Radio frequency signal recognition method and system based on multi-scale attention feature fusion
Patent Information
- Application Number
- CN202311590478.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-11-27
- Publication Date
- 2026-10-16
- Estimated Expiration
- 2043-11-27
AI Technical Summary
但是，基于单网络模型的射频信号识别方案，一般只能提取到时间序列的单一特征，因此该类方案存在特征丢失的缺陷
[0031] The radio frequency signal recognition method and system based on multi-scale attention feature fusion provided by this invention obtains the local dependencies of radio frequency signals through an LSTM network, obtains the global features of radio frequency signals through a TCN network, and finally achieves feature fusion through MS-CAM. Therefore, this invention can not only realize the recognition of radio frequency signals, but also has higher reliability and better accuracy.
Smart Images
Figure CN117556230B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of communication technology, specifically relating to a radio frequency signal recognition method and system based on multi-scale attention feature fusion. Background Technology
[0002] With the development of economy and technology and the improvement of people's living standards, wireless devices have been widely used in people's production and life, bringing endless convenience. Therefore, ensuring the security of wireless devices and wireless communication has become a key concern.
[0003] Wireless devices are inherently open, making wireless signals a frequent target for large-scale malicious attacks. Physical layer security is the most fundamental aspect of wireless transmission security. Radio frequency fingerprints are inherent characteristics of the transmitting device hardware; due to variations in circuit component manufacturing, this characteristic is unique and can distinguish between different transmission devices.
[0004] Research has shown that using deep learning for radio frequency (RF) signal identification is feasible. Currently, commonly used deep learning-based RF signal identification schemes are generally based on single-network models. However, single-network model-based RF signal identification schemes can typically only extract single features from the time series, thus suffering from feature loss. While dual-network model-based RF signal identification schemes exist, they do not consider the causal properties of the time series during global feature extraction; furthermore, as the input sequence length increases, very deep convolutional neural networks are required, making it difficult to effectively extract global features; additionally, these networks do not consider the importance of features at different scales, and fusing multi-scale features through simple linear operations easily leads to the loss of key features. Summary of the Invention
[0005] One of the objectives of this invention is to provide a highly reliable and accurate radio frequency signal recognition method based on multi-scale attention feature fusion.
[0006] The second objective of this invention is to provide a system for implementing the radio frequency signal recognition method based on multi-scale attention feature fusion.
[0007] The radio frequency signal recognition method based on multi-scale attention feature fusion provided by this invention includes the following steps:
[0008] S1. Acquire existing wireless communication radio frequency signal data;
[0009] S2. Process the signal data obtained in step S1 to obtain a training set;
[0010] S3. Construct an initial model for radio frequency signal recognition based on LSTM network, TCN network and MS-CAM network;
[0011] S4. Using the training set obtained in step S2, train the initial radio frequency signal recognition model constructed in step S3 to obtain the trained radio frequency signal recognition model;
[0012] S5. Using the radio frequency signal recognition model obtained in step S4, perform radio frequency signal recognition based on multi-scale attention feature fusion.
[0013] Step S2, which involves processing the signal data acquired in step S1, specifically includes the following steps:
[0014] The signal data obtained in step S1 is processed; the data processing includes signal filtering, signal segmentation, signal amplitude normalization, signal start point detection, and signal orthogonal IQ combining.
[0015] Step S3, which involves constructing an initial model for radio frequency signal identification based on LSTM, TCN, and MS-CAM networks, specifically includes the following steps:
[0016] The initial model for radio frequency signal recognition includes an LSTM network, a TCN network, and an MS-CAM network;
[0017] Among them, the LSTM network is used to learn the local dependencies of radio frequency signals using gating mechanisms and memory cells; the TCN network is used to extract global features of time series using causal convolution and dilated convolution; and the MS-CAM network is used to construct the interdependencies of multi-scale features of radio frequency signals, thereby fusing global features and local dependencies of time series.
[0018] Step S3, which involves constructing an initial model for radio frequency signal identification based on LSTM, TCN, and MS-CAM networks, specifically includes the following steps:
[0019] The model includes a TCN branch, an LSTM branch, a feature fusion module, and a fully connected layer module;
[0020] The TCN branch uses causal convolution and dilated convolution to extract global features of radio frequency signals; the LSTM branch uses gating mechanisms and memory cells to learn the local dependencies of radio frequency signals; the results of the TCN and LSTM branches are fused through a feature fusion module, and finally the radio frequency signals are classified and identified through a fully connected layer module.
[0021] The feature fusion module learns the importance of multi-scale features through global and local channels, realizes effective fusion of multi-scale features, and completes accurate measurement of radio frequency signals.
[0022] The fully connected layer module acts as a classifier;
[0023] The input to the model is a preprocessed one-dimensional radio frequency time-domain signal;
[0024] The TCN branch consists of 12 residual submodules. Each residual submodule uses two identical one-dimensional convolutions with a kernel size of 8, a stride of 1, and an inflation factor of 2. The padding size of adjacent input neurons in the i-th layer is 2. i -1 The i-th layer's historical perception range for the input sequence is r. i For r i =r i-1 +(8-1)×2 i-1 One-dimensional convolution maintains the same size for the input and output data. Finally, the output sequence is processed by feature mapping through a fully connected layer to obtain a global feature sequence of size N×1×256; N represents the number of input data samples.
[0025] The hidden layer state dimension of the LSTM branch is 256, which represents the number of nodes used to remember and store past states; after prediction and output for each input node, a local feature sequence of size N×1×256 is obtained.
[0026] The feature fusion module takes global and local feature sequences as inputs and performs weighted fusion of the input features based on an attention mechanism to obtain a more accurate fused feature sequence of size N×1×256.
[0027] During network training, the Adam optimizer and the following loss function are used to optimize the model parameters;
[0028]
[0029] In the formula, L is the loss function value; N is the sample size of the input data; M is the number of categories; y ic y is a 0-1 variable, where y is the actual class of sample i when i is c. ic =1, otherwise y ic =0; p ic Let be the predicted probability that sample i belongs to category c.
[0030] This invention also provides a system for implementing the radio frequency signal recognition method based on multi-scale attention feature fusion, comprising a data acquisition module, a data processing module, a model building module, a model training module, and a signal recognition module; the data acquisition module, data processing module, model building module, model training module, and signal recognition module are connected in series; the data acquisition module is used to acquire existing wireless communication radio frequency signal data and upload the data to the data processing module; the data processing module is used to process the signal data according to the received data to obtain a training set and upload the data to the model building module; the model building module is used to construct an initial radio frequency signal recognition model based on LSTM network, TCN network, and MS-CAM network according to the received data and upload the data to the model training module; the model training module is used to train the constructed initial radio frequency signal recognition model using the training set according to the received data to obtain a trained radio frequency signal recognition model and upload the data to the signal recognition module; the signal recognition module is used to perform radio frequency signal recognition based on multi-scale attention feature fusion using the obtained radio frequency signal recognition model according to the received data.
[0031] The radio frequency signal recognition method and system based on multi-scale attention feature fusion provided by this invention obtains the local dependencies of radio frequency signals through an LSTM network, obtains the global features of radio frequency signals through a TCN network, and finally achieves feature fusion through MS-CAM. Therefore, this invention can not only realize the recognition of radio frequency signals, but also has higher reliability and better accuracy. Attached Figure Description
[0032] Figure 1 This is a schematic diagram of the method flow of the present invention.
[0033] Figure 2 This is a time-domain waveform diagram of the 2.4G module in an embodiment of the method of the present invention.
[0034] Figure 3 This is a schematic diagram comparing the recognition rate under different signal-to-noise ratios in an embodiment of the method of the present invention.
[0035] Figure 4 This is a schematic diagram illustrating the change in the data recognition rate of the 2.4G module in an embodiment of the method of the present invention.
[0036] Figure 5 This is a schematic diagram of the functional modules of the system of the present invention. Detailed Implementation
[0037] like Figure 1 The diagram shown is a flowchart of the method of the present invention: The radio frequency signal recognition method based on multi-scale attention feature fusion disclosed in this invention includes the following steps:
[0038] S1. Acquire existing wireless communication radio frequency signal data;
[0039] S2. Process the signal data obtained in step S1 to obtain the training set; specifically, this includes the following steps:
[0040] The signal data acquired in step S1 is processed; the data processing includes signal filtering, signal segmentation, signal amplitude normalization, signal start point detection, and signal orthogonal IQ combining.
[0041] S3. Construct an initial model for radio frequency signal identification based on LSTM, TCN, and MS-CAM networks; specifically including the following steps:
[0042] The initial model for radio frequency signal recognition includes an LSTM network, a TCN network, and an MS-CAM network;
[0043] Among them, the LSTM network is used to learn the local dependencies of radio frequency signals using gating mechanisms and memory cells; the TCN network is used to extract global features of time series using causal convolution and dilated convolution; and the MS-CAM network is used to construct the interdependencies of multi-scale features of radio frequency signals, thereby fusing global features and local dependencies of time series.
[0044] In practical implementation, the initial model for radio frequency signal identification includes the following:
[0045] The model includes a TCN branch, an LSTM branch, a feature fusion module, and a fully connected layer module;
[0046] The TCN branch uses causal convolution and dilated convolution to extract global features of radio frequency signals; the LSTM branch uses gating mechanisms and memory cells to learn the local dependencies of radio frequency signals; the results of the TCN and LSTM branches are fused through a feature fusion module, and finally the radio frequency signals are classified and identified through a fully connected layer module.
[0047] The feature fusion module learns the importance of multi-scale features through global and local channels, realizes effective fusion of multi-scale features, and completes accurate measurement of radio frequency signals.
[0048] The fully connected layer module acts as a classifier;
[0049] The input to the model is a preprocessed one-dimensional radio frequency time-domain signal;
[0050] The TCN branch consists of 12 residual submodules. Each residual submodule uses two identical one-dimensional convolutions with a kernel size of 8, a stride of 1, and an inflation factor of 2. The padding size of adjacent input neurons in the i-th layer is 2. i -1The i-th layer's historical perception range for the input sequence is r. i For r i =r i-1 +(8-1)×2 i-1 One-dimensional convolution maintains the same size for the input and output data. Finally, the output sequence is processed by feature mapping through a fully connected layer to obtain a global feature sequence of size N×1×256; N represents the number of input data samples.
[0051] The hidden layer state dimension of the LSTM branch is 256, which represents the number of nodes used to remember and store past states; after prediction and output for each input node, a local feature sequence of size N×1×256 is obtained.
[0052] The feature fusion module takes global and local feature sequences as inputs and performs weighted fusion of the input features based on an attention mechanism to obtain a more accurate fused feature sequence of size N×1×256.
[0053] To improve network training stability, accelerate the gradient descent process, and alleviate gradient oscillations, thereby improving network classification performance, the Adam optimizer is used during network training, and the following loss function is employed to optimize model parameters.
[0054]
[0055] In the formula, L is the loss function value; N is the sample size of the input data; M is the number of categories; y ic y is a 0-1 variable, where y is the actual class of sample i when i is c. ic =1, otherwise y ic =0; p ic Let i be the predicted probability that sample i belongs to category c;
[0056] S4. Using the training set obtained in step S2, train the initial radio frequency signal recognition model constructed in step S3 to obtain the trained radio frequency signal recognition model;
[0057] S5. Using the radio frequency signal recognition model obtained in step S4, perform radio frequency signal recognition based on multi-scale attention feature fusion.
[0058] The method of the present invention will be further described below with reference to an embodiment:
[0059] To verify the performance of the proposed model, signal acquisition was performed on 18 2.4G module communication devices in an open environment.
[0060] The signals used in the experiment were acquired from wireless devices in a real communication environment. The acquisition system consisted of a radio peripheral (USRP B210) and a GNU Radio. The acquisition module included 18 different types of nRF24L01 IoT transmitters, operating in the ISM 2.4GHz band. Each group contained 3000 samples at a sampling frequency of 2.4GHz. The results of signal simulation using Python are shown below. Figure 2 Table 1 shows the parameters of the acquisition equipment used, and Table 2 shows the data set information for the 2.4G module.
[0061] Table 1. Meaning of Data Acquisition Equipment Parameters
[0062] Center frequency 2.4GHz signal length 4096 (≈0.1ms) sampling frequency 40MHz bandwidth 10MHz Transmission gain 30dB Sampling distance 1m
[0063] Table 2. Dataset Information Illustration
[0064] 2.4G module 18 3000 500 63000
[0065] Signal classification under additive white Gaussian noise (ADG) is a typical problem in practical communication. To verify the classification performance of the model under different noise levels, additive white Gaussian noise was added to the test data of the 2.4G module. The recognition results under different SNRs are as follows: Figure 3 As shown, when the SNR is greater than -5dB, the recognition rate of the method of the present invention is greater than 95%. Compared with other recognition algorithms, the method of the present invention uses a dual-channel attention fusion mechanism to reduce the influence of noise on the fusion features, and has better noise resistance and robustness.
[0066] To verify the model's applicability, a 2.4G module dataset collected by USRP was used for comparison with different classification models, including single-model networks such as CNN, TCN, RNN, LSTM, and GRU, dual-model networks including LSTM-FCN, and the model of this invention. The recognition accuracy was verified as follows: Figure 4 As shown, the average recognition accuracy of the 2.4G module is 97.56%. Compared with other radio frequency signal classification methods, the method of this invention has higher recognition accuracy and stability. This indicates that the method of this invention considers features at different scales, allowing the model to more effectively extract and fuse complex multi-scale feature information of radio frequency signals, thereby improving recognition performance.
[0067] like Figure 5The diagram shows the functional modules of the system of this invention: The system for implementing the radio frequency signal recognition method based on multi-scale attention feature fusion disclosed in this invention includes a data acquisition module, a data processing module, a model building module, a model training module, and a signal recognition module; the data acquisition module, data processing module, model building module, model training module, and signal recognition module are connected in series; the data acquisition module is used to acquire existing wireless communication radio frequency signal data and upload the data to the data processing module; the data processing module is used to process the signal data according to the received data to obtain a training set and upload the data to the model building module; the model building module is used to construct an initial radio frequency signal recognition model based on LSTM network, TCN network, and MS-CAM network according to the received data and upload the data to the model training module; the model training module is used to train the constructed initial radio frequency signal recognition model using the training set according to the received data to obtain a trained radio frequency signal recognition model and upload the data to the signal recognition module; the signal recognition module is used to perform radio frequency signal recognition based on multi-scale attention feature fusion using the obtained radio frequency signal recognition model according to the received data.
Claims
1. A radio frequency signal recognition method based on multi-scale attention feature fusion, comprising the following steps: S1. Acquire existing wireless communication radio frequency signal data; S2. Process the signal data obtained in step S1 to obtain the training set; S3. Construct an initial model for radio frequency signal identification based on LSTM, TCN, and MS-CAM networks; including the following steps: The initial model for radio frequency signal recognition includes an LSTM network, a TCN network, and an MS-CAM network. in, LSTM networks are used to learn local dependencies of radio frequency signals using gating mechanisms and memory cells; TCN networks are used to extract global features of time series signals using causal convolution and dilated convolution. MS-CAM networks are used to construct the interdependencies of multi-scale features of radio frequency signals, thereby fusing global and local dependencies of time series signals. In practice, the model includes a TCN branch, an LSTM branch, a feature fusion module, and a fully connected layer module. The TCN branch uses causal convolution and dilated convolution to extract global features of radio frequency signals; the LSTM branch uses gating mechanisms and memory cells to learn the local dependencies of radio frequency signals; the results of the TCN and LSTM branches are fused through a feature fusion module, and finally the radio frequency signals are classified and identified through a fully connected layer module. The feature fusion module learns the importance of multi-scale features through global and local channels, realizes effective fusion of multi-scale features, and completes accurate measurement of radio frequency signals. The fully connected layer module acts as a classifier; The input to the model is a preprocessed one-dimensional radio frequency time-domain signal; The TCN branch consists of 12 residual submodules. Each residual submodule uses two identical one-dimensional convolutions with a kernel size of 8, a stride of 1, and a dilation factor of 2. The padding size of adjacent input neurons in the i-th layer is... The i-th layer's historical perception range of the input sequence for One-dimensional convolution maintains the same input and output data size. Finally, the output sequence undergoes feature mapping processing through a fully connected layer to obtain a sequence of size [size missing]. The global feature sequence; Indicates the sample size of the input data; The hidden layer state dimension of the LSTM branch is 256, representing the number of nodes used to remember and store past states; after prediction and output for each input node, a value of size is obtained. Local feature sequences; The feature fusion module takes global and local feature sequences as input. It uses an attention mechanism to weightedly fuse the input features, resulting in more accurate feature information with a size of [value missing]. fusion feature sequence; The Adam optimizer is used during network training, and the following loss function is used to optimize the model parameters; In the formula, L is the value of the loss function; The sample size of the input data; Number of categories; For a 0-1 variable, when the actual class of sample i is c. ,otherwise ; Let i be the predicted probability that sample i belongs to category c; S4. Using the training set obtained in step S2, train the initial radio frequency signal recognition model constructed in step S3 to obtain the trained radio frequency signal recognition model; S5. Using the radio frequency signal recognition model obtained in step S4, perform radio frequency signal recognition based on multi-scale attention feature fusion.
2. The radio frequency signal recognition method based on multi-scale attention feature fusion according to claim 1, characterized in that... Step S2, which involves processing the signal data acquired in step S1, specifically includes the following steps: The signal data obtained in step S1 is processed; the data processing includes signal filtering, signal segmentation, signal amplitude normalization, signal start point detection, and signal orthogonal IQ combining.
3. A system for implementing the radio frequency signal recognition method based on multi-scale attention feature fusion as described in claim 1 or 2, characterized in that... It includes a data acquisition module, a data processing module, a model building module, a model training module, and a signal recognition module; these modules are connected in series. The data acquisition module acquires existing wireless communication radio frequency signal data and uploads it to the data processing module. The data processing module processes the received signal data to obtain a training set and uploads it to the model building module. The model building module constructs an initial radio frequency signal recognition model based on LSTM, TCN, and MS-CAM networks using the received data and uploads it to the model training module. The model training module is used to train the initial radio frequency signal recognition model based on the received data and the training set, so as to obtain the trained radio frequency signal recognition model and upload the data to the signal recognition module. The signal recognition module is used to perform radio frequency signal recognition based on multi-scale attention feature fusion, using the obtained radio frequency signal recognition model based on the received data.