CN112307927A - Research on recognition of MPSK signal in non-cooperative communication based on BP network - Google Patents

Abstract

The present invention proposes a BP-GA neural network model for intra-class recognition of MPSK signals in non-cooperative communication systems. Firstly, according to the characteristics of MPSK signal, six features based on time domain and frequency domain are selected as the input samples of the model; a BP neural network with two hidden layers is designed as the classifier for modulation recognition, and the genetic algorithm is used to analyze the neural network. The parameters are optimized; finally, in order to reduce the sensitivity of the model to the signal-to-noise ratio, the order of the signal-to-noise ratio of the training samples is disrupted and then used as the input of the network model for network training. Compared with the existing modulation recognition algorithm based on BP neural network, the recognition accuracy of MPSK signal under low signal-to-noise ratio is improved. In addition, the present invention has strong achievability and high recognition accuracy, and can be well applied to related projects of non-cooperative communication systems.

Description

Research on recognition of MPSK signal in non-cooperative communication based on BP network

technical field

The invention relates to a related algorithm of a neural network and a related theory of signal processing, and belongs to the field of communication signal processing and artificial intelligence.

Background technique

With the commercial use of 5G, the research on next-generation mobile communication technology has been launched, and new communication technologies such as satellite communication, spread spectrum communication, and frequency hopping communication have also begun to be widely used. Wireless communication technology has become the fastest growing and most widely used. One of the most advanced communication technologies, followed by a more and more complex electromagnetic environment, and the modulation methods of signals also show a diversified development trend, and in non-cooperative communication systems such as radio spectrum resource supervision and modern electronic warfare, application The automatic modulation identification technology of the signal is a very critical step.

Early identification of signal modulation was achieved through manual interpretation of measured parameters, and therefore heavily relied on the skill level and work experience of the operator. In April 1969, Weaver C S et al. published the first paper on automatic identification of modulation modes in Stanford University Technical Report - "Using Pattern Recognition Technology to Realize Automatic Classification of Modulation Types", thus enabling automatic signal modulation identification. door.

Classical modulation recognition techniques can be divided into two categories: 1. Maximum Likelihood (ML) methods based on hypothesis testing, and 2. Pattern recognition methods based on feature extraction. The former can be viewed as a multiple hypothesis testing problem. It is to theoretically deduce the test statistic (usually using likelihood ratio function) of the intercepted signal under the condition of background interference, find an appropriate threshold, and then make a decision under the minimum Bayesian cost criterion. The latter uses the method of feature extraction to realize modulation mode identification, and it needs to select an appropriate classifier for classification and identification. The task of the classifier is to classify a given input pattern represented by a feature vector into an appropriate pattern category according to a certain criterion, complete the mapping from the feature space to the decision space, and finally give the recognition result. Commonly used classifiers are: decision tree classifier, nearest neighbor (KNN) classifier, Bayesian classifier, support vector machine (SVM), random forest, etc.

Nowadays, with the continuous development of artificial intelligence technologies such as deep learning and machine learning, the application of neural networks as classifiers in the field of modulation recognition has become a major research trend. There are two main types of neural network models used in the field of modulation recognition: 1. Convolutional Neural Network (CNN), 2. Backpropagation (BP) Neural Network. CNN mainly consists of convolutional layers, pooling layers and fully connected layers. The main function of the convolutional layer is to extract the features of the image, the main function of the pooling layer is downsampling, and the main function of the fully connected layer is classification. Convolutional Neural Networks are mainly used in the field of image recognition, so CNNs are usually used together with the relevant graphic features of the signal (such as constellation diagrams) when applying CNNs to modulation mode recognition. In 2019, Weng Jianxin et al. designed a CNN-LSTM parallel network, which directly uses the co-directional component and the quadrature component as input data, without the need to manually design feature parameters, reducing the influence of human factors. The algorithm has a relatively low signal-to-noise ratio. good recognition performance. In 2019, Siyang Zhou et al. proposed a robust automatic modulation recognition method based on convolutional neural network in view of the lack of generalization of the current modulation recognition model, which can recognize 15 kinds of signals, and under low signal-to-noise ratio. Recognition accuracy is also good. In 2020, Chen Changmei et al. proposed an improved convolutional neural network structure that can classify seven different modulation signals. When the signal-to-noise ratio is not less than 5dB, the recognition rate can reach 97.99%. When it is less than 9dB, the recognition rate can reach 100%. The BP neural network includes three parts: input layer, hidden layer and output layer. It mainly solves the weight learning problem of the hidden layers of the multi-layer neural network by back-propagating the error. The application of BP neural network to the field of modulation recognition is usually combined with the instantaneous characteristics of the signal. In 2016, Wang Yi and others trained the neural network through the variable gradient BP correction algorithm to improve the convergence speed and shorten the training time. When the signal-to-noise ratio was 10dB, the recognition rate reached 95%. In 2019, Wu Xiquan and others proposed a signal modulation recognition algorithm based on BP neural network. When the signal-to-noise ratio is 0dB, the recognition rate can reach more than 85%. In 2019, Yuan Meng et al. used the BP neural network algorithm to automatically identify six common digital modulation signals. When the signal-to-noise ratio was greater than 10dB, the accuracy rate reached more than 98%.

Based on the current application of neural network in modulation recognition, it can be seen that the current application of ANN in modulation recognition can be divided into two categories: one is the modulation recognition algorithm based on convolutional neural network, which identifies the modulation recognition of the signal. The problem is converted into an image recognition problem, which does not require any prior information of the signal, and the difficulty lies in the extraction of the graphical features of the signal; the other is the modulation recognition algorithm based on deep neural network (such as BP neural network), which is to use ANN Identifying as a classifier, the difficulty lies in the optimization of parameters during network training.

SUMMARY OF THE INVENTION

Based on the problem that the recognition accuracy of the commonly used neural network model is not high enough under the low signal-to-noise ratio in the signal modulation recognition discussed above, the present invention provides a BP neural network as the basic network model. The network parameters are optimized, and finally a method for improving the accuracy of phase shift keying (MPSK) signal recognition under low signal-to-noise ratio is realized. See Figure 1 for the system flow design diagram of the present invention.

The characteristics of the BP neural network based on genetic algorithm (GA) of the present invention are as follows:

In order to overcome the shortcoming that the BP neural network model is easy to fall into the local minimum value and the shortcomings of the existing neural network algorithm for signal modulation identification, the present invention optimizes the BP neural network by using the GA algorithm, and proposes a new intra-class identification method for MPSK signals. The BP-GA network model. The model preprocesses the features of the signal and uses a BP neural network with two hidden layers for training, and uses the GA algorithm to solve the problem of slow convergence of the BP neural network and easy to fall into local optimum. The invention greatly reduces the sensitivity of the network model to the signal-to-noise ratio, and improves the recognition rate of the PSK signal under low signal-to-noise ratio. This method has been applied to scientific research projects.

The algorithm for intra-class identification of MPSK signal based on BP neural network according to the present invention comprises the following steps:

1) Extract three instantaneous features of MPSK signal and three features based on high-order cumulants as the input of the network model;

2) Construct a BP neural network model with two hidden layers as a classifier for identifying MPSK signals;

3) Use the GA algorithm to optimize the BP network parameters, and use the output error of the network model as the fitness function to obtain the optimal BP neural network parameters;

4) Transfer the optimal BP neural network parameters into the network model, and use the six signal features as the input of the network model to train the BP network model;

5) For the output result of the network model, it is necessary to perform inverse normalization processing, and then the final recognition result can be obtained through judgment;

6) Integrate the modules in the above steps into a program, when a set of MPSK signal values are input, the recognition result can be output directly.

The above-mentioned step 1) uses three instantaneous features of the MPSK signal and three features based on high-order cumulants, specifically including: the instantaneous features of the signal mainly include instantaneous frequency, instantaneous amplitude and instantaneous phase, based on these three signals of the signal. Instantaneous characteristics, the five commonly used transient characteristics of signals are: the maximum value of the normalized instantaneous amplitude spectral density of the zero center, the standard deviation of the absolute value of the zero-centered normalized instantaneous amplitude, and the standard of the nonlinear component of the instantaneous phase of the zero-centered non-weak signal segment. Deviation, the standard deviation of the absolute value of the instantaneous phase nonlinear component of the zero-centered non-weak signal segment, and the standard deviation of the normalized instantaneous frequency absolute value of the zero-centered non-weak signal segment. In view of the characteristics of the MPSK signal, the present invention selects the standard deviation of the instantaneous phase nonlinear component of the zero-center non-weak signal segment, the standard deviation of the absolute value of the instantaneous phase nonlinear component of the zero-center non-weak signal segment, and the zero-center non-weak signal segment normalization. The standard deviation of the absolute value of the normalized instantaneous frequency is used as a partial feature sample.

The high-order cumulant of the signal belongs to the frequency domain characteristics of the signal, which can represent the high-order statistical characteristics of the signal. Order statistics have good anti-fading properties. The present invention selects three high-order cumulant-based features as part of the feature samples to train the network.

In the above-mentioned step 2), a corresponding BP neural network model is designed for the features extracted from the MPSK signal. Since the selected set of eigenvalues has a total of 6 features, the input layer of the BP neural network contains 6 nodes; since this modelist performs intra-class recognition for MPSK (M=2, 4, 8, 16) signals, the BP neural network The output layer of the neural network contains 4 nodes; the number of nodes in the middle two hidden layers is found through experimental simulation analysis: when the number of nodes in the first hidden layer is set to 40 and the number of nodes in the second hidden layer is set to 10 , the recognition effect is the best, and the specific BP neural network model design invention diagram is shown in Figure 2.

The above-mentioned step 3) in order to speed up the convergence speed of the BP neural network and avoid falling into the local optimum, the genetic algorithm is used to optimize the parameters of the network, and the specific algorithm optimization strategy is as follows:

The genetic algorithm selected in the present invention is a parallel random search optimization method simulating the natural genetic mechanism and the theory of biological evolution. The algorithm selects the output error of the neural network as the fitness function. First, the network parameters need to be encoded as population individuals, and then individuals are screened through selection, crossover and mutation operations in genetics, so that individuals with good fitness values are retained. Individuals with poor fitness values are eliminated, and the new group inherits the information of the previous generation and is superior to the previous generation. This cycle is repeated until the conditions are met.

Among them, the selection operation refers to selecting individuals from the old group to the new group. The probability of an individual being selected is related to its fitness value. The smaller the fitness value, the greater the probability of being selected. The crossover operation means that the chromosomes of two individuals in the population randomly select one or more positions to exchange to generate a new individual. The schematic diagram of the specific crossover operation is shown in FIG. 3 . Mutation operation refers to selecting an individual's chromosome from the population, and mutating one point to generate a better individual. The schematic diagram of the specific mutation operation is shown in Figure 4. The crossover probability and mutation probability set in the present invention are respectively 0.3 and 0.1.

Step 4) is to transfer the BP neural network parameters obtained through genetic algorithm optimization in step 3) into the BP neural network model, and in the case of optimal network parameters, use the six eigenvalues extracted in step 1) to pair the network. The model is trained.

Step 5) is to further process the results obtained by the network model - inverse normalization, threshold judgment, so as to obtain the final identification result.

Step 6) is to obtain a trained BP neural network model with the help of the above four steps, and test the MPSK signal recognition effect of the BP neural network classifier by inputting the test sample into the network model.

The main effect of the present invention is to perform intra-class identification for MPSK signals in non-cooperative communication systems under low SNR, reduce the sensitivity of network model to SNR, and improve the accuracy of signal identification under low SNR. details as follows:

The performance metrics for evaluating the model are P (precision rate), R (recall rate) and F (comprehensive performance index F value). Precision rate and recall rate are widely used in the fields of information retrieval and statistical classification. quality plays an important role. P (accuracy rate) represents the proportion of accurate predictions in the system, and measures the precision rate in the system. R (recall rate) represents the ratio of true predictions to all positive samples in the system, and measures the recall rate in the system. Let TP represent the number of positive samples predicted as positive samples, and FP represent the number of false positives predicted from negative samples as positive samples. FN represents the number of false positives that are predicted to be negative samples. The relationship between the three and P, R is shown in formulas (1) and (2). But in general, there is a certain contradiction between the precision rate and the recall rate, and there will be a phenomenon that the precision rate is high, but the recall rate is low, and vice versa. Therefore, at this time, it is necessary to comprehensively consider the accuracy rate and recall rate indicators, that is, use the F-measure evaluation method. The relationship between F-measure and P (precision) and R (recall) is shown in Equation 3.

Description of drawings

Fig. 1 is the system flow design diagram of the present invention

Fig. 2 is the schematic diagram of the BP neural network model designed by the present invention

Fig. 3 is a schematic diagram of crossover operation of genetic algorithm in the present invention

Fig. 4 is the schematic diagram of mutation operation of genetic algorithm in the present invention

Detailed ways

The present invention is called BP-GA model by combining genetic algorithm and BP neural network. The identification model of the present invention is mainly divided into: a feature extraction module and a classifier module composed of a BP-GA model. First, analyze the above modules. The specific steps are as follows:

Step 1: Use the mathematical calculation function to extract the feature of the signal;

Preprocessing part: Calculate the instantaneous amplitude, instantaneous phase, instantaneous frequency and high-order cumulant of the received signal. The specific calculation formula is as follows:

Let the signal be x(t), then the complex analytical formula of the signal can be obtained as

s(t)=x(t)+jy(t) (1)

where y(t) is the Hilbert transform of the signal x(t), i.e.,

From the complex analytical formula of the signal, its instantaneous amplitude and instantaneous phase can be obtained as:

The instantaneous frequency of the signal is the derivative of the instantaneous phase, so for discrete signals, the instantaneous frequency can be calculated by the difference of the instantaneous phase, namely:

For a stationary complex random process X(k) with zero mean, its mixing distance is defined as:

M _pq = E{[x(k)] ^pq [x ^* (t)] ^q } (6)

Where * denotes complex conjugate operation and E{·} is the mathematical expectation. The various higher-order cumulants of the signal can be expressed as:

Feature calculation part: The six feature values of the signal used in the present invention are obtained based on the calculation of the above instantaneous features of the signal and the high-order accumulators as follows:

(1) Feature 1

为零中心瞬时非线性相位。Among them, where C represents the number of sequences of non-weak signals (the signal amplitude is greater than the decision threshold _at ),

Zero-centered instantaneous nonlinear phase.

(2) Feature 2

(3) Feature 3

为零中心归一化瞬时幅度，R_b为码元速率，

为零中心频率。in,

zero-centered normalized instantaneous amplitude, R _b is the symbol rate,

zero center frequency.

(4) Feature 4

(5) Feature 5

(6) Feature 6

Data set processing part: Calculate the number of eigenvalue groups of the above six eigenvalues of each signal under different binary numbers and different signal-to-noise ratios. The number of eigenvalue groups is 2600, and the input dimension of this data set is 10400×6. The input samples are randomly scrambled to reduce the sensitivity of the network model to signal noise; the data set is divided reasonably, the data is uniformly normalized and standardized, the input dimension is unified, and the ratio of training samples to test samples is specified as 9: 1.

Step 2: Genetic algorithm optimizes the BP network model;

(1) Set the relevant parameters of the BP neural network model and the genetic algorithm;

In the present invention, two hidden layers are set in the BP neural network, and the number of hidden layer nodes is set to 40 and 10 respectively; the number of network training times is set to 1000, the learning rate is set to 0.01, and the training target is set to 0.0001; since the input sample and expected output are determined, Therefore, the number of nodes in the input layer and output layer of the neural network is determined, which are 6 and 4 respectively.

The relevant parameters of the genetic algorithm in the present invention are set as follows: the population size is set to 30, and the evolutionary algebra is set to 50; the upper and lower boundary values of the population are respectively set to -3 and 3; the probability of crossover and mutation are respectively set to: 0.3 and 0.1; The fitness function of the genetic algorithm is the error value between the actual output and the expected output of the BP neural network, and its expression is:

Among them, y _out is the expected output, and y _real is the actual output. In the present invention, the smaller the value of the fitness function, the better.

(2) Optimizing BP neural network parameters

After the BP neural network model is built, the initial parameters of the network (threshold and weight of each layer) are input into the genetic algorithm, the parameters are first encoded, and then the selection, crossover and mutation operations of the genetic algorithm are used to iterate continuously. The value of the parameter is changed, and the fitness value of each individual is calculated after each iteration to determine the individuals that are eliminated and left. Finally, when the maximum number of iterations is reached or the fitness value meets the requirements, the optimal BP neural network parameters are obtained.

Step 3: Train the network model;

Steps 1 and 2 respectively complete the two modules of feature extraction and network parameter optimization of the present invention, and then the extracted features can be used as input samples to train the BP neural network using the optimal network model parameters.

Since the extracted samples are arranged in the order from low to high signal-to-noise ratio under different signal-to-noise ratios, the obtained samples need to be preprocessed before training, that is, the order of the samples is randomly shuffled to reduce the network model. Sensitivity of the model to signal noise. After that, 90% of the feature values are selected as training samples to train the model. Stop network training when the maximum number of iterations is reached or the error reaches the desired target. What is obtained at this time is the trained BP neural network model, which can realize the intra-class recognition of MPSK signals.

Step 4: Process the output result of the network model.

Since the input data set is normalized when processing the data set, after the output of the BP neural network model is obtained, the output result needs to be de-normalized, and then the obtained model output result is judged , the result obtained at this time is the final recognition result. The determination threshold set in the present invention is 0.5. When the result is greater than 0.5, the recognition result is determined as 1, otherwise the recognition result is determined as 0.

Step 5: Test the model.

Input the test sample into the trained BP neural network model to get the recognition result. Afterwards, specific values are calculated according to the performance metrics of the above-mentioned model—accuracy rate, recall rate, and comprehensive performance metrics, so as to evaluate the performance of the model.

The invention mainly realizes the intra-class identification for MPSK signals in non-cooperative communication systems by means of BP neural network and genetic algorithm. The network model building and training and testing functions and the functions involved in the basic realization of the genetic algorithm have been open sourced. The present invention proposes a BP-GA network model architecture and uses the genetic algorithm to improve the convergence speed of the network model and avoid the network falling into local optimum. The invention can improve the recognition accuracy of MPSK signal under low signal-to-noise ratio, and at the same time reduce the sensitivity of the network model to signal noise.

Claims (5)

1. This patent proposes a BP neural network model for intra-class recognition of MPSK signals in non-cooperative communication systems. The accuracy of this model under low signal-to-noise ratio is better than previous models. It has been used in actual projects. The algorithm model structure of this patent mainly includes the following steps: 1) Extract three instantaneous features of MPSK signal and three features based on high-order cumulants as the input of the network model; 2) Construct a BP neural network model with two hidden layers as a classifier for identifying MPSK signals; 3) Use the GA algorithm to optimize the BP network parameters, and use the output error of the network model as the fitness function to obtain the optimal BP neural network parameters; 4) Transfer the optimal BP neural network parameters into the network model, and use the six signal features as the input of the network model to train the BP network model; 5) For the output result of the network model, it is necessary to perform inverse normalization processing, and then the final recognition result can be obtained through judgment; 6) Integrate the modules in the above steps into a program, when a set of MPSK signal values are input, the recognition result can be output directly. 2. Based on the feature extraction module described in claim 1, it is characterized in that: based on step 1) by analyzing the feature of MPSK signal, three features based on instantaneous features and three features based on high-order cumulants are selected as feature samples. The features extracted by the predecessors are only a simple superposition of a single type of features or features obtained through complex operations, and the features are not representative enough and the time complexity is too high. The features used in this patent combine the time-domain features and frequency-domain features of the target signal, and the features are more representative and can better reflect the characteristics of PSK signals of different systems. The specific feature extraction includes: Instantaneous characteristics: The commonly used transient characteristics of signals mainly include the instantaneous frequency, the instantaneous amplitude and the instantaneous phase. Due to the difference in phase between PSK signals of different system numbers, this patent selects the characteristics of the PSK signal based on the instantaneous phase and the instantaneous frequency to calculate the characteristic. The three instantaneous features are as follows: (1) Feature 1

Among them, where C represents the number of sequences of non-weak signals (the signal amplitude is greater than the decision threshold _at ),

Zero-centered instantaneous nonlinear phase. (2) Feature 2

(3) Feature 3

in,

zero-centered normalized instantaneous amplitude, R _b is the symbol rate,

zero center frequency. High-order cumulant: The high-order cumulant of the signal is the frequency domain characteristic of the signal. In view of the insensitivity of the high-order cumulant to some Gaussian white noise, this patent obtains three based on the high-order cumulant through the operation of the high-order cumulant. Quantitative frequency domain features. The specific feature calculation formula is as follows: (4) Feature 4

(5) Feature 5

(6) Feature 6

The above steps can obtain the characteristics of the target signal, and some processing is required to obtain the final data set. First, in order to reduce the sensitivity of the network model to noise, the obtained features need to be subjected to SNR disorder processing; after that, the samples need to be normalized to obtain the final input data set. In this patent, the number of eigenvalue groups of the above-mentioned six eigenvalues of each signal under different system numbers and different signal-to-noise ratios is 2600, and the final input dimension of the data set of this patent is 10400×6. 3. The BP neural network model according to claim 1, characterized in that: based on step 2) the BP neural network model of this patent is composed of an input layer, two hidden layers and an output layer. The feature data extracted from the signal can determine that the number of nodes in the input layer of the BP neural network model in this patent is 6; the number of nodes in the output layer of the BP neural network model in this patent can be determined from the number of types of target signals to be 4; the remaining two The number of nodes in each hidden layer Through the previous simulation experiments, the model performance indicators for measuring different numbers of nodes are compared, and the number of hidden layer nodes of the BP neural network in this patent is finally determined to be 40 and 10. In the present invention, the activation function of the hidden layer of the BP neural network uses the hyperbolic tangent function, and its expression is:

The activation function of the output layer uses the linear function purelin, and its expression is: y=purelin(x)=x (8) The training function uses the trainlm function. For the medium-scale BP neural network convergence speed block, and because the algorithm avoids the direct calculation of the Hessian matrix, the training calculation amount is small. This function uses the Levenberg-Marquardt algorithm, which can be implemented by modifying the parameters to achieve the advantages of combining the Gauss-Newton algorithm and the gradient descent method, that is, reducing the parameter value when the descent is too fast to make it closer to the Gauss-Newton method. ; Increase the parameter value to make it closer to gradient descent when the descent is too slow. 4. the present invention adds genetic algorithm in BP neural network to carry out the optimization of network parameters, it is characterized in that: based on step 3) genetic algorithm used in this patent is mainly encoded by individual, these three kinds of genetic operations of selection, crossover, variation, The fitness function calculation consists of these three parts. First, the initialization parameters of the BP neural network model are input into the algorithm to encode the parameters. The coding method selected in this patent is real number coding; then the fitness value of each individual is calculated, and the fitness function selected in this patent is the output error of the BP neural network, and its expression is:

Among them, y _out is the expected output, and y _real is the actual output. In the present invention, the smaller the value of the fitness function, the better. Therefore, after each iteration, the best individual fitness value will be found as the fitness value of the population. In the genetic operation, the roulette method is used for the selection operation, and the fitness value is an important metric for selection; in the crossover operation and mutation operation, individuals are selected for crossover or mutation according to the set probability. At the same time, this patent sets the boundary value of the individuals in the population, and checks whether the operation is effective when performing each genetic operation, so as to ensure the validity of the individuals in the population. 5. according to claim 1,2,3,4 for the BP neural network model of the identification in the class of MPSK signal in the non-cooperative communication system, it is characterized in that: step 4), 5) are the core steps of network model training Specifically, the input data set obtained in step 1) is put into the BP-GA network model obtained through steps 2) and 3) to train the network model. Then through the steps of inverse normalization, threshold judgment, etc., the output results of the network model are further processed, and finally the identification results are obtained. Finally, according to step 6), put the test sample into the trained sample to obtain the recognition result, and calculate the relevant parameters according to the performance index of the model to measure the performance of the model. The invention proposes a BP network model framework and uses the genetic algorithm to improve the convergence speed of the network model and avoid the network from falling into local optimum at the same time. The invention can improve the recognition accuracy of MPSK signal under low signal-to-noise ratio, and at the same time reduce the sensitivity of the network model to signal noise.

Images