^{1}

^{1}

^{2}

^{1}

^{2}

Based on the characteristics of time domain and frequency domain recognition theory, a recognition scheme is designed to complete the modulation identification of communication signals including 16 analog and digital modulations, involving 10 different eigenvalues in total. In the in-class recognition of FSK signal, feature extraction in frequency domain is carried out, and a statistical algorithm of spectral peak number is proposed. This paper presents a method to calculate the rotation degree of constellation image. By calculating the rotation degree and modifying the clustering radius, the recognition rate of QAM signal is improved significantly. Another commonly used method for calculating the rotation of constellations is based on Radon transform. Compared with the proposed algorithm, the proposed algorithm has lower computational complexity and higher accuracy under certain SNR conditions. In the modulation discriminator of the deep neural network, the spectral features and cumulative features are extracted as inputs, the modified linear elements are used as neuron activation functions, and the cross-entropy is used as loss functions. In the modulation recognitor of deep neural network, deep neural network and cyclic neural network are constructed for modulation recognition of communication signals. The neural network automatic modulation recognizer is implemented on CPU and GPU, which verifies the recognition accuracy of communication signal modulation recognizer based on neural network. The experimental results show that the communication signal modulation recognizer based on artificial neural network has good classification accuracy in both the training set and the test set.

Modulation recognition of communication signals has a wide range of application requirements in modern wireless communications [

The method of wavelet transform [

Based on deep neural network and artificial feature engineering, a communication signal modulation recognizer is constructed. In this paper, a concrete design scheme of neural network is proposed firstly, and then, a communication signal modulation recognitor is implemented on CPU and GPU, and the experimental performance is given. In addition, on the basis of studying the different features, a kind of quasirecognition signal set recognition scheme is designed. A statistical algorithm of spectral peak number for FS and K class recognition is proposed. In order to solve the problem of constellation rotation caused by phase mismatch of QAM signal during downconversion, an improvement was made. Finally, the whole recognition scheme is verified by simulation experiment.

In order to realize the combination of artificial feature and deep neural network, the 20-dimensional artificial feature vector was constructed by designing and selecting the appropriate feature parameters, and the four-layer fully connected BP neural network was used to classify the artificial feature vector. Then, time-frequency analysis of the signals is carried out and appropriate preprocessing methods are selected. The modulation recognition algorithm based on deep learning is studied by using the AlexNet model and the Inception-ResNet-V2 model, respectively, in order to realize the end-to-end modulation recognition. Finally, based on the Indulge-ResNet-V2 model with better recognition effect, the existing model is improved by designing the combined model and the triple network model, and new signals outside the preset range are identified to expand the application range of the existing model, as shown in Figure

General framework of modulation recognition of artificial feature engineering deep neural network.

In the construction of artificial feature vectors, extracting characteristic parameters with characterization ability from the received signals is the first step and also a key link in the realization of modulation recognition. In the recognition of MASK signal, MFSK signal, MPSK signal, and MQAM signal, three new transform domain features are designed, and on this basis, a 20-dimensional artificial feature vector is constructed, which contains six temporal instantaneous features, five high-order cumulant features, and nine transform domain features.

Different characteristic parameters can represent different signal modulation modes from different perspectives. The difference of transient characteristics in time domain provides the intensity and stability of signal temporal waveform richness and instantaneous amplitude distribution, while high-order cumulative features can suppress the influence of Gaussian white noise and enhance the recognition effect at low SNR. The characteristics of the transform domain provide the information of discrete spectral line distribution, normalized spectral density, and instantaneous amplitude spectral density of the signal. The mind mapping of all feature parameters in the artificial feature vector is shown in Figure

Modulation identification diagram of deep neural network for all feature parameters in artificial feature vectors.

MASK, MQAM, MPSK, and MFSK signals vary in amplitude to different degrees. The envelope of MPSK signal is slightly undulating, while the envelope of MFSK signal is not undulating. These signals are different in terms of time domain waveform richness, intensity of instantaneous amplitude distribution and stability of amplitude information, etc. Six time domain instantaneous features are adopted, and the specific extraction algorithm of each characteristic parameter is as follows:

Using the zero center of the signal to normalize the instantaneous amplitude, the characteristic parameters are obtained

Firstly, the zero center normalized instantaneous amplitude of the signal is calculated by using the instantaneous amplitude of the signal.

Then, a digital low-pass filter is used to smooth the filter by removing some burrs and amplitude mutations, and then, the parameters can be obtained by calculating the mean value and variance. If the filtered number sequence is

Calculate the standard deviation coefficient and the expectation of absolute value of the two normalized instantaneous amplitudes of the zero center of the signal. The calculation expression is as follows

Calculate the standard deviation of the absolute value of the normalized instantaneous amplitude of the zero center of the differential signal, and obtain the characteristic parameters

The characteristic parameters can be obtained by calculating the variance coefficient of the envelope of the differential signal. Its calculation expression is as follows

The analysis of signal in time domain and transformation domain only describes the same existence from different perspectives. The transformation of perspectives often brings new information. Some signals that are difficult to see the features in time domain will be easier to see the features after being converted to transformation domain. Therefore, the characteristic parameters of the transformation domain also play an important role in modulation identification and can provide information on discrete spectral line distribution, normalized spectral density, instantaneous amplitude spectral density, and other aspects of the signal.

Many studies use high-order cumulants or clustering algorithms to identify QAM signals in class. For QAM signal, using clustering recognition is the most direct method, but the cluster computing complexity is much higher than the level of higher-order cumulant, and higher-order cumulants also have some disadvantages; FS cumulant characteristic value K signal exactly makes impossible using higher-order cumulant characteristic value of FSK signal recognition, so the combination of the two methods is in this paper. Since high-order cumulants have lower operational complexity, it is preferred to use high-order cumulant features for in-class recognition of QAM signals. When the recognition cannot be completed by using high-order cumulant features, clustering is carried out. The theoretical values of high-order cumulants of the four QAM signals of 8QAM, 16QAM, 32QAM, and 64QAM are shown in Table

Theoretical value of higher-order cumulants of QAM signals.

Signal | 8QAM | 16QAM | 32QAM | 64QAM |
---|---|---|---|---|

IC201 | ||||

IC211 | ||||

IC401 | ||||

IC411 |

The simulation shows that 16QAM and 64QAM cannot be distinguished effectively even with eighth-order cumulants. Therefore, after the signal set

In the case of phase asynchronization during mixing, as the constellation diagram rotates, the range of constellation points sampled at regular intervals is smaller than that in the case of phase synchronization, so the radius of clustering should be reduced accordingly. However, as the radius of subtraction clustering is set as a fixed value, the recognition rate of signals is reduced.

As can be seen in Figure

Recognition rate of subtraction clustering when the mixing phase difference is 45 degrees.

With the improvement of computer processing speed and storage capacity, the design and implementation of CNN has gradually become a trend. Communication signal modulation pattern recognition method based on deep neural network is used to recognize modulation signal. Firstly, the received modulation signal is normalized and the time-frequency feature image is preprocessed to generate the training set and test set required for network training. Secondly, a classifier for communication signal modulation pattern recognition, namely, CNN, is designed and built. The training set is input into CNN for training, and the CNN network model is obtained. Finally, the identified modulation signal is preprocessed to generate a test set in the dataset, and the training set is input into the CNN network model to identify the modulation mode of the communication signal. This method takes the time-frequency domain graph as input and the signal modulation mode as output. The specific algorithm flow chart is shown in Figure

Flow chart of communication signal modulation pattern recognition based on deep neural network.

Time-frequency analysis represents nonstationary signals as two-dimensional functions are related to time and frequency, which can be analyzed and processed intuitively. It is a kind of important method for processing nonstationary signals. CWD is selected for time-frequency analysis of communication signal modulation mode, and then, the grayscale feature image generation algorithm is used to generate the training set and test set required by the network. Training set and testing set include 2ASK, 2FSK, 2PSK, five kinds of AM and FM modulation signal time-frequency characteristics of the image, SNR range for

After the design and construction of CNN’s network structure is completed, some network super parameters, such as learning rate and weight attenuation coefficient, need to be set. The setting of super parameters in this network is shown in Table

Network super parameter settings.

Super parameter | Set the value |
---|---|

Vector | 0.0001 |

Attenuation coefficient of weight | 0.00006 |

Number of iterations | 60 |

Batch size | 129 |

After setting the super parameters, the following network model training can be carried out. During training, 30% of the training set is used as the validation set to verify the performance of the network. Through the training of the network, Figure

The relation between the loss value of validation set and the accuracy and the number of iterations.

As can be seen from Figure

The first step is data preparation: in the deep neural network modulation identifier, the input is directly the original communication signal sampling data, and there is no need for feature extraction and preprocessing, but the appropriate dimension transformation must be carried out to adapt to the deep network input. Note that our three deep networks are input dimensions that are all (None, 1024, 1), which is a three-dimensional tensor: the first dimension None represents the number of samples and does not need to be specified. The second dimension, 1024, means the length of our communication signal is 1024. Dimension 1 means there is only one channel. As for the label calibration of modulation mode, label unique thermal vectorization, and the division of training set and test set, it is completely consistent with the description in the implementation of modulation recognizer of deep neural network.

The second step is to build the model: based on the Keras and TensorFlow libraries, three kinds of deep neural networks are cooperatively built on the CPU and GPU computing platform according to the above network architecture description.

The third step is training model and test: with good training, dataset is divided into three deep neural network training, respectively, calculated separately on each training wheels on deep modulation recognizer network training set and testing set of accuracy and losses, after the network convergence, to test the three networks in different SNR test data accuracies.

The number of neurons in the input layer and output layer of the neural network can be clearly obtained by considering the characteristic vector dimension of the signal and the modulation mode types of the signal to be recognized. The signal features we used in the analysis included 5 spectral features and 12 cumulant features, so each signal could be described by a 17-dimensional feature vector, so we set the number of neurons in the input layer as 17. Since there are seven types of modulation signals we need to recognize, the number of neurons in the output layer is set as 7. For the setting of the number of hidden layer neurons, we might as well consider setting a little larger. Because when the greater the number of hidden layer neurons, the greater the capacity of the neural network model, it is more easy to fitting the training data; that is to say, the performance of the neural network on the training set will be better, but it also means that the model had the greater risk than the fitting; we can have a very good avoidance over fitting strategy that is regularization. We can not only apply L2 regularization to the weight of the network but also randomly inactivate the neurons in the two hidden layers. Based on this, we can set the number of neurons in the hidden layer as 40 and 32 and set the random inactivation probability of neurons in the hidden layer as

Specific structure of feedforward neural network.

Network layer | Input dimensions | Output dimensions | Number of parameters |
---|---|---|---|

Densel | (None, 18) | (None, 43) | 725 |

Dropout (0.8) | (None, 41) | (None, 42) | 2 |

Den set | (None, 41) | (None, 31) | 1332 |

Dropout (0.7) | (None, 31) | (None, 31) | 1 |

Dense3 | (None, 30) | (None, 6) | 234 |

The first is the ReLU activation function, which is used for the activation of hidden layer neurons. ReLU activation function is not only simple to calculate but also can well alleviate the problem of gradient disappearance, which can greatly accelerate the convergence speed of neural network and improve the performance of neural network. Although the “dead zone” phenomenon exists in the ReLU function, it can be effectively alleviated by appropriate parameter initialization strategy.

The second is the softmax activation function, which acts on the output nerve layer neurons. Softmax function can map the output to the probability of the communication signal corresponding to each modulation mode, so that the output of the neural network can form a probability distribution vector, which can be used to train the neural network modulation recognizer for the classification task with the cross-entropy loss function.

As for the selection of the loss function, considering that the modulation pattern recognition of communication signals is essentially a multiclassification problem, the loss function can be explicitly designed as the cross-entropy loss function. Because the cross-entropy loss function is used as a measure of the difference between two probability distributions, the output layer using softmax activation function can ensure the output is a probability distribution vector dimensions for the output category number, so for real label samples, it must also be converted to a probability distribution of the same dimension vector to calculate cross office according to the loss. Hot codinq can achieve this goal. Therefore, we can consider the real taqs data of hot codinq and convert it into multiple types of multidimensional vector modulation modes. In this hot vector alone, the modulated signal corresponds to the model. One dimension of the corresponding vector is 1, and the other dimensions are 0. The specific corresponding relationship is shown in Table

Signal modulation mode label and unique heat vector corresponding table.

Modulation mode label | Unique heat coding vector |
---|---|

0 | |

1 | |

2 | |

3 | |

4 | |

5 | |

6 |

The first step is data preparation: in the deep neural network modulation identifier, its input is the feature vector extracted from the original communication signal sampling data, and the feature vector is preprocessed by the maximum and minimum normalization. The label calibration of the modulation method, the unique thermal vectorization of the label, and the division of the training set and the test set are analyzed.

The second step is to build the model: based on the Keras and TensorFlow libraries, this four-layer deep neural network is built on the computing platform of CPU and GPU according to the network structure described.

The third step is to train the model and test: the divided training dataset is used for training the neural network, and then, the test dataset is used to test the performance of the network.

In the feature extraction part, a signal sampling sequence consisting of 40960 sampling points is intercepted. After feature extraction, 20-dimensional feature vectors can be obtained. Each simulation signal ranges from 1 SdB to 20 dB. The features are extracted with different signal-to-noise ratios. It can be found that comparing the two single model and modulation recognition model based on ergonomic 20-dimensional features, the combined model has achieved better results. At -5 dB, Inception-ResNet significantly improves the recognition effect of the V2 model by 93% under the condition of low signal-to-noise ratio. And the recognition effect is slightly better than modulation recognition based on 20-dimensional human body characteristics. 1000 samples are generated every 2 dB, with a ratio of 3 : 1 : 1 By machine extraction, training set, verification set, and test set are added, respectively.

The combined model is based on the migrated Inception-ResNet-V2 model. Figure

Recognition accuracy of different models for 18 kinds of simulation signals.

It can be seen that the addition of 8 artificial feature parameters enables the combined model to provide a more effective feature expression for the signal, and it is effective to use the model to assemble the features online in the modulation identification. Combined model is compatible with the artificial features and advantages; based on time-frequency diagram of the automatic feature in the case of low signal-to-noise ratio (SNR), artificial features play an important role, and when the signal-to-noise ratio is not less than 7 dB, using the migrated Inception-ResNet-V2 model can be the signal of time-frequency diagram that obtained the recognition accuracy higher than 95%; accordingly, sectional model has achieved above 98% accuracy.

Figure

Confusion matrix of the combined model for 18 kinds of simulation signals at 11 dB.

The data of 4ASK, 2FSK, 4FSK, 8FSK, 8PSK, 8QAM, 16QAM, 32QAM, and 64QAM signals are generated by MATLAB. The generation method is as follows: 258 random baseband codes are generated, the codes are formed by the rising cosine filter, the roll drop coefficient is 0.6, the corresponding modulation is carried out according to the selected modulation type, and the carrier frequency is 100 MHz. After AM signal is generated, a low-pass filter with a cut-off frequency of 8 kHz is used for filtering. After filtering, 100 MHz carrier is used for modulation, and the depth of modulation is random value in the range of

The code that calls modulation recognition reads the IQ data file and performs modulation recognition, using 8195 data points for each identification. If it is identified as QAM signal, the entire IQ data file is read for identification. The recognition rate of each signal is shown in Table

Recognition rate of offline test.

Modulation method | Recognition rate ( | Recognition rate ( |
---|---|---|

4ASK | 97.4% | 99.7% |

2FSK | 94.3% | 98.3% |

4FSK | 91.5% | 98.5% |

8FSK | 89.1% | 93.2% |

8PSK | 95.7% | 98.8% |

8QAM | 92.2% | 96.3% |

16QAM | 6.9% | 97.1% |

32QAM | 94.5% | 98.2% |

64QAM | 91.3% | 95.6% |

AM | 91.6% | 95.3% |

Test methods and test results of 16 signals are shown in Table

Test recognition rate of 16 kinds of communication signals.

Modulation method | Test way | Signal source | Recognition rate |
---|---|---|---|

AM | Offline test | MATLAB to produce | 95.3% |

FM | Online test | FM radio | 92.6% |

1JSB | Online test | Signal generator | 98.6% |

LSB | Online test | Signal generator | 99.1% |

2ASK | Online test | Signal generator | 96.7% |

4ASK | Offline test | MATLAB to produce | 99.3% |

2PSK | Online test | Signal generator | 98.4% |

4PSK | Online test | Signal generator | 92.2% |

8PSK | Offline test | MATLAB to produce | 98.5% |

2FSK | Offline test | MATLAB to produce | 98.1% |

4FSK | Offline test | MATLAB to produce | 98.4% |

8FSK | Offline test | MATLAB to produce | 93.7% |

8QAM | Offline test | MATLAB to produce | 96.3% |

16QAM | Offline test | MATLAB to produce | 97.3% |

32QAM | Offline test | MATLAB to produce | 98.2% |

64QAM | Offline test | MATLAB to produce | 95.7% |

Deep nucleus size is usually odd, commonly

Relation between the accuracy of verification set and the number of training iterations for different deep cores.

As can be seen from Figure

The training process of the deep neural network communication signal modulation recognitor is shown in Figure

Accuracy performance curve of deep neural network on training set and verification set.

With the increase of the number of network training rounds, the accuracy of the deep neural network in the training set and the test set gradually improved and tended to be stable, and the modulation recognizer in the training set and the test set reached a high level of accuracy, both of which exceeded 98%. This means that our neural network has been fully and effectively trained and no fitting phenomenon has occurred.

Through experimental verification, the online splicing features take into account the dual advantages of time-frequency analysis and artificial features, and the combined model design with better recognition effect is completed. Then, after selecting an appropriate feature extraction subnetwork, a triplet network model is built, and the similarity of the input time-frequency graph is calculated by using the Euclidean distance between the feature vectors, and a reasonable sample decision method is designed, which not only improves the matching degree between the input time-frequency graph samples and the Inception-ResNet-V2 model but also improves the matching degree between the input time-frequency graph samples and the Inception-ResNet-V2 model. It also realizes the recognition of the newly modulated signal and expands the application range of the existing model. Finally, the algorithm test and performance analysis based on the above two models are completed by using the real signal and simulation signal under different SNR.

Based on Keras and TensorFlow libraries, the deep neural network communication signal modulation recognizer is implemented on CPU and GPU hardware platform, and the training and testing tasks are completed. Experiments show that the deep neural network has good performance on both the training set and the test set, and can accurately identify the modulation mode of the communication signal with low SNR, which shows the robustness of the neural network. The deep neural network in this chapter had better reached the accuracy level of 97%, and we can see that the modulation recognizer based on deep neural network can achieve a better accuracy level. Modulation recognizers for communication signals are constructed based on two kinds of mainstream deep neural network structures: three kinds of modulation recognizers are constructed based on deep neural network architecture, and their performance differences are compared; based on the cyclic neural network architecture, we construct two modulation identifiers and compare their performance differences. The experimental results show that the modulation recognizer based on deep neural network has a good accuracy performance on the simulation dataset. For the modulation signal identification problem adopted in this paper, we try to collect more signal modulation modes to research more signal characteristics, so that the algorithm can identify more communication signal modulation modes.

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.

This study was supported by the National Natural Science Foundation of China: 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 a real-time and accurate method for measuring 2D temperature distribution in magnetic nanoparticle-mediated hyperthermia (Grant 61803346).