Double-branch automatic modulation identification method based on cross signal enhancement

Abstract

The invention belongs to the technical field of wireless communication signal identification, and particularly discloses a double-branch automatic modulation identification method based on cross signal enhancement. The method comprises the following steps: firstly, acquiring a modulation signal, preprocessing the modulation signal, and converting the modulation signal from I / Q data into A / P data; then, building a modulation identification model; the modulation recognition model comprises a cross enhancement module, a channel feature fusion module, a time sequence context feature capture module and a classifier module; the cross enhancement module performs cross enhancement on I and Q channel data of the modulation signal and extracts I / Q features; the channel feature fusion module performs deep fusion on the I / Q features and the A / P features to obtain channel fusion enhanced features; a time sequence context feature capture module extracts time sequence context features from the channel fusion enhancement features, the time sequence context features are input into a classifier for classification, and modulation types are obtained; and finally, training and optimizing the modulation identification model, and using the optimized modulation identification model for identification of modulation signals. According to the method, the unbalance phenomenon between I / Q channels can be effectively reduced, and the recognition accuracy is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of wireless communication signal recognition, and in particular is a double-branch automatic modulation recognition method based on cross signal enhancement. Background Art

[0002] Automatic modulation recognition (AMR) is a key technology of cognitive radio, which aims to use the prior knowledge of communication signals to identify the modulation type. It plays an important role in spectrum perception, wireless interference identification, channel selection and management, etc.

[0003] Traditional AMR methods mainly include likelihood-based methods (LB) and feature-based methods (FB). The LB method regards AMR as a hypothesis testing problem. This method can achieve high recognition accuracy, but its computational complexity and weak robustness to model mismatch limit its real-time and low-cost application in actual communication scenarios. The FB method extracts expert features of the signal and inputs them into the machine learning model for classification. In comparison, this method has lower computational complexity and usually produces suboptimal solutions. However, the features extracted by the FB method are independent of the learning model, the generated features lack representativeness, and the algorithm performance depends largely on the training data set, resulting in poor robustness of the model.

[0004] Compared with traditional methods, the AMR method based on deep learning uses neural networks to autonomously learn signal features, avoiding the feature design process and achieving better recognition accuracy. At present, signal features usually have three types of representation: I / Q data features, A / P data features and image features. There are usually amplitude and phase errors between the I channel and the Q channel of the I / Q data feature, resulting in an imbalance in the I / Q channel, so the I / Q channel needs to be calibrated for imbalance. Traditional I / Q channel imbalance calibration is mainly divided into two types: analog domain and digital domain. Analog domain calibration is mainly achieved by improving the process level of the device and optimizing the circuit layout, which is usually limited by the process level. Digital domain calibration is completed through the back-end digital signal processing algorithm, such as using the Gram-Schmidt algorithm or the I / Q imbalance compensation method based on adjacent channel detection. These methods have the problem of complex calculation. The existing AMR method based on deep learning directly inputs the I / Q signal into the model, does not fully utilize the inherent interaction of the I / Q channel, and ignores the problem of I / Q channel imbalance.

[0005] In addition, existing deep learning-based AMR models usually use a single type of features or multimodal features (sequence features + image features) as input, which can lead to problems such as insufficient feature extraction or complex data preprocessing.

[0006] In view of the above problems, the present invention proposes a dual-branch automatic modulation recognition method based on cross signal enhancement. Summary of the invention

[0007] In view of the deficiencies in the prior art, the technical problem to be solved by the present invention is to provide a dual-branch automatic modulation recognition method based on cross signal enhancement.

[0008] The present invention solves the technical problem by adopting the following technical solution: A dual-branch automatic modulation recognition method based on cross signal enhancement comprises the following steps: Step 1: Obtain the modulated signal and pre-process the modulated signal to convert the modulated signal from I / Q data to A / P data; Step 2: Build a modulation recognition model; the modulation recognition model includes four modules: cross-enhancement, channel feature fusion, temporal context feature capture and classifier; the cross-enhancement module cross-enhances the I and Q channel data of the modulation signal to extract I / Q features; the channel feature fusion module deeply fuses the I / Q features and A / P features to obtain channel fusion enhancement features; the temporal context feature capture module extracts the temporal context features from the channel fusion enhancement features, and the temporal context features are input into the classifier for classification to obtain the modulation type; The cross enhancement module includes an I channel feature extraction branch and a Q channel feature extraction branch with the same architecture but different weights; for the I channel feature extraction branch, the I channel data of the I / Q modulated signal is sequentially subjected to a one-dimensional convolution layer and a reshape to obtain an I channel feature; the I channel feature is subjected to a global average pooling layer to obtain an I channel global average pooling feature, and the I channel global average pooling feature is sequentially subjected to a fully connected layer, a ReLU activation function, a fully connected layer, and a Sigmoid activation function to obtain an I channel weight coefficient; similarly, the Q channel data of the I / Q modulated signal is subjected to a Q channel feature extraction branch to obtain a Q channel weight coefficient; after multiplying the I channel feature with the Q channel weight coefficient, a residual connection is performed with itself to obtain an I channel feature of information interaction; similarly, the Q channel feature of information interaction is obtained; the I channel feature of information interaction is spliced ​​with the Q channel feature to obtain an I / Q feature; Step 3: Train and optimize the modulation recognition model, and use the optimized modulation recognition model to recognize the modulated signal.

[0009] Furthermore, in the channel feature fusion module, the A / P modulated signal passes through a two-dimensional convolution layer to obtain an intermediate feature, and the intermediate feature is residually connected with itself after a squeezing and excitation attention mechanism to obtain an A / P feature; the I / Q feature passes through a two-dimensional convolution layer and is concatenated with the A / P feature in the channel dimension to obtain a channel fusion feature; the channel fusion feature passes through a squeezing and excitation attention mechanism, a two-dimensional convolution layer, a discard layer, a squeezing and excitation attention mechanism in turn to obtain a channel fusion enhanced feature.

[0010] Furthermore, in the temporal context feature capture module, the channel fusion enhanced features are reshaped and then pass through the LSTM layer, the deformation LSTM layer and the LSTM layer in sequence to obtain the temporal context feature.

[0011] Furthermore, both LSTM layers contain 128 LSTM units, and the number of alternating rounds of the deformation LSTM layer is 5.

[0012] Compared with the prior art, the present invention has the following beneficial effects: 1. The cross-enhancement module extracts the I channel weight coefficient and the Q channel weight coefficient respectively, cross-multiplies the I channel weight coefficient with the Q channel feature, and cross-multiplies the Q channel weight coefficient with the I channel feature to realize the interaction of different channel signals, and makes full use of the interactive information between the I and Q channel data to enhance the signals of different channels to compensate for the difference between amplitude and phase, and performs imbalance calibration on the I / Q channels, thereby reducing the imbalance phenomenon between the I / Q channels.

[0013] 2. The channel feature fusion module extracts A / P features and then splices them with I / Q features in the channel dimension to achieve deep feature fusion, making full use of the complementarity between I / Q and A / P data to extract more detailed and rich signal features. In the fusion process, the squeeze and incentive attention mechanism is used to further extract features and enhance features, thereby enhancing the feature expression ability.

[0014] 3. For sequence signals, their features usually show long-term dependencies and rich temporal information, while the traditional LSTM model has limitations in extracting temporal features: the current input and hidden state are considered independent of each other and interact only within the gated unit. This interaction lacks effective use of contextual information before the gated unit is calculated, which may lead to the loss of some key information. Therefore, the temporal context feature capture module of the present invention uses the alternating stacking structure of LSTM and Mogrifier LSTM to gradually extract and enhance the temporal context information in the modulated signal. The current input and hidden state are adjusted to each other by Mogrifier LSTM before the gated unit is calculated, so as to extract the temporal context information more comprehensively, capture the temporal features of the modulated signal more effectively, and enable the model to better learn the dynamic changes of the modulated signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is the overall flow chart; Figure 2 It is the structural diagram of the modulation recognition model; Figure 3 It is the recognition accuracy curve of different models under different signal-to-noise ratios; Figure 4 It is the confusion matrix of the modulation recognition model for different modulation types when the signal-to-noise ratio is 12dB; Figure 5 It is the confusion matrix of the CGDNN model for different modulation types when the signal-to-noise ratio is 12dB; Figure 6 It is the confusion matrix of the DAE model for different modulation types when the signal-to-noise ratio is 12dB; Figure 7 It is the confusion matrix of the LSTM2 model for different modulation types when the signal-to-noise ratio is 12dB; Figure 8 It is the confusion matrix of the MCLDNN model for different modulation types when the signal-to-noise ratio is 12dB; Fig. 9 It is the confusion matrix of the PET-CGDNN model for different modulation types when the signal-to-noise ratio is 12dB. DETAILED DESCRIPTION

[0016] Specific embodiments are given below in conjunction with the accompanying drawings. The specific embodiments are only used to further describe the technical solutions of the present invention in detail and are not intended to limit the protection scope of the present application.

[0017] The present invention provides a dual-branch automatic modulation recognition method based on cross signal enhancement (hereinafter referred to as method, see Figures 1 to 9 ), including the following steps: Step 1: Obtain the modulated signal and pre-process the modulated signal, convert the modulated signal from I / Q data to A / P data, and several modulated signal sequences form a data set; The modulated signal is stored as I / Q data, and the I / Q modulated signal is expressed as: (1) In the formula, represents an I / Q modulated signal, Indicates the signal length, 2 indicates two channels, I and Q. and Respectively represent the I channel data and Q channel data of the I / Q modulated signal, represents the imaginary part; Convert I / Q modulated signals to A / P modulated signals , the conversion formula is: (2) In the formula, and They represent the amplitude and phase of the A / P modulated signal respectively. The amplitude represents the size and amplitude of the signal, and the phase represents the periodic deviation of the signal.

[0018] Step 2: Build a modulation recognition model; Figure 2 As shown in the figure, the modulation recognition model includes four modules: cross enhancement (CE module), channel feature fusion (CFF module), time series context feature capture (TCC module) and classifier (Classifier module). The cross enhancement module cross-enhances the I channel data and Q channel data of the modulated signal to enhance the interaction information between the two channels and obtain the I / Q features; the channel feature fusion module extracts the A / P features and deeply fuses the I / Q features and A / P features in the channel dimension to obtain the channel fusion enhancement feature (CFF Fea); the time series context feature capture module extracts the key information of the time dimension from the channel fusion enhancement feature to obtain the time series context feature (TCC Fea); the classifier identifies the time series context feature to obtain the modulation type.

[0019] The cross-enhancement module includes an I channel feature extraction branch and a Q channel feature extraction branch with the same architecture but different weights. For the I channel feature extraction branch, the I channel data of the I / Q modulated signal After feature extraction through a one-dimensional convolutional layer, the I channel feature is obtained through shape reshaping. ;I channel characteristics After downsampling through the global average pooling layer, the global average pooling feature of channel I is obtained. ; I channel global average pooling feature After passing through the fully connected layer, ReLU activation function, fully connected layer and Sigmoid activation function in sequence, the I channel weight coefficient is obtained Similarly, the input of the Q channel feature extraction branch is the Q channel data of the I / Q modulated signal , the Q channel weight coefficient is obtained by the Q channel feature extraction branch ; Set the I channel characteristics and Q channel weight coefficient After multiplication, residual connection is performed to obtain the I channel feature of information interaction ; Correct the imbalance of amplitude and phase by cross-talking I and Q channel information, so that they have similar amplitude and corresponding phase relationship, which helps to restore the orthogonal relationship between the I / Q channels of the modulated signal; add the corrected features to the original features through residual connection, so that the model can better focus on the most relevant channel in the features, which helps to reduce the impact of irrelevant or noisy features and improve the expressiveness of the model. Similarly, the Q channel features of information interaction are obtained , respectively expressed as: (3) (4) In the formula, Indicates Q channel characteristics; The I channel features of information interaction are spliced ​​with the Q channel features to obtain the output of the cross enhancement module, namely the I / Q features .

[0020] A / P data is more suitable for representing time domain information because it can more comprehensively capture the amplitude and phase changes of the signal. I / Q data represents the signal in the form of complex numbers, which provides support for the model to distinguish nonlinear modulation. Therefore, by integrating these two types of data, the extracted features are more discriminative.

[0021] In the channel feature fusion module, the A / P modulated signal is extracted through a two-dimensional convolution layer to obtain the intermediate features. ; Intermediate features After squeezing and exciting the attention mechanism (SE), it is residually connected with itself to obtain the A / P feature ;I / Q characteristics After the two-dimensional convolution layer, the A / P feature The channel fusion features are obtained by splicing in the channel dimension; the channel fusion features are further extracted and enhanced through the squeeze and incentive attention mechanism, the two-dimensional convolution layer, the dropout layer, and the squeeze and incentive attention mechanism, and the channel fusion enhanced features are obtained, which is the output of the channel feature fusion module. The squeeze and incentive attention mechanism plays a role in feature enhancement and further enhances the feature expression ability.

[0022] In the temporal context feature capture module, the channel fusion enhancement feature is reshaped and then passes through the LSTM layer, the deformation LSTM layer (Mogrifier LSTM) and the LSTM layer in sequence to obtain the temporal context feature, which is the output of the module. In this embodiment, both LSTM layers contain 128 LSTM units. The first LSTM layer is used to preliminarily extract the temporal features and capture the dynamic changes in the modulated signal, and the second LSTM layer is used to further refine the temporal features; the number of alternating rounds of the deformation LSTM layer is 5, which is used to interactively adjust the input and hidden states before the LSTM unit is processed, so as to more fully capture the context information of the temporal features.

[0023] The classifier includes three fully connected layers, the first two fully connected layers are followed by a SELU activation function and a dropout layer, and the last fully connected layer is followed by a softmax activation function; the temporal context features are classified by the classifier, and the modulation category of the modulated signal is output.

[0024] Step 3: Divide the data set into a training set and a validation set, use the training set to train the modulation recognition model, use the validation set to verify and optimize the trained modulation recognition model, and use the optimized modulation recognition model for the recognition of modulated signals.

[0025] The modulation recognition model was trained and validated using the Adam optimization algorithm and the categorical cross entropy loss function. The initial learning rate was set to 0.001. During the training process, if the validation loss did not decrease within a time span of 5 consecutive epochs, the learning rate was multiplied by 0.5 to improve the training efficiency. If the validation loss did not decrease after 30 cycles, the modulation recognition model with the smallest validation loss was saved as the optimized modulation recognition model.

[0026] In order to verify the performance of the modulation recognition model and simulate real environmental conditions, experiments were conducted on the public dataset RML2016.10a. The RML2016.10a dataset has a total of 220,000 samples, including 8 digital modulation schemes: 8PSK, BPSK, CPFSK, GFSK, 4-PAM, 16-QAM, 64-QAM and QPSK, and 3 analog modulation schemes: AM-DSB, AM-SSB and WBFM. The signal-to-noise ratio range is -20dB~18dB, with an interval of 2dB each time.

[0027] In order to verify the effectiveness of the method of the present invention, the modulation recognition model (DB-CLDNN) of the present invention is compared with the current advanced AMR algorithms, including MCLDNN+AE-IEDNN [1] ,DAE [2] , PET-CGDNN [3],CGDNet [4] , MCLDNN [5] 、MCNET [6] and LSTM2 [7] , the test results are shown in Table 1. Figure 3 It is the recognition accuracy curve of different models under different signal-to-noise ratios.

[0028]

[0029] It can be seen from Table 1 that the highest recognition accuracy and average recognition accuracy of the modulation recognition model of the present invention are higher than those of other models. Figure 3 It can be seen that after 0 dB, the modulation recognition model of the present invention performs well, and the recognition accuracy rate significantly optimizes other models. At 12 dB, the highest recognition accuracy rate is 94.59%. Figure 4~Figure 9 They are the confusion matrices of the modulation recognition model of the present invention, CGDNN, DAE, LSTM2, MCLDNN, and PET-CGDNN for different modulation types when the signal-to-noise ratio is 12dB. By comparison, it can be seen that the modulation recognition model of the present invention performs better when processing easily confused modulation types (such as 16QAM and 64QAM, AM-DSB and WBFM), with the recognition accuracy rates of AM-DSB and WBFM reaching 74% and 79% respectively, and the recognition accuracy rates of 16-QAM and 64-QAM reaching 97% and 96% respectively.

[0030] The above test results show that the modulation recognition model of the present invention has better recognition accuracy and the best classification performance. This is because the present invention uses a cross-enhancement module to realize the interaction of I and Q channel signals, and fully utilizes the interactive information between I and Q channel data to enhance the signals of different channels to reduce the imbalance between I / Q channels; the channel feature fusion module realizes the deep fusion of I / Q features and A / P features, and fully utilizes the complementarity between I / Q and A / P data to extract more detailed and rich signal features; the temporal context feature capture module uses the alternating stacking structure of LSTM and Mogrifier LSTM to gradually extract and enhance the temporal context information in the modulated signal, and more effectively captures the temporal features of the modulated signal.

[0031] The sources of each model are: [1] F. Zhang, C. Luo, J. Xu, Y. Luo, “An Autoencoder-Based I / QChannel Interaction Enhancement Method for Automatic Modulation Recognition.”IEEE Trans. Veh. Technol, vol. 72, no. 7, pp. 9620–9625, 1 July 2023, doi:10.1109 / tvt.2023.3248625. [2] Z. Ke, H. Vikalo, “Real-Time Radio Technology and ModulationClassification via an LSTM Auto-Encoder.” IEEE Trans. Wireless Commun, vol.21, no. 1, pp. 370–382, Jan. 2022, doi:10.1109 / twc.2021.3095855. [3] F Zhang, C Luo, J Xu, Y Luo, “An Efficient Deep Learning Modelfor Automatic Modulation Recognition Based on Parameter Estimation andTransformation.” IEEE Commun Lett, vol. 25, no. pp. 3287–3290, 10, Oct. 2021,doi:10.1109 / lcomm.2021.3102656. [4] JN. Njoku, ME. Morocho-Cayamcela and W. Lim, “CGDNet: EfficientHybrid Deep Learning Model for Robust Automatic Modulation Recognition,” IEEENetworking Letters, vol. 3, no. 2, pp. 47–51, June 2021, doi:10.1109 / lnet.2021.3057637. [5] Xu J, Luo C, Parr G, Luo Y, "A spatiotemporal multi-channel learning framework for automatic modulation recognition". EEE Wirel CommunLe, vol. 9, no. 10, pp 1629-1632, October. 2020, doi: 10.1109 / LWC.2020.2999453. [6] T. Huynh-The, CH. Hua, QV. Pham and DS. Kim “MCNet: An EfficientCNN Architecture for Robust Automatic Modulation Classification,” IEEE CommunLett, vol. 24, no. 4, Apr. 2020, pp. 811815, doi:10.1109 / lcomm.2020.2968030. [7] S. Rajendran, W. Meert, D. Giustiniano, V. Lenders and S. Pollin, "Deep Learning Models for Wireless Signal Classification with DistributedLow-Cost Spectrum Sensors," IEEE Trans. Cognit. Commun, vol. 4, no. 3, pp.433–445, Sept. 2018, doi:10.1109 / tccn.2018.2835460. Any matters not described in the present invention are applicable to the prior art.

Claims

1. A dual-branch automatic modulation recognition method based on cross signal enhancement, characterized in that: The method comprises the following steps: Step 1: Obtain the modulated signal and pre-process the modulated signal to convert the modulated signal from I / Q data to A / P data; Step 2: Build a modulation recognition model; the modulation recognition model includes four modules: cross-enhancement, channel feature fusion, temporal context feature capture and classifier; the cross-enhancement module cross-enhances the I and Q channel data of the modulation signal to extract I / Q features; the channel feature fusion module deeply fuses the I / Q features and A / P features to obtain channel fusion enhancement features; the temporal context feature capture module extracts the temporal context features from the channel fusion enhancement features, and the temporal context features are input into the classifier for classification to obtain the modulation type; The cross enhancement module includes an I channel feature extraction branch and a Q channel feature extraction branch with the same architecture but different weights; for the I channel feature extraction branch, the I channel data of the I / Q modulated signal is sequentially subjected to a one-dimensional convolution layer and a reshape to obtain an I channel feature; the I channel feature is subjected to a global average pooling layer to obtain an I channel global average pooling feature, and the I channel global average pooling feature is sequentially subjected to a fully connected layer, a ReLU activation function, a fully connected layer, and a Sigmoid activation function to obtain an I channel weight coefficient; similarly, the Q channel data of the I / Q modulated signal is subjected to a Q channel feature extraction branch to obtain a Q channel weight coefficient; after multiplying the I channel feature with the Q channel weight coefficient, a residual connection is performed with itself to obtain an I channel feature of information interaction; similarly, the Q channel feature of information interaction is obtained; the I channel feature of information interaction is spliced ​​with the Q channel feature to obtain an I / Q feature; Step 3: Train and optimize the modulation recognition model, and use the optimized modulation recognition model to recognize the modulated signal.

2. The dual-branch automatic modulation recognition method based on cross signal enhancement according to claim 1 is characterized in that: In the channel feature fusion module, the A / P modulated signal passes through a two-dimensional convolution layer to obtain an intermediate feature. The intermediate feature is residually connected with itself after the squeezing and excitation attention mechanism to obtain the A / P feature; the I / Q feature passes through a two-dimensional convolution layer and is concatenated with the A / P feature in the channel dimension to obtain the channel fusion feature; the channel fusion feature passes through the squeezing and excitation attention mechanism, the two-dimensional convolution layer, the discard layer, the squeezing and excitation attention mechanism in turn to obtain the channel fusion enhanced feature.

3. The dual-branch automatic modulation recognition method based on cross signal enhancement according to claim 1 or 2, characterized in that: In the temporal context feature capture module, the channel fusion enhanced features are reshaped and then pass through the LSTM layer, the deformation LSTM layer and the LSTM layer in sequence to obtain the temporal context features.

4. The dual-branch automatic modulation recognition method based on cross signal enhancement according to claim 3 is characterized in that: Both LSTM layers contain 128 LSTM units, and the number of alternating rounds of the deformation LSTM layer is 5.

Images