CN114254680B

CN114254680B - Deep learning network modulation identification method based on multi-feature information

Info

Publication number: CN114254680B
Application number: CN202210188206.6A
Authority: CN
Inventors: 王贵; 宁刚玲
Original assignee: Chengdu Dagong Bochuang Information Technology Co ltd
Current assignee: Chengdu Dagong Bochuang Information Technology Co ltd
Priority date: 2022-02-28
Filing date: 2022-02-28
Publication date: 2022-05-17
Anticipated expiration: 2042-02-28
Also published as: CN114254680A

Abstract

The invention relates to a deep learning network modulation identification method based on multi-feature information, which belongs to the technical field of wind turbine generator fault diagnosis and combines the instantaneous amplitude and the instantaneous phase of a modulation signal with an I/Q signal from the modulation signal, thereby enriching the data representation form of each modulation mode and realizing the complementation between different types of data features. An efficient network structure based on the depth separable convolution block and the LSTM is also designed, and an attention mechanism is introduced, so that potential space-time characteristics in the modulation signal can be mined. Experiments on a standard data set show the superiority of the algorithm, and the method shows that for modulation identification tasks, particularly under various interference conditions, a plurality of types of data sources can provide a plurality of observation views for a model, and the difficulty of modulation identification is reduced. In combination with network characteristics, designing a better model structure is also crucial to modulation identification.

Description

Deep learning network modulation identification method based on multi-feature information

Technical Field

The invention belongs to the technical field of signal modulation, and particularly relates to a deep learning network modulation identification method based on multi-feature information.

Background

At present, most of modulation recognition algorithms based on a neural network directly send I/Q signals or instantaneous characteristics into a well-designed network, and then decision judgment is carried out through the neural network, so that recognition of a modulation mode is realized. In a complex and variable electromagnetic environment, potential characteristics of the original I/Q signal may be deficient, and the performance of the model is limited to a certain extent only by providing the I/Q signal or a single transient characteristic.

Therefore, at present, a deep learning network modulation identification method based on multi-feature information needs to be designed to solve the above problems.

Disclosure of Invention

The invention aims to provide a deep learning network modulation recognition method based on multi-feature information, which is used for solving the technical problems in the prior art, most of the current modulation recognition algorithms based on a neural network directly send I/Q signals or instantaneous features into a well-designed network, and then carry out decision judgment through the neural network, thereby realizing the recognition of a modulation mode. In a complex and variable electromagnetic environment, potential characteristics of the original I/Q signal may be deficient, and the performance of the model is limited to a certain extent only by providing the I/Q signal or a single transient characteristic.

In order to achieve the purpose, the technical scheme of the invention is as follows:

the deep learning network modulation identification method based on the multi-feature information comprises the following steps:

s1: under the influence of electromagnetic environment and various channel interferences, combining instantaneous characteristics of the I/Q modulation signal with an original signal, namely multi-characteristic fusion;

s2: on the basis of the step S1, inputting the data after multi-feature fusion into an Attention-SCNN structure for integration;

s3: on the basis of the step S2, an optimized recurrent neural network-LSTM is introduced, the output of the integration of the Attention-SCNN structure is sent to the recurrent neural network-LSTM, the past and future associated information of the data is extracted, and further the comprehensive time sequence characteristics are captured.

Further, step S1 is specifically as follows:

firstly, performing instantaneous feature extraction on an I/Q modulation signal, wherein the instantaneous feature extraction comprises an instantaneous amplitude value and an instantaneous phase value; since the I/Q modulation signal is divided into I, Q two quadrature signals, I channel signal is x (t) and Q channel signal is y (t), the instantaneous amplitude a (t) and the instantaneous phase p (t) can be obtained by the following equations:

(2)

(3)

then, normalizing the obtained instantaneous amplitude value and instantaneous phase value, and splicing and fusing the original I/Q signal and the instantaneous characteristic value;

the original I/Q signal dimension is 2 × N, 2 represents an I path and a Q path, N represents the number of sampling points, and the data dimension after the instantaneous amplitude and phase information are fused is changed into 4 × N.

Further, step S2 is specifically as follows:

the data after multi-feature fusion is firstly input into an Attention-SCNN structure, the Attention-SCNN structure is composed of a plurality of depth separable convolution blocks and a point-by-point convolution layer, in the SepaConvBlock, the depth separable convolution layer is taken as a basic layer, the convolution process is divided into two steps of depth convolution and point-by-point convolution,

in the SepaConvBlock structure, firstly, a 1 × 1 convolution layer is utilized to realize cross-channel interaction and information integration, the number of channels is controlled, and then the BatchNormal is utilized to relieve the influence of internal data distribution deviation and accelerate network training; next, inputting the result of batch normalization into two depth separable convolution layers, wherein the convolution kernel size is 2 x 2, and the number of output channels is 32; and introducing a channel attention mechanism; modeling the dependency relationship among the channels to self-adaptively adjust the characteristic response value among each channel; the attention mechanism comprises two parts of compression and excitation, wherein the compression part compresses the feature mapping U to a channel descriptor z epsilon Rc by using a global average pooling technology, wherein the c < th > element of z can be calculated by the following formula:

(4)

in the formula, each element of z has a global receptive field of a feature mapping U, H represents the height of each channel feature mapping graph, W represents the width of each channel feature mapping graph, then the series of feature weight vectors are activated, and the two full-connection layers and the nonlinear activation function are used for controlling the channel weight value to be between 0 and 1; the final scale operation represents multiplying the learned channel weights by the corresponding channel features to obtain a calibrated feature map ũ; in SepaConvBlock, the feature mapping graph after scale operation is connected with the previous input through cross-layer connection, and the purpose of residual error learning is achieved;

then, two attention operations are added to each SepaConvBlock, and then the obtained feature maps are integrated by stacking 4 sepaconvblocks and using 1 × 1 convolution layers.

Further, step S3 is specifically as follows:

introducing an optimized recurrent neural network-LSTM, and selectively memorizing effective information by adopting a gating mechanism;

the output of the Attention-SCNN is sent to a bidirectional LSTM, the number of nodes is 64, so that the network extracts the past and future associated information of the data and captures the comprehensive time sequence characteristics;

adding an attention mechanism in the LSTM, namely extracting an output vector positioned in the middle of the bidirectional LSTM, using a full connection layer as projection to obtain a query vector, performing dot product operation on the query vector and the output, and performing normalization operation by using a softmax function to obtain an attention weight vector; finally, performing dot product operation on the attention distribution value and the output of the BilSTM to obtain a final feature vector; next, mapping the feature vector to a separated hypothesis space by using two full connection layers, wherein the activation function adopts 'relu', and a dropout strategy is added; finally, utilizing a full connection layer with the number of nodes as the number of modulation types to be identified, wherein the activation function of the full connection layer adopts softmax, and obtaining the probability distribution of the modulation signals corresponding to each type; wherein the category corresponding to the maximum value is the recognition result.

A storage medium having stored thereon a computer program which, when executed, performs a deep learning network modulation recognition method based on multi-feature information as described above.

An electronic device comprises a processor and a memory, wherein the memory is used for storing executable commands of the processor, and the processor executes the executable commands to realize a deep learning network modulation identification method based on multi-feature information.

Compared with the prior art, the invention has the beneficial effects that:

one of the beneficial effects of the scheme is that a modulation recognition algorithm combining multi-feature information is provided, starting from a modulation signal, the instantaneous amplitude and the instantaneous phase of the modulation signal are combined with an I/Q signal, the data representation form of each modulation mode is enriched, and the complementation between different types of data features can be realized. An efficient network structure based on the depth separable convolution block and the LSTM is also designed, and an attention mechanism is introduced, so that potential space-time characteristics in the modulation signal can be mined. Experiments on a standard data set show the superiority of the algorithm, and the method shows that for modulation identification tasks, particularly under various interference conditions, a plurality of types of data sources can provide a plurality of observation views for a model, and the difficulty of modulation identification is reduced. In combination with network characteristics, designing a better model structure is also crucial to modulation identification.

Drawings

Fig. 1 is a schematic diagram of a depth separable convolution process according to an embodiment of the present application.

Fig. 2 is a schematic diagram of the recognition accuracy of the added transient features and the non-added transient features according to the embodiment of the present application.

FIG. 3 is a schematic diagram showing the influence of different SepaConvBlock numbers on the experimental results under 0-18dB in the examples of the present application.

Detailed Description

The technical solutions in the embodiments of the present invention are clearly and completely described below with reference to fig. 1 to 3 of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

At present, most of modulation recognition algorithms based on a neural network directly send I/Q signals or instantaneous characteristics into a well-designed network, and then decision judgment is carried out through the neural network, so that recognition of a modulation mode is realized. In a complex and variable electromagnetic environment, potential characteristics of the original I/Q signal may be deficient, and the performance of the model is limited to a certain extent only by providing the I/Q signal or a single transient characteristic. Therefore, in order to enrich the characteristic expression forms of various modulation signals and improve the algorithm robustness so as to facilitate the model to extract modulation characteristics from multiple dimensions and multiple views, a multi-characteristic information modulation identification algorithm is provided, instantaneous characteristics and I/Q signals are combined and sent into a designed model, and experimental results show that the provided method can obtain higher and more stable modulation identification accuracy and has higher feasibility.

Firstly, a deep learning network modulation identification method based on multi-feature information is provided.

In order to overcome the influence of a complex electromagnetic environment in the communication process and more effectively extract high-order characteristics of a modulation signal, a modulation identification algorithm based on multi-characteristic information is provided, an I/Q modulation signal and instantaneous characteristic information are combined, the characteristic information of the modulation signal is enriched, and an efficient neural network model is designed to automatically extract hidden characteristics of various modulation signals and realize high-precision identification of various modulation signals.

In real life, most effective information belongs to low-frequency information and cannot be directly used as transmission signals, and in order to effectively transmit the information in a long distance, a modulation technology is needed to carry out frequency spectrum shifting on baseband signals, so that stable and efficient transmission of the baseband signals is guaranteed. The received signal model y (t) for the receiver is as follows:

(1)

where y (t) represents the actual received I/Q modulated signal, s (t) represents the modulated signal, h (t) represents the channel impulse response, and n (t) represents the additive white gaussian noise in the channel.

The modulation signal contains abundant characteristic information, and how to mine and extract effective information is a very key step. The traditional expert feature extraction method is to perform a series of mathematical operations on an original signal to finally obtain a corresponding feature value, and then analyze the feature value, wherein the feature differences are obvious and can be regarded as high refining of the signal, but under the influence of complex electromagnetic environment and various channel interferences, the situation that a calculation result is greatly different from an ideal condition easily occurs, so that the modulation type identification is difficult. The instantaneous amplitude and phase value of the modulation signal are simple to calculate and contain rich information, so that in order to more fully mine the characteristics of the I/Q modulation signal and avoid the loss of effective information caused by multiple complex calculations, the instantaneous characteristics of the modulation signal and the original signal are combined and then sent to a neural network for automatic characteristic extraction so as to improve the recognition rate of various modulation modes.

The I/Q modulated signal is first subjected to instantaneous feature extraction, including instantaneous amplitude and instantaneous phase values. Since the I/Q modulation signal is divided into I, Q two quadrature signals, I channel signal is x (t) and Q channel signal is y (t), the instantaneous amplitude a (t) and the instantaneous phase p (t) can be obtained by the following equations:

(2)

(3)

then, the obtained instantaneous amplitude and phase value are normalized, and then the original I/Q signal and the instantaneous characteristic value are spliced and fused. The original I/Q signal dimension is 2 × N, 2 represents an I path and a Q path, N represents the number of sampling points, and the data dimension after the instantaneous amplitude and phase information are fused is changed into 4 × N.

And inputting the signal into a designed neural network model to extract potential characteristic information of different types of modulation signals, and finally obtaining probability distribution of each type of the signal, wherein the type with the maximum probability value is a final modulation type identification result.

Many researchers have been dedicated to develop an efficient neural network model to extract the features of the radio signals and perform the recognition task. The excellent network model plays an important role in the task of modulation recognition. The model proposed herein comprises two parts, by cascading Attention-SCNN and bidirectional LSTM with Attention mechanism, in order to expect the model to extract spatio-temporal feature information inside the modulated signal from multi-azimuth, multi-dimensions.

Attention-SCNN: after fusion of multi-feature information, original signal data is expanded from I/Q dual-channel data to multi-dimensional information 4 x N combining I/Q signals and instantaneous features, so that feature information of modulation signals is richer, a neural network model can extract more useful features conveniently, and high-precision identification of various modulation signals in a complex electromagnetic environment is realized.

The fused data is firstly input into an Attention-SCNN structure designed by us, and the structure mainly comprises a plurality of depth separable volume blocks (SepaConvBlock) and a point-by-point convolution layer. In order to reduce the number of parameters of the model and improve the efficiency of network feature extraction, in SepaConvBlock, we use a depth separable convolution layer as a basic layer, compared with the common convolution, the convolution process is mainly divided into two steps of depth convolution and point-by-point convolution as shown in fig. 1, and the convolution process is widely used in a lightweight network structure.

In the SepaConvBlock structure, firstly, 1 × 1 convolution layer is used for realizing cross-channel interaction and information integration, the number of channels is controlled, and then the BatchNormal is used for relieving the influence of internal data distribution deviation and accelerating network training. The batch normalization results are then input into two depth separable convolution layers, where the convolution kernel size is 2 x 2 and the number of output channels is 32. In consideration of the inconsistent influence of different channels of the feature map, in order to further improve the utilization rate of useful information and the network performance, a channel attention mechanism is also introduced.

And modeling the dependency relationship between the channels to adaptively adjust the characteristic response value between each channel. The implementation of this attention mechanism mainly includes two parts of compression and excitation, the compression part is to compress the feature mapping U to the channel descriptor z e Rc by using the global average pooling technique (GAP), wherein the c-th element of z can be calculated by the following formula:

(4)

in the formula, each element of z has a global receptive field of a feature map U, H represents the height of each channel feature map, W represents the width of each channel feature map, and then the series of feature weight vectors are activated mainly through two full-connection layers and a nonlinear activation function, wherein the sigmoid activation function controls the channel weight value to be between 0 and 1. The final scale operation represents multiplying the learned channel weights by the corresponding channel features to obtain a calibrated feature map ũ. In SepaConvBlock, the feature mapping graph after scale operation is connected with the previous input through cross-layer connection, the purpose of residual learning is achieved, and the degradation of network performance can be further avoided.

Considering that the quantity of parameters of the depth separable convolution layer is small, in order to strengthen the search of the model for an ideal feature space and improve the spatial feature extraction capability of the model, two times of similar attention operations are added in each SepaConvBlock, then, the obtained feature maps are integrated by stacking 4 SepaConvBlock layers, and then, the 1 × 1 convolution layer is used for conveniently sending the feature maps into a bidirectional LSTM structure to perform mining of time sequence features.

Time sequence feature extraction: because the modulation signal belongs to a time sequence signal, the recurrent neural network has great success and wide application in the aspect of time sequence data processing. Therefore, in order to further extract potential time sequence characteristics of the modulation signal, an optimized recurrent neural network-LSTM is also introduced into the designed model, a gating mechanism is mainly adopted, effective information can be selectively memorized, and the problem of long sequence dependence in the neural network can be effectively solved.

In the model proposed herein, by feeding the output of the Attention-SCNN into a bidirectional LSTM, the number of nodes is 64, so that the network can extract past and future association information of data, capturing more comprehensive timing characteristics.

And a large amount of information is stored in the bidirectional LSTM, and in order to reduce the influence of redundant information and enable the model to pay attention to more effective characteristic information, an attention mechanism is added in the LSTM so as to expect to filter out irrelevant information and improve the efficiency of the model. The method mainly comprises the steps of extracting an output vector positioned in the middle of a bidirectional LSTM, using a full connection layer as projection to obtain a query vector, then performing dot product operation with the output, and performing normalization operation by using a softmax function to obtain an attention weight vector. And finally, performing dot product operation on the attention distribution value and the output of the BilSTM to obtain a final feature vector. Next, the feature vector is mapped to an easy-to-separate hypothesis space by using two full connection layers, wherein the activation function adopts 'relu', and a dropout strategy is added to prevent an overfitting phenomenon. And finally, utilizing a full connection layer with the number of nodes as the number of modulation types to be identified, wherein the activation function of the full connection layer adopts softmax, and obtaining the probability distribution of the modulation signals corresponding to each type. Wherein the category corresponding to the maximum value is the recognition result.

The specific implementation case is as follows:

the method takes a radio modulation signal standard data set RML2016.10a as a research object, compares the proposed modulation identification algorithm with other related documents, analyzes the modulation identification accuracy of different algorithms under different signal-to-noise ratios, further discusses the influence of network depth on an experimental result, compares the influence of adding transient characteristics with the influence of not adding transient characteristics on the experiment, and explains the effectiveness of the algorithm proposed by the method according to the experimental result.

In the present case, the standard data set rml2016.10a generated by GNU Radio, published in 2016, is used, which comprises a total of 11 common types of modulation signals, such as: PSK, QPSK, 8PSK, 16QAM, 64QAM, BFSK, CPFSK, WB-FM, AM-SSB, AM-DSB and PAM 4. The signal-to-noise ratio of each modulation type is distributed in an interval of-18-20 dB (the interval is 2 dB), the number of each type of modulation signals under each signal-to-noise ratio is 1000, the size of a signal sample is 2 x 128, and 220000 samples are obtained in total. In the data generation process, besides introducing noise, channel transmission influences such as sampling rate offset, center frequency offset, multipath fading and the like are considered so as to approach a real electromagnetic environment, and the method has high research value.

The experimental training equipment adopts an Nvidia GeForce RTX 2080 GPU, and the memory is 8G. The neural network framework selects Keras with tenserflow as the back end. The blocksize in the training process is set to 64, Adam selected by the optimizer is used, the initial learning rate is 0.001, and a learning rate attenuation strategy is introduced, namely when the accuracy of the verification set is not improved in 8 periods, the learning rate is attenuated to one tenth of the original accuracy. The data set division is to randomly divide a training set, a verification set and a test set according to each type of modulation signals corresponding to each signal-to-noise ratio in a ratio of 7:1: 2. And simultaneously, an early stopping mechanism is introduced, namely whether the accuracy of the observation and verification set is improved or not in 10 periods is realized, and if the accuracy is not improved, the training is stopped.

For comparative evaluation of the algorithms proposed herein, we compared with previously popular algorithms, named VTCNN2, CLDNN, GRU, LSTM-AP, respectively. For the whole experiment, the identification accuracy of different modulation types under different signal-to-noise ratios is mainly discussed, the influence of the number of SepaConvBlock on the identification effect is also analyzed, the identification effects of adding no transient feature and fusing the transient feature are compared, and the superiority of the algorithm is demonstrated from multiple aspects and multiple angles.

Because the modulation signal is greatly interfered by noise and the signal distortion is serious in the stage of low signal-to-noise ratio, the accuracy rate of the algorithm provided by the invention is similar to that of other algorithms. When the signal-to-noise ratio is higher than 0dB, the identification accuracy of the algorithm is obviously higher than that of other algorithms, the average accuracy of 0 to 18dB reaches 91.5%, and the same ratio is improved by 1.1% -17.7%, as shown in Table 1. When the signal-to-noise ratio is greater than 4dB, the recognition accuracy is higher than 92% and can reach 93% at most, which explains the robustness of the algorithm provided by the invention. The algorithm starts from the modulation signal, fully considers various feature expression forms of the modulation signal, combines simple transient features with abundant information quantity and I/Q data, utilizes the respective advantages of a convolutional neural network and a cyclic neural network, effectively extracts potential features of each modulation type from time sequence and space dimensions in multiple directions, and then realizes high-precision identification of various modulation signals.

Various other existing modulation identification models are in a confusion matrix generated when SNR =4 dB; generally, the horizontal axis represents the predicted modulation type, the vertical axis represents the actual modulation type, the diagonal line represents the identification accuracy of each modulation type, and the darker the color, the higher the accuracy. After investigation, it can be seen that for all models, the two most difficult types of modulation schemes to distinguish between them are AM-DSB and WBFM, because they both belong to analog modulation schemes and are generated by sampling an analog sound signal that has a quiet period, which makes identification difficult. For 16QAM and 64QAM, they both belong to quadrature amplitude modulation and have overlapped constellation mapping form, which makes the model have certain difficulty in identifying with lower signal-to-noise ratio. Compared with other existing algorithms, the model can greatly relieve the problem, reduces the difficulty in identifying and mixing the two modulation modes, improves the identification precision of multi-system quadrature amplitude modulation (MQAM), shows that the algorithm has higher robustness, and can effectively improve the performance of a modulation identification task.

When the signal-to-noise ratio is in a higher stage, the characteristics contained in the modulation signal are clearer, so that the model extraction is easier. Wherein the various algorithms identify the specific accuracy of each modulation type at a high signal-to-noise ratio of 18 dB. When SNR =18dB, the average accuracy rate of the algorithm is relatively high and reaches 92.7%, and for the types with difficult modulation identification, such as 16QAM, 64QAM, WBFM and the like, the accuracy rate of the model reaches the highest and has obvious improvement amplitude. The results show the superiority of the algorithm, show that the time sequence and spatial characteristics in the modulation signals are effectively mined by combining multi-characteristic information, the effect of the modulation recognition task can be improved, and the recognition accuracy of certain modulation categories which are easy to be confused is improved.

In order to explore the role of transient features in the modulation identification task, the comparative experiment was performed mainly by changing the form of input data. I.e. directly sending the I/Q signal into the designed network for modulation identification, and combining the instantaneous characteristic and the I/Q signal and then sending into the same network for modulation identification. As can be seen from the graph 2, the modulation recognition accuracy rate after the instantaneous characteristics are fused is higher than that of the I/Q signals which are directly sent to a model for recognition within 0-18 dB. The method discloses the effectiveness of the combination of multi-feature information on automatic modulation identification, and reflects that potential features carried by the I/Q signal per se may have certain defects on modulation identification. The combination of the I/Q signals and the instantaneous characteristics enriches the characteristic representation form of each modulation mode, and facilitates the neural network to clear the internal relation of each modulation mode, thereby improving the accuracy of modulation identification.

In the deep learning field, the performance of neural network models is usually affected by the depth of the network. Therefore, to explore the effect of network depth on automatic modulation recognition, the network depth is mainly controlled by changing the designed number of sepaconvblocks in the model. As can be seen from FIG. 3, the identification accuracy of the model generally tends to increase first and then decrease with the change of the number of SepaConvBlock between 0 and 18 dB. The mean accuracy of 0-18dB was highest at 91.5% when the number of sepaconvblocks was 4 (all sepaconvblocks in the model were 4, not specifically described elsewhere herein). This shows that the network depth has a certain influence on modulation identification, but an excessively complex model may cause an overfitting phenomenon, and the model capacity should be reasonably controlled when the model is designed.

In summary, a modulation identification algorithm combining multi-feature information is provided, starting from a modulation signal itself, instantaneous amplitude and instantaneous phase of the modulation signal are combined with an I/Q signal, a data representation form of each modulation mode is enriched, and complementation between different types of data features can be achieved. An efficient network structure based on the depth separable convolution block and the LSTM is also designed, and an attention mechanism is introduced, so that potential space-time characteristics in the modulation signal can be mined. Experiments on a standard data set show the superiority of the algorithm, and the method shows that for modulation identification tasks, particularly under various interference conditions, a plurality of types of data sources can provide a plurality of observation views for a model, and the difficulty of modulation identification is reduced. In combination with network characteristics, designing a better model structure is also crucial to modulation identification. Generally speaking, many problems to be solved still exist in the current modulation identification field, for example, how to improve the accuracy of modulation identification under a low signal-to-noise ratio, and how to accurately identify the modulation mode of a modulation signal when the modulation signal has a certain degree of aliasing, are worth further research.

The above are preferred embodiments of the present invention, and all changes made according to the technical scheme of the present invention that produce functional effects do not exceed the scope of the technical scheme of the present invention belong to the protection scope of the present invention.

Claims

1. The deep learning network modulation identification method based on the multi-feature information is characterized by comprising the following steps of:

s3: introducing an optimized recurrent neural network-LSTM on the basis of the step S2, sending the output of the integration of the Attention-SCNN structure to the recurrent neural network-LSTM, extracting the past and future associated information of the data, and further capturing the comprehensive time sequence characteristics;

step S1 is specifically as follows:

firstly, performing instantaneous feature extraction on an I/Q modulation signal, wherein the instantaneous feature extraction comprises an instantaneous amplitude value and an instantaneous phase value; since the I/Q modulation signal is divided into I, Q two orthogonal signals, let the I channel signal be x (t) and the Q channel signal be y (t), the instantaneous amplitude a (t) and the instantaneous phase p (t) can be obtained by the following equations:

(2)

(3)

the original I/Q signal dimension is 2 × N, 2 represents an I path and a Q path, N represents the number of sampling points, and the data dimension after the instantaneous amplitude and phase information are fused is changed into 4 × N;

step S2 is specifically as follows:

the data after multi-feature fusion is firstly input into an Attention-SCNN structure, wherein the Attention-SCNN structure is composed of a plurality of depth separable volume blocks and a point-by-point convolution layer, the depth separable volume blocks are SepaConvBlock, in the SepaConvBlock, the depth separable volume layers are taken as a basic layer, and the convolution process is divided into two steps of depth convolution and point-by-point convolution;

(4)

2. The deep learning network modulation recognition method based on multi-feature information as claimed in claim 1, wherein step S3 is as follows:

the output of the Attention-SCNN is sent to a bidirectional LSTM (BiLSTM), the number of nodes of the bidirectional LSTM is 64, so that the network extracts the past and future associated information of the data and captures the comprehensive time sequence characteristics;

then adding an attention mechanism in the BilSTM, namely extracting an output vector positioned in the middle of the BilSTM, using a full-connection layer as projection to obtain a query vector, performing dot product operation on the query vector and the output, and performing normalization operation by using a softmax function to obtain an attention weight vector; finally, performing dot product operation on the attention weight vector and the output of the BilSTM to obtain a final feature vector; next, mapping the feature vector to a separated hypothesis space by using two full connection layers, wherein the activation function adopts 'relu', and a dropout strategy is added; finally, utilizing a full connection layer with the number of nodes as the number of modulation types to be identified, wherein the activation function of the full connection layer adopts softmax, and obtaining the probability distribution of the modulation signals corresponding to each type; wherein the category corresponding to the maximum value is the recognition result.

3. A storage medium having stored thereon a computer program which, when executed, performs a method of deep learning network modulation recognition based on multi-feature information according to claim 1 or 2.

4. An electronic device, comprising a processor and a memory, wherein the memory is used for storing executable commands of the processor, and the processor implements a deep learning network modulation recognition method based on multi-feature information according to claim 1 or 2 by executing the executable commands.