Communication signal modulation identification method based on deep learning
Through deep learning combined with CNN and LSTM networks, the time-frequency characteristics of communication signals are extracted, and the problem of signal modulation recognition in complex electromagnetic environments is solved, and the rapid and accurate recognition is achieved in complex backgrounds and adapted to spectrum environments of multiple frequency bands.
Patent Information
- Application Number
- CN202411302521.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-18
- Publication Date
- 2024-12-13
AI Technical Summary
The existing communication signal modulation recognition methods are difficult to accurately identify new modulation patterns in complex electromagnetic environments. Traditional methods rely on manual feature extraction and cannot adapt to rapidly changing spectrum environments, and lack general signal modulation recognition models.
Using a deep learning-based method, combining convolutional neural network (CNN) and long and short-term memory network (LSTM), time-frequency features are extracted through short-term Fourier transform, and a communication signal modulation identification network model combining CNN and LSTM is constructed. Automatically extract signal features for identification.
It improves the accuracy and efficiency of signal modulation methods, reduces the dependence on manual experience, and can quickly and accurately identify different modulation methods in complex contexts, adapting to spectrum environments of multiple frequency bands.
Smart Images
Figure CN119135489A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a communication signal modulation recognition method, and in particular to a communication signal modulation recognition method based on deep learning. Background Art
[0002] With the rapid development of communication technology, the increase in the number of mobile devices and communication transmission rates, the increasing number and complexity of modulation styles in communication systems, and the emergence of various new application scenarios, the electromagnetic environment has become increasingly complex, and traditional modulation recognition technology has gradually become difficult to meet application needs. At present, deep learning technology has achieved significant performance improvements in many fields such as computer vision, natural language processing, image processing, and satellite communications. It is conceivable that combining deep learning with modulation recognition will make the development of modulation recognition a step further. Therefore, it is of great significance to study signal modulation recognition using deep learning technology.
[0003] The modulation recognition method of the signal can be divided into two types: manual recognition and automatic recognition, depending on whether manual parameter extraction is required. With the development of communication technology, the types of communication signal modulation are increasing, and the modulation algorithms are becoming more and more complicated. Digital modulation has become the main position, and manual recognition is gradually declining, and automatic recognition has gradually replaced it.
[0004] In the automatic recognition method of communication signal modulation, although the application of neural network models in the field of modulation recognition has achieved certain results, these methods all assume that the training and test data are extracted from a common closed set from the same label and feature space, that is, closed set recognition. The premise of deep neural networks is to collect enough types and a large amount of signal data. However, it is difficult to collect enough samples for some signals, especially in the military field. A more realistic scenario is open set recognition (OSR). Once an unknown signal that does not appear in the training set appears in the test set, the neural network can only classify it into the known class with the highest probability score, and cannot regard it as unknown. Therefore, in the face of increasingly complex electromagnetic environments and endless new modulation styles, the closed set recognition model has gradually become inapplicable. How to accurately judge new modulation styles in an open environment has become a hot topic of current research. In 2016, A. Bendale et al. first proposed the OpenMax open set recognition algorithm and proposed a "distance vector" as a standard for distinguishing between categories. The recognition accuracy rate can reach 92% when tested on the MINST public data set. In 2017, GE Z et al. used generative adversarial networks (GANs) to generate fake data of unknown categories for the network to learn features, which improved the performance of the original openmax algorithm by 3%-5%. In 2019, Hao Yunfei et al. proposed a reconstruction discriminant network model based on GAN for one-dimensional signal data. When the signal-to-noise ratio is greater than 0dB, the recognition rate of signals with known and unknown modulation modes is greater than 93%. In 2022, Qin Bowei et al. designed a data reconstruction and extreme value theory generative adversarial network model to solve the problem of low accuracy of open set recognition of modulated signals under lightweight models, which significantly reduced the complexity of the algorithm.
[0005] With the continuous deepening of communication technology exploration, especially in non-cooperative communication scenarios such as radio spectrum monitoring, signal modulation recognition has become increasingly critical. The difficulty in identifying communication signal modulation methods lies in the fact that the feature parameters required to extract different digital modulation methods are different, and as the number of modulation methods increases, the recognition system needs to continuously improve and modify the algorithm. Based on the above difficulties, there is currently no universal signal modulation recognition model, and traditional recognition methods will develop completely different recognition systems based on recognition requirements, and feature extraction is highly dependent on personal experience. With the day-to-day deterioration of the spectrum environment in each frequency band, how to quickly and accurately identify signal modulation methods in complex background environments is an important research direction at present.
[0006] The automatic modulation recognition method based on deep learning and convolutional neural network can directly obtain hidden feature information from the signal time-frequency domain data, without the need for manual feature extraction, thus reducing the difficulty of recognition. The present invention combines the deep learning method to propose a communication signal modulation recognition method based on deep learning. Summary of the invention
[0007] At present, the signal recognition method using the general signal modulation mode recognition model is to formulate completely different recognition systems according to the recognition requirements, and the extraction of signal features depends on personal experience. With the day-to-day deterioration of the spectrum environment in each frequency band, how to quickly and accurately recognize the signal modulation mode in a complex background environment is an important research direction at present. Therefore, the present invention proposes a communication signal modulation recognition method based on deep learning.
[0008] The above purpose is achieved through the following technical solutions:
[0009] A communication signal modulation recognition method based on deep learning is implemented by the following steps:
[0010] Step 1: Determine that the communication signal dataset is the public dataset RML2016.10A;
[0011] Step 2: Use short-time Fourier transform to extract the time-frequency characteristics of the communication signal, and perform localized analysis on the communication signal in time and frequency. Each point in the time-frequency diagram represents the characteristics of the signal at the corresponding time and frequency.
[0012] Step three, construct a communication signal modulation recognition network model combining a convolutional neural network CNN and a long short-term memory network LSTM, and use the time-frequency features extracted by short-time Fourier transform in step one as the input of the network model, and the output of the network model is the modulation mode of the communication signal.
[0013] Furthermore, the RadioML2016.10a data set used in step 1 specifically includes 8 digital modulation modes: 8PSK, BPSK, CPFSK, GFSK, PAM4, 16QAM, 64QAM, QPSK; and 3 analog modulation modes: AM-DSB, AM-SSB, and WBFM.
[0014] Furthermore, the process of extracting the time-frequency characteristics of the communication signal by using short-time Fourier transform described in step 2 is specifically:
[0015] Short Time Fourier Transform, or STFT for short, is a general audio signal processing method that describes the sinusoidal frequency and phase content of a local part of a signal as it changes over time. It uses a window to decompose a longer signal into shorter signal segments, and then performs Fourier transform on each segment of the signal. The STFT pair is expressed as follows:
[0016]
[0017] Where x[k] is the original time domain signal, g[k] is the analysis window of length L, and the STFT of x[k] is interpreted as the Fourier transform of x[k]·g[km];
[0018] STFT is used to reflect the local time-frequency characteristics of the signal and jointly represent the time-frequency domain information. STFT uses a fixed-resolution window function g[k] in the time-frequency analysis process. The time-domain resolution and frequency-domain resolution need to satisfy the Heisenberg inequality (uncertainty principle), as shown in the following formula:
[0019]
[0020] Where Δt is the signal time width and Δf is the signal bandwidth.
[0021] Furthermore, the process of constructing a communication signal modulation recognition network model combining a convolutional neural network CNN and a long short-term memory network LSTM described in step 3 is specifically: based on the convolutional neural network CNN and the long short-term memory network LSTM dual network, the spatial and temporal features of the signal are jointly extracted, and the features are identified through a classifier; the convolutional neural network CNN in the fusion network structure includes: a convolutional layer, a pooling layer, a convolutional layer, a pooling layer, and two fully connected layers, and the long short-term memory network LSTM includes: a convolutional layer, a convolutional layer, an LSTM layer and a fully connected layer, and the output of the fully connected layer of the convolutional neural network CNN is feature fused with the output of the fully connected layer of the long short-term memory network LSTM, and the feature is input to the Softmax layer for prediction;
[0022] The convolutional neural network (CNN) network comprises four layers. The convolutional layer processes the input data, extracts features and increases the dimension of the feature map. After the dimension increases, the image size becomes larger, and then the pooling process is performed. The pooling layer processes each small area, retaining only representative features and reducing the data scale. After that, the features are transformed through the fully connected layer, and the obtained local feature maps are globally integrated. Finally, the prediction result is obtained by the Softmax layer.
[0023] The weight sharing of convolutional neural network (CNN) reduces the number of parameters in the convolution kernel, that is, the parameters of the filter remain unchanged during the traversal operation of the image; automatic feature extraction is based on an adaptive method to extract abstract features of the image; convolutional neural network (CNN) autonomously extracts representative features based on the input raw data, and the receptive field can present the local feature correlation of the feature map;
[0024] The long short-term memory network LSTM uses h and c to represent the transmission state, and x to represent the current input. It selectively learns data features and updates the transmission state through three gates. The specific principles of the forget gate, input gate, and output gate are as follows;
[0025] Under the action of the forget gate, the data first selectively learns some features, f t =σ(U f h t-1 +W f x t );
[0026] The transmission state c is updated under the action of the input gate:
[0027] i t =σ(U i h t-1 +W i x t );
[0028] c t =tanhσ(U g h t-1 +W g x t )⊙i t +c t-1 ⊙f t
[0029] In the output gate part, the transmission state ht and output are obtained through the activation function:
[0030] o t =σ(U o h t-1 +W o x t )
[0031] h t =tanh(c t )⊙o t
[0032] The long short-term memory network LSTM adjusts the opening and closing degrees of the forget gate, the input gate, and the output gate to achieve targeted reception or output of data.
[0033] The beneficial effects of the present invention are:
[0034] The present invention constructs a deep learning method of a communication signal modulation recognition network model combining a convolutional neural network CNN and a long short-term memory network LSTM to perform modulation recognition on the signal. Compared with the traditional pattern recognition method, the neural network can automatically learn sample features from a large number of samples instead of using manually designed expert features. A network model is built using a multi-layer neural network, and the network is trained to study the optimal parameter settings, explore the internal connections between data, and finally use the extracted features for classification and recognition to improve the accuracy of classification. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 is a flow chart of the method of the present invention;
[0036] Figure 2 It is a diagram of the fusion network structure involved in the present invention;
[0037] Figure 3 It is the structure of the convolutional neural network CNN involved in the present invention;
[0038] Figure 4 It is a diagram of the LSTM unit structure involved in the present invention. DETAILED DESCRIPTION
[0039] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Specific implementation method one:
[0041] A communication signal modulation recognition method based on deep learning in this embodiment, such as Figure 1 As shown, the method is implemented by the following steps:
[0042] Step 1: Determine that the communication signal dataset is the public dataset RML2016.10A;
[0043] Step 2: Use short-time Fourier transform to extract the time-frequency characteristics of the communication signal, and perform localized analysis on the communication signal in time and frequency. Each point in the time-frequency diagram represents the characteristics of the signal at the corresponding time and frequency. Since the time domain and frequency domain characteristics of different signals are usually different, even if two signals are similar at a certain time or frequency, they may still be different at other times or frequencies. Therefore, the time-frequency characteristics obtained by short-time Fourier transform of two different signals are different, and the time-frequency diagram can uniquely represent a certain signal.
[0044] Step three, construct a communication signal modulation recognition network model combining a convolutional neural network CNN and a long short-term memory network LSTM, convert the signal modulation recognition method into an image classification method, and use the time-frequency features extracted by short-time Fourier transform in step one as the input of the network model. The output of the network model is the modulation mode of the communication signal. Specific implementation method 2:
[0046] Different from the first specific implementation mode, the present implementation mode is a communication signal modulation identification method based on deep learning. The RadioML2016.10a data set used in step one specifically includes 8 digital modulation modes: 8PSK, BPSK, CPFSK, GFSK, PAM4, 16QAM, 64QAM, QPSK; and 3 analog modulation modes: AM-DSB, AM-SSB, and WBFM. Specific implementation method three:
[0048] Different from the second specific implementation mode, in the present implementation mode, a method for modulation identification of communication signals based on deep learning, the process of extracting the time-frequency features of the communication signals by using short-time Fourier transform described in step 2 is specifically:
[0049] Short Time Fourier Transform, or STFT for short, is a powerful general audio signal processing method and a typical linear time-frequency analysis method. STFT is related to FT and is a natural extension of FT. It describes the sinusoidal frequency and phase content of the local part of the signal as the signal changes over time, solving the problem of signal non-stationarity. From a mathematical point of view, STFT is a series of FTs of windowed signals. It uses a window to decompose a longer signal into shorter signal segments, and then Fourier transforms each segment of the signal separately. The STFT pair is expressed as follows:
[0050]
[0051] Where x[k] is the original time domain signal, g[k] is the analysis window of length L. As can be seen from the above formula, the STFT of x[k] is interpreted as the Fourier transform of x[k]·g[km];
[0052] STFT is used to reflect the local time-frequency characteristics of the signal and jointly represent the time-frequency domain information; however, STFT uses a fixed-resolution window function g[k] in the time-frequency analysis process. A narrow-width window can produce better time-domain resolution, but will produce poor frequency-domain resolution. In addition, the time-domain resolution and frequency-domain resolution cannot be arbitrarily small and need to satisfy the Heisenberg inequality (uncertainty principle), as shown in the following formula:
[0053]
[0054] Where Δt is the signal time width and Δf is the signal bandwidth. Specific implementation method four:
[0056] Different from the third specific implementation mode, the present implementation mode is a method for communication signal modulation recognition based on deep learning. The process of constructing a communication signal modulation recognition network model combining a convolutional neural network CNN and a long short-term memory network LSTM as described in step three is specifically: in practical applications, it is necessary not only to know the global information of the signal, but also to require accurate local information. However, the convolution operation, activation function, and pooling operation in the CNN network will lose some information. This information may be invalid or may be the key to solving the problem. If only the CNN network is used, the timing characteristics of the original data will be lost to a certain extent after the original data undergoes the convolution operation; if only the LSTM network is used, although the timing characteristics of the original data are extracted, the accuracy of the original data image is reduced to a certain extent. Therefore, the design of the present invention is based on the dual network of the convolutional neural network CNN and the long short-term memory network LSTM to jointly extract the spatial and temporal features of the signal, and identify it through a classifier; the fusion network structure is as follows: Figure 2 As shown in the figure, the convolutional neural network CNN includes: convolutional layer, pooling layer, convolutional layer, pooling layer, and 2 fully connected layers. The long short-term memory network LSTM includes: convolutional layer, convolutional layer, LSTM layer and fully connected layer. The output of the fully connected layer of the convolutional neural network CNN is feature fused with the output of the fully connected layer of the long short-term memory network LSTM, and input to the Softmax layer for prediction.
[0057] The convolutional neural network (CNN) network contains four layers. The convolutional layer processes the input data, extracts features and increases the dimension of the feature map. After the dimension increases, the size of the image will also increase. The image is pooled after the convolution operation. The pooling layer processes each small area and only retains the most representative features. Therefore, the pooling layer helps to reduce the data scale and improve the computing efficiency. After that, the previous features are transformed through the fully connected layer, and the obtained local feature maps are globally integrated. Finally, the prediction result is obtained by the Softmax layer. The structure of the convolutional neural network (CNN) is as follows: Figure 3 As shown;
[0058] The characteristics of CNN in information processing are reflected in weight sharing, automatic feature extraction and receptive field. The significance of weight sharing is to reduce the number of parameters in the convolution kernel, that is, the parameters of the filter remain unchanged during the traversal operation of the image. Automatic feature extraction is based on an adaptive method to extract abstract features of the image. CNN autonomously extracts representative features based on the input raw data. The receptive field can present the local feature correlation of the feature map. When identifying modulated signals, this method can extract more abstract features from the signal;
[0059] The LSTM network is an optimization and deformation based on the recurrent neural network. It overcomes the disadvantage of the recurrent neural network that it only has short-term memory and can handle long-term memory problems. Therefore, the network performs well in solving time-related problems and is widely used in natural language processing and time-series signal processing. Its structure is as follows: Figure 4 shown.
[0060] The long short-term memory network LSTM uses h and c to represent the transmission state, and x to represent the current input. It selectively learns data features and updates the transmission state through three gates. The specific principles of the forget gate, input gate, and output gate are as follows;
[0061] Under the action of the forget gate, the data first selectively learns some features, f t =σ(U f h t-1 +W f x t );
[0062] The transmission state c is updated under the action of the input gate:
[0063] i t =σ(U i h t-1 +W i x t );
[0064] c t =tanhσ(U g h t-1 +W g x t )⊙i t +c t-1 ⊙f t
[0065] In the output gate part, the transmission state ht and output are obtained through the activation function:
[0066] o t =σ(U o h t-1 +W o x t )
[0067] h t =tanh(c t )⊙o t
[0068] The long short-term memory network LSTM adjusts the opening and closing degree of the forget gate, input gate, and output gate to achieve targeted reception or output of data. This method has obvious adaptability and flexibility. There are three such "gates" LSTM. Therefore, the control method based on "gates" improves the efficiency of responding to the external environment and makes the network more expressive.
[0069] The embodiments of the present invention disclose preferred embodiments, but are not limited thereto. A person skilled in the art can easily understand the spirit of the present invention based on the above embodiments and make different extensions and changes. However, as long as they do not deviate from the spirit of the present invention, they are all within the protection scope of the present invention.
Claims
1. A communication signal modulation recognition method based on deep learning, characterized in that: The method is implemented by the following steps: Step 1: Determine that the communication signal dataset is the public dataset RML2016.10A; Step 2: Use short-time Fourier transform to extract the time-frequency characteristics of the communication signal, and perform localized analysis on the communication signal in time and frequency. Each point in the time-frequency diagram represents the characteristics of the signal at the corresponding time and frequency. Step three, construct a communication signal modulation recognition network model combining a convolutional neural network CNN and a long short-term memory network LSTM, and use the time-frequency features extracted by short-time Fourier transform in step one as the input of the network model, and the output of the network model is the modulation mode of the communication signal.
2. The communication signal modulation recognition method based on deep learning according to claim 1, characterized in that: The RadioML2016.10a dataset used in step 1 specifically includes 8 digital modulation modes: 8PSK, BPSK, CPFSK, GFSK, PAM4, 16QAM, 64QAM, QPSK; and 3 analog modulation modes: AM-DSB, AM-SSB, and WBFM.
3. The communication signal modulation recognition method based on deep learning according to claim 2, characterized in that: The process of extracting the time-frequency characteristics of the communication signal by using short-time Fourier transform described in step 2 is specifically: Short Time Fourier Transform, or STFT for short, is a general audio signal processing method that describes the sinusoidal frequency and phase content of a local part of a signal as it changes over time. It uses a window to decompose a longer signal into shorter signal segments, and then performs Fourier transform on each segment of the signal. The STFT pair is expressed as follows: Where x[k] is the original time domain signal, g[k] is the analysis window of length L, and the STFT of x[k] is interpreted as the Fourier transform of x[k]·g[km]; STFT is used to reflect the local time-frequency characteristics of the signal and jointly represent the time-frequency domain information. STFT uses a fixed-resolution window function g[k] in the time-frequency analysis process. The time-domain resolution and frequency-domain resolution need to satisfy the Heisenberg inequality (uncertainty principle), as shown in the following formula: Where Δt is the signal time width and Δf is the signal bandwidth.
4. The communication signal modulation recognition method based on deep learning according to claim 3 is characterized in that: The process of constructing a communication signal modulation recognition network model combining a convolutional neural network CNN and a long short-term memory network LSTM as described in step 3 is specifically: based on the dual networks of the convolutional neural network CNN and the long short-term memory network LSTM, the spatial and temporal features of the signal are jointly extracted, and the features are identified through a classifier; the convolutional neural network CNN in the fusion network structure includes: a convolutional layer, a pooling layer, a convolutional layer, a pooling layer, and two fully connected layers, and the long short-term memory network LSTM includes: a convolutional layer, a convolutional layer, an LSTM layer and a fully connected layer. The output of the fully connected layer of the convolutional neural network CNN is feature fused with the output of the fully connected layer of the long short-term memory network LSTM, and the feature is input to the Softmax layer for prediction; The convolutional neural network (CNN) network comprises four layers. The convolutional layer processes the input data, extracts features and increases the dimension of the feature map. After the dimension increases, the image size becomes larger, and then the pooling process is performed. The pooling layer processes each small area, retaining only representative features and reducing the data scale. After that, the features are transformed through the fully connected layer, and the obtained local feature maps are globally integrated. Finally, the prediction result is obtained by the Softmax layer. The weight sharing of convolutional neural network (CNN) reduces the number of parameters in the convolution kernel, that is, the parameters of the filter remain unchanged during the traversal operation of the image; automatic feature extraction is based on an adaptive method to extract abstract features of the image; convolutional neural network (CNN) autonomously extracts representative features based on the input raw data, and the receptive field can present the local feature correlation of the feature map; The long short-term memory network LSTM uses h and c to represent the transmission state, and x to represent the current input. It selectively learns data features and updates the transmission state through three gates. The specific principles of the forget gate, input gate, and output gate are as follows; Under the action of the forget gate, the data first selectively learns some features, f t =σ(U f h t-1 +W f x t ); The transmission state c is updated under the action of the input gate: i t =σ(U i h t-1 +W i x t ); c t =tanhσ(U g h t-1 +W g x t )⊙i t +c t-1 ⊙f t In the output gate part, the transmission state ht and output are obtained through the activation function: the t =σ(U o h t-1 +W o x t ) h t =tanh(c t )⊙o t The long short-term memory network LSTM adjusts the opening and closing degrees of the forget gate, the input gate, and the output gate to achieve targeted reception or output of data.