USRP modulation signal identification method based on migration deep learning

Abstract

The invention relates to the technical field of software radio transmission, and discloses a USRP modulation signal identification method based on migration deep learning, a wireless signal is transmitted through a transmitting end, after a USRP receiving end receives the signal, the signal is received and stored through a signal acquisition program, and the USRP modulation signal identification method based on migration deep learning is realized. Providing a required signal training set and a signal test set for a CNN model for modulation signal identification, wherein the signal training set comprises a pre-training data set and a transfer learning training data set; pre-training a convolutional neural network model for modulation signal identification through the generated pre-training data set; pre-training a convolutional neural network model for modulation signal identification through the generated transfer learning training data set; deploying the trained transfer learning convolutional neural network model to a signal test set, and verifying a modulation signal identification result; according to the invention, the modulation signal identification problem in wireless communication is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of software radio transmission, and in particular to a USRP modulation signal recognition method based on migration deep learning. Background Art

[0002] Modulation recognition technology has important application value in both commercial and military fields. Modulation recognition technology can be divided into modulation recognition based on likelihood ratio and modulation recognition based on signal features. Modulation recognition based on likelihood ratio uses the principle of statistical inference. By constructing a likelihood ratio model of the modulation type, comparing the likelihood ratios of known modulation types, and comparing the signal to be identified with the likelihood ratios of various modulation types, the modulation type of the unknown signal is determined; modulation recognition based on signal features extracts the characteristic parameters or characteristic sequences of the signal (including spectral characteristics, power spectral density, phase change, instantaneous amplitude, modulation index, etc.), and then determines the modulation type of the signal based on these characteristics.

[0003] Because the likelihood ratio-based method needs to accurately establish the likelihood ratio model of the modulation type, it may face the challenge of model construction for complex modulation types; the signal feature-based method is more flexible and can adapt to different modulation types according to different feature parameters, but it requires appropriate feature extraction and decision rule design, so the modulation recognition technology based on signal features is usually more widely used. The traditional modulation recognition technology based on signal features first designs and extracts signal features, and then designs classification rules to classify the modulation format of signal features. Such manually designed signal features and classification rules usually have limited recognition accuracy in complex channel environments. In recent years, the development of deep learning has promoted its widespread application in computer vision, natural language processing and other fields, and has achieved remarkable results. Therefore, deep learning is introduced in modulation recognition technology. Automatically learn classification rules from feature data to improve classification accuracy. Deep learning algorithms can realize automatic learning of signal features and classification rules in the field of modulation recognition, thereby further improving the performance of modulation recognition.

[0004] Compared with traditional schemes, the modulation signal scheme based on deep learning has its advantages, especially in signal feature processing. At the same time, this scheme also has shortcomings. It requires effective training of convolutional neural networks, which puts forward certain requirements on the scale of training sets. For scenarios with small-scale data training, its recognition performance is subject to certain limitations. In addition, when the channel environment changes, the trained model cannot meet the performance requirements of modulation signal recognition. In this case, a modulation signal recognition scheme based on deep transfer learning is proposed to solve these scenario defects. Summary of the invention

[0005] The purpose of the present invention is to solve the above problems and design a USRP modulation signal recognition method based on migration deep learning.

[0006] The present invention provides a USRP modulation signal recognition method based on migration deep learning, and the USRP modulation signal recognition method based on migration deep learning comprises the following steps:

[0007] Step 1, signal data acquisition: The wireless signal is transmitted through the transmitter. After receiving the signal, the USRP receiver receives and stores the signal through the signal acquisition program, and provides the required signal training set and signal test set for the CNN model of modulation signal recognition, wherein the signal training set includes a pre-training data set and a transfer learning training data set;

[0008] Step 2: Pre-training of the convolutional neural network model: Pre-training the convolutional neural network model for modulated signal recognition using the generated pre-training data set;

[0009] Step 3: Transfer learning training of convolutional neural network model: pre-train the convolutional neural network model for modulated signal recognition using the generated transfer learning training data set;

[0010] Step 4: Deploy the trained transfer learning convolutional neural network model to the signal test set to verify the modulated signal recognition results.

[0011] Optionally, in the first implementation of the present invention, the transmitting end RF parameters in step 1 include at least RF antenna, center frequency, transmission gain, and modulation mode, and the USRP receiving end RF parameters include at least RF antenna, center frequency, and receiving gain.

[0012] Optionally, in a second implementation of the present invention, in step 1, the USRP receiving end starts to receive wireless information from the receiving antenna, passes through the RF receiving module, undergoes analog-to-digital conversion, and down-conversion processing to receive a modulated signal.

[0013] Optionally, in a third implementation of the present invention, the pre-training data set in step 1: the classification signal set used in the channel environment I is Ω4 = {BPSK, 8PSK, 4QAM, 16QAM}, the signal sequence length is L = 100, and the average signal-to-noise ratio of the sequence is uniformly distributed between [-10dB, 20dB], and the number of each type of signal is N I =100000, the pre-training data set is set to receive N signals of each type, with a total of 4N signal sequences. During training, 90%×4N are taken as the training set, and 10%×4N are used for cross-validation. The test set has 1000 signals of each type, and the SNR range is [-10dB, 20dB]. Tests are performed at each interval and the average is taken to obtain the modulation signal recognition rate.

[0014] Optionally, in a fourth implementation of the present invention, in step 1, the transfer learning training data set: in channel environment II, the number of each type of signal in the transfer learning training set is set to N II ,In the test set, there are 1000 signals of each type, the SNR range is [-10dB, 20dB], and the test is performed at every 2dB interval. The average result is taken to obtain the modulation signal recognition rate.

[0015] Optionally, in the fifth implementation of the present invention, step 3 specifically includes: the USRP receiver collects signal data sets under two types of different channel environments, uses a transfer learning convolutional neural network to identify modulated signals, and performs transfer training on the convolutional layer and the fully connected layer based on the convolutional neural network.

[0016] Optionally, in a sixth implementation of the present invention, a transfer learning convolutional neural network uses two convolution layers with a convolution step of 1. The first convolution layer has 64 kernels, the size of the first convolution kernel is 2×4, the excitation function is ReLU, a bias term is used and initialized to zero, the weight initialization method is Glorot uniform distribution initialization, the number of convolution kernels of the second convolution layer is 16, the size of the convolution kernel is 1×4, and the other settings are the same as the first convolution layer.

[0017] Optionally, in a seventh implementation of the present invention, the transfer learning convolutional neural network fully connected neural network includes three fully connected layers, the first two layers are set to 64 and 16 neurons respectively, the activation function is ReLU, the weight is initialized with Glorot uniform distribution, and regularization terms and weight constraints are used. The output layer activation function uses Softmax.

[0018] In the technical solution provided by the present invention, transmission is performed through the transmitting end of the wireless signal, and after receiving the signal, the USRP receiving end receives and stores the signal through a signal acquisition program, so as to provide the required signal training set and signal test set for the CNN model of modulation signal recognition. In this process, the signal training set includes a pre-training data set and a transfer learning training data set. First, the convolutional neural network model for modulation signal recognition is pre-trained through the generated pre-training data set to improve the initial performance of the model. Then, the convolutional neural network model for modulation signal recognition is further pre-trained through the generated transfer learning training data set to adapt to the specific modulation signal recognition task. Finally, the trained transfer learning convolutional neural network model is deployed on the signal test set, and the modulation signal recognition result is verified to ensure the accuracy and reliability of the model. The present invention successfully verifies that the signal modulation recognition technology based on deep learning can effectively solve the modulation signal recognition problem in wireless communication through a communication test environment built based on USRP software radio, and demonstrates its great potential and advantages in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the following detailed description of the preferred embodiment.The drawings are only for the purpose of illustrating the preferred embodiments and are not to be construed as limiting the invention.

[0020] Figure 1 A schematic diagram of a USRP modulation signal recognition method based on transfer deep learning provided by an embodiment of the present invention;

[0021] Figure 2 A functional block diagram of a transmitter provided in an embodiment of the present invention;

[0022] Figure 3 A functional block diagram of a receiving end provided by an embodiment of the present invention;

[0023] Figure 4 A schematic diagram of the deep learning neural network structure of a CNN provided in an embodiment of the present invention;

[0024] Figure 5 A schematic diagram of a transmitter panel configuration according to an embodiment of the present invention;

[0025] Figure 6 A schematic diagram of a receiving terminal panel setting provided in an embodiment of the present invention;

[0026] Figure 7 A USRP signal acquisition program diagram provided by an embodiment of the present invention;

[0027] Figure 8A comparison diagram of transfer learning and other solutions provided in an embodiment of the present invention when transfer training samples are limited;

[0028] Fig. 9 A comparison chart of transfer learning and other solutions when there are sufficient transfer training samples provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0029] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, device, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0030] Modulation signal recognition refers to the process of identifying the modulation method of the received signal in wireless communication. Modulation is the process of converting a digital or analog signal into a signal form suitable for transmission on a wireless channel, while modulation signal recognition is to determine its modulation method by analyzing the characteristics of the modulated signal after receiving it. In wireless communication, modulation signal recognition is very important to ensure correct signal demodulation and high-quality communication, and is an important technical link in wireless communication systems.

[0031] For ease of understanding, the specific process of the embodiment of the present invention is described below. The USRP modulation signal recognition method based on transfer deep learning provided by the embodiment of the present invention specifically includes the following steps:

[0032] Step 1, signal data acquisition: The wireless signal is transmitted through the transmitter. After receiving the signal, the USRP receiver receives and stores the signal through the signal acquisition program, and provides the required signal training set and signal test set for the CNN model of modulation signal recognition, wherein the signal training set includes a pre-training data set and a transfer learning training data set;

[0033] In this embodiment, regarding the pre-training data set: the classification signal set used in the channel environment I is Ω4 = {BPSK, 8PSK, 4QAM, 16QAM}, the signal sequence length is L = 100, and the average signal-to-noise ratio of the sequence is uniformly distributed between [-10dB, 20dB], and the number of each type of signal is N I=100000. The training set is set to receive N signals of each type, with a total of 4N signal sequences. During training, 90%×4N are taken as the training set, 10%×4N are used for cross-validation, and the test set has 1000 signals of each type, with an SNR range of [-10dB, 20dB]. Tests are performed at each interval and the average is taken to obtain the modulation signal recognition rate.

[0034] About the transfer learning training data set: In channel environment II, the number of each type of signal in the transfer learning training set is set to N II ,In the test set, there are 1000 signals of each type, the SNR range is [-10dB, 20dB], and the test is performed at every 2dB interval. Finally, the average result is taken to obtain the modulation signal recognition rate.

[0035] Step 2: Pre-training of the convolutional neural network model: Pre-training the convolutional neural network model for modulated signal recognition using the generated pre-training data set;

[0036] Step 3: Transfer learning training of convolutional neural network model: pre-train the convolutional neural network model for modulated signal recognition using the generated transfer learning training data set;

[0037] Step 4: Deploy the trained transfer learning convolutional neural network model to the signal test set to verify the modulated signal recognition results.

[0038] In this embodiment, LabVIEW graphical programming is performed on the host computer to realize a transceiver system that can switch multiple modulation signals; after receiving the signal, the USRP receiver receives and stores the signal through the signal acquisition program, and provides the required signal training set and signal test set for the CNN model of modulation signal recognition; deep learning is used to recognize the modulation signal and train the CNN model; the trained CNN model is deployed to the test set data, and after the training is completed, the trained deep learning model is tested and verified. The trained modulation signal recognition module is deployed to the actual application scenario for modulation recognition of the receiver. Through the communication test environment built based on the USRP software radio, it is verified that the signal modulation recognition technology based on deep learning can well solve the modulation signal recognition problem in wireless communication.

[0039] In this embodiment, LabVIEW (Laboratory Virtual Instrument Engineering Workbench) is a graphical programming language, which is mainly used in fields such as data acquisition, instrument control and industrial automation. LabVIEW, with its unique graphical programming method, enables engineers and scientists to quickly develop complex measurement and control systems. The core of LabVIEW is a programming model based on data flow. Users build programs by dragging and dropping graphical functions and structures. These graphical elements are called virtual instruments (VIs for short). LabVIEW provides a wealth of libraries and tools to support various data acquisition hardware devices, such as data acquisition cards, GPIB devices, serial devices, etc. In addition, LabVIEW also supports interfaces with other programming languages, such as C, Python, etc., so that users can flexibly expand its functions. Another notable feature of LabVIEW is its powerful data visualization function. Users can easily create various charts and graphical interfaces to display data and results in real time. This makes LabVIEW very popular in test, measurement and control systems. In addition, LabVIEW also provides a wealth of data analysis and signal processing tools to help users process and analyze data; LabVIEW is a powerful graphical programming tool that is widely used in engineering, scientific research and industrial automation. Its unique programming method and powerful functions enable users to efficiently develop complex measurement and control systems.

[0040] In this embodiment, the core functions of the software radio, including frequency conversion, analog-to-digital conversion (A / D) and digital-to-analog conversion (D / A) and RF driver, are all implemented by the hardware platform USRP RIO. In addition, all other functions of the software radio are designed and implemented through software, which means that, except for the most basic transceiver function, all extended functions need to be designed and programmed by themselves. This design method makes the software radio extremely flexible and scalable, and can be customized and optimized according to different application scenarios and requirements.

[0041] In this embodiment, some main physical parameters of USRP RIO are as follows: USRO RIO device has 2×2 multi-input / multi-output RF transceiver, which can realize dual transmission and dual reception. Moreover, the maximum transmission signal power of RF bandwidth transceiver is 30dBm, which can receive or transmit signals from 1.2GHz to 6GHz, and the maximum instantaneous bandwidth can reach 40MHz. The maximum bidirectional point-to-point bandwidth of USRP RIO device is 830MB / s, and its digital back-end interface is PCle×4 bus, and its speed can reach 800MB / s.

[0042] In this embodiment, the software part of the experiment uses two programming languages, Labview2015 and Python3.7, to design and debug the program; Labview2015 is used to build the user interface and perform preliminary data processing, while Python3.7 is used to implement more complex algorithms and deep learning models. By utilizing the powerful library support of Python3.7, a convolutional neural network (CNN) model of deep transfer learning is designed to process and analyze signal data; in terms of signal acquisition, a universal software radio peripheral (USRP) device is used to receive wireless signals. Through USRP, the required signal data can be efficiently collected and a detailed data set containing modulated signals can be generated. This data set provides rich training samples for subsequent deep learning model training; in addition, on the basis of Python3.7, the RF transceiver driver provided by the software itself is expanded, and a series of functions that meet the requirements of the present invention are added. The expansion of these functions enables better control of the USRP device and more accurate signal acquisition and processing. Through these improvements, the software was successfully combined with the hardware platform USRP to realize the design and implementation of wireless signal transmission communication; in the implementation process of the entire invention, the advantages of Labview2015 and Python3.7 were fully utilized, combined with the powerful performance of the USRP hardware platform, to ensure the efficiency and accuracy of signal acquisition, processing and transmission. Through this combination of software and hardware, not only the flexibility and scalability of the experiment are improved, but also a solid foundation for further research and development in the future is laid. Figure 1 As shown in the figure, in the experimental environment, a transmitter and a receiver are set up. After the signal is transmitted and received through the wireless communication process of USRP, LabVIEW collects data on the received modulated signal to form a data set. Using deep learning technology, the data set is provided to the convolutional neural network for training and learning, and offline modulation signal recognition and processing is performed.

[0043] The experimental program is mainly divided into four parts: transmitter and receiver, data set extraction module, and CNN recognition module.

[0044] Transmitter and receiver experimental setup:

[0045] In the process of conducting data communication experiments, transmitters and receivers play a vital role. In order to meet the experimental requirements, a USRP (Universal Software Radio Peripheral) device is used, which has high flexibility and configurability. On the USRP device, multiple antennas can be installed and configured to achieve multiple communication modes; specifically, two single antennas are configured on this USRP device, which serve as transmitting antennas and receiving antennas respectively. Although these two antennas are physically located on the same device, in the experiment, the data communication between them is completed through a wireless channel. This setting makes it possible to simulate the process of transmitting data in a real wireless environment; in this way, the performance of the wireless communication system can be effectively simulated and tested under laboratory conditions. This self-transmitting and self-receiving SISO (single input single output) configuration not only simplifies the complexity of the experimental equipment, but also provides an in-depth understanding of the characteristics of the wireless channel. In this way, the performance of the wireless communication system can be better evaluated and optimized to ensure that the expected effect can be achieved in practical applications.

[0046] Taking the transmitter and receiver as an example, Figure 2 This is the functional block diagram of the transmitter. Figure 3 The functional block diagram of the receiver is Figure 2 It can be seen that at the transmitting end, starting from the signal source, after source coding, channel coding, signal modulation, USRP up-conversion, digital-to-analog conversion (DAC), it is sent to the wireless channel through the RF transmitting module. In order to modify the RF parameters of the transmitting end in real time, such as center frequency, gain, etc., an external expansion interface is added to the RF transmitting module to realize the call of the transmission parameter modification module; Figure 3 It can be seen that at the receiving end, wireless information is received from the receiving antenna, passes through the RF receiving module, and then undergoes analog-to-digital conversion (CAD) and down-conversion processing to receive the modulated signal. The RF receiving module adds an external expansion interface to achieve real-time modification of the RF parameters of the receiving end.

[0047] Dataset extraction module settings:

[0048] The wireless signal is transmitted through the transmitter. After receiving the signal, the USRP receiver receives and stores the signal through the signal acquisition program, providing the required signal training set and signal test set for the CNN model of modulated signal recognition.

[0049] CNN recognition module settings:

[0050] Figure 4 Finalize the convolutional neural network for the deep learning neural network structure of CNN, and select appropriate network parameters and training parameters for the deep learning framework. The specific settings are:

[0051] 1. Feature extraction network design: Two convolution layers are used, and the convolution step size is 1. A two-layer convolution layer scheme with 64 kernels in the first layer is adopted. The size of the first convolution kernel is 2×4, the excitation function is ReLU, the bias term is used and initialized to zero, the weight initialization method is Glorot uniform distribution initialization, and the convolution kernel does not consider the completion method, regularization term and weight constraint; the number of convolution kernels in the second convolution layer is 16, the size of the convolution kernel is 1×4, and the rest of the settings are the same as the first convolution layer.

[0052] 2. Recognition network design: The subsequent fully connected neural network includes three fully connected layers. The first two layers are set to 64 and 16 neurons respectively, and the activation function is ReLU; the weight initialization adopts the Glorot uniform distribution method, which is also called Xavier uniform distribution. In this way, the weight can maintain a small variance at the time of initialization, so that the signal can maintain a relatively balanced propagation in each layer of the neural network. In addition, in order to prevent overfitting and ensure the generalization ability of the model, a regularization term is introduced in the training process. The regularization term constrains the size of the weight by adding an additional penalty term to the loss function, thereby avoiding excessive weights and causing the model to be too complex. Common regularization methods include L1 regularization and L2 regularization, among which L2 regularization is also called weight decay. In this way, the complexity of the model can be effectively controlled and its performance on unknown data can be improved; the output layer activation function adopts Softmax.

[0053] The specific training parameters for CNN are: using Adam optimization method and multi-category cross entropy loss function. The learning rate is set to 10 -4 , learning rate update related parameters β1=0.9,β2=0.999; fuzzy factor is set to 10 -8 , the decay value of the learning rate after update is 0.

[0054] The experimental process of this embodiment is as follows:

[0055] To ensure efficient operation of the wireless communication system, set the signal transmission and reception parameters and collect signal data. Figure 5 and Figure 6, which are the adjustable RF parameters of the transmitter and the receiver respectively. The RF parameters of the transmitter are set as follows: RF antenna, center frequency, transmission gain, modulation mode, etc., where the RF antenna ensures that the signal can be effectively transmitted. The center frequency is the core frequency point of signal transmission. The adjustment of the transmission gain determines the transmission strength of the signal. The modulation mode directly affects the transmission quality and efficiency of the signal; the RF parameters of the receiver are set as follows: RF antenna, center frequency, receiving gain, etc., where the RF antenna ensures that the signal can be effectively received. The center frequency needs to be consistent with the transmitter to ensure the correct reception of the signal. The adjustment of the receiving gain determines the sensitivity and range of the signal reception. By setting the configuration, wireless communication transmission of different modulation signals under different signal-to-noise ratios can be achieved, which not only improves the stability and reliability of the communication system, but also ensures that efficient signal transmission can be maintained in various complex environments.

[0056] The user starts to prepare for communication. There are 8 types of modulation signals for transmission between transceivers: 2ASK, 4ASK, BPSK, 4PSF.

[0057] Perform signal data acquisition: Figure 7 It is a USRP signal acquisition program. The wireless signal is transmitted through the transmitter. After receiving the signal, the USRP receiver receives and stores the signal through the signal acquisition program, providing the required signal training set and signal test set for the CNN model of modulation signal recognition.

[0058] Pre-train the CNN model: Use the generated pre-training dataset to pre-train the CNN model for modulated signal recognition proposed in this paper;

[0059] Conduct transfer learning training for the CNN model: Use the generated transfer learning training dataset to pre-train the CNN model for modulated signal recognition proposed in this paper;

[0060] The trained CNN model is deployed to the test set data to verify the modulation signal recognition results.

[0061] Figure 8 The experimental results are shown in the figure for comparison between transfer learning and other schemes when the transfer training samples are limited. Through experimental analysis, it can be found that when the transfer training samples are limited, the method based on deep transfer learning is lower than the scheme based on average likelihood ratio test, but it still has recognition performance advantages. Compared with the method of retraining CNN, which has basically no effect and the performance of feature processing of instantaneous features combined with full neural network, the machine learning scheme based on CNN deep transfer learning algorithm shows its advantages in the case of limited transfer training samples, and can still have relatively excellent modulation signal recognition performance under low data training.

[0062] Fig. 9 The experimental results of transfer learning and other schemes are shown in the figure. When there are sufficient transfer training samples, the scheme based on CNN deep transfer learning and retraining CNN achieve basically the same modulation signal recognition performance, and the modulation signal recognition performance under high signal-to-noise ratio is close to the peak performance, that is, the traditional ALRT modulation signal recognition performance. In contrast, the scheme based on instantaneous features of modulation signals and full neural network (TF-NN) is slightly weaker than the above methods. The reason is that the signal feature extraction effect of the convolutional layer structure in CNN is better than the feature extraction effect of instantaneous features.

[0063] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A USRP modulation signal recognition method based on transfer deep learning, characterized in that: The USRP modulation signal recognition method based on migration deep learning includes the following steps: Step 1, signal data acquisition: The wireless signal is transmitted through the transmitter. After receiving the signal, the USRP receiver receives and stores the signal through the signal acquisition program, and provides the required signal training set and signal test set for the CNN model of modulation signal recognition, wherein the signal training set includes a pre-training data set and a transfer learning training data set; Step 2: Pre-training of the convolutional neural network model: Pre-training the convolutional neural network model for modulated signal recognition using the generated pre-training data set; Step 3: Transfer learning training of convolutional neural network model: pre-train the convolutional neural network model for modulated signal recognition using the generated transfer learning training data set; Step 4: Deploy the trained transfer learning convolutional neural network model to the signal test set to verify the modulated signal recognition results.

2. The USRP modulation signal recognition method based on migration deep learning as claimed in claim 1, characterized in that: In step 1, the transmitting end RF parameters at least include RF antenna, center frequency, transmitting gain, and modulation mode, and the USRP receiving end RF parameters at least include RF antenna, center frequency, and receiving gain.

3. The USRP modulation signal recognition method based on migration deep learning as claimed in claim 1, characterized in that: In step 1, the USRP receiver starts to receive wireless information from the receiving antenna, passes through the RF receiving module, undergoes analog-to-digital conversion, and down-converts to receive the modulated signal.

4. The USRP modulation signal recognition method based on migration deep learning as claimed in claim 1, characterized in that: Pre-training data set in step 1: the classification signal set used in channel environment I is Ω4 = {BPSK, 8PSK, 4QAM, 16QAM}, the signal sequence length is L = 100, and the average signal-to-noise ratio of the sequence is uniformly distributed between [-10dB, 20dB], and the number of each type of signal is N I =100000, the pre-training data set is set to receive N signals of each type, with a total of 4N signal sequences. During training, 90%×4N are taken as the training set, and 10%×4N are used for cross-validation. The test set has 1000 signals of each type, and the SNR range is [-10dB, 20dB]. Tests are performed at each interval and the average is taken to obtain the modulation signal recognition rate.

5. The USRP modulation signal recognition method based on migration deep learning as claimed in claim 1, characterized in that: Transfer learning training data set in step 1: In channel environment II, the number of each type of signal in the transfer learning training set is set to N II ,In the test set, there are 1000 signals of each type, the SNR range is [-10dB, 20dB], and the test is performed at every 2dB interval. The average result is taken to obtain the modulation signal recognition rate.

6. The USRP modulation signal recognition method based on migration deep learning as claimed in claim 1, characterized in that: Step 3 specifically includes: the USRP receiver collects signal data sets under two different channel environments, uses a transfer learning convolutional neural network to identify modulated signals, and performs transfer training on the convolutional layer and the fully connected layer based on the convolutional neural network.

7. The USRP modulation signal recognition method based on migration deep learning as claimed in claim 6, characterized in that: The transfer learning convolutional neural network uses two convolutional layers with a convolution step of 1. The first layer of the two convolutional layers has 64 kernels, the size of the first convolution kernel is 2×4, the activation function is ReLU, the bias term is used and initialized to zero, the weight initialization method is Glorot uniform distribution initialization, the number of convolution kernels in the second convolutional layer is 16, the size of the convolution kernel is 1×4, and the other settings are the same as the first convolutional layer.

8. The USRP modulation signal recognition method based on migration deep learning as claimed in claim 6, characterized in that: Transfer learning convolutional neural network The fully connected neural network includes three fully connected layers. The first two layers are set to 64 and 16 neurons respectively. The activation function is ReLU. The weights are initialized with Glorot uniform distribution, and regularization terms and weight constraints are used. The output layer activation function uses Softmax.

Images