A LoRa non-stationary radio frequency fingerprint feature extraction method and related device

Abstract

A LoRa non-stationary RF fingerprint feature extraction method and related device include: collecting physical layer signals from different LoRa device models, preprocessing the collected one-dimensional signals to obtain corresponding non-stationary signals; constructing a RF fingerprint feature extraction model for LoRa non-stationary signals based on a fractional-domain wavelet scattering network to extract RF fingerprint feature coefficients; inputting the RF fingerprint feature coefficients into a corresponding classification network, and using a convolutional neural network to accurately identify and classify feature vectors. The proposed fractional-domain wavelet scattering network can efficiently extract RF feature information under non-stationary conditions, significantly reducing the amount of RF feature data and improving model learning efficiency.

Description

Technical Field

[0001] The present invention belongs to the technical field of radio frequency feature extraction, and in particular relates to a LoRa non-stationary radio frequency fingerprint feature extraction method and related devices. Background Art

[0002] In recent years, with the large-scale deployment of low-cost and low-power smart LoRa devices in the Internet of Things, the Internet of Things can provide low-cost, anytime, anywhere network access. The network access of smart terminal devices first requires the completion of the device's legitimacy detection. The traditional device legitimacy authentication process often relies on high-level encryption solutions. Figure 2 In traditional IoT applications, systems like LoRa, Bluetooth, and NB-IoT are often limited by computing resources and energy consumption, making complex high-level encryption schemes infeasible. Therefore, implementing reliable and lightweight authentication for connected devices in scenarios with massive device connectivity is a key research direction for future industrial IoT, autonomous driving, and smart home applications. Because the upper-layer protocols of smart devices like LoRa and Bluetooth support upper-layer encryption authentication, which requires high computational complexity and encryption overhead, and is vulnerable to attacks such as identity spoofing, developing an ultra-reliable, ultra-low computational complexity physical layer authentication solution based on the physical layer is an effective approach for addressing the legitimacy of future large-scale smart devices. Radio frequency fingerprint-based individual identification technology is an efficient approach for identifying individual devices by extracting unique hardware defect signatures (such as I / Q mismatch, amplifier nonlinearity, and carrier offset) from physical layer signals. Because the defect signatures of each hardware module are unique, the accumulated signatures of each module create a unique device signature for each individual. This RF signature can then be used to uniquely identify and distinguish the device. However, traditional machine learning methods require large amounts of precisely labeled data samples to train network models on RF signature information, which in turn places a significant demand on computing resources. Furthermore, traditional machine learning models use the collected physical layer signal as input and the corresponding classification results as output. The RF signature extraction process operates in a black-box mode, lacking clear system interpretability. Furthermore, given the core issue of wireless transmission signals being inevitably affected by noise, traditional network models confuse this with RF signatures during extraction, failing to mitigate the impact of noise. This, in turn, raises concerns about the impact of noise on device recognition accuracy. Summary of the Invention

[0003] The purpose of the present invention is to provide a LoRa non-stationary radio frequency fingerprint feature extraction method and related devices to solve the problem that the traditional network model confuses it with the radio frequency feature extraction and cannot reduce the impact of noise on it, thus also bringing about some problems about the impact of noise on device recognition accuracy.

[0004] To achieve the above object, the present invention adopts the following technical solutions:

[0005] A LoRa non-stationary radio frequency fingerprint feature extraction method, comprising:

[0006] The physical layer signals of different LoRa devices are collected, and the collected one-dimensional signals are preprocessed to obtain the corresponding non-stationary signals;

[0007] A radio frequency fingerprint feature extraction model for LoRa non-stationary signals is constructed based on the fractional domain wavelet scattering network to extract the radio frequency fingerprint feature coefficients;

[0008] The RF fingerprint feature coefficients are input into the corresponding classification network, and the convolutional neural network is used to achieve accurate recognition and classification of the feature vector.

[0009] Furthermore, the LoRa device collects physical layer signals:

[0010] The operating bandwidth of the LoRa device is represented by B, and the corresponding minimum operating frequency is The maximum operating frequency is The corresponding LoRa signal instantaneous frequency is expressed as:

[0011] f(t)=f L +kt(0≤t≤T)

[0012] in, Represents the frequency change speed of the signal, T represents each symbol period; the instantaneous phase of the signal is obtained by integrating the instantaneous frequency of the signal:

[0013]

[0014] The basic Upchirp baseband signal is obtained by orthogonal modulation of the signal:

[0015]

[0016] Among them, A represents the signal amplitude; LoRa signal realizes SF-bit information modulation by setting different degrees of frequency offset of the basic signal; SF represents the expansion factor of the signal, which is 2 for LoRa signal. SF Different frequency offsets are used to complete the corresponding information modulation; the final modulated signal is expressed as

[0017]

[0018] Where Δf represents the signal frequency offset, and x(t) represents the complete modulated signal. After passing through different modules, the signal is sent to the wireless space through the antenna. On the receiving side, the signal is expressed as

[0019] y(t)=H*G(x(t))+n

[0020] Here, H represents the channel impulse response, G represents the device RF characteristic information, and n represents the Gaussian noise introduced during the transmission process.

[0021] Furthermore, the signal y(t) on the receiving side contains the radio frequency fingerprint feature information of the wireless channel, noise and device. During the LoRa signal transmission process, the preamble is composed of a fixed basic Upchirp signal. The preamble signal at the receiving end is obtained through signal synchronization and symbol segmentation, and the unique radio frequency feature information of the device is further extracted from the signal. Short-time Fourier transform time-frequency transformation is used to analyze the non-stationary characteristics of the signal. Different windows and lengths are used to realize the time-frequency transformation of the signal, as shown below.

[0022]

[0023] Where y[n] represents the input signal, h[n] represents the sliding window in the corresponding STFT; the received signal is normalized.

[0024] Furthermore, a radio frequency fingerprint feature extraction model for LoRa non-stationary signals is constructed based on the fractional domain wavelet scattering network:

[0025] The fractional domain Fourier transform is the basic transform in DFSNet, as shown below

[0026]

[0027] Where d represents the dimension of the signal f(t), Meet the following conditions

[0028]

[0029] Among them, α i represents the rotation angle of the fractional Fourier transform, For the corresponding transform coefficient, based on the fractional domain Fourier transform, the corresponding basis function is replaced by the wavelet kernel function to obtain the fractional domain wavelet transform

[0030]

[0031] Among them, the wavelet kernel function ψ α,λ,t (κ) satisfies the following conditions

[0032]

[0033] Where λ and t represent the corresponding scale transformation and time factor respectively. When α = π / 2, the above formula will become the traditional wavelet transform.

[0034] Furthermore, the characteristic coefficients of the input time-frequency feature map are calculated using the fractional domain wavelet scattering network. The basic structure of the fractional domain wavelet scattering network is consistent with the traditional convolutional neural network. Both are multi-layer linear connections, and nonlinear changes and operations are performed at different network nodes. The calculation results of the fractional domain wavelet scattering network are divided into two categories: general features and subtle feature information. The general signal corresponding to the input time-frequency feature is expressed as

[0035]

[0036] Where J represents the scale factor, Represents the calculated overview feature information; the corresponding detail feature information is expressed as

[0037]

[0038] Where K represents different wavelet basis rotation angles, further simplifying the traditional layered convolution calculation process, as shown below

[0039]

[0040]

[0041] in, represents the kernel function of the low-frequency filter characteristics, Represents a wavelet kernel function with bandpass filtering characteristics.

[0042] Furthermore, the RF fingerprint feature coefficients are input into the corresponding classification network: the input signal f(t) is convolved with different wavelet kernel functions in the first layer of the network to obtain the corresponding scattering coefficients, i.e., the RF fingerprint feature information; the calculation process corresponding to the second layer of the scattering network is to continue to decompose and calculate the coefficients calculated in the first layer. The nonlinear operation process of the fractional domain wavelet scattering network consists of two nonlinear operations, the fractional domain wavelet filter and the module value. The coefficients of each layer in the DFSNet network model are composed of S α [l (m) ]f(t) is calculated, and the fractional domain wavelet scattering network consists of three layers, where the output of each layer is expressed as

[0043]

[0044] The information nodes transmitted on the network transmission path are represented as

[0045]

[0046] where l (m) The representative is m The transmission path of the fractional-domain wavelet scattering network is shown, which represents the calculation results related to the path.

[0047] Furthermore, after calculating the corresponding scattering coefficients, they are combined into a unified feature vector as the input of the classification network. The classification network is mainly composed of a one-dimensional residual network, which includes two basic modules: ID and Conv modules. For the residual network body, a filter with a size of 7×1 is first used to perform preliminary convolution calculations after the input layer, followed by a maximum pooling layer and 7 different ID and Conv modules to perform in-depth information extraction and classification of the input signal. A one-dimensional AlexNet network is introduced for comparison with the traditional multi-layer perceptron network, where the multi-layer perceptron network is composed of three convolutional neural networks with sizes of 128×1, 128×2, and 128×2.

[0048] Furthermore, a LoRa non-stationary radio frequency fingerprint feature extraction system includes:

[0049] The data acquisition module is used to collect physical layer signals from different types of LoRa devices and pre-process the collected one-dimensional signals to obtain corresponding non-stationary signals;

[0050] The extraction module is used to build a radio frequency fingerprint feature extraction model for LoRa non-stationary signals based on the fractional domain wavelet scattering network and extract the radio frequency fingerprint feature coefficients;

[0051] The recognition and classification module is used to input the radio frequency fingerprint feature coefficients into the corresponding classification network and use the convolutional neural network to achieve accurate recognition and classification of feature vectors.

[0052] Furthermore, a computer device includes a memory, a processor, and a computer program stored in the memory and runnable on the processor, wherein the processor implements the steps of a LoRa non-stationary radio frequency fingerprint feature extraction method when executing the computer program.

[0053] Furthermore, a computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a LoRa non-stationary radio frequency fingerprint feature extraction method.

[0054] Compared with the prior art, the present invention has the following technical effects:

[0055] The present invention aims to extract and identify the radio frequency fingerprint feature information of LoRa devices, and realizes efficient extraction of non-stationary radio frequency fingerprint feature information through a fractional domain wavelet scattering network, and then realizes accurate identification of the feature vectors corresponding to large-scale LoRa devices through a classification network. The fractional domain wavelet scattering network introduces the fractional domain wavelet transform and uses its dynamic adaptive scaling kernel function to realize accurate extraction of general feature signals and subtle feature signals in non-stationary signals, which greatly reduces the dimension of the input signal of the classification network. At the same time, the fractional domain deformation formed by the presence of channel noise will be strictly controlled within a certain range, thereby greatly improving the accuracy of the radio frequency feature information calculated by the fractional domain wavelet scattering network. Therefore, under non-stationary transmission conditions, the use of the fractional domain wavelet scattering network can realize efficient extraction of radio frequency fingerprint feature information, enhance the interpretability of the system, and control the influence of noise alignment within a certain range while greatly reducing the amount of feature data. It has the following advantages:

[0056] First: The proposed fractional domain wavelet scattering network can solve the problem of efficient extraction of RF feature information under non-stationary conditions, greatly reduce the amount of RF feature data, and improve model learning efficiency.

[0057] Second: The fractional domain wavelet scattering network can control the noise in the signal within a certain range, minimizing the impact of noise on RF fingerprint characteristics.

[0058] Third: The scattering coefficient calculated based on the scattering network represents clear general features and subtle features, which can clearly explain the characteristic information contained in the signal and improve the interpretability of the network model.

[0059] Fourth: The fractional domain wavelet scattering network can efficiently extract the radio frequency fingerprint feature information corresponding to the LoRa device, and can complete the convergence learning of the network under the conditions of a small number of samples. Therefore, this framework is more suitable for radio frequency feature extraction and individual identification under small sample conditions.

[0060] Furthermore, the above-mentioned network model can be deployed and applied in conjunction with the signal acquisition module. By completing the RF fingerprint feature extraction and further classification network model training of the collected signals, the invalid calculation of massive redundant data is avoided, the interpretability of the network model is enhanced, and the system efficiency is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 LoRa radio frequency fingerprint feature recognition model structure based on mixed fractional domain wavelet scattering network.

[0062] Figure 2 Typical architecture of LoRa application scenarios.

[0063] Figure 3Scattering coefficient based on fractional domain wavelet scattering network when J=2L=3M=2.

[0064] Figure 4 LoRa radio frequency fingerprint feature extraction and recognition system based on fractional domain wavelet scattering network.

[0065] Figure 5 LoRa device recognition results based on fractional domain wavelet scattering network; (a) Recognition accuracy under different sample numbers in indoor scenarios; (b) Recognition accuracy under different sample numbers in outdoor scenarios.

[0066] Figure 6 Confusion matrix of the recognition results of the hybrid network model; (a) Recognition confusion map when the number of training samples is 1000; (b) Recognition confusion map when the number of training samples is 5000. DETAILED DESCRIPTION

[0067] The present invention is further described below with reference to the accompanying drawings:

[0068] The present invention aims to extract and identify the radio frequency fingerprint feature information of LoRa devices. It uses a fractional domain wavelet scattering network to achieve efficient extraction of non-stationary radio frequency fingerprint feature information, and then uses a classification network to achieve accurate identification of the feature vectors corresponding to large-scale LoRa devices. The fractional domain wavelet scattering network introduces the fractional domain wavelet transform and uses its dynamic adaptive scaling kernel function to achieve accurate extraction of general feature signals and subtle feature signals in non-stationary signals, greatly reducing the dimension of the input signal of the classification network. At the same time, the fractional domain deformation formed by the presence of channel noise will be strictly controlled within a certain range, thereby greatly improving the accuracy of the radio frequency feature information calculated by the fractional domain wavelet scattering network. Therefore, under non-stationary transmission conditions, the use of the fractional domain wavelet scattering network can achieve efficient extraction of radio frequency fingerprint feature information, enhance the interpretability of the system, and control the influence of noise alignment within a certain range while greatly reducing the amount of feature data.

[0069] The present invention provides a LoRa non-stationary radio frequency fingerprint feature extraction method and system, which uses a fractional domain wavelet scattering network to efficiently extract and analyze the LoRa radio frequency fingerprint feature information hidden in the non-stationary signal, thereby improving the individual recognition accuracy of large-scale LoRa devices in a wireless environment and constructing a device legitimacy assessment framework based on physical layer signals.

[0070] The present invention is specifically implemented through the following technical solutions:

[0071] The first step is to collect physical layer signals from 25 different types of LoRa devices, and then obtain a complete physical signal frame by segmenting and extracting the collected one-dimensional signals.

[0072] In the second step, a radio frequency fingerprint feature extraction model for LoRa non-stationary signals is constructed based on the fractional domain wavelet scattering network to complete the extraction of radio frequency fingerprint feature coefficients.

[0073] The third step is to input the RF fingerprint feature coefficients calculated by the scattering network into the corresponding classification network, and use the convolutional neural network to achieve accurate recognition and classification of the feature vector. The above related content is introduced in detail below.

[0074] First, the signal generation, RF feature generation mechanism, and signal frame synchronization and segmentation issues involved in the signal acquisition process are described in detail. A LoRa information transmission block mainly consists of three parts: preamble, start identifier, and data payload block. The basic symbol composition is a linear Chirp signal, where the frequency of the Chirp signal gradually increases from the lowest frequency to the highest operating frequency. The operating bandwidth of the LoRa system can be represented by B, and the corresponding minimum operating frequency is The maximum operating frequency is Therefore, the corresponding LoRa signal instantaneous frequency can be expressed as:

[0075] f(t)=f L +kt (0≤t≤T)

[0076] in, Represents the frequency change speed of the signal, and T represents each symbol period. By integrating the instantaneous frequency of the signal, the instantaneous phase of the signal can be further obtained:

[0077]

[0078] By orthogonalizing the signal, the basic Upchirp baseband signal can be further obtained:

[0079]

[0080] Among them, A represents the signal amplitude. LoRa signal realizes SF-bit information modulation by setting different degrees of frequency offset on the basic signal. Among them, SF represents the expansion factor of the signal. For LoRa signal, there are 2 SF Different frequency offsets are used to complete the corresponding information modulation. Therefore, the final modulated signal can be expressed as

[0081]

[0082] Where Δf represents the signal frequency offset, and x(t) represents the complete modulated signal. After passing through different modules, the signal is sent to the wireless space through the antenna. On the receiving side, the signal can be further expressed as

[0083] y(t)=H*G(x(t))+n

[0084] Among them, H represents the channel impulse response, G represents the device RF feature information, and n represents the Gaussian noise introduced during the transmission process. Therefore, the signal y(t) on the receiving side contains the wireless channel, noise, and the RF fingerprint feature information contained in the device. How to accurately extract the RF feature information G of the device and use it to achieve accurate identification of large-scale LoRa devices. During the LoRa signal transmission process, the preamble is composed of a fixed basic Upchirp signal. The preamble signal at the receiving end is obtained through signal synchronization and symbol segmentation, and then the unique RF feature information of the device is further extracted from the signal. Since the LoRa signal uses frequency modulation, non-stationary features are introduced during the signal transmission process. In order to better characterize the non-stationary characteristics of the signal, time-frequency transform is needed for analysis. Short-time Fourier transform (STFT) is a time-frequency analysis method implemented by windowed Fourier transform. Different windows and lengths can be used to implement time-frequency transform of the signal, as shown below

[0085]

[0086] Where y[n] represents the input signal, and h[n] represents the corresponding sliding window in the STFT. To uniformly process the RF signature information of signals with different transmit powers, the received signal must be normalized to prevent the network model from learning the corresponding transmit power parameters and ignoring the actual RF fingerprint information.

[0087] Secondly, the scattering network (DFSNet) based on fractional domain wavelet can extract the high and low frequency information hidden in the time-frequency signal, while significantly reducing the size of the feature vector corresponding to the signal. The fractional domain Fourier transform is the basic transformation in DFSNet, as shown below

[0088]

[0089] Where d represents the dimension of the signal f(t), Meet the following conditions

[0090]

[0091] Among them, α i represents the rotation angle of the fractional Fourier transform, is the corresponding transformation coefficient. According to the above formula, when the rotation angle α=π / 2, the above fractional Fourier transform will degenerate into the traditional Fourier transform. Based on the fractional domain Fourier transform, the corresponding basis function can be replaced by the wavelet kernel function to further obtain the fractional domain wavelet transform.

[0092]

[0093] Among them, the wavelet kernel function ψ α,λ,t (κ) must satisfy the following conditions

[0094]

[0095] Where λ and t represent the corresponding scale transformation and time factor respectively. When α = π / 2, the above formula will become the traditional wavelet transform. The characteristic coefficients of the input time-frequency feature map are calculated using the fractional domain wavelet scattering network. The basic structure of the fractional domain wavelet scattering network is consistent with the traditional convolutional neural network. Both are multi-layer network linear connections, but nonlinear changes and operations are performed at different network nodes. The calculation results of the fractional domain wavelet scattering network can basically be divided into two categories: general features and subtle feature information. The general signal corresponding to the input time-frequency feature can be expressed as

[0096]

[0097] Where J represents the scale factor, represents the calculated overview feature information. The corresponding detail feature information can be further expressed as

[0098]

[0099] Where K represents different wavelet basis rotation angles. The above calculation process can further simplify the traditional layered convolution calculation process, as shown below

[0100]

[0101]

[0102] in, represents the kernel function of the low-frequency filter characteristics, Represents a wavelet kernel function with bandpass filtering characteristics

[0103] Finally, if Figure 1 The figure shows the network structure of the LoRa radio frequency fingerprint feature recognition model based on the mixed fractional domain wavelet scattering network. (a) shows the corresponding fractional domain wavelet scattering network calculation process. The input signal f(t) is convolved with different wavelet kernel functions in the first layer of the network to obtain the corresponding scattering coefficient, which is the radio frequency fingerprint feature information. The rightmost part is the general signal, and the rest is the corresponding subtle feature information. The calculation process corresponding to the second layer of the scattering network is to continue to decompose and calculate the coefficients calculated in the first layer, as shown in Figure 2. Figure 1The nonlinear operation process of the fractional domain wavelet scattering network shown in the figure mainly consists of two nonlinear operations: the fractional domain wavelet filter and the modulus value. The coefficients of each layer in the DFSNet network model are composed of S α [l (m) ]f(t) is calculated. The figure shows a fractional domain wavelet scattering network composed of three layers, where the output of each layer can be further expressed as

[0104]

[0105] The information nodes transmitted on the network transmission path can be further represented as

[0106]

[0107] where l (m) represents the mth fractional domain wavelet scattering network transmission path, which represents the calculation results related to the path. Figure 3 The figure shows the characteristic part represented by the scattering coefficients of different computing nodes in the scattering network. The figure shows the normalized time-frequency characteristics obtained by inverse restoration of the corresponding scattering coefficients. It can be seen from the figure that the scattering coefficients belonging to the overview characteristics represent the overall change trend in the time-frequency diagram, while the detail characteristics fully represent the RF fingerprint feature information contained in the signal. Therefore, in the process of extracting RF fingerprint feature information, the corresponding overview feature information and detail feature information can be obtained through the above calculations. After calculating the corresponding scattering coefficients, they are combined into a unified feature vector as the input of the classification network. Figure 3 As shown, the classification network primarily consists of a one-dimensional residual network, which includes two basic modules: the ID and Conv modules. The residual network performs a preliminary convolution operation after the input layer using a 7×1 filter. This is followed by a max-pooling layer and seven different ID and Conv modules to extract and classify the input signal. To compare the effectiveness of the proposed model, a one-dimensional AlexNet network is introduced alongside a traditional multilayer perceptron network. The multilayer perceptron network consists of three convolutional neural networks (CNNs) with sizes of 128×1, 128×2, and 128×2.

[0108] The present invention aims to extract and identify subtle feature information corresponding to massive LoRa IoT devices. Based on the fractional domain wavelet scattering network, the invention utilizes predefined wavelet kernel functions and nonlinear operations to complete the fusion of fractional domain wavelet transform and scattering network, thereby improving the efficiency of device RF fingerprint feature information extraction and classification network learning, while ensuring the interpretability of the entire network model.

[0109] The LoRa radio frequency fingerprint feature extraction system model based on fractional domain wavelet scattering network is as follows Figure 4 As shown in the figure, the entire system consists of three main components. The first is the LoRa transmitter, which primarily performs a series of signal processing steps, including I / Q modulation, upconversion, and power amplification of the LoRa baseband signal. During this process, the RF fingerprint characteristics of each module are embedded into the corresponding physical layer signal. The second component is the LoRa gateway. On the receiving side, synchronization and data segmentation are used to obtain the corresponding LoRa signal preamble. The preamble consists of 8 symbols and the data length is 8192. After obtaining the corresponding physical layer signal, alignment and normalization are required to prevent the subsequent network model from learning the power instead of the corresponding RF characteristics. The third component is the LoRa RF feature extraction and classification network based on the fractional domain wavelet scattering network. The time-frequency feature map is fed into the corresponding scattering network to calculate different signal overview and detail features, thereby forming a complete feature vector. The classification network then deeply mines and learns the feature vectors of different devices to accurately identify individual devices based on the feature vectors. During the calculation of the feature vector, the corresponding overview feature signals and detail feature signals can be fully obtained. Therefore, compared with traditional machine learning models, this model has better interpretability.

[0110] The above section introduces the basic theory of extracting RF fingerprint vital signs information using a hybrid fractional domain wavelet scattering network. In order to fully evaluate the performance of RF fingerprint feature extraction based on the fractional domain wavelet scattering network, Figure 4 As shown in the figure, physical layer air interface data is collected from 25 different LoRa devices. Eight complete preamble sequences of length 8192 are obtained for each frame through data synchronization and signal segmentation. Each preamble consists of a standard Upchirp signal. After passing through the various RF modules of the device, nonlinear characteristics will be embedded in the signal waveform. Therefore, the receiving side can obtain the corresponding RF fingerprint feature information of the device from the received physical layer signal. The configuration information of all LoRa devices is as follows:

[0111] Table 1 LoRa device related parameter settings

[0112] parameter set up Carrier frequency 915MHz Expansion Factor SF 7 Working bandwidth BW 125kHz Number of preamble symbols 8 Symbol period 2SF / BW Sampling rate 1MHz Transmit power 20dBm Bitrate 4 / 5

[0113] Table 1 shows the operating parameters of the Pycom LoRa device. The corresponding LoRa signal receiver is the Ettus B210 USRP. During the signal transmission process, each signal block is transmitted for 20 seconds, totaling 10 transmission blocks. According to the parameters shown in Table 1, each preamble symbol contains 1024 discrete values. In order to accurately determine the starting position of the symbol in the received signal, the ideal preamble sequence can be used to calculate the correlation value with the received signal to obtain the corresponding peak point to determine the signal starting position. The correlation calculation process can be expressed as

[0114]

[0115] where y ideal [n] represents the ideal preamble sequence, y start The starting position of the preamble symbol can accurately determine the symbol cutting position. According to the parameter setting table, the length of each symbol is 1024, and the sum of all lengths of the 8 preamble codes is 8192 valid discrete signal values. Then, the short-time Fourier transform (STFT) can be used to obtain a time-frequency diagram of each symbol with a size of 200×200. In order to comprehensively evaluate the RF fingerprint feature extraction model based on the fractional domain wavelet scattering network, data collection and experimental evaluation were carried out indoors and outdoors. Each device completes the collection of 6000 valid samples in different environments, and then all samples are divided into two parts, the test set and the training set, at a ratio of 1:5 for subsequent related calculation processes. During the training and testing process of the model, in order to obtain better RF feature representation coefficients, the rotation angle of the fractional domain wavelet is set to (5π / 20,π / 2), and the corresponding scattering coefficients are calculated using the fractional domain wavelet scattering network to form the device RF feature vector, which is then used as the input of the classification network for in-depth mining of RF feature information. As Figure 5As shown, different network models exhibit varying device recognition accuracy performance with different training set sizes. DFSNet achieves approximately 50.1% device recognition accuracy in indoor environments with a training set size of 100, representing a 42.51% performance improvement over the baseline network model. As the training dataset size increases, DFSNet's performance continues to improve, maintaining its advantage over both baseline models, ultimately reaching approximately 96.7% recognition accuracy. In outdoor scenarios, DFSNet achieves approximately 55.2% recognition accuracy with a training set size of 100, exceeding the recognition rate in indoor scenarios. This is primarily due to the greater presence of Loss of Sight (LOS) channel components in outdoor propagation environments, while the increased presence of multipath reflections in indoor transmission environments significantly degrades RF fingerprint information. When the training dataset size increases to 5,000, DFSNet achieves approximately 98.5% device recognition accuracy. Furthermore, because DFSNet better represents the eigenvalues of non-stationary signals, the proposed network model achieves superior device recognition performance in both environments. Secondly, DFSNet can achieve better recognition accuracy than other network models in the case of small sample training sets, so it is easier to be deployed in actual application scenarios.

[0116] like Figure 6 The figure shows the device recognition confusion diagram corresponding to the DFSNet network model under different training set sizes. Figure (a) shows the confusion diagram of the recognition results of device samples in the test set when the training sample is 1000. It can be seen from the figure that all devices are misjudged when the number of training samples is small. The main reason is that the data volume is insufficient at this time, and the classification network cannot fully learn the distinguishing boundaries of the RF feature vector in the high-dimensional space.

[0117] The LoRa non-stationary RF fingerprint feature extraction method and network model described in this invention can automatically complete the acquisition, feature extraction, and device classification of physical layer signals from relevant LoRa-related devices. While providing complete interpretability of RF fingerprint features, this invention significantly reduces the size of feature vectors, improves the learning efficiency of the network model, and effectively solves the problem of RF fingerprint feature extraction from non-stationary signals. Furthermore, all algorithms and processing steps involved in this invention can be integrated into a general-purpose ARM / FPGA hardware platform, laying the foundation for practical industrial applications.

[0118] In another embodiment of the present invention, a LoRa non-stationary radio frequency fingerprint feature extraction system is provided, which can be used to implement the above-mentioned LoRa non-stationary radio frequency fingerprint feature extraction method. Specifically, the system includes:

[0119] The data acquisition module is used to collect physical layer signals from different types of LoRa devices and pre-process the collected one-dimensional signals to obtain corresponding non-stationary signals;

[0120] The extraction module is used to build a radio frequency fingerprint feature extraction model for LoRa non-stationary signals based on the fractional domain wavelet scattering network and extract the radio frequency fingerprint feature coefficients;

[0121] The recognition and classification module is used to input the radio frequency fingerprint feature coefficients into the corresponding classification network and use the convolutional neural network to achieve accurate recognition and classification of feature vectors.

[0122] The module division in the embodiments of the present invention is illustrative and represents only one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional modules in various embodiments of the present invention may be integrated into a single processor, exist physically as separate modules, or two or more modules may be integrated into a single module. The integrated modules may be implemented in either hardware or software functional modules.

[0123] In another embodiment of the present invention, a computer device is provided, comprising a processor and a memory, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc., which are the computing core and control core of the terminal, and are suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in a computer storage medium to implement a corresponding method flow or corresponding function; the processor described in the embodiment of the present invention can be used for the operation of a LoRa non-stationary radio frequency fingerprint feature extraction method.

[0124] In another embodiment of the present invention, the present invention further provides a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device for storing programs and data. It is understandable that the computer-readable storage medium here can include both built-in storage media in the computer device and, of course, extended storage media supported by the computer device. The computer-readable storage medium provides a storage space that stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space. These instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the LoRa non-stationary radio frequency fingerprint feature extraction method in the above embodiment.

[0125] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0126] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0127] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0128] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A LoRa non-stationary radio frequency fingerprint feature extraction method, characterized in that: include: The physical layer signals of different LoRa devices are collected, and the collected one-dimensional physical layer signals are preprocessed to obtain the corresponding non-stationary signals; A radio frequency fingerprint feature extraction model for LoRa non-stationary signals is constructed based on the fractional domain wavelet scattering network to extract the radio frequency fingerprint feature coefficients; The RF fingerprint feature coefficients are input into the corresponding classification network, and the convolutional neural network is used to achieve accurate recognition and classification of the feature vectors; Short-time Fourier transform is used to analyze the non-stationary characteristics of the signal; different windows and lengths are used to realize the time-frequency transformation of the signal, as shown below Where y[m] represents the input signal, and STFT is short-time Fourier transform; A radio frequency fingerprint feature extraction model for LoRa non-stationary signals is constructed based on the fractional domain wavelet scattering network: The fractional domain Fourier transform is the basic transformation in DFSNet. The transformation process is as follows Where d represents the dimension of the signal f(t), Meet the following conditions Among them, α i represents the rotation angle of the fractional Fourier transform, For the corresponding transform coefficient, based on the fractional domain Fourier transform, the corresponding basis function is replaced by the wavelet kernel function to obtain the fractional domain wavelet transform Among them, the wavelet kernel function ψ α,λ,t (κ) satisfies the following conditions Where λ and t represent the scale and time factors in the corresponding fractional domain wavelet transform respectively. When α = π / 2, the above formula will become a traditional wavelet transform.

2. A LoRa non-stationary radio frequency fingerprint feature extraction method according to claim 1, characterized in that, Collect physical layer signals from large-scale LoRa devices: The operating bandwidth of the LoRa device is represented by B, and the corresponding minimum operating frequency is The maximum operating frequency is The corresponding LoRa signal instantaneous frequency is expressed as: f(t)=f L +kt(0≤t≤T) in, Represents the frequency change speed of the signal, T represents each symbol period; the instantaneous phase of the signal is obtained by integrating the instantaneous frequency of the signal: The basic Upchirp baseband signal is obtained by orthogonal modulation of the signal: Among them, A represents the signal amplitude; LoRa signal realizes SF-bit information modulation by setting different degrees of frequency offset of the basic signal; SF represents the expansion factor of the signal, which is 2 for LoRa signal. SF Different frequency offsets are used to complete the corresponding information modulation; the final modulated signal is further expressed as Where Δf represents the signal frequency offset, and x(t) represents the complete modulated signal. After passing through different modules, the signal is sent to the wireless space through the antenna. On the receiving side, the signal is expressed as y(t) = H*G(x(t)) + n Here, H represents the channel impulse response, G represents the device RF characteristic information, and n represents the Gaussian noise introduced during the transmission process.

3. a kind of LoRa non-stationary radio frequency fingerprint feature extraction method according to claim 2, is characterized in that, The signal y(t) on the receiving side contains the radio frequency fingerprint feature information of the wireless channel, noise and the device. During the LoRa signal transmission process, the preamble code is composed of a fixed basic Upchirp signal. The preamble code signal at the receiving end is obtained through signal synchronization and symbol segmentation, and the unique radio frequency feature information of the device is further extracted from the signal.

4. A LoRa non-stationary radio frequency fingerprint feature extraction method according to claim 1, characterized in that, The characteristic coefficients of the input time-frequency feature map are calculated using the fractional domain wavelet scattering network. The basic structure of the fractional domain wavelet scattering network is consistent with the traditional convolutional neural network. Both are multi-layer linear connections, and nonlinear changes and operations are performed at different network nodes. The calculation results of the fractional domain wavelet scattering network are divided into two categories: general features and subtle feature information. The general signal corresponding to the input time-frequency feature is expressed as Where J represents the scale factor, Represents the calculated overview feature information; the corresponding detail feature information is expressed as Where K represents different wavelet basis rotation angles, further simplifying the traditional layered convolution calculation process, as shown below in, represents the kernel function of the low-frequency filter characteristics, Represents a wavelet kernel function with bandpass filtering characteristics.

5. A LoRa non-stationary radio frequency fingerprint feature extraction method according to claim 1, characterized in that, The RF fingerprint feature coefficient is input into the corresponding classification network: the input signal f(t) is convolved with different wavelet kernel functions in the first layer of the network to obtain the corresponding scattering coefficient, i.e., the RF fingerprint feature information; the calculation process corresponding to the second layer of the scattering network is to continue to decompose and calculate the coefficients calculated in the first layer. The nonlinear operation process of the fractional domain wavelet scattering network consists of two nonlinear operations: the fractional domain wavelet filter and the module value. The coefficients of each layer in the DFSNet network model are composed of S α [l (m) ]f(t) is calculated, and the fractional domain wavelet scattering network consists of three layers, where the output of each layer is expressed as The information nodes transmitted on the network transmission path are represented as where l (m) represents the mth scattering network transmission path, and shows the calculation results related to the decomposed path.

6. A LoRa non-stationary radio frequency fingerprint feature extraction method according to claim 1, characterized in that, After calculating the corresponding scattering coefficients, they are combined into a unified feature vector as the input of the classification network. The classification network is mainly composed of a one-dimensional residual network, which contains two basic modules: ID and Conv modules. For the residual network body, a filter of size 7×1 is first used to perform preliminary convolution calculations after the input layer, followed by a maximum pooling layer and 7 different ID and Conv modules to perform in-depth information extraction and classification of the input signal. At the same time, a one-dimensional AlexNet network is introduced for comparison with a traditional multi-layer perceptron network, where the multi-layer perceptron network is composed of three convolutional neural networks of sizes 128×1, 128×2, and 128×2.

7. A LoRa non-stationary radio frequency fingerprint feature extraction device, characterized in that: include: The data acquisition module is used to collect physical layer signals from different types of LoRa devices and pre-process the collected one-dimensional signals to obtain corresponding non-stationary signals; The extraction module is used to build a radio frequency fingerprint feature extraction model for LoRa non-stationary signals based on the fractional domain wavelet scattering network and extract the radio frequency fingerprint feature coefficients; The recognition and classification module is used to input the radio frequency fingerprint feature coefficients into the corresponding classification network and use the convolutional neural network to achieve accurate recognition and classification of feature vectors.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the LoRa non-stationary radio frequency fingerprint feature extraction method as described in any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of a LoRa non-stationary radio frequency fingerprint feature extraction method as described in any one of claims 1 to 6 are implemented.

Images