CN117768278A - Modulation identification method based on hybrid complex neural network - Google Patents
Modulation identification method based on hybrid complex neural network Download PDFInfo
- Publication number
- CN117768278A CN117768278A CN202311627715.5A CN202311627715A CN117768278A CN 117768278 A CN117768278 A CN 117768278A CN 202311627715 A CN202311627715 A CN 202311627715A CN 117768278 A CN117768278 A CN 117768278A
- Authority
- CN
- China
- Prior art keywords
- signal
- training
- neural network
- complex
- parameter
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 63
- 238000000034 method Methods 0.000 title claims abstract description 43
- 238000012549 training Methods 0.000 claims abstract description 63
- 238000012360 testing method Methods 0.000 claims abstract description 28
- 238000004891 communication Methods 0.000 claims abstract description 27
- 238000000605 extraction Methods 0.000 claims abstract description 27
- 238000012795 verification Methods 0.000 claims abstract description 25
- 238000005070 sampling Methods 0.000 claims abstract description 5
- 230000006870 function Effects 0.000 claims description 50
- 230000009466 transformation Effects 0.000 claims description 37
- 230000004913 activation Effects 0.000 claims description 20
- 238000004364 calculation method Methods 0.000 claims description 17
- 238000012545 processing Methods 0.000 claims description 15
- 230000002457 bidirectional effect Effects 0.000 claims description 13
- 238000013136 deep learning model Methods 0.000 claims description 13
- 230000003044 adaptive effect Effects 0.000 claims description 12
- 238000004422 calculation algorithm Methods 0.000 claims description 12
- 238000005457 optimization Methods 0.000 claims description 12
- 230000008569 process Effects 0.000 claims description 9
- 238000010606 normalization Methods 0.000 claims description 7
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 3
- 238000010200 validation analysis Methods 0.000 claims description 3
- 241000764238 Isis Species 0.000 claims 1
- 239000011159 matrix material Substances 0.000 description 11
- 230000000694 effects Effects 0.000 description 8
- 125000004122 cyclic group Chemical group 0.000 description 5
- 238000010586 diagram Methods 0.000 description 5
- 238000001228 spectrum Methods 0.000 description 5
- 230000006835 compression Effects 0.000 description 4
- 238000007906 compression Methods 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 4
- 238000003062 neural network model Methods 0.000 description 4
- 238000013527 convolutional neural network Methods 0.000 description 3
- 238000013461 design Methods 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 3
- 238000012544 monitoring process Methods 0.000 description 3
- 230000002411 adverse Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 239000000654 additive Substances 0.000 description 1
- 230000000996 additive effect Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000003491 array Methods 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 239000002699 waste material Substances 0.000 description 1
Landscapes
- Digital Transmission Methods That Use Modulated Carrier Waves (AREA)
Abstract
The invention discloses a modulation identification method based on a hybrid complex neural network, which belongs to the technical field of radio communication and comprises the following steps: extracting signals in the data set according to a preset proportion by adopting a method without repeated random sampling, and dividing the signals into a training set, a verification set and a test set; the method comprises the steps of replacing a feature extraction module in PET-CGDNN with a complex feature extraction module to construct a modulation recognition model based on a hybrid complex neural network; training the modulation recognition model by using a training set; using the verification set to verify the modulation recognition model applying each round of model parameters, and selecting the model parameter with the highest recognition accuracy on the verification set as the optimal model parameter; and inputting the test set into a modulation recognition model applying optimal model parameters, and outputting a prediction category. The invention solves the technical problem that the real neural network does not fully utilize the internal connection between the in-phase/quadrature time sequence data, and realizes the automatic signal modulation identification under the condition that stable prior information cannot be obtained.
Description
Technical Field
The invention belongs to the field of signal modulation recognition in the technical field of radio communication, and relates to a modulation recognition method based on a hybrid complex neural network.
Background
Under the condition that stable prior information cannot be obtained, how to effectively and automatically identify, modulate and classify the signals has very wide application value in the satellite communication field, such as the fields of communication safety, signal monitoring and management, radio spectrum management and the like. Such as: in the field of communication security, automatic modulation recognition can help detect and recognize illegal signals or malicious interfering signals in a network; by analyzing and identifying the signals in real time, corresponding safety measures can be timely adopted to protect the stability and safety of the communication system. In the field of signal monitoring and management, automatic modulation identification may be used to monitor and manage various signals in a communication network; by identifying different modulation modes, the state monitoring, signal analysis and fault investigation of the communication network can be realized, and the operation efficiency and management level of the communication network are improved. In the field of radio spectrum management, automatic modulation identification can help to effectively manage the radio spectrum; by identifying signals of different modulation modes, dynamic allocation and utilization of frequency spectrum can be realized, and waste of frequency spectrum resources is reduced.
In recent years, an automatic modulation recognition method based on deep learning is applied to the field of modulation recognition because of the advantages of no need of manually extracting features and high recognition accuracy. But in general most of the collected signals are complex, such as electromagnetic, light wave, radio frequency signals. For example, in practical cases, the modulated signal in the communication field is in complex form and is composed of in-phase/quadrature time-series data, i.e., I-path and Q-path. Where the phase component represents the time course or position difference and the amplitude represents the energy or power of the wave. Although literature on deep learning applications in communications is relatively rich, few examples consider complex representations of signal properties.
A complex neural network is an artificial neural network whose input and network parameters are complex. The complex neural network can directly process the phase and amplitude information and can completely extract wave information, so that the complex neural network can improve the data expression capacity, is more suitable for processing an information processing model related to waves than a real neural network, and provides a new solution for processing the problems with complex characteristics or calculation in a complex domain. At present, although research on a complex neural network in signal processing has been carried out, the complex neural network applied to the field of signal modulation identification is still lacking.
Such as: Y.Tu, Y.Lin, C.Hou, and S.Mao, "Complex-Valued Networks for Automatic Modulation Classification," IEEE Transactions on Vehicular Technology, article vol.69, no.9, pp.10085-10089,Sep 2020,doi:10.1109/tvt.2020.3005707, give Complex convolution in Complex networks, complex batch normalization, complex weight initialization and Complex dense layer module design, and use the radio ML2016.10A dataset to conduct comparative studies on three different Complex neural network models and their corresponding real neural network models for the above modules, confirming superior performance of Complex neural network applications in signal automatic modulation recognition. In s.kim, h. -y.yang, and d.kim, "Fully Complex Deep Learning Classifiers for Signal Modulation Recognition in Non-Cooperative Environment," IEEE Access, vol.10, pp.20295-20311,2022 2022,doi:10.1109/access.2022.3151980, a complex convolutional neural network and a complex residual neural network for automatic signal modulation recognition are proposed, which propose complex neural networks with faster learning speed, higher optimality and better generalization capability than real neural networks of the same network layer, especially in recognition of phase-related modulation schemes such as PSK, APSK and QAM. Although the above complex neural network model has better recognition effect than the real neural network set by the equal network layer in automatic modulation recognition, the performance of the complex neural network model is lower than that of the real neural network with superior performance at present. For example J.Krzyston, R.Bhattacharjea and A.Stark, "Complex-Valued Convolutions for Modulation Recognition using Deep Learning," in 2020IEEE International Conference on Communications Workshops (ICC workbench), 6/2020 2020,Dublin,Ireland:IEEE,pp.1-6, doi:10.1109/ICC workbench 49005.2020.9145469, it is proposed to implement the computation of Complex convolution layers in real neural networks by means of a real convolution function in linear combination with two columns of real arrays, the number of model parameters implemented by this method is small, the computational complexity is low, but it only performs Complex processing on the convolution layers.
The efficient deep learning model (PET-CGDNN) of parameter estimation and transformation in F.Zhang, C.Luo, J.Xu, and y.luo, "An Efficient Deep Learning Model for Automatic Modulation Recognition Based on Parameter Estimation and Transformation," IEEE Communications Letters, vol.25, no.10, pp.3287-3290,10/2021 2021,doi:10.1109/lcomm.2021.3102656 consists of parameter estimation, parameter transformation and feature extraction modules consisting of convolutional neural networks and gating recursion units. Because the feature extraction module is a real neural network and consists of a real convolutional neural network and a real unidirectional gating recursion unit, the inherent relation between in-phase/quadrature time sequence data is not fully utilized, and therefore the modulation recognition performance of the network is still to be further mined and improved.
Disclosure of Invention
Therefore, in order to fully utilize the internal relation between in-phase/quadrature time sequence data and improve the feature extraction capability of an automatic modulation recognition model and further improve the recognition performance of the model, the invention improves on the basis of a high-efficiency deep learning model (PET-CGDNN) for parameter estimation and transformation in documents F.Zhang, C.Luo, J.Xu, and Y.Luo, "An Efficient Deep Learning Model for Automatic Modulation Recognition Based on Parameter Estimation and Transformation," IEEE Communications Letters, vol.25, no.10, pp.3287-3290,10/2021 2021,doi:10.1109/LCOMM.2021.3102656, and expands a real network extraction layer in a hybrid neural network to a complex domain, thereby providing a modulation recognition method based on the hybrid complex neural network. The model can realize higher modulation recognition accuracy under the condition of relatively less network model parameter quantity; the method is suitable for automatic modulation and identification of the signal under the condition that stable prior information cannot be obtained, solves the technical problem that the real neural network does not fully utilize the internal connection between in-phase/quadrature time sequence data, can achieve the effect of higher modulation and identification accuracy, and has the characteristics of fewer model parameters and good generalization.
The aim of the invention is realized by the following technical scheme:
the invention discloses a modulation identification method based on a hybrid complex neural network, which comprises the following steps:
firstly, extracting signals in a data set in the wireless communication field by adopting a method without repeated random sampling according to a preset proportion, and dividing the signals into a training set, a verification set and a test set;
step two, a feature extraction module in a high-efficiency deep learning model (PET-CGDNN) for parameter estimation and transformation is replaced by a complex feature extraction module, and a modulation recognition model based on a hybrid complex neural network is constructed by the complex feature extraction module, a parameter inverse transformation module and a feature recognition module in the high-efficiency deep learning model for parameter estimation and transformation;
training a modulation recognition model based on a hybrid complex neural network by using a training set based on a self-adaptive moment estimation optimization algorithm and a cross entropy loss function, and storing model parameters obtained after each training; using a verification set to verify a modulation recognition model based on a mixed complex neural network applying model parameters of each round, and selecting the model parameter with the highest recognition accuracy on the verification set as an optimal model parameter;
and step four, inputting the test set into a modulation recognition model based on a mixed complex neural network applying optimal model parameters, and outputting a prediction type.
The first step specifically comprises the following steps:
randomly selecting N in a computer programming language python according to a preset proportion for signals in a wireless communication field data set by using random Training/testing/measuring Taking non-repeated integers as index values of the samples;
according to the index value of the sample, extracting the sample with the appointed index from the wireless communication field data set, and dividing the sample into a training set, a verification set or a test set sample according to the preset proportion of the sample;
wherein N is Training/testing/measuring The calculation method of (2) is as follows:
wherein N is Training/testing/measuring Representing the number of samples of the training set, the validation set or the test set; r is (r) Training device Representing the dividing proportion of the training set; r is (r) Verification Representing the dividing proportion of the verification set; r is (r) Measuring Representing the test set split ratio.
The second step specifically comprises the following steps:
s1, inputting an input signal y in a data set in the wireless communication field into a leveling layer flat and a first Dense layer Dense-Linear in a high-efficiency deep learning model for parameter estimation and transformation to obtain a phase parameter phi;
s2, carrying out parameter inverse transformation on the input signal y and the phase parameter phi, and outputting an output signal after inverse transformationObtaining a parameter inverse transformation module;
s3, outputting the signalThe spatial characteristics Z of the signal are obtained by inputting a first complex convolution layer CConv (75 filters and a kernel size of 2×8) and a first complex activation function CReLU 1 ；
To spatial feature Z 1 The second complex convolution layer (25 filters and 2×5 kernel size) and the second complex activation function are input for further compression to obtain the extracted features Z of the signal 2 ；
Will extract feature Z 2 Inputting a plurality of bidirectional gating cyclic unit layers CBiGRU (the number of units is set to 64), and extracting the time characteristic H of the signal t Obtaining a plurality of feature extraction modules;
s4, time characteristic H t Inputting a characteristic recognition module consisting of a second Dense layer Dense-Linear (the number of hidden units of the second Dense layer is the number of modulation categories) and a normalization function softmax to obtain output signal prediction category probability;
and constructing a modulation recognition model based on the hybrid complex neural network by the parameter inverse transformation module, the complex feature extraction module and the feature recognition module.
In step S1, the input signal y is:
wherein y is I Is an in-phase component; y is Q Is a quadrature component; j is an imaginary unit, and the value is Is the real part of the discrete sample signal; />Is the imaginary part of the discrete sample signal; y [ l ]]Representing the input signal stored in discrete form, the sample length is L.
In step S2, the input signal y and the phase parameter phi are subjected to parameter inverse transformation, and the output signal after inverse transformation is outputThe method of (2) is as follows:
wherein,is the real part of the discrete sample signal; />Is the imaginary part of the discrete sample signal; y [ l ]]Representing the input signal stored in discrete form, the sample length being L; e is the base of natural logarithm operation; j is an imaginary unit, and the value is +.> Is a phase parameter.
In step S3, spatial feature Z 1 And extracting feature Z 2 The calculation method of (1) specifically comprises:
1) Output signalA first spatial feature Z of the signal obtained by CConv processing of the first complex convolution layer; wherein,
wherein Z is DL To output signalsIs stored in discrete form of an input signal y l]Data obtained by processing the two-dimensional real convolution Conv2d with 75 filters and a kernel size of 2 multiplied by 8;
2) Inputting the first spatial feature Z into a first complex activation function CReLU to obtain the spatial feature Z of the signal 1 The method comprises the following steps:
wherein CReLU (Z 1 ) A spatial feature Z obtained for the first spatial feature Z through a first complex activation function CReLU 1 ；Is the real part characteristic of the first spatial characteristic Z; />Is an imaginary feature of the first spatial feature Z; />To->Substituting the activation function to process; />To->Substituting the activation function to process; i is an imaginary unit.
By mixing the above Z DL Replaced by spatial features Z 1 The two-dimensional real convolution Conv2d processing with 75 filters and 2 multiplied by 8 kernel size is replaced by the two-dimensional real convolution Conv2d processing with 25 filters and 2 multiplied by 5 kernel size, the calculation steps are repeated, and the extracted feature Z of the signal is obtained by further compression 2 。
In step S3, time feature H t The calculation method of (1) comprises the following steps:
the plurality of bidirectional gating circulating unit layers CBiGRU consist of a plurality of forward gating circulating units and a plurality of backward gating circulating units; the concrete calculation formula of CBiGRU is as follows:
wherein,the system is a plurality of forward gating circulating units and is in a hidden state at the current forward moment; x is x t The input information is the mode of the real part characteristic and the imaginary part characteristic output by the previous layer of network; />Is a hidden state from the front to the previous moment; />The system is a plurality of backward gating circulating units and is in a hidden state at the current backward moment; h t For the hidden state of the bidirectional current moment, the hidden state of the forward current moment is +.>And hidden state of backward current time>The time characteristics of the signal extracted in S3 are the constitution.
Wherein, the expression of the plural forward/backward gating cyclic units CGRU is as follows:
wherein x is t Representing input information, which is the extracted feature Z of the previous layer network output 2 A modulus of real and imaginary characteristics of (a);the real part characteristic is output for the previous layer network; />The virtual part characteristics output by the previous layer of network are obtained; z t And r t Representing an update gate and a reset gate, respectively; sigma is a normalization function; w (W) z And W is r Updating the gate weight parameter and resetting the gate weight parameter respectively;h t-1 and h t The candidate hidden state of the front/back current moment, the hidden state of the front/back front previous moment and the hidden state of the front/back current moment are respectively represented; w is the latest gate weight parameter.
In the third step, based on the adaptive moment estimation optimization algorithm and the cross entropy loss function, the specific steps of training the modulation recognition model based on the hybrid complex neural network by using the training set include:
(1) Inputting signals in the training set into a modulation recognition model based on a hybrid complex neural network, outputting a prediction probability, and taking a class label corresponding to the maximum prediction probability value as a prediction class label;
(2) Inputting the real class label and the predicted class label of the signal into a cross entropy loss function, calculating a loss value, and updating network parameters of a modulation recognition model based on a hybrid complex neural network based on the loss value;
the method for calculating the cross entropy loss function comprises the following steps:
Loss＝-∑t i log(p i )；
where Loss represents the value of the cross entropy Loss function, t i A true class label for the signal; p is p i A category label for prediction;
(3) Repeating the steps (1) and (2), training a modulation recognition model based on a hybrid complex neural network by using an adaptive moment estimation optimization algorithm, and setting a learning rate initial value; when the value of the cross entropy loss function is not reduced within the preset iteration times, multiplying the learning rate by a preset multiple; and when the learning rate reaches a preset value, finishing training, and finishing training the modulation recognition model through the constructed training set.
The adaptive moment estimation optimization algorithm is implemented by an adaptive moment estimation optimizer created by using a torch.optim.adam function in python and is used for dynamically adjusting the learning rate of each parameter, so that a modulation recognition model based on a hybrid complex neural network can obtain better convergence performance.
The beneficial effects of the invention are as follows:
(1) The complex bidirectional gating cycle unit layer provided by the invention can better learn each sample in the time sequence, is used for extracting the time characteristics of in-phase/quadrature time sequence data, and effectively reduces the number of model parameters on the premise of ensuring the identification accuracy effect.
(2) The modulation recognition model based on the hybrid complex neural network provided by the invention consists of a parameter inverse transformation module, a complex feature extraction module and a feature recognition module. The parameter inverse transformation module can reduce noise and interference from a channel received by a signal and adverse effects on automatic modulation identification caused by imperfect hardware design; the complex feature extraction module can fully utilize the internal correlation in the in-phase/quadrature time sequence data to effectively extract the spatial features and the time features of the signals, so that the modulation recognition accuracy is improved, and the model has generalization by using a complex neural network.
Drawings
The invention is described in further detail below with reference to the drawings and examples.
Fig. 1 is a schematic diagram of a hybrid complex neural network composed of a parametric inverse transform module, a complex feature extraction module and a feature recognition module according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of recognition accuracy of MC, MCLDNN, PET-CGDNN and CV-PET-CBiGDNN models on a RML2016.10a dataset provided by an embodiment of the present invention;
FIG. 3 is a schematic diagram of recognition accuracy of MC, MCLDNN, PET-CGDNN and CV-PET-CBiGDNN models on a RML2018.01a dataset provided by an embodiment of the present invention;
FIG. 4 is a plot of the confusion matrix for the MC model on the RML2016.10a dataset when the signal-to-noise ratio is 0dB, provided by an embodiment of the present invention;
FIG. 5 is a graph of a confusion matrix for the MCLDNN model on the RML2016.10a dataset when the signal-to-noise ratio is 0dB, provided by an embodiment of the present invention;
FIG. 6 is a graph of a confusion matrix for the PET-CGDNN model on the RML2016.10a dataset when the signal-to-noise ratio is 0dB, provided by an embodiment of the present invention;
FIG. 7 is a graph of a confusion matrix for the CV-PET-CBiGDNN model on the RML2016.10a dataset when the signal-to-noise ratio is 0dB, provided by an embodiment of the present invention;
FIG. 8 is a plot of the confusion matrix for the MC model on the RML2018.01a dataset when the signal-to-noise ratio is 10dB, provided by an embodiment of the present invention;
FIG. 9 is a graph of a confusion matrix for the MCLDNN model on the RML2018.01a dataset when the signal-to-noise ratio is 10dB, provided by an embodiment of the invention;
FIG. 10 is a graph of a confusion matrix for the PET-CGDNN model on the RML2018.01a dataset when the signal-to-noise ratio is 10dB, provided by an embodiment of the present invention;
FIG. 11 is a graph of a confusion matrix for the CV-PET-CBiGDNN model on the RML2018.01a dataset provided by an embodiment of the present invention when the signal to noise ratio is 10 dB.
Detailed Description
Example 1
As shown in fig. 1, a first embodiment of the present invention provides a modulation recognition method based on a hybrid complex neural network, which includes:
firstly, extracting signals in a data set in the wireless communication field by adopting a method without repeated random sampling according to a preset proportion, and dividing the signals into a training set, a verification set and a test set;
step two, a feature extraction module in a high-efficiency deep learning model (PET-CGDNN) for parameter estimation and transformation is replaced by a complex feature extraction module, and a modulation recognition model based on a hybrid complex neural network is constructed by the complex feature extraction module, a parameter inverse transformation module and a feature recognition module in the high-efficiency deep learning model for parameter estimation and transformation;
training a modulation recognition model based on a hybrid complex neural network by using a training set based on a self-adaptive moment estimation optimization algorithm and a cross entropy loss function, and storing model parameters obtained after each training; using a verification set to verify a modulation recognition model based on a mixed complex neural network applying model parameters of each round, and selecting the model parameter with the highest recognition accuracy on the verification set as an optimal model parameter;
and step four, inputting the test set into a modulation recognition model based on a mixed complex neural network applying optimal model parameters, and outputting a prediction type.
The first step specifically comprises the following steps:
randomly selecting N in a computer programming language python according to a preset proportion for signals in a wireless communication field data set by using random Training/testing/measuring Taking non-repeated integers as index values of the samples;
according to the index value of the sample, extracting the sample with the appointed index from the wireless communication field data set, and dividing the sample into a training set, a verification set or a test set sample according to the preset proportion of the sample;
wherein N is Training/testing/measuring The calculation method of (2) is as follows:
wherein N is Training/testing/measuring Representing the number of samples of the training set, the validation set or the test set; r is (r) Training device Representing the dividing proportion of the training set; r is (r) Verification Representing the dividing proportion of the verification set; r is (r) Measuring Representing the test set split ratio.
The second step specifically comprises the following steps:
s1, inputting an input signal y in a data set in the wireless communication field into a leveling layer flat and a first Dense layer Dense-Linear in a high-efficiency deep learning model for parameter estimation and transformation to obtain a phase parameter phi;
s2, carrying out parameter inverse transformation on the input signal y and the phase parameter phi, and outputting an output signal after inverse transformationObtaining a parameter inverse transformation module;
s3, outputting the signalThe spatial characteristics Z of the signal are obtained by inputting a first complex convolution layer CConv (75 filters and a kernel size of 2×8) and a first complex activation function CReLU 1 ；
To spatial feature Z 1 The second complex convolution layer (25 filters and 2×5 kernel size) and the second complex activation function are input for further compression to obtain the extracted features Z of the signal 2 ；
Will extract feature Z 2 Inputting a plurality of bidirectional gating cyclic unit layers CBiGRU (the number of units is set to 64), and extracting the time characteristic H of the signal t Obtaining a plurality of feature extraction modules;
s4, time characteristic H t Inputting a characteristic recognition module consisting of a second Dense layer Dense-Linear (the number of hidden units of the second Dense layer is the number of modulation categories) and a normalization function softmax to obtain output signal prediction category probability;
and constructing a modulation recognition model based on the hybrid complex neural network by the parameter inverse transformation module, the complex feature extraction module and the feature recognition module.
In step S1, the input signal y is:
wherein y is I Is an in-phase component; y is Q Is a quadrature component; j is an imaginary unit, and the value is Is the real part of the discrete sample signal; />Is the imaginary part of the discrete sample signal; y [ l ]]Representing the input signal stored in discrete form, the sample length is L.
In step S2, the input signal y and the phase parameter phi are subjected to parameter inverse transformation, and the output signal after inverse transformation is outputThe method of (2) is as follows:
wherein,is the real part of the discrete sample signal; />Is the imaginary part of the discrete sample signal; y [ l ]]Representing the input signal stored in discrete form, the sample length being L; e is self-containedA base of a log-log operation; j is an imaginary unit, and the value is +.> Is a phase parameter.
In step S3, spatial feature Z 1 And extracting feature Z 2 The calculation method of (1) specifically comprises:
1) Output signalA first spatial feature Z of the signal obtained by CConv processing of the first complex convolution layer; wherein,
wherein Z is DL To output signalsIs stored in discrete form of an input signal y l]Data obtained by processing the two-dimensional real convolution Conv2d with 75 filters and a kernel size of 2 multiplied by 8;
2) Inputting the first spatial feature Z into a first complex activation function CReLU to obtain the spatial feature Z of the signal 1 The method comprises the following steps:
wherein CReLU (Z 1 ) Is the firstSpatial feature Z obtained by passing spatial feature Z through first complex activation function CReLU 1 ；Is the real part characteristic of the first spatial characteristic Z; />Is an imaginary feature of the first spatial feature Z; />To->Substituting the activation function to process; />To->Substituting the activation function to process; i is an imaginary unit.
By mixing the above Z DL Replaced by spatial features Z 1 The two-dimensional real convolution Conv2d processing with 75 filters and 2 multiplied by 8 kernel size is replaced by the two-dimensional real convolution Conv2d processing with 25 filters and 2 multiplied by 5 kernel size, the calculation steps are repeated, and the extracted feature Z of the signal is obtained by further compression 2 。
In step S3, time feature H t The calculation method of (1) comprises the following steps:
the plurality of bidirectional gating circulating unit layers CBiGRU consist of a plurality of forward gating circulating units and a plurality of backward gating circulating units; the concrete calculation formula of CBiGRU is as follows:
wherein,the system is a plurality of forward gating circulating units and is in a hidden state at the current forward moment; x is x t The input information is the mode of the real part characteristic and the imaginary part characteristic output by the previous layer of network; />Is a hidden state from the front to the previous moment; />The system is a plurality of backward gating circulating units and is in a hidden state at the current backward moment; h t For the hidden state of the bidirectional current moment, the hidden state of the forward current moment is +.>And hidden state of backward current time>The time characteristics of the signal extracted in S3 are the constitution.
Wherein, the expression of the plural forward/backward gating cyclic units CGRU is as follows:
wherein x is t Representing input information, which is the extracted feature Z of the previous layer network output 2 A modulus of real and imaginary characteristics of (a);the real part characteristic is output for the previous layer network; />The virtual part characteristics output by the previous layer of network are obtained; z t And r t Representing an update gate and a reset gate, respectively; sigma is a normalization function; w (W) z And W is r Updating the gate weight parameter and resetting the gate weight parameter respectively;h t-1 and h t The candidate hidden state of the front/back current moment, the hidden state of the front/back front previous moment and the hidden state of the front/back current moment are respectively represented; w is the latest gate weight parameter.
In the third step, based on the adaptive moment estimation optimization algorithm and the cross entropy loss function, the specific steps of training the modulation recognition model based on the hybrid complex neural network by using the training set include:
(1) Inputting signals in the training set into a modulation recognition model based on a hybrid complex neural network, outputting a prediction probability, and taking a class label corresponding to the maximum prediction probability value as a prediction class label;
(2) Inputting the real class label and the predicted class label of the signal into a cross entropy loss function, calculating a loss value, and updating network parameters of a modulation recognition model based on a hybrid complex neural network based on the loss value;
the method for calculating the cross entropy loss function comprises the following steps:
Loss＝-∑t i log(p i )；
where Loss represents the value of the cross entropy Loss function, t i A true class label for the signal; p is p i A category label for prediction;
(3) Repeating the steps (1) and (2), training a modulation recognition model based on a hybrid complex neural network by using an adaptive moment estimation optimization algorithm, and setting a learning rate initial value; when the value of the cross entropy loss function is not reduced within the preset iteration times, multiplying the learning rate by a preset multiple; and when the learning rate reaches a preset value, finishing training, and finishing training the modulation recognition model through the constructed training set.
The adaptive moment estimation optimization algorithm is implemented by an adaptive moment estimation optimizer created by using a torch.optim.adam function in python and is used for dynamically adjusting the learning rate of each parameter, so that a modulation recognition model based on a hybrid complex neural network can obtain better convergence performance.
The beneficial effects of the invention are as follows:
(1) The complex bidirectional gating cycle unit layer provided by the invention can better learn each sample in the time sequence, is used for extracting the time characteristics of in-phase/quadrature time sequence data, and effectively reduces the number of model parameters on the premise of ensuring the identification accuracy effect.
(2) The modulation recognition model based on the hybrid complex neural network provided by the invention consists of a parameter inverse transformation module, a complex feature extraction module and a feature recognition module. The parameter inverse transformation module can reduce noise and interference from a channel received by a signal and adverse effects on automatic modulation identification caused by imperfect hardware design; the complex feature extraction module can fully utilize the internal correlation in the in-phase/quadrature time sequence data to effectively extract the spatial features and the time features of the signals, so that the modulation recognition accuracy is improved, and the model has generalization by using a complex neural network.
For the convenience of understanding by those skilled in the art, the technical scheme provided by the invention is now described in detail:
the equivalent complex baseband signal model containing in-phase/quadrature (I/Q) components of the present application is:
where y (L) represents the received signal stored in discrete I/Q form, the sample length is L, a (L) represents the wireless channel gain, x (L) represents the transmitted signal, and n (L) represents complex additive white gaussian noise. To facilitate communication signal data processing and modulation identification, the received signal may be expressed as:
wherein y is I And y Q In-phase component and quadrature component, j is an imaginary unit, and the value isIn addition, the magnitude y of y A And may also be viewed as a representation containing I/Q channel information, which may be expressed as:
the test is verified by adopting an IQ signal data set RML2016.10a and an IQ signal data set RML2018.01a which are generated by simulation of a GNU Radio software platform and shown in table 1, and the table 1 is detailed information of the data set RML2016.10a and the data set RML 2018.01a. The hardware and software configurations used for the experiments are shown in table 2; the experimental part model training hyper-parameter setting is shown in table 3, the initial value of the learning rate is 0.001, and when the verification loss is not reduced in 10 iteration times, the verification loss is multiplied by a coefficient of 0.5; other hyper-parameters are default values.
TABLE 1
TABLE 2
TABLE 3 Table 3
In the experiment, the accuracy is used for measuring the performance of the generalization capability of the network model, and the identification accuracy can be calculated as follows:
experiments were performed on rml2016.10a and rml2018.01a datasets, respectively.
Firstly, non-repeated random sampling is adopted, and the preset proportion is 6:2:2 and 4:3:3 respectively;
then constructing a modulation recognition model based on a hybrid complex neural network:
(1) Obtaining an estimated phase parameter phi by inputting an input signal y into a combination of a flattening layer flat and a Dense layer Dense-Linear and co-training with a subsequent model;
(2) Then the signal y and the phase parameter phi are subjected to parameter transformation calculation to obtain an output signal after inverse transformation
(3) Will output a signalInputting a first complex convolution layer CConv (75 filters and 2×8 kernel sizes) and a complex activation function CReLU in the complex feature extraction layer to extract spatial features of signals; then inputting a second complex convolution layer (25 filters and 2×5 kernel sizes) and complex activation functions to further compress the extracted features; then inputting the extracted features into a plurality of bidirectional gating cyclic unit layers CBiGRU (the number of units is set to 64) to extract the time features of the signals;
(4) The features obtained in (3) are input into a feature recognition layer (composed of Dense layers of Dense-Linear and softmax, the number of hidden units of Dense layers is 11 and 24 respectively) to obtain a prediction probability.
And inputting the training samples obtained by division into a constructed modulation recognition model, and outputting the training sample prediction probability. Inputting the signal real class labels and the prediction probability into a cross entropy loss function, calculating a loss value, and updating network parameters based on the loss value; verifying the model after each round of training by using the verification samples obtained by dividing; after 200 rounds of iterative training, selecting the model parameter with the highest recognition accuracy in the verification sample as the final model parameter, and finally inputting the test sample into the final model to obtain a model recognition accuracy result.
In order to verify the generalization and effectiveness of the modulation recognition model constructed by the invention, a classical automatic modulation recognition model is selected to provide reference comparison, wherein the modulation recognition model comprises MC, MCLDNN and PET-CGDNN.
Table 4 shows the recognition accuracy of CV-PET-CBiGDNN model and other reference models on different data sets. The average accuracy is expressed as the recognition accuracy obtained by averaging all modulation patterns under all signal-to-noise ratios, namely-20-18 dB or-20-30 dB, and the recognition accuracy under the highest signal-to-noise ratio is respectively expressed as the recognition accuracy obtained under the condition that the signal-to-noise ratio is 20dB (RML 2016.10 a) and the signal-to-noise ratio is 30dB (RML 2018.01 a).
TABLE 4 Table 4
* Where A represents the RML2016.10a dataset and B represents the RML2018.01a dataset.
Fig. 2 and 3 show the recognition accuracy of each model at different signal-to-noise ratios.
As can be seen from table 4, fig. 2 and fig. 3, the average accuracy of the CV-PET-cbigmann model was highest on the rml2016.10a dataset compared to the other reference models; on the rml2018.01a dataset, compared with other reference models, the average accuracy of the CV-PET-cbigmann model is 1.05% lower than that of the MCLDNN model with the highest numerical value, but the model parameter quantity is reduced by 85% compared with that of the MCLDNN model.
Figures 4 to 7 show the confusion matrix plots of the proposed CV-PET-CBiGDNN with other reference models MC, MCLDNN and PET-CGDNN on the rml2016.01a dataset when the signal to noise ratio is 0 dB. Figures 8 to 11 show the confusion matrix plots for the various models on the rml2018.10a dataset at a signal to noise ratio of 10 dB. Wherein predicted label represents a predicted class label and true label represents a true class label; the darker the color of the color block in the confusion matrix diagram is, the higher the identification accuracy is, the more obvious the diagonal relation in the diagram is, the fewer the misjudgment is, and the better the identification effect is.
As is clear from FIGS. 4 and 11, the diagonal relationship between CV-PET-CBiGDNN is relatively clear, indicating that the recognition effect is optimal. And its good performance on both datasets suggests that the model has some generalization.
The foregoing is merely illustrative of the present invention, and the present invention is not limited thereto, and any person skilled in the art will readily recognize that variations or substitutions are within the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims (10)
1. A modulation identification method based on a hybrid complex neural network is characterized by comprising the following steps:
firstly, extracting signals in a data set in the wireless communication field by adopting a method without repeated random sampling according to a preset proportion, and dividing the signals into a training set, a verification set and a test set;
step two, replacing a feature extraction module in the efficient deep learning model of parameter estimation and transformation with a complex feature extraction module, and constructing a modulation recognition model based on a hybrid complex neural network by the complex feature extraction module, a parameter inverse transformation module and a feature recognition module in the efficient deep learning model of parameter estimation and transformation;
training a modulation recognition model based on a hybrid complex neural network by using a training set based on a self-adaptive moment estimation optimization algorithm and a cross entropy loss function, and storing model parameters obtained after each training; using a verification set to verify a modulation recognition model based on a mixed complex neural network applying model parameters of each round, and selecting the model parameter with the highest recognition accuracy on the verification set as an optimal model parameter;
and step four, inputting the test set into a modulation recognition model based on a mixed complex neural network applying optimal model parameters, and outputting a prediction type.
2. The method according to claim 1, wherein the first step specifically includes:
in the computer programming language pythonRandomly selecting N for signals in the data set of the wireless communication field according to a preset proportion by using random. Choice function in numpy library Training/testing/measuring Taking non-repeated integers as index values of the samples;
according to the index value of the sample, extracting the sample with the appointed index from the wireless communication field data set, and dividing the sample into a training set, a verification set or a test set sample according to the preset proportion of the sample;
wherein N is Training/testing/measuring The calculation method of (2) is as follows:
wherein N is Training/testing/measuring Representing the number of samples of the training set, the validation set or the test set; r is (r) Training device Representing the dividing proportion of the training set; r is (r) Verification Representing the dividing proportion of the verification set; r is (r) Measuring Representing the test set split ratio.
3. The method according to claim 1, wherein the second step specifically includes:
s1, inputting an input signal y in a data set in the wireless communication field into a leveling layer and a first dense layer in a high-efficiency deep learning model for parameter estimation and transformation to obtain a phase parameter phi;
s2, carrying out parameter inverse transformation on the input signal y and the phase parameter phi, and outputting an output signal after inverse transformationObtaining a parameter inverse transformation module;
s3, outputting the signalInputting the first complex convolution layer and the first complex activation function to obtain the spatial feature Z of the signal 1 ；
To spatial feature Z 1 The input second complex convolution layer and the second complex activation function are further compressedExtracted features Z to the signal 2 ；
Will extract feature Z 2 Inputting a plurality of bidirectional gating circulating unit layers, and extracting time characteristics H of signals t Obtaining a plurality of feature extraction modules;
s4, time characteristic H t Inputting a characteristic recognition module consisting of a second dense layer and a normalization function to obtain output signal prediction class probability;
and constructing a modulation recognition model based on the hybrid complex neural network by the parameter inverse transformation module, the complex feature extraction module and the feature recognition module.
4. A method according to claim 3, wherein in step S1, the input signal y is:
wherein y is I Is an in-phase component; y is Q Is a quadrature component; j is an imaginary unit, and the value isIs the real part of the discrete sample signal; />Is the imaginary part of the discrete sample signal; y [ l ]]Representing the input signal stored in discrete form, the sample length is L.
5. A method as claimed in claim 3, characterized in that in step S2 the input signal y and the phase parameter Φ are subjected to a parametric inverse transformation, outputting an inverse transformed output signalThe method of (2) is as follows:
wherein,is the real part of the discrete sample signal; />Is the imaginary part of the discrete sample signal; y [ l ]]Representing the input signal stored in discrete form, the sample length being L; e is the base of natural logarithm operation; j is an imaginary unit, and the value is +.>Is a phase parameter.
6. A method according to claim 3, wherein in step S3, the spatial signature Z 1 And extracting feature Z 2 The calculation method of (1) specifically comprises:
1) Output signalA first spatial feature Z of the signal obtained through the processing of the first complex convolution layer; wherein,
wherein Z is DL To output signalsIs stored in discrete form of an input signal y l]Data obtained by processing the two-dimensional real convolution Conv2d with 75 filters and a kernel size of 2 multiplied by 8;
2) Inputting the first spatial feature Z into a first complex activation function to obtain the spatial feature Z of the signal 1 The method comprises the following steps:
wherein CReLU (Z 1 ) A spatial feature Z obtained by a first complex activation function for the first spatial feature Z 1 ；Is the real part characteristic of the first spatial characteristic Z; />Is an imaginary feature of the first spatial feature Z; />To->Substituting the activation function to process; />To->Substituting the activation function to process; i is an imaginary unit.
By mixing the above Z DL Replaced by spatial features Z 1 The two-dimensional real convolution Conv2d process with 75 filters and 2×8 kernel size is replaced by the two-dimensional real convolution Conv2d process with 25 filters and 2×5 kernel size, and the above calculation is repeatedStep, further compressing to obtain the extracted feature Z of the signal 2 。
7. A method according to claim 3, wherein in step S3, the time characteristic H t The calculation method of (1) comprises the following steps:
the plurality of bidirectional gating circulating unit layers consist of a plurality of forward gating circulating units and a plurality of backward gating circulating units; the specific calculation formula of the complex bidirectional gating circulation unit layer is as follows:
wherein,the system is a plurality of forward gating circulating units and is in a hidden state at the current forward moment; x is x t The input information is the mode of the real part characteristic and the imaginary part characteristic output by the previous layer of network; />Is a hidden state from the front to the previous moment; />The system is a plurality of backward gating circulating units and is in a hidden state at the current backward moment; h t For the hidden state of the bidirectional current moment, the hidden state of the forward current moment is +.>And hidden state of backward current time>The time characteristics of the signal extracted in S3 are the constitution.
8. The method of claim 7, wherein the expression of the plurality of forward/backward gating loop units is as follows:
wherein x is t Representing input information, which is the extracted feature Z of the previous layer network output 2 A modulus of real and imaginary characteristics of (a);the real part characteristic is output for the previous layer network; />The virtual part characteristics output by the previous layer of network are obtained; z t And r t Representing an update gate and a reset gate, respectively; sigma is a normalization function; w (W) z And W is r Updating the gate weight parameter and resetting the gate weight parameter respectively; />h t-1 And h t The candidate hidden state of the front/back current moment, the hidden state of the front/back front previous moment and the hidden state of the front/back current moment are respectively represented; w is the latest gate weight parameter.
9. The method of claim 1, wherein in step three, the specific step of training the modulation recognition model based on the hybrid complex neural network using the training set based on the adaptive moment estimation optimization algorithm and the cross entropy loss function comprises:
(1) Inputting signals in the training set into a modulation recognition model based on a hybrid complex neural network, outputting a prediction probability, and taking a class label corresponding to the maximum prediction probability value as a prediction class label;
(2) Inputting the real class label and the predicted class label of the signal into a cross entropy loss function, calculating a loss value, and updating network parameters of a modulation recognition model based on a hybrid complex neural network based on the loss value;
the method for calculating the cross entropy loss function comprises the following steps:
Loss＝-∑t i log(p i )；
where Loss represents the value of the cross entropy Loss function, t i A true class label for the signal; p is p i A category label for prediction;
(3) Repeating the steps (1) and (2), training a modulation recognition model based on a hybrid complex neural network by using an adaptive moment estimation optimization algorithm, and setting a learning rate initial value; when the value of the cross entropy loss function is not reduced within the preset iteration times, multiplying the learning rate by a preset multiple; and when the learning rate reaches a preset value, finishing training, and finishing training the modulation recognition model through the constructed training set.
10. The method of claim 1 or 9, wherein the adaptive moment estimation optimization algorithm is implemented by an adaptive moment estimation optimizer created using a torch.optim.adam function in python for enabling a better convergence performance by dynamically adjusting the learning rate of each parameter, such that a modulation recognition model based on a hybrid complex neural network is obtained.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311627715.5A CN117768278B (en) | 2023-11-30 | 2023-11-30 | Modulation identification method based on hybrid complex neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311627715.5A CN117768278B (en) | 2023-11-30 | 2023-11-30 | Modulation identification method based on hybrid complex neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117768278A true CN117768278A (en) | 2024-03-26 |
| CN117768278B CN117768278B (en) | 2024-06-14 |
Family
ID=90317308
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311627715.5A Active CN117768278B (en) | 2023-11-30 | 2023-11-30 | Modulation identification method based on hybrid complex neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117768278B (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20190319659A1 (en) * | 2018-04-11 | 2019-10-17 | Booz Allen Hamilton Inc. | Method of constructing and training a neural network |
| CN110766070A (en) * | 2019-10-22 | 2020-02-07 | 北京威信通信息技术股份有限公司 | Sparse signal identification method and device based on cyclic self-encoder |
| CN113553988A (en) * | 2021-08-03 | 2021-10-26 | 同济大学 | Analog signal identification method based on complex neural network and attention mechanism |
| CN115186712A (en) * | 2022-07-15 | 2022-10-14 | 中国人民解放军国防科技大学 | Modulated signal identification method and system |
| WO2023019601A1 (en) * | 2021-08-16 | 2023-02-23 | 苏州大学 | Signal modulation recognition method for complex-valued neural network based on structure optimization algorithm |
| CN116680608A (en) * | 2023-05-08 | 2023-09-01 | 哈尔滨工程大学 | Signal modulation identification method based on complex graph convolutional neural network |
-
2023
- 2023-11-30 CN CN202311627715.5A patent/CN117768278B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20190319659A1 (en) * | 2018-04-11 | 2019-10-17 | Booz Allen Hamilton Inc. | Method of constructing and training a neural network |
| CN110766070A (en) * | 2019-10-22 | 2020-02-07 | 北京威信通信息技术股份有限公司 | Sparse signal identification method and device based on cyclic self-encoder |
| CN113553988A (en) * | 2021-08-03 | 2021-10-26 | 同济大学 | Analog signal identification method based on complex neural network and attention mechanism |
| WO2023019601A1 (en) * | 2021-08-16 | 2023-02-23 | 苏州大学 | Signal modulation recognition method for complex-valued neural network based on structure optimization algorithm |
| CN115186712A (en) * | 2022-07-15 | 2022-10-14 | 中国人民解放军国防科技大学 | Modulated signal identification method and system |
| CN116680608A (en) * | 2023-05-08 | 2023-09-01 | 哈尔滨工程大学 | Signal modulation identification method based on complex graph convolutional neural network |
Non-Patent Citations (1)
| Title |
|---|
| FUXIN ZHANG等: "An Efficient Deep Learning Model for Automatic Modulation Recognition Based on Parameter Estimation and Transformation", IEEE COMMUNICATIONS LETTERS, 5 August 2021 (2021-08-05) * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117768278B (en) | 2024-06-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zhang et al. | An efficient deep learning model for automatic modulation recognition based on parameter estimation and transformation | |
| Zeng et al. | Spectrum analysis and convolutional neural network for automatic modulation recognition | |
| Pan et al. | Specific emitter identification based on deep residual networks | |
| Wu et al. | Robust automatic modulation classification under varying noise conditions | |
| Ma et al. | Automatic modulation classification under non-Gaussian noise: A deep residual learning approach | |
| CN112364729A (en) | Modulation identification method based on characteristic parameters and BP neural network | |
| Like et al. | Signal classification in fading channels using cyclic spectral analysis | |
| CN109543643B (en) | Carrier signal detection method based on one-dimensional full convolution neural network | |
| CN112347871B (en) | Interference signal modulation identification method for communication carrier monitoring system | |
| Wu et al. | DSLN: Securing Internet of Things through RF fingerprint recognition in low-SNR settings | |
| CN114881093B (en) | Signal classification and identification method | |
| Zhang et al. | Modulated autocorrelation convolution networks for automatic modulation classification based on small sample set | |
| Yang et al. | One-dimensional deep attention convolution network (ODACN) for signals classification | |
| CN114422311A (en) | Signal modulation identification method and system combining deep neural network and expert prior characteristics | |
| Sun et al. | Digital signal modulation recognition algorithm based on vggnet model | |
| Han et al. | Communication emitter individual identification via 3D‐Hilbert energy spectrum‐based multiscale segmentation features | |
| Tan et al. | Specific emitter identification based on software-defined radio and decision fusion | |
| CN116628566A (en) | Communication signal modulation classification method based on aggregated residual transformation network | |
| Wei et al. | Rf fingerprint extraction method based on ceemdan and multidomain joint entropy | |
| Zha et al. | Intelligent identification technology for high‐order digital modulation signals under low signal‐to‐noise ratio conditions | |
| Ghauri | KNN based classification of digital modulated signals | |
| CN109309538A (en) | A kind of frequency spectrum sensing method, device, equipment, system and storage medium | |
| CN117768278B (en) | Modulation identification method based on hybrid complex neural network | |
| Hazza et al. | Robustness of digitally modulated signal features against variation in HF noise model | |
| Xing et al. | Method to reduce the signal‐to‐noise ratio required for modulation recognition based on logarithmic properties |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |