spectrum of the signal Perform instantaneous power spectrum transformation IPS on the communication signal S(t), and obtain the instantaneous power spectrum expression of the signal as follows: Where: IPS(t,f) is the instantaneous power spectrum (IPS) transform of signal S(t); t is the time variable; f is the frequency variable; τ is the time delay; S * is the conjugate of S; h(τ) is the window function, satisfying h(0)=1, where h(0) represents the value of the window function h(τ) at 0; Step 3: Extract instantaneous frequency and instantaneous amplitude According to the instantaneous power spectrum expression of the signal, the expressions in the time domain and frequency domain obtained by the properties of the instantaneous power spectrum are derived: Wherein, S(t) represents the communication signal, S(f) represents the Fourier transform of S(t); Instantaneous frequency f i (t) The expression derived from the above formula is: Instantaneous frequency f i The energy density at (t) represents the amplitude of the communication signal, and the instantaneous amplitude is obtained by the intensity of the energy value at the instantaneous frequency; Step 4: Instantaneous phase extraction Select the nonlinear phase as the phase characteristic sequence, and transform the communication signal S(t) in the time domain through Hilbert transform to obtain the analytical signal Z(t)=H[S(t)]=x(t)+j*y(t). And perform the unwinding operation to obtain the unwinding phase: Then remove the linear phase component caused by the carrier frequency to obtain the nonlinear phase sequence: In the formula, Z(t) is the analytical signal after Hilbert transform; x(t) is the real part of the analytical signal; y(t) is the imaginary part of the analytical signal; φ(i) is the nonlinear phase; f c represents the carrier frequency; t is the time variable; f s represents the sampling frequency; i represents the sampling point; Step 5: Sampling and reducing the dimension of feature sequences The instantaneous amplitude sequence and instantaneous frequency sequence of the signal with a length of N obtained in step 3 and the instantaneous phase sequence with a length of N obtained in step 4 are subjected to multiple dimensionality reduction operations of sampling every other 1, so that the long sequence with a length of N is reduced to within 2000 sampling points, where N is a natural number; Step 6: Extract the instantaneous feature sequence as a data set Select a multi-level amplitude keying signal MASK, a multi-level frequency shift keying signal MFSK, and a multi-level phase shift keying signal MPSK, and randomly select M of each modulation signal, where M is a natural number, repeat steps one to five for each communication signal, and respectively obtain M corresponding instantaneous amplitude, instantaneous frequency, and instantaneous phase sequences after dimensionality reduction, save the M instantaneous amplitude sequences in a .csv file format to form an instantaneous amplitude sequence set, save the M instantaneous frequency sequences in a .csv file format to form an instantaneous frequency sequence set, save the M instantaneous phase sequences in a .csv file format to form an instantaneous phase sequence set, and the sequence data corresponding to the same sequence number in each data set comes from the same signal; Step 7: Dataset Division The three instantaneous feature sequence sets, namely, the instantaneous amplitude sequence set, the instantaneous frequency sequence set and the instantaneous phase sequence set, are divided into a training set and a validation set according to a set ratio; Step 8. Convolutional Neural Network-Long Short-Term Memory Network CNN-LSTM Model Construction Construct a three-branch parallel fusion convolutional neural network-long short-term memory network CNN-LSTM model. Using the data in the training set and the validation set, the three branches input the amplitude feature sequence, frequency feature sequence and phase feature sequence corresponding to the same sequence number one by one for synchronous training and validation, and obtain the optimal convolutional neural network-long short-term memory network CNN-LSTM model; Step 9: Apply the convolutional neural network-long short-term memory network CNN-LSTM model to classify and identify modulation categories Perform steps 1 to 5 on the multi-level amplitude keying signal MASK, multi-level frequency shift keying signal MFSK or multi-level phase shift keying signal MPSK that has been pre-classified and identified, and obtain the corresponding instantaneous amplitude sequence, instantaneous frequency sequence and instantaneous phase sequence after dimensionality reduction, all of which are saved in the reserved corresponding sequence set in .csv file format as a test set; The .CSV file of the instantaneous amplitude sequence, instantaneous frequency sequence and instantaneous phase sequence of the communication signal that has been pre-classified and identified is used as input and input into the optimal convolutional neural network-long short-term memory network CNN-LSTM model obtained in step eight to identify the modulation category of the communication signal.
2. The signal modulation identification method based on instantaneous characteristics according to claim 1 is characterized in that: The method for constructing the three-branch parallel fusion convolutional neural network-long short-term memory network CNN-LSTM model in step eight is specifically as follows: ① Each of the three branches first selects a one-dimensional convolutional neural network for processing one-dimensional data and a long short-term memory network specifically for processing time series, that is, a multi-layer convolutional neural network-long short-term memory network CNN-LSTM serial network structure is selected; ② The three branches use the same network structure and are set up in sequence. Each branch first customizes the number of one-dimensional convolutional layers, the number of pooling layers, and the number of long short-term memory networks (LSTMs), and then sets the number and size of convolution kernels in each convolutional layer, the parameters of the pooling layer, and the parameters of the long short-term memory layer; ③ Connect the three set branches using the concatenate function, and splice the features extracted from the three network branches; ④ Set the number of fully connected layers and the parameters of each layer in the convolutional neural network-long short-term memory network CNN-LSTM and input the result after splicing features in step ③; ⑤ In the output layer of the convolutional neural network-long short-term memory network CNN-LSTM, the softmax function is used to output the multi-classification task neural network, and the modulation type of the output communication signal is the classification recognition result.
3. The signal modulation identification method based on instantaneous characteristics according to claim 2 is characterized in that: In step eight, the specific steps of using the data in the training set and the validation set to input the amplitude feature sequence, frequency feature sequence and phase feature sequence corresponding to the same sequence number in three branches one by one for synchronous training and validation are as follows: ① Setting the initial parameters of the three-branch parallel fusion convolutional neural network-long short-term memory network CNN-LSTM and using the training set and the validation set to simultaneously train the neural network, the initial parameters include the number of iterations epoch, the learning rate, the gradient descent function and the loss function; ② After the training is completed, the convolutional neural network-long short-term memory network CNN-LSTM generates two loss value curves corresponding to the training set and the validation set, and fits the two loss value curves. If underfitting or overfitting occurs, jump to step ③, and if there is no underfitting or overfitting, jump to step ⑤; ③ Reset the number and parameters of convolutional layers, pooling layers, and long short-term memory layers of each branch, and reset the number of epochs, learning rate, gradient descent function, and loss function of the convolutional neural network-long short-term memory network CNN-LSTM; ④ Repeat steps ① and ② for the convolutional neural network-long short-term memory network CNN-LSTM with new parameters set in step ③, using the training set and the validation set to train the convolutional neural network-long short-term memory network CNN-LSTM until the two loss value curves corresponding to the training set and the validation set are fitted and no underfitting or overfitting occurs; ⑤Save the convolutional neural network-long short-term memory network CNN-LSTM as the optimal convolutional neural network-long short-term memory network CNN-LSTM model.
