A method and apparatus for intelligent identification of signal transmission protocols for radio monitoring

Abstract

This invention discloses an intelligent identification method and apparatus for signal transmission protocols in radio monitoring, relating to the field of radio detection, to improve the robustness, versatility, and practicality of signal protocol identification. After normalizing the signal power, this invention performs initial classification based on the signal's time-domain characteristics, classifying it into discrete and continuous types. For discrete types, further classification is performed based on the signal's higher-order spectral characteristics and time-frequency characteristics; for continuous types, further classification is performed based on the signal's time-frequency characteristics, higher-order spectral characteristics, and autocorrelation characteristics. This invention enhances the robustness, versatility, and practicality of signal protocol identification.

Description

Technical Field

[0001] This invention relates to the field of radio detection, and in particular to a method and apparatus for intelligent identification of signal transmission protocols for radio monitoring. Background Technology

[0002] With the rapid development of wireless communication technology, the number of wireless access devices has also increased significantly. Generally, communication between wireless access devices follows the specific frequency range and modulation scheme designed by each protocol. If a signal of a different protocol type suddenly appears, such as a change in frequency band or modulation scheme, it may indicate abnormal activity in the scenario, such as unauthorized access, malicious attacks, or interference. Therefore, identifying the protocol of the received signal is crucial for ensuring the security of wireless communication. The rapid development of wireless communication protocols presents an even greater challenge to efficiently monitoring large-scale protocol signals.

[0003] Most known signal protocol identification methods focus on application-layer protocol identification, typically by analyzing features from partial signal parameters. For example, Chinese patent document CN108712397A proposes a deep learning-based communication protocol identification method that extracts features from the signal's I / Q sequence and feeds them into a convolutional recognition network to identify the signal's communication protocol type. However, this type of method fails to consider the changes in signal parameters caused by channel variations. This can lead to drastic changes in signal parameters even if the protocol remains unchanged, resulting in discrepancies between the identification results and reality. Furthermore, the similarity in parameters or modulation schemes among certain protocols can also prevent this type of identification method from accurately identifying the signal protocol. Therefore, this type of identification method has poor generalization ability.

[0004] In addition, some studies have shown that signal protocols can be identified using the time-frequency characteristics of signals. For example, Chinese patent document CN118646687A proposes a method for identifying IoT communication protocols based on a combination of time and frequency domains. This method identifies IoT communication protocols by combining the time-frequency characteristics of protocol signals. However, the samples used in such methods are almost entirely from simulated QAM (Quadrature Amplitude Modulation) data. They do not consider the impact of real protocol signals from different signal sources and various noises and interferences that may exist in complex aerial scenarios on the algorithm, resulting in poor practicality of such methods.

[0005] Another approach involves establishing protocol feature vectors using various parameters of the protocol signal, including modulation scheme and frequency. These vectors are then vectorized by calculating entropy, and Signal Vector Machines (SVMs) are used for signal protocol classification and identification. For example, Chinese patent document CN117596193A proposes a neural network-based message protocol identification method and system, which uses message sequence features for filtering and analysis to identify the communication protocol of data packets. However, such methods heavily rely on the accuracy of signal parameter identification. In complex aerial scenarios, various types of interference can lead to error accumulation, severely reducing protocol identification accuracy and resulting in poor robustness. Summary of the Invention

[0006] The purpose of this invention is to provide a method and apparatus for intelligent identification of signal transmission protocols for radio monitoring, addressing all or part of the problems mentioned above, so as to improve at least one of the performance characteristics of robustness, versatility and practicality of signal protocol identification.

[0007] The technical solution adopted in this invention is as follows:

[0008] A method for intelligent identification of signal transmission protocols for radio monitoring, comprising:

[0009] S1. Perform power normalization on the received signal S(n) to obtain the unity power signal S. a (m);

[0010] S2, based on S a The time-domain characteristics of S(n) determine whether S(n) is a time-discrete signal or a time-continuous signal;

[0011] S3. For time-discrete signals, based on S a The higher-order spectral characteristics of S(n) determine whether S(n) is {GSM}, {802.11b}, {CDMA2000} or belongs to {NR, TD-LTE, 802.11a / g / n}; if S(n) belongs to {NR, TD-LTE, 802.11a / g / n}, then based on S a The time-frequency characteristics of S(n) determine whether S(n) is {NR}, {TD-LTE} or {802.11a / g / n};

[0012] S4. For time-continuous signals, based on S a The time-frequency characteristics of S(n) determine whether S(n) is a frequency-hopping signal or a non-frequency-hopping signal; if S(n) is a frequency-hopping signal, then based on S(n)... a The higher-order spectral characteristics of S(n) determine whether S(n) is {BR / EDR} or {BLE}; if S(n) is a non-frequency hopping signal, then based on S... aThe autocorrelation characteristics of S(n) determine whether S(n) is {FD-LTE} or {Noise}.

[0013] To address all or some of the aforementioned problems, the present invention also provides an apparatus comprising a processor and a storage medium, wherein the storage medium stores computer instructions, and the processor, when executing the computer instructions, can perform the aforementioned intelligent identification method for signal transmission protocols for radio monitoring.

[0014] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0015] This invention identifies the protocol type of received signals at multiple levels based on the characteristics of different signal protocols. Compared to methods that directly extract signal parameters (e.g., I / Q sequences) for identification, this improves the versatility of the identification method. Furthermore, this invention fully considers the multimodal characteristics of the signal and deeply integrates the signal protocol standard to identify the signal protocol type, enhancing the robustness of the identification method. Moreover, this invention is directly applied to real air interface signals acquired by the receiver, taking into account the impact of actual air interface signals from different transmitting devices in complex air scenarios, thus verifying the practicality of the identification method. Attached Figure Description

[0016] The present invention will be described by way of example and with reference to the accompanying drawings, wherein:

[0017] Figure 1 This is a flowchart of a method for intelligent identification of signal transmission protocols for radio monitoring, according to an embodiment of this application.

[0018] Figure 2 This is a signal flow graph of one embodiment of the intelligent identification method for signal transmission protocols for radio monitoring provided in this application.

[0019] Figure 3 This is a flowchart of the time-domain identification process provided in the embodiments of this application.

[0020] Figure 4 This is a structural diagram of the MobileNet image recognition network provided in the embodiments of this application.

[0021] Figure 5 This is a waveform diagram after taking the modulus value of 10 signals, including GSM and CDMA2000, in the time domain of this application embodiment.

[0022] Figure 6 This is a time-domain I / Q waveform and square spectrum feature diagram of the GSM signal in the embodiments of this application.

[0023] Figure 7 These are the square spectrum feature maps corresponding to the four different transmission rates of the 802.11b protocol in the embodiments of this application.

[0024] Figure 8 This is a time-domain I / Q waveform and fourth-order spectrum feature map of the CDMA2000 signal in an embodiment of this application.

[0025] Figure 9 This is a time-frequency characteristic map of NR, TD-LTE, and 802.11a / g / n in the embodiments of this application.

[0026] Figure 10 These are the time-frequency characteristic maps of Bluetooth signals and the square spectrum characteristic maps of BLE and BR / EDR signals in the embodiments of this application.

[0027] Figure 11 This is an autocorrelation feature map obtained by performing autocorrelation calculation on FD-LTE and pure noise signals in the embodiments of this application.

[0028] Figure 12 This is a confusion matrix diagram for identifying nine common protocol signals and air interface noise in the embodiments of this application. Detailed Implementation

[0029] All features disclosed in this specification, or all steps in all disclosed methods or processes, may be combined in any way, except for mutually exclusive features and / or steps.

[0030] Any feature disclosed in this specification (including any appended claims and abstract) may be replaced by other equivalent or similar features, unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is merely one example of a series of equivalent or similar features.

[0031] To address the shortcomings of existing signal protocol identification methods in terms of versatility, robustness, and practicality, this application provides a method and apparatus for intelligent identification of signal transmission protocols for radio monitoring, thereby improving the performance of signal protocol identification methods in one or more aspects of versatility, robustness, and practicality.

[0032] In some embodiments, such as Figure 1 As shown in the embodiments of this application, the intelligent identification method for signal transmission protocols for radio monitoring includes the following steps:

[0033] S1. Perform power normalization on the received signal S(n) (where n represents the sampling point) to obtain the unity power signal S. a (n).

[0034] The received signal S(n) is the signal that needs to be identified as the protocol type. This received signal is the protocol signal in a real non-cooperative air scenario, and contains pure noise and interference signals.

[0035] As an optional implementation, the received signal S(n) is normalized using the following formula:

[0036]

[0037] Where L is the number of sampling points of S(n), that is, the length of the received signal S(n).

[0038] S2, based on S a The time-domain characteristics of S(n) determine whether S(n) is a discrete-time signal or a continuous-time signal.

[0039] Because some protocols use the TDMA standard or have frame structures with significant time intervals, this type of signal is necessarily discontinuous in time, specifically manifested as variations in signal amplitude. Other signals, however, that do not involve this characteristic, remain continuous in time. Therefore, the initial classification of S(n) between discrete-time and continuous-time signals can be obtained based on the signal's time-domain characteristics.

[0040] As an optional implementation method, such as Figure 2 As shown, step S2 includes:

[0041] S2.1, Obtain S a The I / Q sequence (n) is then divided into multiple I / Q sequence segments.

[0042] The so-called I / Q sequence is the signal S A (n) In step S2.1, the I / Q sequence is divided into multiple I / Q sequence segments according to the set length, and the last I / Q sequence segment is reserved for later use.

[0043] S2.2. The I / Q sequence is fed into the stacked dilated convolution module and the first fully connected module for processing to obtain the first feature.

[0044] In some feasible implementations, such as Figure 3 As shown, the stacked dilated convolution module includes, in sequence, a one-dimensional convolutional layer Conv1d, an activation function GELU, a one-dimensional max pooling layer MaxPool1d, a one-dimensional convolutional layer Conv1d, an activation function GELU, a one-dimensional max pooling layer MaxPool1d, a one-dimensional convolutional layer Conv1d, an activation function GELU, a one-dimensional max pooling layer MaxPool1d, a one-dimensional convolutional layer Conv1d, an activation function GELU, and a one-dimensional max pooling layer MaxPool1d.

[0045] The first fully connected layer mentioned above consists of a linear layer, a ReLU activation function, and another linear layer.

[0046] S2.3. The last I / Q sequence segment is fed into the temporal feature convolution module for processing to obtain the second feature.

[0047] In some feasible implementations, such as Figure 3 As shown, the aforementioned temporal feature convolutional module sequentially includes a one-dimensional convolutional layer Conv1d, a one-dimensional average pooling layer AvgPool1d, a KAN network layer KANLayer, a one-dimensional batch normalization layer BatchNorm1d, an activation function ReLU, a one-dimensional convolutional layer Conv1d, a one-dimensional average pooling layer AvgPool1d, a KAN network layer KANLayer, an LSTM network layer, an activation function ReLU, a one-dimensional batch normalization layer BatchNorm1d, a KAN network layer KANLayer, and an activation function ReLU.

[0048] The KAN network layer described above is a neural network architecture, representing an application of the Kolmogorov-Arnold representation theorem to neural networks. The Kolmogorov-Arnold representation proves a mathematical theorem that a multivariate continuous function can be written as a nested superposition of single-variable continuous functions, specifically as follows:

[0049]

[0050] In the formula, Φ q It is called an external function, φ p,q These are called internal functions, and both are continuous univariate functions. The explanation of the above equation is common knowledge and will not be discussed in detail here. The above equation shows that in the real number domain, only summation is the only truly multivariate function; all other functions can be characterized by the summation of univariate functions. Therefore, KAN networks have powerful fitting capabilities. When processing time series analysis, KAN networks learn complex relationships and nonlinear patterns in the data, resulting in higher accuracy and interpretability in handling complex function fitting and solving partial differential equations.

[0051] In addition, the LSTM (Long Short-Term Memory) layer mentioned above is used for the analysis of time series data. It can effectively capture the temporal dependencies in the data. It includes forget gate, input gate and output gate. The relevant explanations are also common knowledge.

[0052] like Figure 3 As shown, the stacked dilated convolution module, the first fully connected module, the temporal feature convolution module, and the second fully connected module together constitute a temporal signal recognition network based on LSTM (Long-Short Term Memory Network).

[0053] S2.4. The first and second features are fused and fed into the second fully connected module to obtain the classification results of S(n) in time-discrete and time-continuous signals.

[0054] In some feasible implementations, the first feature and the second feature are spliced ​​and fused together before being sent to a second fully connected module for classification. The structure of the second fully connected module is the same as that of the first fully connected module described above.

[0055] Based on the above classification, the implementation based on S a The time-domain characteristics of S(n) preliminarily classify the received signal S(n) into either a time-discrete signal or a time-continuous signal, narrowing the range of signal protocol type identification. In the embodiments of this application, the possible protocol types of the received signal include {GSM, 802.11a / g / n, 802.11b, CDMA2000, TD-LTE, NR, BR / EDR, BLE, FD-LTE, Noise}. After the above preliminary classification, the received signal S(n) either belongs to the time-discrete signal {GSM, 802.11a / g / n, 802.11b, CDMA2000, TD-LTE, NR} or the time-continuous signal {BR / EDR, BLE, FD-LTE, Noise}.

[0056] S3. For time-discrete signals, based on S a The higher-order spectral characteristics of S(n) determine whether S(n) is {GSM}, {802.11b}, {CDMA2000} or belongs to {NR, TD-LTE, 802.11a / g / n}; if it belongs to {NR, TD-LTE, 802.11a / g / n}, then based on S a The time-frequency characteristics of (m) determine whether S(n) is {NR}, {TD-LTE}, or {802.11a / g / n}.

[0057] For ease of differentiation in the following text, S will be used as the prefix. a1 (n) represents a time-discrete signal belonging to {GSM, 802.11a / g / n, 802.11b, CDMA2000, TD-LTE, NR}, denoted by S. a2 (n) represents a time-discrete signal belonging to {802.11a / g / n, CDMA2000, TD-LTE, NR}, denoted by S. a3 (n) represents a time-discrete signal belonging to {802.11a / g / n, TD-LTE, NR}.

[0058] In this embodiment of the application, for time-discrete signals, based on S a1The higher-order spectral characteristics of S(n) determine whether S(n) is {GSM}, {802.11b}, {CDMA2000}, or belongs to {NR, TD-LTE, 802.11a / g / n}. As an optional implementation, this process includes:

[0059] S3.1, based on S a1 The square spectrum characteristics of S(n) determine whether S(n) is {GSM}, {802.11b} or belongs to {CDMA2000, NR, TD-LTE, 802.11a / g / n}.

[0060] In some feasible implementations, in step S3.1, signal S is performed. a1 (n) Fast Fourier Transform of the quadratic power and modulus processing to extract signal S. a1 The square spectrum feature F of (n) a1 (k), represented as:

[0061]

[0062] Then the squared spectral feature F a1 (k) Input to the MobileNet image recognition network, which receives data from F... a1 Discrete spectral line information extracted from the feature image of (k) is used to classify S(n). Since the GSM protocol signal uses GMSK modulation with a modulation coefficient h = 0.5, two distinct discrete spectral lines will appear on either side of the image in the square spectrum feature, which can be used for classification. That is, if F... a1 If two discrete spectral lines appear on both sides of the square spectrum feature image of S(k), then the protocol type of S(n) is {GSM}. The 802.11b protocol signal uses four different modulation methods depending on the rate, including DSSS direct-sequence spread spectrum and CCK keying modulation. These methods produce a distinct discrete spectral line at the zero frequency in the square spectrum feature, which can be used for classification. Therefore, in step S3.1, the final identification result of S(n) by the MobileNet image recognition network is {GSM}, {802.11b}, or {CDMA2000, NR, TD-LTE, 802.11a / g / n}.

[0063] As an optional implementation method, such as Figure 4 As shown, the MobileNet image recognition network consists of a deep convolutional module a, a deep convolutional module b, and a classification module:

[0064] The depthwise convolution module a sequentially includes a two-dimensional convolutional layer Conv2d, a two-dimensional convolutional layer Conv2d, a two-dimensional convolutional layer Conv2d, a two-dimensional max pooling layer Maxpool2d, and an activation function ReLU.

[0065] The depthwise convolution module b consists of a 2D convolutional layer Conv2d, a 2D batch normalization layer BatchNorm2d, and an activation function ReLU.

[0066] The classification module consists of a two-dimensional average pooling layer (Avgpool2d), a flatten layer, a linear layer, an activation function (ReLU), another linear layer (ReLU), and yet another linear layer (Linear).

[0067] S3.2 If S(n) belongs to {CDMA2000, NR, TD-LTE, 802.11a / g / n}, then based on S a2 The fourth power spectral characteristics of S(n) determine whether S(n) is {CDMA2000} or belongs to {NR, TD-LTE, 802.11a / g / n}.

[0068] In some feasible implementations, in step S3.2, signal S is performed. a2 (n) Fourth-order Fast Fourier Transform and modulus processing to extract the fourth-order spectral features F of the signal. a2 (k), represented as:

[0069]

[0070] Then the fourth power spectral feature F a2 (k) Input to the MobileNet image recognition network, and according to the MobileNet image recognition network from F a2The discrete spectral line information extracted from the feature image of (k) is used to classify S(n). Since the CDMA2000 forward link waveform uses QPSK modulation technology, there are no obvious features in the quadratic spectral features, but three discrete spectral lines appear in the fourth spectral features, and one obvious discrete spectral line is located at zero frequency. Therefore, it can be used to identify CDMA2000 protocol signals. In step S3.2, the final classification result of S(n) by the MobileNet image recognition network is {CDMA2000} or {NR, TD-LTE, 802.11a / g / n}.

[0071] In addition, step S3 also includes:

[0072] S3.3 If S(n) belongs to {NR, TD-LTE, 802.11a / g / n}, then based on S a3 The time-frequency characteristics of S(n) determine whether S(n) is {NR}, {TD-LTE}, or {802.11a / g / n}.

[0073] As an optional implementation, by modifying S a3 (n) Perform a short-time Fourier transform to obtain its time-frequency characteristics S. α3 (m,f) is represented as:

[0074] Where m and f represent time and frequency, respectively, g(n) is a window function, and R is a constant.

[0075] In some feasible implementations, according to S α3(m,f) plots a time-frequency feature map, which is then fed into the MobileNet image recognition network to classify S(n) based on both time and frequency domain information. 802.11a uses OFDM modulation, and its frame structure design includes a fixed 0.1ms time interval when continuously transmitting data packets. Therefore, within a fixed-length time-frequency feature map, 802.11a can be identified using this feature. 802.11g uses the same modulation scheme as 802.11a, but it operates in the 2.4GHz band, while 802.11a operates in the 5GHz band. 802.11n uses higher-speed, higher-throughput OFDM modulation. Therefore, these three protocol types can be grouped together for identification. For TD-LTE, a radio frame is 10ms long and contains 10 subframes. Each subframe contains two time slots and has three types: uplink subframes, downlink subframes, and special subframes. LTE has seven fixed subframe configuration methods during time-division multiplexing, including DSUUU DSUUU, DSUUD DSUUD, and DSUUU DDDDD. In other words, TD-LTE has flexible transmission intervals in the time dimension, which is significantly different from the 802.11 series of protocols. Furthermore, the length of each uplink subframe is 1ms, which is much longer than the time interval designed for the 802.11 protocol. This time-domain characteristic allows for the differentiation between 802.11a / g / n and TD-LTE protocol types. Finally, the physical resources designed for the NR protocol include the SSB synchronization block necessary for signal transmission. This block consists of four parts: the primary synchronization signal (PSS), the secondary synchronization signal (SSS), the demodulation reference signal (DMRS), and the physical broadcast channel (PBCH). It occupies four OFDM symbols in the time domain and 240 subcarriers in the frequency domain. Its special characteristic lies in the PSS, which is located on subcarriers 56 to 182 of the 0th OFDM symbol in the time-frequency resource, with all other subcarrier positions empty. The 240 subcarriers on the 1st to 3rd OFDM symbols are generally fully occupied. Therefore, based on this frequency domain characteristic, the SSB synchronization block can be found on the time-frequency feature map to determine the type of the NR protocol signal. In step S3.3, the final identification result of S(n) by the MobileNet image recognition network is {NR}, {TD-LTE}, or {802.11a / g / n}.

[0076] S4. For time-continuous signals, based on S a The time-frequency characteristics of S(n) determine whether S(n) is a frequency-hopping signal or a non-frequency-hopping signal; if S(n) is a frequency-hopping signal, then based on S(n)... a The higher-order spectral characteristics of S(n) determine whether S(n) is {BR / EDR} or {BLE}; if S(n) is a non-frequency hopping signal, then based on S... aThe autocorrelation characteristics of S(n) determine whether S(n) is {FD-LTE} or {Noise}.

[0077] For ease of distinction in the following text, S will be used. a4 (n) represents a time-continuous signal belonging to {BR / EDR, BLE, FD-LTE, Noise}, denoted by S. a5 (n) represents a time-continuous signal belonging to the frequency-hopping signal family, denoted by S. a6 (n) represents a time-continuous signal that is a non-frequency-hopping signal.

[0078] Step S4 can be further divided into three processes:

[0079] S4.1 For time-continuous signals, based on S a4 The time-frequency characteristics of S(n) determine whether S(n) is a frequency-hopping signal or a non-frequency-hopping signal.

[0080] As an optional implementation, for signal S a4 (n) Perform a short-time Fourier transform to obtain its time-frequency characteristics S. α4 (m,f) is represented as:

[0081]

[0082] In some feasible implementations, according to S α4 (m,f) Draw a time-frequency feature map and send it to the MobileNet image recognition network. Based on the time and frequency domain information, S(n) is classified to determine whether it is a frequency-hopping signal or a non-frequency-hopping signal. {BR / EDR, BLE} are both Bluetooth protocols. For Bluetooth, frequency hopping is used within the transmission band, and the main frequency components change over time. Non-frequency-hopping signals do not change over time, so the classification of frequency-hopping and non-frequency-hopping signals can be intuitively performed from the time-frequency feature map. That is, in step S4.1, the final identification result of S(n) by the MobileNet image recognition network is {BR / EDR, BLE} or {FD-LTE, Noise}.

[0083] S4.2 If S(n) is a frequency hopping signal, then based on S a5 The higher-order spectral characteristics of S(n) determine whether S(n) is {BR / EDR} or {BLE}.

[0084] As an optional implementation, in step S4.2, based on S... a5 The square spectrum characteristics of S(n) determine whether S(n) is {BR / EDR} or {BLE}.

[0085] In some feasible implementations, signal S is performed.a5 (n) Fast Fourier Transform of the quadratic power and modulus processing to extract signal S. a5 The square spectrum feature F of (n) a5 (k), represented as:

[0086]

[0087] Then the squared spectral feature F a5 (k) Input to the MobileNet image recognition network, which receives data from F... a5 The discrete spectral line information extracted from the feature image of (k) is used to classify S(n). The Bluetooth protocol includes the classic Bluetooth waveform BR / EDR and the Bluetooth Low Energy waveform BLE. Both use GFSK modulation, but the BLE signal is similar to GSM modulation, with the same modulation coefficient h = 0.5. Therefore, within the squared spectral feature, the BLE signal also has two obvious discrete spectral lines located on both sides of the image. However, the modulation coefficient of the classic Bluetooth BR / EDR waveform is between 0.28 and 0.35, and there are no such obvious discrete spectral lines. This feature can be fully utilized to classify the two. In step S4.2, the final recognition result of S(n) by the MobileNet image recognition network is {BR / EDR} and {BLE}.

[0088] S4.3 If S(n) is a non-frequency hopping signal, then based on S a6 The autocorrelation characteristics of S(n) determine whether S(n) is {FD-LTE} or {Noise}.

[0089] For non-frequency hopping signals, only two protocol types remain: {FD-LTE} and {Noise}. As an optional implementation, for signal S... a6 (n) Perform autocorrelation calculation to obtain S a6 The autocorrelation characteristic R of (n) s (τ), represented as:

[0090]

[0091] Where τ is the delay time, N corr This is the upper limit of the time limit.

[0092] Since FD-LTE signals use OFDM modulation, guard intervals are typically added in OFDM systems to reduce inter-symbol interference. In practice, a cyclic prefix (CP) method is generally used. At the transmitter, the data segment at the tail of the OFDM symbol is copied into the guard interval to form a cyclic prefix (CP). This ensures that each subcarrier maintains a complete cycle within the FFT time window during CP removal and FFT processing at the receiver. This copied guard interval can be detected using autocorrelation features, and compared to the random pure noise signal {Noise}, it can be used to classify the two. Therefore, in step S4.3, the final recognition result of S(n) by the MobileNet image recognition network is {FD-LTE} and {Noise}.

[0093] It should be noted that the LSTM-based signal recognition network and MobileNet image recognition network mentioned in the embodiments of this application are both trained using data samples before use. The training method is a well-known technology in the field, so it is not described in detail in the embodiments of this application.

[0094] Furthermore, this application also provides an apparatus comprising a processor and a storage medium storing computer instructions. When the processor executes the computer instructions, it can perform the intelligent identification method for signal transmission protocols for radio monitoring described above.

[0095] This application also experimentally verifies the proposed method. In this experiment, the algorithm's performance was verified using 10 common airborne protocol signal types, including 2G mobile communication standard (GSM), 3G mobile communication system (CDMA2000), 4G mobile communication system (Time Division Duplex: TD-LTE, Frequency Division Duplex: FD-LTE), 5G mobile communication system (NR), classic Bluetooth waveform (BR / EDR), Bluetooth Low Energy waveform (BLE), wireless LAN protocol suite (802.11a, 802.11b, 802.11g, 802.11n), and pure noise. The signal datasets used in this application are all from actual complex airborne communication scenarios, acquired by a Siyi 3943B spectrum monitoring receiver with an antenna. The signals originate from various transmitting devices, fully considering the practicality of this algorithm. When acquiring signals, the Siyi 3943B spectrum monitoring receiver needs to select an appropriate demodulation bandwidth based on the bandwidth of different protocol signals. The demodulation bandwidth and the sampling rate have a certain mapping relationship, as shown in Table 1.

[0096] Table 1. Relationship between demodulation bandwidth and sampling rate

[0097] Demodulation bandwidth Sampling rate 300kHz 1.4MHz 500kHz 2MHz 800kHz 3.5MHz 1000KHz 4MHz 1250KHz 5.6MHz 1500KHz 7MHz 2000KHz 7MHz 5000KHz 14MHz 8000KHz 28MHz 10000KHz 28MHz 12500kHz 28MHz 15000kHz 28MHz 20000KHZ 28MHz

[0098] When constructing the dataset of all protocol signals, the different bandwidth sizes of each protocol in the actual scenario were taken into account. Therefore, each time an air signal is received, in order to avoid mutual interference between different protocol signals, only one protocol needs to be received within the demodulation bandwidth, and the different protocol signals come from different frequency bands. The parameters are shown in Table 2.

[0099] Table 2 Frequency Band Parameters for Receiving Over-the-Air Protocol Signals

[0100]

[0101]

[0102] According to the specific implementation of the embodiments of this application, this application can identify a total of 9 types of protocol signals and 1 type of empty noise. Among them, WIFI-802.11a / g / n is classified as one type, WIFI-802.11b as another type, classic Bluetooth BR / EDR and Bluetooth Low Energy (BLE) as yet another type, and TD-LTE and FD-LTE as yet another type for identification. The number of signal sampling points L used in this application to identify the protocol signal type is 16384.

[0103] In the experimental verification of the LSTM-based time-domain signal recognition network, the signal S fed into the network... a (n) The number of sampling points L = 16384, the sequence dimension fed into the stacked dilated convolution module is 2 × 16384, and the sequence dimension fed into the temporal feature convolution module is 2 × 512. The parameters of the LSTM-based temporal signal recognition network are shown in Table 3. The output dimension l = 2 of the linear layer of the classification module represents the probability that the recognition result may be {GSM, 802.11a / g / n, 802.11b, CDMA2000, TD-LTE, NR} and {BR / EDR, BLE, FD-LTE, Noise}. In addition, a symbol is added after the layer to distinguish the layers with the same structure in the same module.

[0104] Table 3. Parameters of LSTM-based temporal signal recognition network

[0105]

[0106]

[0107]

[0108] In the experimental verification of the MobileNet image recognition network, signal S was extracted. a1 (n), S a5 The quadratic spectrum of (n), the extracted signal S a2The fourth power spectrum of (n), the extracted signal S a3 (n), S a4 The time-frequency characteristics of (n) and the extracted signal S a6 The number of signal sampling points used for the autocorrelation feature of (n) is L = 16384, and the image dimension of the input network is 64×64×3. The network parameters of the MobileNet image recognition network are shown in Table 4, where the final output dimension of the linear layer of the classification module is l = L, which varies depending on different input requirements. For signal S a1 (n), the final output dimension l=3, representing the probability that the recognition result may be {GSM}, {802.11b}, {CDMA2000, NR, TD-LTE, 802.11a / g / n}; for signal S a2 (n), the final output dimension l=2, representing the probability that the recognition result may be {CDMA2000}, {NR, TD-LTE, 802.11a / g / n}; for signal S a3 (n), the final output dimension l=3, representing the probability that the recognition result may be {NR}, {TD-LTE}, {802.11a / g / n}; for signal S a4 (n), the final output dimension l=2, representing the probability that the recognition result may be {BR / EDR, BLE} and {FD-LTE, Noise}; for signal S a5 (n), the final output dimension l=2, representing the probability that the recognition result may be {BLE} and {BR / EDR}; for signal S a6 (n), the final output dimension l=2, representing the probability that the recognition result may be {FD-LTE} and {Noise}.

[0109] Table 4 MobileNet Image Recognition Network Parameters

[0110]

[0111]

[0112]

[0113] Experiment 1: From Figure 5 As can be seen, the waveforms after taking the time-domain I / Q modulus of the signals GSM, CDMA2000, TD-LTE, NR, 802.11a / g / n, 802.11b, BR / EDR, BLE, FD-LTE, and Noise signals, respectively, can be clearly divided into two categories: time-continuous and time-discontinuous signals.

[0114] Experiment 2: From Figure 6As can be seen from the time-domain I / Q waveform and squared spectrum of the GSM signal, two discrete spectral lines can be clearly seen on both sides of the spectrum.

[0115] Experiment 3: From Figure 7 As can be seen from the square spectrum images of the protocol signals corresponding to the four different transmission rates of the 802.11b protocol, there is always a discrete spectral line at the zero frequency.

[0116] Experiment 4: From Figure 8 As can be seen from the time-domain I / Q waveform and fourth-order spectrum of the CDMA2000 signal, the discrete spectral line located at zero frequency in the middle of the spectrum is clearly visible.

[0117] Experiment 5: From Figure 9 As can be seen from the time-frequency diagrams of NR, TD-LTE, and 802.11a / g / n, the location of the SSB synchronization block within the physical resources of the NR signal and the different time interval lengths of TD-LTE and 802.11a / g / n are clearly visible.

[0118] Experiment 6: From Figure 10 As can be seen, the time-frequency diagram of the frequency hopping signal in the Bluetooth band, the square spectrum of the BLE signal, and the square spectrum of the BR / EDR signal are shown in sequence. The characteristics of the frequency hopping signal and the two discrete spectral lines located on both sides of the BLE square spectrum can be seen.

[0119] Experiment 7: From Figure 11 As can be seen, the images drawn after performing autocorrelation operations on FD-LTE and noise signals respectively clearly show the autocorrelation of OFDM modulation compared to irregular noise.

[0120] Experiment 8: From Figure 12 As can be seen from the identification confusion matrix obtained after testing the actual airborne nine common types of protocol signals and airborne noise signals using the method of this application, the identification accuracy of TD-LTE is over 96%; the identification accuracy of BLE is over 97%; the identification accuracy of CDMA2000, FD-LTE, NR, 802.11a / g / n, and 802.11b is over 98%; and the identification accuracy of GSM, BR / EDR, and Noise is over 99%.

[0121] The signal dataset used in this application is collected from real-world aerial scenarios and can be directly used as a standalone module on the monitoring receiver. Signals are acquired via a spectrum monitoring receiver, and data from multiple transmission protocol models are used as training and testing sets, respectively. By comprehensively considering the complex aerial environment, including noise and interference, highly robust signal features are extracted. Furthermore, deep learning algorithms are combined to fully exploit the signal's latent features. Compared to existing algorithms using simulated datasets, this application not only considers real-world complex scenarios but can also be used on existing monitoring receiver equipment, expanding the functionality of current monitoring receivers and fully validating the practicality of the identification method presented in this application.

[0122] Furthermore, addressing the issues of limited existing wireless transmission protocol identification algorithms and the accumulation of errors in these algorithms, the identification method proposed in this application fully utilizes the characteristics of each protocol, such as frame structure, modulation method, and physical resources. It combines the autocorrelation of the signal with the combination of multimodal features in the time domain, frequency domain, time-frequency domain, and autocorrelation domain. By employing deep learning algorithms, it fully leverages the high robustness of the signal for identification, thereby improving the accuracy of the identification method. In complex scenarios, it exhibits stronger robustness and versatility than existing algorithms.

[0123] This invention is not limited to the specific embodiments described above. The invention extends to any new feature or combination disclosed in this specification, as well as any new method or process step or combination disclosed herein.

Claims

1. A method for intelligent identification of signal transmission protocols for radio monitoring, characterized in that, include: S1, Regarding the received signal Power normalization is performed to obtain a unity power signal. ; S2, based on Determination of time-domain characteristics Is it a discrete-time signal or a continuous-time signal? This includes: S2.1, Obtain The I / Q sequence is then sequentially divided into multiple I / Q sequence segments; S2.2 The I / Q sequence is fed into a stacked dilated convolution module and a first fully connected module for processing to obtain the first feature; S2.

3. The last I / Q sequence segment is fed into the temporal feature convolution module for processing to obtain the second feature; S2.

4. The first feature and the second feature are fused and sent to the second fully connected module for processing to obtain the desired result. Classification results for discrete-time signals and continuous-time signals; S3. For time-discrete signals, based on Determination of high-order spectral characteristics Is it {GSM}, {802.11b}, {CDMA2000}, or does it belong to {NR, TD-LTE, 802.11a / g / n}? If If it belongs to {NR, TD-LTE, 802.11a / g / n}, then it is based on Determination of time-frequency characteristics Is it {NR}, {TD-LTE}, or {802.11a / g / n}? S4. For time-continuous signals, based on Determination of time-frequency characteristics Is it a frequency-hopping signal or a non-frequency-hopping signal? If If it belongs to a frequency hopping signal, then based on Determination of high-order spectral characteristics Is it {BR / EDR} or {BLE}? If If it belongs to a non-frequency hopping signal, then based on Determination of autocorrelation characteristics Is it {FD-LTE} or {Noise}? 2. The intelligent identification method for signal transmission protocols for radio monitoring as described in claim 1, characterized in that, right The method for normalization is as follows: ， in, yes The number of sampling points.

3. The intelligent identification method for signal transmission protocols for radio monitoring as described in claim 1, characterized in that, The stacked dilated convolution module sequentially includes a one-dimensional convolutional layer Conv1d, an activation function GELU, a one-dimensional max pooling layer MaxPool1d, a one-dimensional convolutional layer Conv1d, an activation function GELU, a one-dimensional max pooling layer MaxPool1d, a one-dimensional convolutional layer Conv1d, an activation function GELU, a one-dimensional max pooling layer MaxPool1d, a one-dimensional convolutional layer Conv1d, an activation function GELU, and a one-dimensional max pooling layer MaxPool1d; The temporal feature convolution module sequentially includes a one-dimensional convolutional layer Conv1d, a one-dimensional average pooling layer AvgPool1d, a KAN network layer KANLayer, a one-dimensional batch normalization layer BatchNorm1d, an activation function ReLU, a one-dimensional convolutional layer Conv1d, a one-dimensional average pooling layer AvgPool1d, a KAN network layer KANLayer, an LSTM network layer, an activation function ReLU, a one-dimensional batch normalization layer BatchNorm1d, a KAN network layer KANLayer, and an activation function ReLU. Both the first fully connected layer and the second fully connected layer sequentially include a linear layer (Linear), an activation function (ReLU), and another linear layer (Linear).

4. The intelligent identification method for signal transmission protocols for radio monitoring as described in claim 1, characterized in that, In step S3, based on Determination of high-order spectral characteristics Is it {GSM}, {802.11b}, {CDMA2000}, or does it belong to {NR, TD-LTE, 802.11a / g / n}, including: S3.1, based on Determination of the square spectrum characteristics Is it {GSM}, {802.11b}, or does it belong to {CDMA2000, NR, TD-LTE, 802.11a / g / n}? S3.2, if If it belongs to {CDMA2000, NR, TD-LTE, 802.11a / g / n}, then based on Determination of the fourth power spectral characteristics Is it {CDMA2000} or does it belong to {NR, TD-LTE, 802.11a / g / n}? 5. The intelligent identification method for signal transmission protocols for radio monitoring as described in claim 1, characterized in that, In step S4, based on Determination of high-order spectral characteristics Is it {BR / EDR} or {BLE}, including: based on Determination of the square spectrum characteristics Is it {BR / EDR} or {BLE}? 6. The intelligent identification method for signal transmission protocols for radio monitoring as described in claim 4 or 5, characterized in that, In at least one of steps S3.1, S3.2, and S4, discrete spectral line information is extracted from the square spectral features or fourth-power spectral features using the MobileNet image recognition network to... Classify them.

7. The intelligent identification method for signal transmission protocols for radio monitoring as described in claim 6, characterized in that, The MobileNet image recognition network consists of a deep convolutional module a, a deep convolutional module b, and a classification module; The depthwise convolution module a sequentially includes a two-dimensional convolutional layer Conv2d, a two-dimensional convolutional layer Conv2d, a two-dimensional convolutional layer Conv2d, a two-dimensional max pooling layer Maxpool2d, an activation function ReLU, a two-dimensional convolutional layer Conv2d, a two-dimensional convolutional layer Conv2d, a two-dimensional convolutional layer Conv2d, a two-dimensional max pooling layer Maxpool2d, and an activation function ReLU. The depthwise convolution module b sequentially includes a two-dimensional convolutional layer Conv2d, a two-dimensional batch normalization layer BatchNorm2d, and an activation function ReLU. The classification module includes, in sequence, a two-dimensional average pooling layer (Avgpool2d), a flatten layer (Flatten), a linear layer (Linear), an activation function (ReLU), another linear layer (Linear), and yet another linear layer (Linear).

8. The intelligent identification method for signal transmission protocols for radio monitoring as described in claim 1, characterized in that, The The time-frequency characteristics are obtained by analyzing the time-frequency characteristics of ... It is obtained by performing a short-time Fourier transform.

9. A smart identification device for signal transmission protocols for radio monitoring, comprising a processor and a storage medium, wherein the storage medium stores computer instructions, characterized in that, When the processor executes the computer instructions, it can perform the intelligent identification method for signal transmission protocols for radio monitoring as described in any one of claims 1-8.

Images