A specific protocol signal recognition method based on deep residual network

Abstract

A method for identifying specific protocol signals based on a deep residual network, S1: extending time domain template matching and spectrum template matching to time-frequency template matching, and converting the time-frequency energy of the signal into a grayscale image; S2: constructing a deep residual network and establishing a signal recognition model; S3: using the grayscale image to train the deep residual network model; S4: detecting and identifying the specific communication protocol signal trained during the transmission process through the trained deep residual network model; the present invention provides an extension of time domain template matching and spectrum template matching to time-frequency template matching, and uses a neural network to learn and mine the best time-frequency spectrum template from time-frequency data, designs a shortwave specific communication protocol recognition model with adaptive learning ability, and proposes an automatic feature extraction method based on a deep residual network, which solves the problem of identifying specific protocols under low noise, multipath delay, Doppler frequency deviation, strong interference and strong aliasing conditions.

Description

Technical Field

[0001] The present invention relates to the field of communication technologies, and in particular to a specific protocol signal recognition method based on a deep residual network. Background Art

[0002] Shortwave communication protocol recognition plays an important role in the field of shortwave communication countermeasures and cognitive radio, and plays a vital role in communication interference and target identification. However, how to achieve accurate recognition of the protocol has always been a difficult and hot topic in the field of non-cooperative receivers.

[0003] The existing communication protocol automatic recognition technology mainly covers multiple technical fields such as software radio, modulation recognition and coding recognition. These methods can be roughly divided into two categories: one is protocol recognition based on signal modulation features, and the other is protocol recognition based on bit streams. In the recognition method based on modulation feature extraction, the feature template matching algorithm performs well when there is sufficient prior information, but it is very sensitive to the signal-to-noise ratio and is prone to false alarms under low signal-to-noise ratios. Methods such as support vector machines have been applied in the field of signal recognition and can show good recognition results, but they are sensitive to error boundaries, are not suitable for big data experiments, and lack the necessary probability information.

[0004] In recent years, the development of deep learning technology has provided new ideas for communication protocol recognition. Deep neural networks can automatically extract useful features from large amounts of data and use these features for classification and recognition. In particular, deep residual neural networks (ResNet) have demonstrated excellent performance in many fields such as image processing. However, the application of deep learning technology to communication protocol recognition still faces some technical challenges. How to extract effective features from complex communication signals and how to design a deep neural network model suitable for communication protocol recognition are urgent issues to be solved. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a specific protocol signal recognition method based on a deep residual network. It provides an extension of time domain template matching and spectrum template matching to time-spectrum template matching, and uses a neural network to learn and mine the best time-spectrum graph template from time-frequency data. A shortwave specific communication protocol recognition model with adaptive learning ability is designed, and an automatic feature extraction method based on a deep residual network is proposed to solve the problem of identifying specific protocols under low noise, multipath delay, Doppler frequency deviation, strong interference and strong aliasing conditions.

[0006] In order to achieve the above-mentioned object, the present invention adopts the following technical scheme: a specific protocol signal recognition method based on a deep residual network, S1: extending the time domain template matching and spectrum template matching into time-frequency template matching, and converting the time-frequency energy of the signal into a grayscale image;

[0007] S2: Build a deep residual network and establish a signal recognition model;

[0008] S3: Use grayscale images to train the deep residual network model;

[0009] S4: The trained deep residual network model is used to detect and identify the specific communication protocol signals trained during the transmission process.

[0010] Further, the specific steps of s1 are: using short-time Fourier transform as a time-frequency analysis method, in the process of processing, using a centrally symmetrical sliding window to intercept the observed signal, performing Fourier transform on the signal in the sliding window, and finally obtaining a time-frequency spectrum composed of each segment of the signal;

[0011] For the signal S(t), the short-time Fourier transform of the signal can be defined as:

[0012]

[0013] Where γ(·) represents the window function.

[0014] Furthermore, the steps of S2 are: s2.1. H(x) is marked as a potential mapping, and the stacked network is fitted with a residual mapping H(x)=F(x)-x;

[0015] S2.2. When the dimensions of the input feature map and the output feature map after the residual unit do not match, a convolution kernel of size 1x1 is used to increase the dimension of the input feature map.

[0016] Further, the step of S3 is as follows: s3.1 According to the internal structure of the network and the actual training requirements, the size of the time-frequency graph needs to be consistent during training. Combined with the actual shortwave signal search requirements, during the training process, the size of the time-frequency graph is uniformly set to 320x320;

[0017] S3.2 identifies three types of specific communication protocol signals and one type of noise; the three types of specific communication protocol signals are link11, clover2000 and 2G-ALE, and the noise type is Gaussian white noise;

[0018] S3.3 After each layer of data is convolved to obtain a feature map, it is not directly processed by the activation function, but the data is first batch normalized;

[0019] S3.4 solves the internal covariate shift that occurs during training through batch normalization, allowing the network to converge quickly.

[0020] Furthermore, the steps of s4 are as follows: s4.1 In view of the influence of the signal environment on the network recognition rate, the experimental signal adopts the simulation signal under Link11, CLOVER2000, 2G-ALE system, and the specific parameters of signal generation are realized according to the conditions specified by the protocol. The carrier frequency f i within a certain range Randomly generated, the frequency resolution of the signal time-frequency diagram is f s / N(N=2 m , m=6,7,8,9,10), the signal-to-noise ratio range is -10dB-10dB;

[0021] S4.2 examines the impact of wireless channels on algorithm performance, taking into account the existence of Doppler frequency deviation and multipath delay, and uses the Watterson channel model recommended in the F.1487 standard of ITU-R as a shortwave test channel to test the signal protocol recognition method under the shortwave channel;

[0022] When the specific frame structure information on the S4.3 test time-frequency diagram is partially blocked and affects the network recognition performance, a PSK signal without a frame structure is used to perform time-frequency aliasing on the target signal. The time-frequency aliasing is defined as follows:

[0023] There are only two component signals in the aliased signal, and the signals completely overlap in time. The aliasing degree in the frequency domain is defined as:

[0024]

[0025] Where Δf1, Δf2 are the bandwidths of the signals, Δf 12 is the overlapping bandwidth, and the signal-to-noise ratio is the ratio of the two powers;

[0026] S4.4 Regarding the influence of network parameters on recognition performance, the network is tested on test samples after every 10 iterations of training. In each test, 600 signal samples are randomly selected from the sample data of 3 types of protocols, and the signal-to-noise ratio and aliasing conditions of the signals are random.

[0027] The beneficial effects of the present invention are as follows: compared with the prior art, the present invention has the following advantages: the present invention provides an extension of time domain template matching and spectrum template matching into time-spectrum template matching, and uses a neural network to learn and mine the best time-spectrum graph template from time-frequency data, designs a shortwave specific communication protocol recognition model with adaptive learning capability, and proposes a feature automatic extraction method based on a deep residual network, which solves the problem of identifying specific protocols under low noise, multipath delay, Doppler frequency deviation, strong interference and strong aliasing conditions; the parts not described in detail in the present invention are existing commonly used technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The present invention will be further described below in conjunction with the accompanying drawings:

[0029] Figure 1 The basic framework of the residual unit used in the present invention;

[0030] Figure 2 This is a basic schematic diagram of the deep residual network model used in the present invention. DETAILED DESCRIPTION

[0031] The present invention is further described in detail below in conjunction with embodiments and specific implementation modes:

[0032] Example 1

[0033] S1 expands the time domain template matching and spectrum template matching into time-frequency template matching, converting the time-frequency energy of the signal into a grayscale image;

[0034] In the process of forward design of shortwave communication signals, a series of specific rules are usually set for the signals to ensure the communication quality and facilitate subsequent processing. The establishment of such special rules causes the signals to present different time-frequency visual characteristics;

[0035] The present invention adopts short-time Fourier transform as a time-frequency analysis method. During the processing, a centrally symmetrical sliding window is used to intercept the observed signal, and the signal in the sliding window is subjected to Fourier transform, and finally a time-frequency spectrum composed of each segment of the signal is obtained;

[0036] For the signal S(t), the short-time Fourier transform of the signal can be defined as:

[0037]

[0038] Where γ(·) represents the window function.

[0039] S2 builds a deep residual network and establishes a signal recognition model;

[0040] The idea of ​​the residual network model is: instead of learning an ideal potential mapping directly through each stacked convolutional network, the ideal residual mapping is learned through the network;

[0041] S2.1 As Figure 1 As shown in the figure, it is a basic framework of a residual unit; H(x) is marked as a potential mapping, and the stacked network is fitted with a residual mapping H(x) = F(x) - x. Compared with the original potential mapping, the residual mapping is easier to optimize, solving the problem of network degradation when the number of network layers is increased;

[0042] S2.2 As Figure 2 This is a basic schematic diagram of the deep residual network model designed for the present invention. The dotted line in the figure indicates that when the dimensions of the input feature map and the output feature map after the residual unit do not match, a convolution kernel of size 1x1 is used to increase the dimension of the input feature map.

[0043] S3 uses grayscale images to train the deep residual network model;

[0044] S3.1 Due to the internal structure of the network and the actual training requirements, the size of the time-frequency graph needs to be consistent during training. Combined with the actual shortwave signal search requirements, the size of the time-frequency graph is uniformly set to 320x320 during training;

[0045] S3.2 The present invention mainly identifies three types of specific communication protocol signals. In order to make the network more robust, an additional noise class is added when considering the interference of out-of-class signals;

[0046] S3.3 After each layer of data is convolved to obtain a feature map, it is not directly processed by the activation function, but the data is first batch normalized;

[0047] S3.4 Since the input of each layer will change with the change of the parameters of the previous layer during the training process, that is, the dynamic range of the input data of each layer is different, there will be dimensional effects, and the adjustment of network parameters will become difficult. This phenomenon is called internal covariate shift (Internal Covariate Shift). Therefore, it is necessary to ensure that the dimensions of the input vector of each layer of the network are at the same level. The solution to this problem can be achieved through batch normalization, and batch normalization can use a higher learning rate, which ultimately enables the network to converge quickly.

[0048] S4 detects and recognizes the specific communication protocol signals trained during the transmission process through the trained deep residual network model;

[0049] S4.1 In view of the influence of signal environment on network recognition rate, the experimental signal adopts Link11, CLOVER2000, and simulated signal under 2G-ALE system. The specific parameters of signal generation are realized according to the conditions specified by the protocol. The carrier frequency f i within a certain range Randomly generated, the frequency resolution of the signal time-frequency diagram is fs / N(N=2 m , m=6,7,8,9,10), the signal-to-noise ratio ranges from -10dB to 10dB;

[0050] S4.2 examines the impact of the wireless channel on the algorithm performance, taking into account the existence of Doppler frequency deviation and multipath delay, and uses the recommended Watterson channel model in the F.1487 standard of ITU-R as a shortwave test channel to test the signal protocol identification method of the present invention under the shortwave channel;

[0051] S4.3 tests the influence of partial obstruction of specific frame structure information on the time-frequency diagram on network recognition performance. The present invention uses a PSK signal without a frame structure to perform time-frequency aliasing on the target signal. Since there is no strict definition for the degree of time-frequency aliasing, the present invention defines time-frequency aliasing according to the following method:

[0052] There are only two component signals in the aliased signal, and the signals completely overlap in time. The aliasing degree in the frequency domain is defined as:

[0053]

[0054] Where Δf1, Δf2 are the bandwidths of the signals, Δf 12 is the overlapping bandwidth, and the signal-to-noise ratio is the ratio of the two powers;

[0055] S4.4 Regarding the influence of network parameters on recognition performance, the network is tested on test samples after every 10 iterations of training. In each test, 600 signal samples are randomly selected from the sample data of 3 types of protocols, and the signal-to-noise ratio and aliasing conditions of the signals are random.

[0056] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A specific protocol signal recognition method based on a deep residual network, characterized in that: S1: Extend the time domain template matching and spectrum template matching to time-frequency template matching, and convert the time-frequency energy of the signal into a grayscale image; S2: Build a deep residual network and establish a signal recognition model; S3: Use grayscale images to train the deep residual network model; S4: The trained deep residual network model is used to detect and identify the specific communication protocol signals trained during the transmission process.

2. According to claim 1, a specific protocol signal recognition method based on a deep residual network is characterized in that: The specific steps of s1 are: using short-time Fourier transform as a time-frequency analysis method, in the process of processing, using a centrally symmetrical sliding window to intercept the observed signal, performing Fourier transform on the signal in the sliding window, and finally obtaining a time-frequency spectrum composed of each segment of the signal; For the signal S(t), the short-time Fourier transform of the signal can be defined as: Where γ(·) represents the window function.

3. According to claim 1, a specific protocol signal recognition method based on a deep residual network is characterized in that: The steps of S2 are: s2.

1. Mark H(x) as a potential mapping, and the stacked network is fitted with a residual mapping H(x)=F(x)-x; S2.

2. When the dimensions of the input feature map and the output feature map after the residual unit do not match, a convolution kernel of size 1x1 is used to increase the dimension of the input feature map.

4. According to claim 1, a specific protocol signal recognition method based on a deep residual network is characterized in that: The steps of S3 are as follows: s3.1 According to the internal structure of the network and the actual training requirements, the size of the time-frequency graph needs to be consistent during training. Combined with the actual shortwave signal search requirements, during the training process, the size of the time-frequency graph is uniformly set to 320x320; S3.2 identifies three types of specific communication protocol signals and one type of noise; the three types of specific communication protocol signals are link11, clover2000 and 2G-ALE, and the noise type is Gaussian white noise; S3.3 After each layer of data is convolved to obtain a feature map, it is not directly processed by the activation function, but the data is first batch normalized; S3.4 solves the internal covariate shift that occurs during training through batch normalization, allowing the network to converge quickly.

5. The method for identifying a specific protocol signal based on a deep residual network according to claim 1, characterized in that: The steps of s4 are as follows: s4.1 In view of the influence of the signal environment on the network recognition rate, the experimental signal adopts the simulation signal under Link11, CLOVER2000, and 2G-ALE system. The specific parameters of signal generation are realized according to the conditions specified by the protocol. The carrier frequency f i within a certain range Randomly generated, the frequency resolution of the signal time-frequency diagram is f s / N(N=2 m , m=6,7,8,9,10), the signal-to-noise ratio range is -10dB-10dB; S4.2 examines the impact of wireless channels on algorithm performance, taking into account the existence of Doppler frequency deviation and multipath delay, and uses the Watterson channel model recommended in the F.1487 standard of ITU-R as a shortwave test channel to test the signal protocol recognition method under the shortwave channel; When the specific frame structure information on the S4.3 test time-frequency diagram is partially blocked and affects the network recognition performance, a PSK signal without a frame structure is used to perform time-frequency aliasing on the target signal. The time-frequency aliasing is defined as follows: There are only two component signals in the aliased signal, and the signals completely overlap in time. The aliasing degree in the frequency domain is defined as: Where Δf1, Δf2 are the bandwidths of the signals, Δf 12 is the overlapping bandwidth, and the signal-to-noise ratio is the ratio of the two powers; S4.4 Regarding the influence of network parameters on recognition performance, the network is tested on test samples after every 10 iterations of training. In each test, 600 signal samples are randomly selected from the sample data of 3 types of protocols, and the signal-to-noise ratio and aliasing conditions of the signals are random.

Images