Communication signal automatic modulation recognition method and system based on transfer learning

Abstract

The application provides a communication signal automatic modulation recognition method and system based on transfer learning, comprising the following steps: S1, generating a pre-training sample set and constructing a communication signal recognition pre-training sample set; S2, dividing the pre-training sample set into a training set and a verification set, training a convolutional neural network with the training set and verifying the recognition accuracy with the verification set; S3, continuously collecting and saving real-time communication signals with a signal monitoring device, and collecting and saving the signals into frames according to the antenna frequency band; S4, replacing part of the pre-training sample set with the communication signals, loading the convolutional neural network as the initialization parameters of a complete model, and continuing to train the convolutional neural network; and S5, repeating step S4 until the pre-training sample set is all real-time collected communication signals, and obtaining a communication modulation recognition model. The application can monitor the communication signals in an experimental environment in real time, continuously enrich the training data set, and gradually complete the transfer learning of communication signal modulation recognition.

Description

Technical Field

[0001] This invention relates to the field of communication signal recognition technology, and more specifically, to an automatic modulation recognition method and system for communication signals based on transfer learning. Background Technology

[0002] Modulation identification is a technique that studies and identifies modulation methods for wireless communication signals. Communication signal modulation identification can automatically identify the modulation scheme of a modulated communication signal, thereby detecting various parameters of the modulated signal. It plays a significant role in both military and civilian communication signal processing, and is an indispensable technical means, particularly in fields such as spectrum monitoring and electronic countermeasures.

[0003] Traditional communication signal recognition techniques can be categorized into methods based on Bayesian decision theory and methods based on machine learning theory, based on classification strategies. Both methods achieve good recognition results under ideal conditions, but their recognition rates drop sharply in complex electromagnetic environments and low signal-to-noise ratio (SNR) environments. Due to the superior feature learning capabilities of deep learning, its application to communication signal modulation recognition has attracted widespread attention. Convolutional neural networks, in particular, have demonstrated good performance in complex electromagnetic environments and under various SNR conditions.

[0004] Deep learning-based modulation recognition often achieves higher recognition rates than traditional techniques by modulating trained parameters on training data. However, it often suffers from poor generalization ability, leading to a sharp drop in recognition rates in real-world applications. Therefore, research on communication signal modulation recognition algorithms based on transfer learning is of significant practical importance in addressing this issue.

[0005] Existing deep learning-based modulation recognition algorithms can only achieve a recognition rate of around 90% on a given dataset. When the trained model is deployed in a real-world production environment, the recognition rate often drops significantly. Directly collecting signals from the actual production environment for training is extremely time-consuming, and the quality of communication signals in the actual production environment cannot be guaranteed to be suitable as sample data for training neural networks, making it difficult to train a usable neural network. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an automatic modulation and identification method and system for communication signals based on transfer learning.

[0007] According to the present invention, an automatic modulation and identification method and system for communication signals based on transfer learning is provided, the scheme of which is as follows:

[0008] Firstly, a method for automatic modulation recognition of communication signals based on transfer learning is provided, the method comprising:

[0009] Step S1: Generate a pre-trained sample set and construct a communication signal recognition system based on the pre-trained sample set;

[0010] Step S2: Divide the pre-trained sample set into a training set and a validation set, train the convolutional neural network using the training set, and verify the recognition accuracy using the validation set;

[0011] Step S3: Continuously collect and save real-time communication signals using signal monitoring equipment. The monitoring equipment collects and saves the signals in frames according to the antenna frequency band.

[0012] Step S4: Replace part of the pre-training sample set with the real-time acquired communication signals, merge them into a complete dataset, load the convolutional neural network as the initialization parameters of the complete model, and continue training the convolutional neural network.

[0013] Step S5: Repeat step S4 until the pre-training sample set consists entirely of real-time acquired communication signals. The resulting convolutional neural network is the communication modulation recognition model after transfer learning.

[0014] Preferably, the generation of the pre-training sample set includes: modulating the input signal, transmitting the modulated IQ signal through a damaged channel to obtain a damaged IQ signal, and synthesizing the pre-training sample set from the damaged IQ signal.

[0015] Preferably, step S2 specifically includes: dividing the pre-trained sample set into a training set, a validation set, and a test set in a ratio of 7:2:1; wherein the training set and the validation set are used to train the convolutional neural network and to validate the recognition rate in each round, respectively, and the test sample set is used to test the recognition accuracy of the convolutional neural network after training is completed.

[0016] Preferably, the real-time acquisition of communication signals uses a software radio receiver. The cooperating transmitter continuously transmits analog radio frequency signals with known modulation types. The radio frequency tuner converts the analog radio frequency signals into analog intermediate frequency signals. Then, the A / D converter digitizes the analog intermediate frequency signals, thereby converting them into digital sampling points, which are then sent to the digital down-conversion module.

[0017] Preferably, the digital downconversion module is a single FPGA core, including: a digital mixer, a digital local oscillator, and an FIR low-pass filter;

[0018] The digital mixer and digital local oscillator down-convert the analog intermediate frequency signal to a baseband I / Q signal, which is then passed through an FIR low-pass filter to obtain the final acquired communication signal.

[0019] Secondly, an automatic modulation and identification system for communication signals based on transfer learning is provided, the system comprising:

[0020] Module M1: Generates a pre-trained sample set and constructs a communication signal recognition system for the pre-trained sample set;

[0021] Module M2: Divide the pre-trained sample set into a training set and a validation set, train a convolutional neural network using the training set, and validate the recognition accuracy using the validation set;

[0022] Module M3: Utilizes signal monitoring equipment to continuously collect and save real-time communication signals. The monitoring equipment collects and saves the signals in frames according to the antenna frequency band.

[0023] Module M4: Replaces part of the pre-training sample set with real-time acquired communication signals, merges them into a complete dataset, loads the convolutional neural network as the initialization parameters of the complete model, and continues to train the convolutional neural network.

[0024] Module M5: Repeat module M4 until the pre-training sample set consists entirely of real-time acquired communication signals. The resulting convolutional neural network is the communication modulation recognition model after transfer learning.

[0025] Preferably, the generation of the pre-training sample set includes: modulating the input signal, transmitting the modulated IQ signal through a damaged channel to obtain a damaged IQ signal, and synthesizing the pre-training sample set from the damaged IQ signal.

[0026] Preferably, module M2 specifically includes: dividing the pre-trained sample set into a training set, a validation set, and a test set in a ratio of 7:2:1; wherein the training set and the validation set are used to train the convolutional neural network and to validate the recognition rate of each round, respectively, and the test sample set is used to test the recognition accuracy of the convolutional neural network after the training is completed.

[0027] Preferably, the real-time acquisition of communication signals uses a software radio receiver. The cooperating transmitter continuously transmits analog radio frequency signals with known modulation types. The radio frequency tuner converts the analog radio frequency signals into analog intermediate frequency signals. Then, the A / D converter digitizes the analog intermediate frequency signals, thereby converting them into digital sampling points, which are then sent to the digital down-conversion module.

[0028] Preferably, the digital downconversion module is a single FPGA core, including: a digital mixer, a digital local oscillator, and an FIR low-pass filter;

[0029] The digital mixer and digital local oscillator down-convert the analog intermediate frequency signal to a baseband I / Q signal, which is then passed through an FIR low-pass filter to obtain the final acquired communication signal.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] 1. The communication signal pre-training dataset constructed by the present invention in step S1 takes into account the problems such as multipath fading, frequency offset, and signal-to-noise ratio that may be encountered in the actual signal transmission process. It can realistically simulate the process of actual communication signal generation and transmission, and the signal fidelity is high.

[0032] 2. In step S2 of this invention, each sample data consists of 1024 IQ signals. The number of samples is sufficient and the size is fixed, which is enough to describe the content transmission of a communication signal.

[0033] 3. In step S3, since the number of modulation types of communication signals is limited, a particularly complex neural network structure is not required. The network depth used is 28 layers, including: 1 I / Q signal input layer, 6 convolutional layers, 6 batch normalization layers, 6 ReLU activation layers, 5 max pooling layers, 1 average pooling layer, 1 fully connected layer, 1 regression layer, and 1 modulation category output layer. Using this simple network model, a 91% recognition accuracy can be achieved on the generated sample set.

[0034] 4. This invention utilizes a cooperative signal transmission source and self-developed signal monitoring equipment to monitor communication signals in the experimental environment in real time;

[0035] 5. This invention utilizes communication signals continuously collected by a pre-trained model and monitoring equipment to continuously enrich the training dataset;

[0036] 6. This invention gradually fuses the collected communication signals with the generated communication signals, which can accelerate the convergence speed of the new model and gradually complete the transfer learning of communication signal modulation recognition. Attached Figure Description

[0037] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0038] Figure 1 For signal generation path;

[0039] Figure 2 This is a time-domain plot of the IQ signal;

[0040] Figure 3 The time-frequency domain plot of the IQ signal;

[0041] Figure 4 It is a convolutional neural network model;

[0042] Figure 5 The classification results of the pre-trained model;

[0043] Figure 6 It is a real-time communication signal acquisition system. Detailed Implementation

[0044] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0045] This invention provides an automatic modulation and identification method for communication signals based on transfer learning, the method specifically including:

[0046] Step S1: Generate a pre-trained sample set and construct a communication signal recognition system based on the pre-trained sample set.

[0047] Specifically, the modulation types of the pre-training sample set include: BPSK, QPSK, 8PSK, 16QAM, 64QAM, PAM4, GFSK, CPFSK, BFM, DSB-AM, and SSB-AM. Channel parameters include additive white Gaussian noise, Ricean multipath fading, and clock skew (causing center frequency and sampling time shifts). The input signal is modulated, and the modulated output IQ signal is then transmitted through a damaged channel to obtain a damaged IQ signal. The damaged IQ signal is then used to synthesize the pre-training sample set. The pre-training sample set generation process is detailed in the appendix. Figure 1 As shown.

[0048] The pre-training sample set contains 1024 IQ samples per frame. The time-series and frequency-domain waveforms of the sample signals are shown in the appendix. Figure 2 and Figure 3 As shown.

[0049] In this implementation scheme, the pre-training sample set includes 11 types of modulated communication signals, and each type of modulated communication signal contains 10,000 frames of IQ signals.

[0050] Channel Model: Signal-to-Noise Ratio: Each type of modulated signal contains 10 different signal-to-noise ratios (0dB, 2dB, 4dB, 6dB, 8dB, 10dB, 12dB, 14dB, 16dB, 18dB).

[0051] Rice multipath fading: Assume a delay distribution of [0, 1.8, 3.4] samples, with a corresponding average path gain of [0, -2, -10] dB. The K-factor is 4, and the maximum Doppler shift is 4 Hz, equivalent to a walking speed of 900 MHz.

[0052] Frequency offset: Based on the clock offset factor C and the center frequency Fs, a frequency offset is applied to each frame, measured in parts per million (ppm). In this example, the maximum clock offset is 5 ppm, meaning that assuming a center frequency of 100 MHz, the maximum frequency offset is 500 Hz.

[0053] Step S2: Divide the pre-training sample set into a training set and a validation set. Train the convolutional neural network using the training set and validate the recognition accuracy using the validation set. See the appendix for the neural network model. Figure 4 As shown, the accuracy of the pre-trained model is listed in the appendix. Figure 5 As shown.

[0054] The pre-trained sample set is divided into a training set, a validation set, and a test set in a ratio of 7:2:1. The training set and the validation set are used to train the convolutional neural network and to validate the recognition rate in each round, respectively. The test sample set is used to test the recognition accuracy of the convolutional neural network after training is completed.

[0055] The network has a depth of 28 layers, including: one I / Q signal input layer, six convolutional layers, six batch normalization layers, six ReLU activation layers, five max pooling layers, one average pooling layer, one fully connected layer, one regression layer, and one modulation category output layer. The input layer contains framed I / Q modulated signals. The convolutional block sizes are 16x1, 24x1, 32x1, 48x1, 64x1, and 96x1. The batch normalization and ReLU layers use the most commonly used batch normalization and ReLU rectified functions in neural networks. The first five pooling layers are 1x2 max pooling, and the sixth pooling layer is a 1x32 average pooling layer.

[0056] In this example, the classifier consists of a fully connected layer and a softmax layer. The 1024 IQ signals input per frame are processed through convolution, pooling, and other operations to obtain a 192-dimensional feature vector. This feature vector is then processed through a fully connected layer and a regression layer to obtain the final predicted class.

[0057] As attached Figure 5 As shown in the confusion matrix, the network confuses 16-QAM and 64-QAM frames. This problem is expected because each frame carries only 128 symbols, and 16-QAM is a subset of 64-QAM. The network also confuses QPSK and 8-PSK frames because the constellation diagrams of these modulation types look similar after channel fading and frequency offset induce phase rotation.

[0058] Step S3: Continuously collect and save real-time communication signals using signal monitoring equipment. The monitoring equipment collects and saves signals in frames according to the antenna frequency band, as shown in the appendix. Figure 6 As shown.

[0059] Real-time acquisition of communication signals utilizes a software-defined radio receiver. A cooperative transmitter continuously transmits analog radio frequency (RF) signals of a known modulation type. An RF tuner converts the analog RF signals into analog intermediate frequency (IF) signals. An A / D converter then digitizes these IF signals, converting them into digital sampling points, which are then sent to an auxiliary circuit. Figure 6 The digital downconverter module is indicated by the dashed line.

[0060] The digital downconversion module is a separate FPGA core, mainly consisting of three parts: a digital mixer, a digital local oscillator, and an FIR low-pass filter. The digital mixer and digital local oscillator downconvert the analog intermediate frequency signal to a baseband I / Q signal, which is then passed through the FIR low-pass filter to obtain the final digital baseband signal, which is the final acquired communication modulation signal.

[0061] Step S4: Replace part of the pre-training sample set with the real-time acquired communication signals, merge them into a complete dataset, load the convolutional neural network as the initialization parameters of the complete model, and continue training the convolutional neural network.

[0062] Step S5: Repeat step S4 until the pre-training sample set consists entirely of real-time acquired communication signals. The resulting convolutional neural network is the communication modulation recognition model after transfer learning.

[0063] This invention provides an automatic modulation and recognition method and system for communication signals based on transfer learning. The communication signal pre-training dataset constructed in step S1 takes into account the problems encountered in the actual signal transmission process, such as multipath fading, frequency shift, and signal-to-noise ratio. It can realistically simulate the actual communication signal generation and transmission process with high signal fidelity.

[0064] Through step S2, each sample data consists of 1024 IQ signals. The number of samples is sufficient and the size is fixed, enough to describe the content transmission of a communication signal.

[0065] In step S3, since the number of modulation types in the communication signal is limited, a particularly complex neural network structure is not required. A network depth of 28 layers is used, including: one I / Q signal input layer, six convolutional layers, six batch normalization layers, six ReLU activation layers, five max pooling layers, one average pooling layer, one fully connected layer, one regression layer, and one modulation category output layer. Using this simple network model, a 91% recognition accuracy can be achieved on the generated sample set.

[0066] By utilizing collaborative signal transmission sources and self-developed signal monitoring equipment, communication signals in the experimental environment can be monitored in real time. The communication signals continuously collected using the pre-trained model and monitoring equipment can continuously enrich the training dataset. Gradually fusing the collected communication signals with the generated communication signals can accelerate the convergence speed of the new model and progressively complete the transfer learning of communication signal modulation recognition.

[0067] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0068] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for automatic modulation recognition of communication signals based on transfer learning, characterized in that, Comprising: Step S1: generating a pre-training sample set, and constructing a communication signal recognition pre-training sample set; Step S2: dividing the pre-training sample set into a training set and a validation set, training a convolutional neural network with the training set, and verifying the recognition accuracy with the validation set; Step S3: continuously collecting and saving real-time communication signals using a signal monitoring device, and collecting and saving frames according to the antenna frequency band; Step S4: replacing part of the pre-training sample set with real-time collected communication signals, loading the convolutional neural network as the initialization parameters of the complete model, and continuing to train the convolutional neural network; Step S5: repeating step S4 until the pre-training sample set is completely replaced by real-time collected communication signals, and the final convolutional neural network is the communication modulation recognition model after transfer learning; The real-time collected communication signal uses a software radio receiver, a cooperative transmitting source continuously transmits analog radio frequency signals of known modulation types, a radio frequency tuner converts the analog radio frequency signals into analog intermediate frequency signals, and then an A / D converter digitizes the analog intermediate frequency signals, thereby converting them into digital sampling points, and the sampling points are sent to a digital down conversion module; The digital down conversion module is a separate FPGA core, including a digital mixer, a digital local oscillator, and a FIR low-pass filter; The digital mixer and the digital local oscillator down-convert the analog intermediate frequency signals to baseband I / Q signals, and then pass through the FIR low-pass filter to obtain the final collected communication signal.

2. The method of claim 1, wherein the method is based on transfer learning. The generation of the pre-training sample set includes: input signals are modulated, and then the modulated IQ signals are transmitted through a damaged channel to obtain damaged IQ signals, and the damaged IQ signals are synthesized to obtain the pre-training sample set.

3. A communication signal automatic modulation recognition system based on transfer learning, characterized in that, Comprising: Module M1: generating a pre-training sample set, and constructing a communication signal recognition pre-training sample set; Module M2: dividing the pre-training sample set into a training set and a validation set, training a convolutional neural network with the training set, and verifying the recognition accuracy with the validation set; Module M3: continuously collecting and saving real-time communication signals using a signal monitoring device, and collecting and saving frames according to the antenna frequency band; Module M4: replacing part of the pre-training sample set with real-time collected communication signals, loading the convolutional neural network as the initialization parameters of the complete model, and continuing to train the convolutional neural network; Module M5: repeating module M4 until the pre-training sample set is completely replaced by real-time collected communication signals, and the final convolutional neural network is the communication modulation recognition model after transfer learning; The real-time collected communication signal uses a software radio receiver, a cooperative transmitting source continuously transmits analog radio frequency signals of known modulation types, a radio frequency tuner converts the analog radio frequency signals into analog intermediate frequency signals, and then an A / D converter digitizes the analog intermediate frequency signals, thereby converting them into digital sampling points, and the sampling points are sent to a digital down conversion module; The digital down conversion module is a separate FPGA core, including a digital mixer, a digital local oscillator, and a FIR low-pass filter; The digital mixer and the digital local oscillator down-convert the analog intermediate frequency signal to a baseband I / Q signal, and the final collected communication signal is obtained through a FIR low-pass filter.

4. The system according to claim 3, wherein the system is configured to: The generating of the pre-training sample set comprises: inputting a signal for modulation, subjecting the modulated output IQ signal to a damaged channel transmission to obtain a damaged IQ signal, and synthesizing the pre-training sample set from the damaged IQ signal.

Images