Automatic modulation identification method based on lightweight space-time fusion combination network

Abstract

The invention relates to an automatic modulation identification method based on a lightweight space-time fusion combined network, and belongs to the field of automatic modulation identification of communication signals. According to the method, I / Q components of a to-be-recognized signal are shaped into a time stream and a space stream respectively, the time stream and the space stream are input into a time feature extraction channel and a space feature extraction channel respectively, time features and space features are extracted, then the extracted features are connected and input into a classification module, and finally a recognition result, namely the modulation type of the to-be-recognized signal, is output. According to the lightweight space-time fusion combined network model provided by the invention, by reducing the number of convolution channels and the hidden layer dimension of the LSTM network, the model parameter quantity and the floating point calculation quantity are reduced, and meanwhile, the sufficient extraction of the features is ensured, so that relatively high recognition accuracy is realized. Compared with the prior art, the method has the advantages of being small in parameter quantity, low in calculation amount, high in recognition accuracy and the like, and is suitable for resource-limited equipment.

Description

Technical Field

[0001] The invention belongs to the field of automatic modulation recognition of communication signals and relates to an automatic modulation recognition method based on a lightweight space-time fusion combined network. Background Art

[0002] Automatic Modulation Recognition (AMR) is an important research direction in the field of wireless communications. Its purpose is to automatically identify the modulation type of the received signal. AMR technology is widely used in the fields of communication, military and signal monitoring. Traditional automatic modulation recognition algorithms can be divided into two categories, namely likelihood-based methods and feature-based methods. Likelihood-based methods are based on statistics and usually classify modulation types by maximizing the likelihood function of the received signal under certain assumptions. Likelihood-based methods have high theoretical optimality, especially when the channel conditions are ideal, and can achieve low error probability. However, since it requires complex statistical modeling and calculation of the signal, its computational cost is high for high-dimensional signals or complex scenarios, and its real-time and robustness may be limited. Feature-based methods extract signal features, such as amplitude, phase, instantaneous frequency, waveform, etc., and use these features to build a recognition model to classify the modulation type. Compared with likelihood-based methods, feature-based methods have lower computational complexity and are more suitable for real-time processing in actual scenarios. However, their performance depends largely on the quality of feature design and the selection of classifiers. In the case of low signal-to-noise ratio or complex channel conditions, the performance may decline.

[0003] In recent years, deep learning has developed rapidly and has achieved remarkable results in many traditional fields. For example, convolutional neural networks (CNN) have shown excellent performance in the field of image processing, bringing breakthrough progress to tasks such as image classification, target detection and semantic segmentation. In 2016, O'Shea et al. first introduced CNN, a deep learning algorithm, into automatic modulation recognition to classify the modulation mode of radio signals. The model showed better performance under low signal-to-noise ratio. O'shea et al. also provided a benchmark dataset that comprehensively considered various harsh radio channel environment interference conditions, which facilitated the comparison of automatic modulation recognition algorithms. Unlike traditional feature-based methods that separate feature extraction and classifiers, for automatic modulation recognition algorithms based on deep learning, the deep learning model itself can undertake both feature extraction and classification tasks. CNN network and RNN network models are widely used in deep learning algorithms for automatic modulation recognition. Over time, CNN and RNN models are combined and used in deep learning algorithms for automatic modulation recognition, further improving the recognition accuracy. Wu et al. combined CNN and LSTM networks to form a CLDNN network model to explore temporal and spatial correlations. The input of the model is the I / Q signal component. Zhang et al. considered the I / Q form and A / P form of the signal, and used CNN and LSTM in series to extract features for the two types of signals. The features learned from the two streams interacted in pairs, which increased the diversity of features. Xu et al. considered that the existence of amplitude and phase imbalance deteriorated the orthogonal relationship between the I channel and the Q channel, and proposed an efficient multi-stream structure to process the I / Q multi-channel data of the original modulated signal, and supplemented it with independent I / Q channel input to fully extract features, so that the accuracy of modulation recognition was greatly improved. Although the performance of the network reached an excellent level, the network had a large number of parameters and a large amount of calculation, which was not conducive to deployment on a computing-constrained platform. Ke et al. proposed a learning framework based on LSTM autoencoder to extract stable and robust features, and use the learned features to infer the modulation type. The algorithm designed a compact neural network architecture and was easy to implement on a low-cost computing platform. Although this model solves the problem of large number of model parameters, the network needs to calculate the phase and amplitude of the IQ components and normalize the data, which will bring additional delay and calculation.

[0004] In summary, there are many problems in current modulation recognition: large number of model parameters, large amount of floating-point calculations, low accuracy, and some models require preprocessing of the received I / Q components, which brings additional calculations and delays. Therefore, it is urgent to design a lightweight, high-recognition-accuracy modulation recognition network that does not require additional preprocessing of the received I / Q components. Summary of the invention

[0005] In view of this, the purpose of the present invention is to provide an automatic modulation recognition method based on a lightweight spatiotemporal fusion combined network. The method can significantly reduce the number of model parameters while improving the recognition accuracy under different signal-to-noise ratio conditions. Unlike the prior art, the present invention does not require additional preprocessing operations on the I / Q components, but directly inputs them into the network model for automatic modulation recognition. The method provided by the present invention includes constructing and training a lightweight spatiotemporal fusion combined network model, obtaining a signal I / Q component input model to be identified, and outputting the modulation type after network calculation.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] An automatic modulation recognition method based on a lightweight spatiotemporal fusion combined network comprises the following steps:

[0008] Obtaining I / Q components of the signal to be identified;

[0009] Reshaping the I / Q components into a time stream and a spatial stream respectively;

[0010] Inputting the time stream into a time feature extraction channel to extract time features;

[0011] Inputting the spatial stream into a spatial feature extraction channel to extract spatial features;

[0012] Connecting the temporal features and the spatial features;

[0013] The connected features are input into a classification module to output a recognition result, wherein the recognition result is the modulation type of the signal to be recognized.

[0014] Furthermore, the time feature extraction channel includes:

[0015] One layer of one-dimensional convolution;

[0016] A one-dimensional batch normalization layer;

[0017] Two-layer LSTM network.

[0018] Furthermore, the spatial feature extraction channel includes:

[0019] One layer of 2D convolution;

[0020] One 2D batch normalization layer;

[0021] A convolution block comprising three one-dimensional convolution layers and three one-dimensional batch normalization layers;

[0022] One layer of LSTM network;

[0023] One adaptive global average pooling layer.

[0024] Further, the classification module includes:

[0025] One fully connected layer.

[0026] Further, the I / Q components are respectively shaped into a time stream and a space stream, specifically:

[0027] The I / Q components are connected in series and in parallel respectively to form a time stream and a space stream.

[0028] Furthermore, the time stream is input into the time feature extraction channel, specifically:

[0029] The time stream is subjected to the one-dimensional convolution to obtain feature data with reduced step size dimension;

[0030] The feature data is passed through the two-layer LSTM network to extract time information, and the feature data of the last time step is taken as the input of the concatenate layer.

[0031] Furthermore, the spatial stream is input into a spatial feature extraction channel, specifically:

[0032] The spatial stream is subjected to the two-dimensional convolution to extract spatial features, and then subjected to the convolution block to deeply extract spatial features to obtain feature data;

[0033] The feature data is extracted through the LSTM network, and then reshaped and passed through the adaptive global average pooling layer to obtain feature data, which is then input into the concatenate layer.

[0034] Further, the temporal feature and the spatial feature are connected, specifically:

[0035] The temporal features and spatial features are connected through a concatenate layer.

[0036] Further, the connected features are input into a classification module to output a recognition result, specifically:

[0037] The connected features are classified through the batch normalization layer and the fully connected layer to obtain the recognition result.

[0038] Furthermore, the parameters of the temporal feature extraction channel, the spatial feature extraction channel and the classification module are obtained through training.

[0039] The beneficial effects of the present invention are as follows: the present invention reshapes the I / Q and sends them to the time feature extraction channel and the space feature extraction channel respectively, and then connects the features extracted from the time channel and the space channel using a concatenate layer and then obtains the classification result through a full connection. While fully extracting the time and space features of the I / Q components, the amount of parameters and floating-point calculations are greatly reduced, and a high recognition accuracy is guaranteed. At the same time, the model can directly recognize the I / Q signal received by the receiver without additional preprocessing, thereby reducing additional calculations and delays.

[0040] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0042] Figure 1 This is the structural diagram of the LW-STCLDNN model of the present invention;

[0043] Figure 2 This is the internal structure diagram of the convolution block ConvBlock of the present invention;

[0044] Figure 3 The recognition accuracy of the signal under different signal-to-noise ratios of LW-STCLDNN and MCDLNN models;

[0045] Figure 4 Confusion matrices for LW-STCLDNN and MCDLNN models at snr=4dB. DETAILED DESCRIPTION

[0046] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0047] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0048] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0049] The present invention provides an automatic modulation recognition method based on a lightweight space-time fusion combined network, including constructing and training a lightweight space-time fusion combined network (LW-STCLN) model, obtaining the I / Q component of the modulation signal to be identified and inputting the trained lightweight space-time fusion combined network model, and outputting the identified modulation type. The lightweight space-time fusion combined network training steps are as follows:

[0050] 1. Obtain a data set of I / Q component signals of different modulation types in an environment with characteristic transmission parameters similar to those of the real channel.

[0051] Specifically, for the received radio signal r(t):

[0052] r(t)=g[s(t)]*h(t)+n(t)

[0053] Automatic modulation identification is an intermediate process between receiver detection and demodulation. Its goal is to predict the modulation scheme of the received signal, that is, the function g(□), and then provide modulation information to recover the modulated signal □(□) from the received signal r(t). The I / Q components are the real and imaginary parts of the received signal r(t). The data sets include RadioML2016.10a and RadioML2016.10b, which are generated using the simulated channel in GNU Radio. RadioML2016.10a contains 11 modulation types, namely 8PSK, AM-DSB, AM-SSB, BPSK, CPFSK, GFSK, 4-PAM, 16-QAM, 64-QAM, QPSK, and WBFM. The signal-to-noise ratio of the signal covers -20db to 18db, and there are 220,000 I / Q component signal samples. Compared with RadioML2016.10a, the RadioML2016.10b dataset has one less modulation type, AM-SSB, and has a larger sample size of 1,200,000. Each sample in the two datasets is a 2x128 matrix, indicating that there are two I / Q features and the number of sampling points is 128.

[0054] 2. Encode the different modulation types of the acquired data set, and divide the entire data set into training set, validation set, and test set according to the ratio of 8:1:1.

[0055] Specifically, different modulation types of the acquired data set are encoded. The modulation type encoding results of the RadioML2016.10a data set are as follows: "8PSK":0,"BPSK":1,"CPFSK":2,"GFSK":3,"PAM4":4,"QAM16":5,"QAM64":6,"QPSK":7,"AM-DSB":8,"AM-SSB":9,"WBFM":10. The modulation type encoding results of the RadioML2016.10b data set are as follows: "8PSK":0,"BPSK":1,"CPFSK":2,"GFSK":3,"PAM4":4,"QAM16":5,"QAM64":6,"QPSK":7,"AM-DSB":8,"WBFM":9.

[0056] 3. Build a network model based on lightweight spatiotemporal fusion combination.

[0057] Specifically, the model structure is as follows Figure 1 As shown, the lightweight spatiotemporal fusion combined network model includes a temporal feature extraction channel, a spatial feature extraction channel, and a classification module;

[0058] The temporal feature extraction channel contains one layer of one-dimensional convolution, the convolution kernel size is 2, the number of convolution kernels is 16, one layer of batch normalization layer, two layers of LSTM network layer, the input feature size and hidden layer dimension are both 16; the spatial feature extraction channel contains one layer of two-dimensional convolution, the convolution kernel height is 2, the width is 1, the number of convolution kernels is 16, one layer of convolution block, which contains 3 one-dimensional convolution and 3 batch normalization layers, the convolution kernel size of one-dimensional convolution is 3, the number of convolution kernels is 16, one layer of LSTM layer, the input feature size and output feature size are both 16; the classification module has only one fully connected layer, the number of input features is 32, and the number of output features is the number of modulation types. The eps of all batch normalization layers are set to 1e-5 and the momentum is set to 0.1.

[0059] Batch normalization is a regularization technique used in deep neural networks to reduce internal covariate shift, thereby accelerating training and improving model stability. It normalizes the input of each layer to have zero mean and unit variance, making model training more efficient. For a small batch input {x1,x2,...,x m}, the process of batch normalization is as follows:

[0060] a: Calculate the mean and variance:

[0061]

[0062] Where m is the batch size.

[0063] b: Normalization:

[0064]

[0065] where ε is a small constant that prevents division by zero.

[0066] c: Rescaling and offsetting: Among them, γ and β are learnable parameters, used for scaling and offset respectively, and the parameters are fixed after training. The network uses batch normalization layers for the following purposes:

[0067] a. Accelerate training convergence: By normalizing the input of each layer, the problem of activation values ​​being too large or too small is avoided, making gradient propagation more stable.

[0068] b. Improve model robustness: Reduce the sensitivity of weight initialization to training, so that the model performs well under a wider range of hyperparameter settings.

[0069] c. Reduce overfitting: Although batch normalization itself is not a regularization method, its normalization effect and small batch noise can indirectly play a regularization role.

[0070] d. Reduce sensitivity to learning rate: Batch normalization allows the model to use larger learning rates without causing gradient explosion or gradient vanishing problems.

[0071] 4. Set the initial learning rate, learning rate adjustment strategy, training rounds, batch size, and loss function type.

[0072] Specifically, the Adam optimizer with an initial learning rate of 0.001 and a batch size of 512 is used. If the validation loss does not decrease after 5 epochs during training, the learning rate is adjusted with a decay factor of 0.8. The main purpose is to allow the optimizer to jump out of the local optimum and find a better solution. The training rounds are set to a maximum of 200. If the validation loss does not improve after 20 epochs, the training is stopped. The loss function uses the cross entropy loss entropy function. For a sample x, its true category label is y, and the probability distribution predicted by the model is The cross entropy loss function is defined as:

[0073]

[0074] C is the total number of categories, yi is the encoding of the true category, is the model's predicted probability for the i-th class.

[0075] 5. The I / Q components in the training set are shaped into a lightweight spatiotemporal fusion network model constructed by inputting time stream and space stream. Specifically, I / Q can be expressed by the following formula:

[0076] I k = {I k,0 ,I k,1 ,I k,2 ,…,I k,N-1 ,I k,N}

[0077] Q k = {Q k,0 ,Q k,1 ,Q k,2 ,…,Q k,N-1 ,Q k,N}

[0078] The I / Q components are shaped into time streams and space streams, as shown below:

[0079]

[0080] Stream kt =[I k,1 …I k,N Q k,1 …Q k,N ]

[0081] Where K represents the Kth sample in the data set, and N represents the number of sampling points of the signal. According to the parameters of the open source data set, N is 128, so the shape of the spatial stream tensor is (1, 2, 128), and the shape of the temporal stream tensor is (1, 1, 256). The spatial stream needs to extract spatial features through the spatial channel. The spatial stream (1, 2, 128) first passes through a layer of two-dimensional convolution with a convolution kernel height of 2, a width of 1, and an output channel of 16. Among them, 128 means that there are 128 time steps, and 2 represents that there are two features, I and Q, at each time step. Therefore, a convolution kernel with a height of 2 is used to help the model capture the spatial (feature) dependency in the data, and the output channel 16 is to capture different feature patterns in the input signal. After a layer of two-dimensional convolution, the data shape obtained is feature data of (1, 16, 1, 128). In order to reduce the amount of calculation of the subsequent convolution block and minimize feature loss, a two-dimensional maximum pooling of size (1, 2) is used. After the pooling layer, the dimension of the time step of the feature data is reduced by half. Then enter the convolution block with three layers of one-dimensional convolution (see Figure 2), the convolution kernel size of each one-dimensional convolution is 3, and the output channel is 16; the shape of the feature data does not change after being processed by the convolution block. Its main function is to perform deeper feature extraction on the features extracted after the first layer of convolution, and then extract the temporal dependency in its sequence through a layer of LSTM network. The input feature size and hidden layer dimension are both 16. Since convolution only extracts the spatial features of I / Q, the extracted spatial features are output in the form of time series. Therefore, LSTM is used to further extract spatial features, and the obtained feature data shape is (1, 64, 16). Then the Transpose function is used to exchange dimensions to change the shape to (1, 16, 64). The purpose is to adapt to the adaptive global average pooling, so that the entire time series of each input channel is averaged and pooled, and finally the spatial feature data with a shape of (1, 16) is output. The time stream needs to pass through the time feature extraction channel to extract the time features, so LSTM is mainly used to extract its time dependency. First, in order to keep the input step size of 128 unchanged and extract the time features of multiple feature expressions of the input signal, the time stream passes through a one-dimensional convolution with a convolution kernel size of 2 and an output channel of 16. Then, a two-layer LSTM network is used, with both the input feature size and the hidden layer dimension of 16. By stacking multiple LSTM layers, the model's expressive power is increased, enabling it to capture more complex and multi-level time series dependencies and better extract the time features of the I / Q signal. Finally, the time features extracted by the time channel and the spatial features extracted by the spatial channel are concatenated, with an input shape of (1, 32), and then passed through a fully connected layer with an input feature of 32 and an output feature of 11 to obtain the classification result. Most of the traditional combined networks for automatic modulation recognition directly input the I / Q signals into the convolution to extract their features, and then use the LSTM network to extract features after the convolution has extracted the features. Although the two networks are combined to take advantage of their advantages, the accuracy has not been significantly improved. At the same time, the hidden dimensions of the convolution channel and the LSTM network are large, resulting in huge model parameters and floating-point calculations, which is not conducive to the deployment of resource-constrained devices. The combined network model proposed in the present invention reduces the number of convolution channels and the hidden layer dimension of the LSTM network to reduce the number of model parameters, while fully considering the spatial feature extraction advantages of convolution and the temporal feature advantages of LSTM, shaping the input I / Q components into time streams and space streams, and then sending them to the time channel and space channel for feature extraction respectively, and then concatenating the dual-channel features and sending them to the fully connected classification module for classification, and then reducing the number of model parameters while ensuring that the features are fully extracted to achieve a higher accuracy.

[0082] 6. Calculate the training loss function and back-propagate to update the model parameters until the model converges.

[0083] 7. Input the test set into the trained lightweight spatiotemporal fusion combined network model to obtain the modulation recognition result.

[0084] In order to illustrate the effect of the present invention, a simulation experiment is performed to further illustrate:

[0085] Simulation conditions: NVIDIA GeForce A6000 GPU is used for training, the network model is implemented using pytorch, and the dataset uses the open source datasets RadioML2016.10a and RadioML2016.10b.

[0086] Simulation results: The model of the present invention is used to train on two training sets respectively, calculate the recognition accuracy of the test set, and draw a curve chart of the recognition accuracy changing with the signal-to-noise ratio. In order to highlight the advantages of the present invention, the network of the present invention is compared with the currently excellent combined network benchmark model MCLDNN. The MCLDNN network input adopts three-channel input, which are I / Q parallel, and a separate I path and a separate Q path. The three paths are connected together after convolution, and then pass through two layers of LSTM for time feature extraction, and then pass through two layers of fully connected layers, and finally classify and output the results. Figure 3 (a) shows the curve of recognition accuracy of the LW-STCLN and MCLDNN model test set changing with the signal-to-noise ratio under the RadioML2016.10a dataset. Figure 3 (b) shows the curve of recognition accuracy of the LW-STCLN and MCLDNN model test set as a function of signal-to-noise ratio under the RadioML2016.10b dataset. Figure 4 (a) Figure 4 (b) Figure 4 (c) Figure 4 (d) represents the confusion matrix of the snr = 4dB signal sample in the RadioML2016.10a and RadioML2016.10b data sets in the LW-STCLN and MCLDNN models. At the same time, in order to show the advantages of the present invention over other benchmark models, the parameters, floating point calculation numbers, and recognition accuracy of the present invention are compared with those of CNN1, CNN2, CNN_LSTM, and LSTM models, and the results are shown in Table 1.

[0087] Table 1 Comparison of the parameters, floating-point calculation numbers, and recognition accuracy of the present invention and the current high recognition rate model in the RadioML2016.10a dataset.

[0088] Table 1

[0089]

[0090]

[0091] Depend on Figure 3 (a) and Figure 3 (b) It can be seen that the accuracy of the model of the present invention is basically close to that of the MCLDNN model in both data sets. Figure 4 (a) and Figure 4 (b) It can be seen that the LW-STCLDNN model has confusion with 16-QAM and 64-QAM signals and WBFM signals on the RML2016.10a dataset, but other modulation types can be classified correctly, which is the same problem as the MCLDNN model. Figure 4 (c) and Figure 4 (d) It can be seen that the LW-STCLDNN model has significantly improved the confusion of 16-QAM and 64-QAM signal recognition on the RML2016.10b data set, but the confusion of WBFM recognition still exists. As shown in Table 1, the model of the present invention has significantly reduced the number of parameters and floating-point calculations compared with MCLDNN while ensuring that the accuracy is close to that of the excellent combined network model MCLDNN. At the same time, compared with some other benchmark models, not only the number of parameters and floating-point calculations have certain advantages, but also the recognition accuracy exceeds these benchmark models.

[0092] In summary, the present invention fully combines the advantages of CNN and LSTM networks, extracts time features and spatial features from I / Q signal components, and then performs modulation recognition. Simulation results show that while ensuring the recognition accuracy comparable to that of the current excellent combined network model, the present invention significantly reduces the number of model parameters and floating-point calculations, showing excellent efficiency and practicality. Compared with some other benchmark models, the present invention not only has significant advantages in model complexity, but also achieves further improvement in recognition accuracy.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. An automatic modulation recognition method based on a lightweight spatiotemporal fusion combined network, characterized in that: The following steps are involved: Obtaining I / Q components of the signal to be identified; Reshaping the I / Q components into a time stream and a spatial stream respectively; Inputting the time stream into a time feature extraction channel to extract time features; Inputting the spatial stream into a spatial feature extraction channel to extract spatial features; Connecting the temporal features and the spatial features; The connected features are input into a classification module to output a recognition result, wherein the recognition result is the modulation type of the signal to be recognized.

2. The automatic modulation recognition method based on lightweight spatiotemporal fusion combined network according to claim 1 is characterized in that: The time feature extraction channel includes: One layer of one-dimensional convolution; A one-dimensional batch normalization layer; Two-layer LSTM network.

3. The automatic modulation recognition method based on lightweight spatiotemporal fusion combined network according to claim 1 is characterized in that: The spatial feature extraction channel includes: One layer of 2D convolution; One 2D batch normalization layer; A convolution block comprising three one-dimensional convolution layers and three one-dimensional batch normalization layers; One layer of LSTM network; One adaptive global average pooling layer.

4. The automatic modulation recognition method based on lightweight spatiotemporal fusion combined network according to claim 1 is characterized in that: The classification module comprises: One fully connected layer.

5. The automatic modulation recognition method based on lightweight spatiotemporal fusion combined network according to claim 1 is characterized in that: The shaping of the I / Q components into a time stream and a space stream respectively is specifically as follows: The I / Q components are connected in series and in parallel respectively to form a time stream and a space stream.

6. The automatic modulation recognition method based on lightweight spatiotemporal fusion combined network according to claim 1 is characterized by: The time stream inputs the time feature extraction channel, specifically: The time stream is subjected to the one-dimensional convolution to obtain feature data with reduced step size dimension; The feature data is passed through the two-layer LSTM network to extract time information, and the feature data of the last time step is taken as the input of the concatenate layer.

7. The automatic modulation recognition method based on lightweight spatiotemporal fusion combined network according to claim 1 is characterized by: The spatial stream is input into the spatial feature extraction channel, specifically: The spatial stream is subjected to the two-dimensional convolution to extract spatial features, and then subjected to the convolution block to deeply extract spatial features to obtain feature data; The feature data is extracted through the LSTM network, and then reshaped and passed through the adaptive global average pooling layer to obtain feature data, which is then input into the concatenate layer.

8. The automatic modulation recognition method based on lightweight spatiotemporal fusion combined network according to claim 1 is characterized by: The connecting of the time feature and the space feature is specifically as follows: The temporal features and spatial features are connected through a concatenate layer.

9. The automatic modulation recognition method based on lightweight spatiotemporal fusion combined network according to claim 1 is characterized by: The connected features are input into a classification module to output a recognition result, specifically: The connected features are classified through the batch normalization layer and the fully connected layer to obtain the recognition result.

10. The automatic modulation recognition method based on lightweight spatiotemporal fusion combined network according to claim 1 is characterized by: The parameters of the temporal feature extraction channel, the spatial feature extraction channel and the classification module are obtained through training.

Images