This paper implements a deep learning-based modulation pattern recognition algorithm for communication signals using a convolutional neural network architecture as a modulation recognizer. In this paper, a multiple-parallel complex convolutional neural network architecture is proposed to meet the demand of complex baseband processing of all-digital communication signals. The architecture learns the structured features of the real and imaginary parts of the baseband signal through parallel branches and fuses them at the output according to certain rules to obtain the final output, which realizes the fitting process to the complex numerical mapping. By comparing and analyzing several commonly used time-frequency analysis methods, a time-frequency analysis method that can well highlight the differences between different signal modulation patterns is selected to convert the time-frequency map into a digital image that can be processed by a deep network. In order to fully extract the spatial and temporal characteristics of the signal, the CLP algorithm of the CNN network and LSTM network in parallel is proposed. The CNN network and LSTM network are used to extract the spatial features and temporal features of the signal, respectively, and the fusion of the two features as well as the classification is performed. Finally, the optimal model and parameters are obtained through the design of the modulation recognizer based on the convolutional neural network and the performance analysis of the convolutional neural network model. The simulation experimental results show that the improved convolutional neural network can produce certain performance gains in radio signal modulation style recognition. This promotes the application of machine learning algorithms in the field of radio signal modulation pattern recognition.
Secure and efficient transmission of information is the basic requirement of wireless communication. In the actual communication system, the baseband signal cannot be transmitted directly due to the channel spectrum characteristics and the modulation is usually used to load the code element information carried by the baseband signal to the digital characteristics of the sinusoidal signal and then transmit it through the antenna [
Modulation pattern recognition is one of the key technologies for software radio, communication countermeasures, and illegal spectrum monitoring, which has important military and civil values. In the military field, the future war is an information-driven war and electronic countermeasures are an important part of information warfare and modulation pattern recognition is one of the functions that must be considered in the receiver design process, where the receiver automatically demodulates the intercepted signals through modulation pattern recognition technology. In electronic reconnaissance and jamming, modulation technology is the key to implement precise jamming, which in turn disrupts enemy communications. In civil applications, modulation pattern recognition technology is a key technique to identify the illegal use of frequency bands to prevent spectrum abuse and interference with normal communications [
This thesis focuses on the deep learning-based modulation pattern recognition method for communication signals. The scheme adopts a multibranch CNN architecture to realize the convolutional mapping of the input signal in the complex domain and complete the preprocessing work of signal denoising and channel equalization to improve the input for modulation recognition; it investigates the impact of abstract features learned by CNN and artificially designed expert features, multiple machine learning classification models on modulation recognition, and modulation recognition algorithms; with the help of a general-purpose software radio platform, a variety of modulation signal sequences are collected with the help of a general-purpose software radio platform, the data sets used for training and testing are established, and the algorithms are designed and coded and validated by a deep learning framework and software platform. The first chapter is the introduction part of the thesis, which introduces the background and significance of the thesis and finally gives the research content and structural arrangement of the thesis. Chapter 2 is the related work section, which systematically describes the research status and analyzes the advantages and disadvantages of domestic and foreign technologies in modulation identification, signal denoising, and channel equalization. The third chapter analyzes and studies the communication signal feature processing and explains the specific implementation of the algorithm, and finally, the design study of the modulation identifier is carried out. Chapter 4 is the analysis of the results. By analyzing the performance of the algorithm proposed in this paper and simulation tests, the method can identify the modulation patterns of communication signals well under low signal-to-noise ratio, which proves the feasibility and effectiveness of the method. Chapter 5 summarizes the full text of the work and provides an outlook.
The maximum likelihood hypothesis testing method based on decision theory theoretically ensures that its decision results are optimal under the Bayesian least-cost criterion and can guarantee the performance of the method in a certain low signal-to-noise environment. However, the main drawbacks of the method are as follows: more a priori knowledge is required, such as signal-to-noise ratio, carrier frequency, symbol rate, oversampling multiplier, and other parameters, and secondly, the existence of unknown parameters leads to a complex computational push-to process and high computational complexity, which is difficult to implement in practical production [
In order to take full advantage of the temporal and structural features of the signal, this paper studies CNN and LSTM parallel modulation-style recognition algorithms. In order to further improve the signal recognition performance, the integrated learning algorithm of the heterogeneous basis classifier is also studied [
With the increase in computer processing speed and storage capacity, the design and implementation of CNN has gradually become an exhibition trend [
Flowchart of modulation pattern recognition of communication signal based on the convolutional neural network.
The digital signal (−1, +1) to be transmitted is mapped to
The effect of noise and the channel makes the communication signal amplitude vary very much. If the acquired time domain signal is used directly for time-frequency analysis, it will cause some difficulties in processing the time-frequency map after time-frequency analysis. Therefore, the signal has to be normalized [
A feature is an abstract representation of an object or a class of objects. Relative to objects, features use a set of low-dimensional tensors to express the focal properties of the original object and are the key to distinguishing multiple objects. The features together with the training set data determine the theoretical upper limit of the machine learning task, and the models and algorithms are intended to approximate this limit as closely as possible. Therefore, the selection of features should be cantered on the task [
In the convolutional layer, each convolutional kernel can be considered as a linear system for extracting a certain feature, but before the training of the network, the operational parameters of the whole system are unknown, so the weight parameters of the convolutional kernels are randomly initialized, and the parameters of this system can only be updated by the BP algorithm to continuously optimize by reducing the value of the objective function, and when the training is completed, the system can be used to extract the input features. The cascade of convolutional layers allows the input signal to be mapped by layers of abstraction to obtain the feature vector needed by the classifier.
In the encoder stage, it is downsampled by the pooling layer to compress the size of the output feature map, after the signal is mapped by the convolution layer of CAE, and after several identical operations, it reaches the bottom convolution layer and the output of the bottom layer can be considered as the abstraction of the original input signal
The result obtained after the time-frequency analysis of the modulated signal is a representation of the modulated signal in the time-frequency plane, which cannot be precisely input into the deep neural network model for processing. Therefore, it is necessary to convert the time-frequency map of the signal to generate a digital image first and then use the deep learning algorithm to identify the modulation of the signal [
By generating gray scale feature images of modulated signals with different signal-to-noise ratios, it is found that when the signal-to-noise ratio is low, the modulated signals are more affected by the background noise, which makes them appear more disordered in the gray scale feature images and the feature information becomes somewhat blurred [
For noise, Gaussian white noise can be used, because the noise of general signals is mainly divided into two categories—one is external noise and the other is internal noise of the receiver. The general noise characteristics of the collected signal are very similar to Gaussian white noise, so it is appropriate to replace the internal noise of the receiver with Gaussian white noise. For the amplitude of the noise level, in order to ensure a certain detection probability, the signal-to-noise ratio is required to be greater than 10 db.
Decision trees can be divided into classification trees and regression trees based on the nature of the data labels. When the data labels are continuous values, we call the decision tree in a regression tree; when the data labels are a series of discrete values, it is referred to as a classification tree. Each leaf node of the classification tree represents a classification result, and the branches of the tree are equivalent to the features on which the classification is founded. Since the digital signal modulation identification in this paper is a classification problem, the decision trees discussed refer to classification trees without separate emphasis. The creation of a classification tree can be summarized as follows: training data is input to the decision tree model and new branches are derived from the root node to the leaf nodes using a recursive approach based on the direction of data flow determined by the judgment conditions in the internal nodes until a leaf node is generated. Classification trees can be generated using a variety of algorithms [
Classification trees are generated by discriminating branches according to internal node conditions, which are essentially a feature selection process. In algorithms such as ID3, information entropy is used to perform feature selection [
In the CART algorithm, the concept of information first is not continued but replaced by the GINI value [
CNN consists of a series of convolutional layers, pooling layers, and fully connected layers, of which the convolutional layer is the core of CNN [
The above equation is a convolution operation on a one-dimensional continuous time system, where
In the convolution operation of the input, a large number of feature maps will be obtained; if it was directly input to the next layer, it will make the input signal of the next layer too large. Generally, before the feature map is input to the next layer, the output feature map pooling operation, on the one hand, can reduce the number of parameters and training time; on the other hand, it can reduce redundancy and enhance generalization. The pooling function generally uses the overall statistical features of the neighbouring outputs at a specified location to replace the network’s output at that location. Before the final output, CNN changes the obtained feature map into a one-dimensional form before the classification output [
Communication signals not only have temporal characteristics but also have different constellation maps for different modulation signals. This suggests that communication signals have strong spatial characteristics, so this paper explores the CNN-based modulation pattern recognition method [
After completing the training, the algorithm can integrate the base classifiers by linear weighting. In the first step, the initial value of the weight is determined here. It should be pointed out that
Next,
Modulation recognition algorithm flow.
Simulation analysis shows that the time-frequency map of the signal can describe the modulation pattern characteristics of the communication signal well. In this paper, the time-frequency map is used as the input of the CNN and the size of the time-frequency image is generally
Considering the characteristics of the modulation pattern of the communication signal and the size of the time-frequency map, the CNN contains a convolutional layer and a pooling layer to extract the effective feature vector; the nonlinearity of the network model is provided by the activation function Re LU used after each convolutional layer, and using the Re LU function as the activation function can suppress the gradient disappearance or explosion that occurs during the training of the network; the last layer is the final layer and is a fully connected layer used to integrate local features to obtain global features of the input data; finally, the global features are classified and identified by using the SOFTMAX activation function [
In this paper, convolutional kernels of size lama are used extensively, which is equivalent to convolutional kernels to map the real and imaginary parts of the input signal
This paper proposes to verify the communication modulation awareness algorithm in a real environment by building a wireless communication transceiver platform. NI-USRP is based on the public version of the software radio platform USRP Radio by NI Instruments, and some of the external circuitry is modified. NI-USPR hardware has a common software-defined radio (SDR) architecture, and in its FPGA digital signal processing logic, the communication transmitter modulates user data into digital baseband data and the output becomes analogy baseband signal
The test set of 10,000 data samples in the radio signal data set was tested on the classic convolutional neural network and the improved convolutional neural network, and the final prediction results were obtained. The modulation pattern recognition accuracy results of different algorithms are shown in Figure
Modulation pattern recognition accuracy of different algorithms.
The running time of the CNN algorithm with different numbers of convolutional layers is given in Figure
Running time of algorithms with different numbers of convolution layers.
Figure
CNN algorithm recognition performance.
The individual signal recognition rates of the CNN algorithm show the seven signals when
Single signal recognition rate of the CNN algorithm with a 6 dB signal-to-noise ratio.
For the simulation test, a test set containing 5 modulated signals of 2ASK, 2FSK, 2PSK, AM, and FM was used, of which the number of samples in the test set was 3000. The average recognition accuracy of the five types of signals from −10 dB to 10 dB is tested, and the test results are presented in Figure
Comparison of average recognition accuracy and recognition accuracy.
The database required for the network is produced by preprocessing the measured signals with the same number of samples as in the simulation test, which is 3000, and then tested. The test results are shown in Figure
Measured signal test results.
The accuracy of the training set and the accuracy of the validation set change with the number of iterations, and the loss value of the training set and the loss value of the validation set change with the number of iterations. Figure
The relationship between the training set loss value and the validation set loss value and the number of iterations.
For continuous phase modulation signal, this paper considers the GMSK signal, which is a commonly used communication modulation signal with high spectrum utilization and high noise and channel interference immunity. Unlike modulation types such as MPSK and MQAM, GMSK signals do not have fixed theoretical constellation points in the complex plane, as shown in Figure
Example of the time domain before and after the recovery of the GMSK signal.
In this paper, the problem of communication signal modulation pattern recognition based on deep learning is studied. Firstly, the mechanism of communication signal generation and the related theory of deep learning are introduced to provide the theoretical basis for the identification of modulated signals. Secondly, from the time-frequency domain of the signal, multiple time-frequency analysis methods are compared to select the time-frequency analysis method that better characterizes the modulated signal and the feature image generation algorithm is used to convert the time-frequency image into a database that can be used by the neural network. Finally, through the design of the modulation recognizer based on the convolutional neural network and the analysis of the performance of the convolutional neural network, a convolutional neural network model is established to realize the fast and accurate recognition of modulation patterns of communication signals in a complex electromagnetic environment and it can have good recognition effect under the condition of low signal-to-noise ratio. Compared with classical convolutional neural network modulation pattern recognition, recognition accuracy is improved; the convolutional neural network can be improved by changing the network layer structure, using a sequential convolutional module structure or using small convolutional kernel to extract the fine details in the radio signal; the characteristics of the radio signal will be clearer, compared with classical convolutional neural network modulation pattern recognition; the improved convolutional neural network not only shortens in training time but also improves the recognition accuracy. Theoretical analysis and computer simulation results show that the proposed algorithm based on the machine learning decision theory for communication signal modulation and recognition is practical, effective, and easy to implement and has the value of application in practical engineering.
Data sharing is not applicable to this article as no datasets were generated or analyzed during the current study.
Informed consent was obtained from all individual participants included in the study references.
We declare that there is no conflict of interest.
The National Natural Science Foundation of China, for the research on ferromagnetic resonance temperature imaging based on superparamagnetic nanoparticles, Grant 61773018; research on low complexity coding model and method based on 3D-HEVC, Grant 61771432; research on information acquisition and processing method of temperature field in superparamagnetic nanoparticle targeted, Grant 61374014; and research real-time and accurate method for measuring 2D temperature distribution in magnetic nanoparticle-mediated hyperthermia, Grant 61803346, supported the study.