A method and system for radio frequency fingerprint recognition of narrowband IoT transceivers

Abstract

This invention discloses a method and system for radio frequency fingerprinting of narrowband IoT transceivers. The method involves a receiver acquiring wireless messages sent by the transceiver to be identified, obtaining in-phase / orthogonal discrete sample sequences, locating and extracting preamble sample segments, performing a short-time Fourier transform on the preamble sample segments to generate a spectrogram energy matrix, and extracting compressed time-frequency features through singular value decomposition. The main component of the carrier frequency offset is obtained based on the cross-correlation peak position shift between rising and falling chirps, and the carrier frequency offset correction component is obtained based on the phase drift of adjacent rising chirps. The compressed time-frequency features and the carrier frequency offset estimation results are concatenated into a fused feature vector, which is then input into a convolutional neural network classification model to output the device category and category confidence. This invention reduces the dependence on channel location features and decreases model input overhead, making it suitable for online identification of narrowband IoT devices.

Description

Technical Field

[0001] This invention relates to the fields of IoT wireless communication security and radio frequency fingerprint recognition technology, specifically to a narrowband IoT transceiver radio frequency fingerprint recognition method and system. Background Technology

[0002] With the continuous deployment of IoT devices in scenarios such as industrial monitoring, smart buildings, energy management, agricultural production, and public facility management, a massive number of low-power terminals are accessing networks and interacting with data via wireless links. These terminals are typically characterized by small size, low power consumption, limited computing power, and dispersed deployment locations. The reliable verification of their access identities is crucial for business data collection, remote control command issuance, and network operation security. For large-scale narrowband IoT networks, relying solely on application-layer accounts, keys, or fixed numbers for identity management can easily lead to problems such as unauthorized node access, data source confusion, and increased security management costs when terminals are counterfeited, credentials are copied, or the communication environment changes.

[0003] Existing IoT device identification technologies largely rely on encryption authentication, protocol field verification, or fingerprint recognition based on wireless channel status. Encryption authentication methods place certain demands on terminal computing resources and energy consumption, resulting in high deployment costs in low-power narrowband devices. Channel-feature-based identification methods are easily affected by changes in device location, antenna orientation, obstruction environments, and deployment conditions. When the terminal moves or the network is redeployed, the original identification model often needs to re-collect samples and update fingerprint records. For narrowband communication devices, the available signal bandwidth and feature dimensions for identification are limited. Directly adopting the radio frequency identification approach from broadband communication systems can easily lead to problems such as insufficient feature stability, poor cross-scenario adaptability, and high model input overhead.

[0004] Therefore, in narrowband IoT communication scenarios, a technical solution is needed that can combine the signal characteristics of conventional wireless messages for device identification. This solution should leverage relatively stable differences in transmitter hardware to form a basis for device identification without increasing the additional authentication burden on terminals, and reduce dependence on changes in device deployment location and communication environment. Simultaneously, it needs to balance feature representation capabilities, model computational overhead, and online identification efficiency to meet the real-time access identification requirements of resource-constrained IoT gateways or edge receivers.

[0005] In view of this, the present invention provides a method and system for radio frequency fingerprint recognition of narrowband Internet of Things transceivers to solve the above problems. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for radio frequency fingerprint recognition of narrowband IoT transceivers, which solves the problems of insufficient recognition stability, high redeployment costs, and large online recognition overhead caused by the influence of changes in device location, channel environment, and terminal computing resources on existing narrowband IoT device identification methods.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] In a first aspect, the present invention provides a method for radio frequency fingerprint recognition of a narrowband Internet of Things transceiver, comprising the following steps:

[0009] The receiving end collects the wireless messages sent by the narrowband IoT transceiver to be identified, performs down-conversion, sampling and digitization processing on the wireless messages, and obtains a complex-form in-phase / orthogonal discrete sample sequence.

[0010] Packet start detection and preamble localization are performed on the in-phase / orthogonal discrete sample sequence. Preamble sample segments containing rising chirps and falling chirps are extracted, and amplitude normalization and time alignment are performed on the preamble sample segments. Short-time Fourier transform is performed on the preamble sample segments to generate a spectrogram energy matrix.

[0011] The rising chirp and falling chirp are selected in the preamble sample segment respectively. The sampling point offset is determined according to the offset of the cross-correlation peak position between the rising chirp and the falling chirp. The carrier frequency offset principal component is obtained according to the ratio between the sampling point offset and the number of sampling points corresponding to a single chirp, combined with the channel bandwidth.

[0012] Extract the complex sample vectors corresponding to adjacent rising chirps from the preamble sample fragment, obtain the average phase difference based on the inner product phase of the complex sample vectors corresponding to adjacent rising chirps, and obtain the carrier frequency offset correction component according to the correspondence between the average phase difference, the spreading factor and the channel bandwidth.

[0013] The carrier frequency offset principal component and the carrier frequency offset correction component are combined to obtain the carrier frequency offset estimation result;

[0014] The compressed time-frequency features are concatenated with the carrier frequency offset estimation results to form a fused feature vector; the fused feature vector is input into a trained convolutional neural network classification model to output the device category and category confidence of the narrowband IoT transceiver to be identified.

[0015] As a preferred embodiment of the present invention, the starting position of the candidate packet is determined based on the energy change of the received signal of the in-phase / orthogonal discrete sample sequence; and the starting position of the preamble in the wireless message is located using the starting position of the candidate packet as a reference.

[0016] Extract a preamble sample segment containing the complete preamble structure according to the preset preamble length; calculate the number of sampling points corresponding to a single chirp based on narrowband communication parameters, and perform boundary correction and time alignment on the preamble sample segment based on the number of sampling points.

[0017] As a preferred embodiment of the present invention, the number of sampling points corresponding to a single chirp is determined according to the receiver sampling rate, spreading factor, and channel bandwidth; the duration of a single chirp is determined according to the spreading factor and channel bandwidth, and the number of sampling points corresponding to a single chirp is determined according to the duration of the single chirp and the receiver sampling rate; the number of sampling points corresponding to a single chirp is used for boundary correction, chirp alignment, cross-correlation calculation, and phase difference estimation of preamble sample segments.

[0018] As a preferred embodiment of the present invention, generating the spectrogram energy matrix includes:

[0019] The preamble sample segment is segmented using a window function, and a short-time Fourier transform is performed on each frame signal to obtain the time-frequency representation result.

[0020] The time-frequency representation result is subjected to modulus squaring to obtain the spectrogram energy matrix, wherein the spectrogram energy matrix is ​​used to characterize the sweep trajectory, local texture, energy boundary and frequency drift features of the preamble chirp on the time axis and frequency axis.

[0021] As a preferred embodiment of the present invention, extracting compressed time-frequency features includes:

[0022] The spectrogram energy matrix is ​​decomposed into a left singular vector matrix, a singular value diagonal matrix, and the conjugate transpose of a right singular vector matrix;

[0023] A preset number of principal singular values ​​are selected in order of their magnitude, or the principal components corresponding to the principal singular values ​​are selected.

[0024] The selected principal singular values ​​or principal components are used as compressed time-frequency features to convert the spectrogram energy matrix into a low-dimensional feature sequence.

[0025] As a preferred embodiment of the present invention, rising chirp segments and falling chirp segments are selected from the preamble sample segments; the cross-correlation result between the rising chirp segments and the falling chirp segments is calculated; and the sampling point offset is determined based on the peak position in the cross-correlation result.

[0026] Based on the proportional relationship between the sampling point offset and the number of sampling points corresponding to a single chirp, the proportional relationship is mapped to the frequency range corresponding to the channel bandwidth, and the carrier frequency offset principal component is obtained by binary division based on the opposite frequency offset directions of rising and falling chirps.

[0027] As a preferred technical solution of the present invention, multiple complex sample vectors corresponding to adjacent rising chirps are extracted from the preamble sample segment; the complex vector inner product is calculated on the complex sample vectors corresponding to adjacent rising chirps to obtain multiple complex vector inner product results.

[0028] The phase of the multiple complex vector inner product results is calculated to obtain the average phase difference; according to the phase accumulation relationship of the average phase difference in a single chirp period, and combined with the spreading factor and channel bandwidth, the carrier frequency offset correction component is obtained; the carrier frequency offset principal component and the carrier frequency offset correction component are added to obtain the carrier frequency offset estimation result.

[0029] As a preferred embodiment of the present invention, the compressed time-frequency features are arranged into a one-dimensional feature sequence;

[0030] The carrier frequency offset estimation result is appended as a hardware frequency offset feature to the end of the one-dimensional feature sequence;

[0031] According to the preset vector length and preset feature order of the input layer of the convolutional neural network classification model, the appended feature sequence is arranged in a fixed order to form a fused feature vector that matches the input layer of the convolutional neural network classification model.

[0032] Wherein: the fused feature vector simultaneously includes the preamble time-frequency structure difference feature and the transmitter hardware frequency offset difference feature.

[0033] As a preferred embodiment of the present invention, the convolutional neural network classification model is trained in the following manner:

[0034] Collect wireless packets sent by multiple registered narrowband IoT transceivers on different collection dates and at different deployment locations;

[0035] For each wireless message, perform preamble sample fragment extraction, spectrogram energy matrix generation, singular value decomposition, carrier frequency offset principal component estimation, carrier frequency offset correction component estimation, and feature concatenation in sequence to obtain the corresponding fused feature vector.

[0036] The fused feature vectors obtained from the same registered narrowband IoT transceiver under different collection dates and deployment locations are configured as the same device label to form a labeled training set.

[0037] A convolutional neural network classification model is trained using the labeled training set, enabling the convolutional neural network classification model to establish a mapping relationship between fused feature vectors and device categories.

[0038] In a second aspect, the present invention provides a narrowband Internet of Things transceiver radio frequency fingerprint recognition system for performing the first aspect, including a receiving radio frequency front end, a processor and a memory, wherein the memory stores program instructions, and the processor executes the program instructions to form the following functional modules;

[0039] The signal acquisition module is used to receive wireless messages sent by the narrowband IoT transceiver to be identified, and to perform down-conversion, sampling and digitization processing on the wireless messages to generate a complex in-phase / orthogonal discrete sample sequence.

[0040] The preamble processing module, connected to the signal acquisition module, is used to perform packet start detection and preamble localization on the in-phase / orthogonal discrete sample sequence, extract preamble sample segments containing rising chirps and falling chirps, and perform amplitude normalization and time alignment on the preamble sample segments.

[0041] The spectrogram construction module, connected to the preamble processing module, is used to perform a short-time Fourier transform on the preamble sample fragment to generate a spectrogram energy matrix.

[0042] A compressed feature extraction module, connected to the spectrogram construction module, is used to perform singular value decomposition on the spectrogram energy matrix and extract compressed time-frequency features from the singular value decomposition results.

[0043] The frequency offset estimation module, connected to the preamble processing module, is used to obtain the main component of carrier frequency offset based on the cross-correlation peak position offset between rising chirps and falling chirps and the proportional relationship between it and the number of sampling points corresponding to a single chirp, and to obtain the carrier frequency offset correction component based on the inner product phase of the complex sample vectors corresponding to adjacent rising chirps, and to combine the main component of carrier frequency offset and the carrier frequency offset correction component to output the carrier frequency offset estimation result.

[0044] The feature fusion module is connected to the compressed feature extraction module and the frequency offset estimation module respectively, and is used to concatenate the compressed time-frequency features with the carrier frequency offset estimation result into a fused feature vector.

[0045] The model training module, connected to the feature fusion module, is used to train a convolutional neural network classification model based on the fused feature vectors and device labels corresponding to multiple registered narrowband IoT transceivers.

[0046] The online identification module is connected to the feature fusion module and the model training module respectively. It is used to input the fused feature vector corresponding to the narrowband IoT transceiver to be identified into the trained convolutional neural network classification model and output the device category and category confidence.

[0047] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0048] This invention extracts preamble samples from regular wireless messages of narrowband IoT transceivers and extracts spectrogram features and carrier frequency offset features based on the fixed signal structure of the preamble. This eliminates the primary reliance on channel features susceptible to location, obstruction, and channel environment factors for device identification. Simultaneously, by performing singular value decomposition on the spectrogram energy matrix, high-dimensional time-frequency image information is compressed into a low-dimensional feature sequence, which is then concatenated with the carrier frequency offset estimation result to form a fused feature vector. This allows the convolutional neural network classification model to simultaneously utilize differences in the preamble's time-frequency structure and transmitter hardware frequency offset differences for device classification. Therefore, this invention improves the stability of narrowband IoT transceiver identification in mobile, redeployment, and cross-date acquisition scenarios without requiring the terminal to send additional authentication messages. It also reduces the model's input dimensionality and computational overhead, making it suitable for online access identification at resource-constrained IoT receivers or gateways. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0050] Figure 1 This is a flowchart of the radio frequency fingerprint recognition method for narrowband IoT transceivers according to the present invention;

[0051] Figure 2 This is a schematic diagram of the convolutional neural network classification model structure of the present invention;

[0052] Figure 3 This is a schematic diagram of the radio frequency fingerprint recognition process of the narrowband IoT transceiver of the present invention;

[0053] Figure 4 This is a framework diagram of the narrowband IoT transceiver radio frequency fingerprint recognition system of the present invention. Detailed Implementation

[0054] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art. The drawings are merely illustrative of this disclosure and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted.

[0055] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more exemplary embodiments. Numerous specific details are provided in the following description to give a full understanding of exemplary embodiments of this disclosure. However, those skilled in the art will recognize that the technical solutions of this disclosure may be practiced with one or more specific details omitted, or methods, components, steps, etc. In other instances, well-known structures, methods, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this disclosure.

[0056] The narrowband IoT transceiver described in this invention can be LoRa, Zigbee, or other narrowband wireless communication devices; the receiver can be a software-defined radio receiver or a narrowband wireless receiver with down-conversion, sampling, and digitization capabilities. The following embodiments use LoRa messages as an example, but should not be construed as limiting the type of communication protocol.

[0057] Example 1

[0058] like Figure 1 As shown, this invention provides a method for radio frequency fingerprint identification of narrowband IoT transceivers. The method involves a receiver collecting regular wireless packets sent by the narrowband IoT transceiver to be identified, and using the time-frequency texture features and frequency offset features in the packet preamble to form a radio frequency fingerprint, thereby completing device category identification and legality determination. The method includes the following steps:

[0059] S101. Collect the wireless messages sent by the narrowband IoT transceiver to be identified, and obtain the in-phase / orthogonal discrete sample sequence.

[0060] The receiving end receives the wireless messages sent by the narrowband IoT transceiver to be identified, and performs down-conversion, sampling, and digitization processing on the received radio frequency signals in sequence to obtain a complex in-phase / orthogonal discrete sample sequence, denoted as x[n]. The in-phase / orthogonal discrete sample sequence retains the amplitude, phase, and frequency change information in the wireless message, serving as the data basis for subsequent preamble localization, spectrogram construction, and carrier frequency offset estimation.

[0061] In this embodiment, the receiving end does not require the narrowband IoT transceiver to be identified to send additional authentication messages, but instead uses its normal business communication messages for identification, so as to reduce terminal-side communication overhead and energy consumption.

[0062] S102. Perform packet initiation detection and preamble localization on the in-phase / orthogonal discrete sample sequence, and extract the preamble sample fragment.

[0063] After the receiver obtains the in-phase / orthogonal discrete sample sequence, it first performs packet start detection based on the energy change of the received signal and the periodic structure of the narrowband modulation signal to determine the candidate packet start position; then, using the candidate packet start position as a reference, it locates the preamble start position of the wireless message and extracts I / Q sample segments containing the complete preamble structure according to the preset preamble length.

[0064] In LoRa messages, the preamble consists of multiple rising and falling chirs. This embodiment can extract a preamble sample segment containing both rising and falling chirs, allowing subsequent spectrogram construction and carrier frequency offset estimation to be based on the signal portion with a fixed structure. The preamble exhibits structural stability compared to the payload data, reducing the impact of payload content differences on the recognition results.

[0065] After extracting the preamble sample fragment, the preamble sample fragment is normalized to reduce the impact of receive gain fluctuations on subsequent feature extraction; and the sample fragment is time-aligned according to the communication parameters so that different message samples participate in subsequent processing under a unified reference position.

[0066] When determining the chirp alignment parameters, the system calculates the duration of a single chirp and the number of sampling points corresponding to a single chirp based on the LoRa channel bandwidth, spreading factor, and receiver sampling rate. The specific formula is as follows:

[0067] ;

[0068] ;

[0069] Where: T is the duration of a single chirp, L_C is the number of sampling points corresponding to a single chirp, BW is the LoRa channel bandwidth, SF is the spreading factor, and F_s is the receiver sampling rate.

[0070] S103. Perform a short-time Fourier transform on the preamble sample segment to generate a spectrogram energy matrix.

[0071] The normalized and time-aligned preamble sample fragments are used as the objects for constructing the spectrogram. For discrete-time signals... The system uses window functions. The signal is divided into frames, and a short-time Fourier transform is performed on each frame to obtain the time-frequency representation:

[0072] ;

[0073] ;

[0074] in: Indicates the time window index. Represents angular frequency. The spectrogram energy matrix is ​​represented. The window function uses either a Hamming window or a Hanning window for framing, and the window length, overlap rate, and frequency resolution can be configured according to the sampling rate and bandwidth. The spectrogram energy matrix characterizes the sweep trajectory of the preamble chirp on the time and frequency axes, reflecting differences in local texture, energy boundaries, trajectory offset, and frequency drift between different transceivers. Since these differences originate from variations in the oscillator, modulation link, and frequency control during the transmitter hardware implementation, they can serve as the basis for the time-frequency characteristics of RF fingerprinting.

[0075] Since the LoRa preamble is composed of fixed-mode chirps, its spectrogram has a relatively stable structural basis in different messages, while hardware differences will manifest as texture distortion, trajectory offset and energy distribution differences, thus transforming the problem into a classification problem for time-frequency images.

[0076] S104. Perform singular value decomposition on the spectrogram energy matrix to extract compressed time-frequency features.

[0077] Since the spectrogram energy matrix is ​​a two-dimensional high-dimensional matrix, directly inputting it into a classification model would increase the model's input dimensionality and computational overhead. This embodiment performs singular value decomposition on the spectrogram energy matrix M:

[0078] ;

[0079] in: For the spectrogram matrix, It is a left singular vector matrix. It is a singular value diagonal matrix. It is the conjugate transpose of the right singular vector matrix.

[0080] The first k principal singular values ​​are selected in order of their magnitude, or the principal components corresponding to the principal singular values ​​are selected to form compressed time-frequency features. In this way, the main energy structure of the spectrogram is preserved, and the original high-dimensional spectrogram matrix is ​​converted into a low-dimensional feature sequence suitable for model input, thereby reducing the storage and computational burden of subsequent convolutional neural network classification models.

[0081] S105. Based on the preamble sample fragment, perform a coarse estimation of the carrier frequency offset to obtain the principal component of the carrier frequency offset.

[0082] The carrier frequency offset is coarsely estimated by utilizing the offset of the cross-correlation peak positions between the rising and falling chirps in the preamble. Specifically, the system selects the rising and falling chirp segments in the preamble, calculates the cross-correlation results between them, and determines the time offset based on the sampling point distance between the cross-correlation peaks.

[0083] Let Lp represent the sampling distance between the cross-correlation peaks of the rising and falling chirps in the preamble, and L_C represent the number of sampling points corresponding to a single chirp. Then the time offset can be expressed as:

[0084] ;

[0085] In the formula, the reason for multiplying by 1 / 2 is that this time offset is reflected on both the rising and falling chirps. Based on this time offset relationship, the principal component of the carrier frequency offset is obtained. :

[0086] ;

[0087] The principal component of the carrier frequency offset is used to determine the approximate range of the frequency offset and to provide a basis for subsequent fine estimation, thereby reducing the risk of ambiguous judgments in phase estimation and resolving the modulo 2π ambiguity problem that may occur in the fine estimation stage.

[0088] S106. Perform fine estimation of carrier frequency offset based on the phase drift of adjacent rising chirps to obtain carrier frequency offset correction components.

[0089] Based on the coarse estimation, the system further utilizes the phase drift between adjacent rising chirps to perform a fine estimation of the carrier frequency offset. The system extracts complex sample vectors corresponding to multiple adjacent rising chirps from the preamble and performs a complex vector inner product calculation on the complex sample vectors of adjacent rising chirps.

[0090] If a carrier frequency offset exists, a cumulative phase difference will occur between adjacent rising chirps. The cumulative phase difference and the carrier frequency offset satisfy the following relationship:

[0091] ;

[0092] ;

[0093] In implementation, a complex vector inner product is calculated for multiple adjacent rising chirps in the leader, and the average phase difference is taken as the estimation result:

[0094] ;

[0095] in: Indicates the first A complex sample vector of increasing chirp, The conjugate transpose of ci is represented; the carrier frequency offset correction component is obtained based on the average phase difference. :

[0096] ;

[0097] in: This represents the carrier frequency offset correction component. Indicates channel bandwidth. Indicates the spreading factor. Indicates the average phase difference;

[0098] The carrier frequency offset principal component and the carrier frequency offset correction component are combined to obtain the carrier frequency offset estimation result. The carrier frequency offset estimation results reflect the frequency deviation characteristics of the transceiver hardware oscillator and the radio frequency link, which are inherent hardware characteristics that are more stable than changes in channel position.

[0099] S107. The compressed time-frequency features and the carrier frequency offset estimation results are concatenated to form a fused feature vector. The compressed time-frequency features obtained in step S104 are arranged into a one-dimensional feature sequence, and the carrier frequency offset estimation results obtained in steps S105 to S106 are added to the end of the one-dimensional feature sequence as hardware frequency offset features to form a fused feature vector.

[0100] In practice, the first k singular values ​​or their corresponding principal components can be selected as compression features to reduce the input size while preserving the main structural information as much as possible.

[0101] ;

[0102] in: to These are the sorted principal singular values ​​or principal components. The final estimated carrier frequency offset is obtained by using the fused feature vector, which simultaneously contains the main time-frequency structure information in the preamble spectrogram and the transmitter hardware frequency offset information for subsequent device identification classification. This method preserves the main energy structure in the spectrogram, compressing each sample from the original high-dimensional time-frequency matrix into a low-dimensional vector, facilitating subsequent CNN training and deployment.

[0103] S108. Input the fused feature vector into the convolutional neural network classification model, and output the device category and category confidence.

[0104] like Figure 2 As shown, the convolutional neural network classification model uses a fused feature vector as input. The convolutional neural network classification model includes an input layer, a convolutional feature extraction layer, a pooling layer, a fully connected layer, and a softmax output layer.

[0105] In one implementation, the convolutional feature extraction layer includes a front-end single-scale one-dimensional convolutional structure and a rear-end multi-scale one-dimensional convolutional structure. The front-end single-scale one-dimensional convolutional structure is used to extract local continuous variation features in the fused feature vector. Each convolutional layer is followed by a batch normalization layer and a ReLU activation layer, and the feature dimension is reduced by a max pooling layer. The rear-end multi-scale one-dimensional convolutional structure uses one-dimensional convolutional kernels of different lengths to extract features in parallel, which is used to obtain short-term local textures, mid-scale frequency shift patterns, and structural differences over a longer range.

[0106] After convolutional feature extraction, the model feeds the output features into a fully connected layer and generates a probability distribution for each registered device category through a Softmax output layer. The system selects the category with the highest probability as the candidate device category and uses the highest probability as the category confidence score. During the training phase, the model calculates the transceiver classification loss based on the difference between the Softmax output and the device label, and updates the model parameters based on the classification loss. During the online recognition phase, the model outputs the device category and category confidence score of the transceiver to be identified based on the probability distribution.

[0107] When the category confidence level meets the preset judgment conditions, the system determines that the narrowband IoT transceiver to be identified belongs to the corresponding registered device; when the category confidence level does not meet the preset judgment conditions, the system determines that the narrowband IoT transceiver to be identified is an unknown device or a suspicious device, and triggers an alarm, isolation or further authentication process.

[0108] S109. During the offline training phase, the system collects wireless packets sent by multiple registered narrowband IoT transceivers on different dates and at different deployment locations, and extracts the fused feature vector corresponding to each wireless packet according to steps S101 to S107. The system combines the fused feature vector with the corresponding device label to form a labeled training set, and uses the labeled training set to train a convolutional neural network classification model to form a device fingerprint model.

[0109] During training, the classification loss can be calculated using the cross-entropy loss function, and the model parameters can be updated using gradient optimization, so that the fused feature vectors of the same transceiver under different collection dates and deployment locations are mapped to the same device category.

[0110] During the online identification phase, the receiving end samples the regular service messages sent by the device to be identified, extracts only the preamble portion, and repeatedly performs spectrogram construction, SVD compression, carrier frequency offset estimation, and feature concatenation to obtain the fused feature vector of the device under test. The system inputs the fused feature vector into a trained convolutional neural network classification model, outputs the device category and category confidence, and completes the identification based on this.

[0111] Example 2

[0112] like Figure 3 As shown, based on Embodiment 1, this embodiment provides a workflow for a narrowband IoT transceiver radio frequency fingerprint recognition method, including a conventional receiving and decoding path, an online detection path, and an offline feature extraction path.

[0113] The conventional receive-decode path is used by the LoRa receiver to perform the decoding process on the received message and output the decoded bits. This path corresponds to the conventional service communication process, enabling the receiver to provide collectable wireless message samples for RF fingerprint recognition while completing the service data decoding.

[0114] The online detection path is used to extract preamble samples from the wireless packets collected by the receiving end, calculate spectrogram features and carrier frequency offset features, and input the fused features into a convolutional neural network classification model to output the LoRa device identification result. This online detection path is set up in parallel with the conventional receive and decode path, without requiring changes to the service payload format of the LoRa packets or requiring the device to be identified to send additional authentication messages.

[0115] The offline feature extraction path is used to collect fingerprints from multiple registered transceivers and train a convolutional neural network classification model using the collected fused features. The trained convolutional neural network classification model is then configured in the online detection path for classifying the transceivers to be identified.

[0116] During the operation, the offline feature extraction path first collects wireless packets sent by multiple registered transceivers on different dates and at different deployment locations, extracts the spectrogram compression features and carrier frequency offset features corresponding to each wireless packet, and forms a fusion feature training set with device labels; then, the fusion feature training set is used to train a convolutional neural network classification model.

[0117] After receiving a regular service message from the transceiver to be identified, the online detection path extracts a preamble sample fragment from the message and generates a fused feature vector using the same processing method as the offline feature extraction path. This fused feature vector is then input into a trained convolutional neural network classification model to obtain the device category and category confidence score. Thus, the system can complete RF fingerprint identification of narrowband IoT transceivers without affecting the conventional LoRa receive and decode path.

[0118] Example 3

[0119] like Figure 4 As shown, based on Embodiment 2, this embodiment provides a narrowband IoT transceiver radio frequency fingerprint recognition system for implementing... Figure 3 The online detection path and offline feature extraction path shown are used in parallel with the conventional reception and decoding path of the LoRa receiver; including a signal acquisition module, a preamble processing module, a spectrogram construction module, a compressed feature extraction module, a frequency offset estimation module, a feature fusion module, a model training module, and an online recognition module.

[0120] The signal acquisition module is used to receive wireless messages sent by the narrowband IoT transceiver to be identified, and to perform down-conversion, sampling, and digitization processing on the wireless messages to generate a complex-form in-phase / orthogonal discrete sample sequence. The signal acquisition module can be implemented by a software-defined radio receiver or by a narrowband wireless receiver with equivalent sampling capability.

[0121] The preamble processing module is connected to the signal acquisition module and is used to perform packet initiation detection and preamble localization on in-phase / orthogonal discrete sample sequences, and to extract I / Q sample segments containing complete preamble structures. The preamble processing module is also used to perform amplitude normalization and time alignment on the extracted I / Q sample segments, and to calculate the number of sampling points corresponding to a single chirp based on bandwidth, spreading factor, and sampling rate, providing a parameter basis for subsequent chirp alignment, cross-correlation calculation, and phase difference estimation.

[0122] The spectrogram construction module is connected to the preamble processing module. It is used to frame the preamble sample segments after amplitude normalization and time alignment, and to perform a short-time Fourier transform on each frame to generate a time-frequency representation. The spectrogram construction module also performs modulus squaring on the time-frequency representation to obtain a spectrogram energy matrix. The spectrogram energy matrix is ​​used to characterize the frequency sweep trajectory, local texture, energy boundaries, and frequency drift features of the preamble chirp.

[0123] The compressed feature extraction module is connected to the spectrogram construction module and is used to perform singular value decomposition on the spectrogram energy matrix, and select principal singular values ​​or principal components from the singular value decomposition results to form compressed time-frequency features. The compressed time-frequency features are used to reduce the dimensionality of the model input while retaining the main structural information of the spectrogram.

[0124] The frequency offset estimation module is connected to the preamble processing module and is used to calculate the carrier frequency offset estimation result based on the preamble sample fragment. The frequency offset estimation module includes a coarse estimation unit and a fine estimation unit.

[0125] The coarse estimation unit is used to select the rising chirp segment and the falling chirp segment in the preamble, calculate the cross-correlation result between them, and obtain the main component of the carrier frequency offset based on the offset of the cross-correlation peak position.

[0126] The fine estimation unit is used to extract complex sample vectors corresponding to multiple adjacent rising chirps, perform complex vector inner product calculation on the complex sample vectors of adjacent rising chirps, and obtain the carrier frequency offset correction component based on the average phase difference of the multiple complex vector inner product results.

[0127] The frequency offset estimation module combines the main component of carrier frequency offset with the carrier frequency offset correction component and outputs the carrier frequency offset estimation result.

[0128] The feature fusion module is connected to both the compressed feature extraction module and the frequency offset estimation module. It arranges the compressed time-frequency features into a one-dimensional feature sequence and appends the carrier frequency offset estimation result to the end of the one-dimensional feature sequence to form a fused feature vector. This fused feature vector simultaneously characterizes the differences in the preamble time-frequency structure and the transmitter hardware frequency offset.

[0129] The model training module is connected to the feature fusion module and is used to receive fused feature vectors and device tags corresponding to multiple registered narrowband IoT transceivers during the offline training phase, and train a convolutional neural network classification model based on the fused feature vectors and device tags. The convolutional neural network classification model includes an input layer, a convolutional feature extraction layer, a pooling layer, a fully connected layer, and a softmax output layer; the convolutional feature extraction layer includes single-scale one-dimensional convolutional structures and multi-scale one-dimensional convolutional structures to extract radio frequency fingerprint features at different scales.

[0130] The online identification module is connected to both the feature fusion module and the model training module. During the online identification phase, it receives the fused feature vector corresponding to the device to be identified and inputs the fused feature vector into the trained convolutional neural network classification model, outputting the device category and category confidence score. The online identification module also generates a legitimate device determination result or a suspicious device determination result based on the category confidence score.

[0131] During system operation, Figure 3 The offline feature extraction path shown is jointly implemented by the signal acquisition module, the preamble processing module, the spectrogram construction module, the compressed feature extraction module, the frequency offset estimation module, the feature fusion module, and the model training module; Figure 3 The online detection path shown is jointly implemented by a signal acquisition module, a preamble processing module, a spectrogram construction module, a compressed feature extraction module, a frequency offset estimation module, a feature fusion module, and an online recognition module. Therefore, the convolutional neural network classification model trained offline is configured in the online recognition module, enabling the system to perform RF fingerprinting on the narrowband IoT transceiver without altering the conventional LoRa receiver decoding path.

[0132] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for radio frequency fingerprint recognition of a narrowband Internet of Things transceiver, characterized in that, Includes the following steps: The receiving end collects the wireless messages sent by the narrowband IoT transceiver to be identified, performs down-conversion, sampling and digitization processing on the wireless messages, and obtains a complex-form in-phase / orthogonal discrete sample sequence. Packet start detection and preamble localization are performed on the in-phase / orthogonal discrete sample sequence. Preamble sample segments containing rising chirps and falling chirps are extracted, and amplitude normalization and time alignment are performed on the preamble sample segments. Short-time Fourier transform is performed on the preamble sample segments to generate a spectrogram energy matrix. Singular value decomposition is performed on the spectrogram energy matrix to extract compressed time-frequency features. The rising chirp and falling chirp are selected from the preamble sample fragments respectively. The sampling point offset is determined according to the offset of the cross-correlation peak position between the rising chirp and the falling chirp. The carrier frequency offset principal component is obtained according to the ratio between the sampling point offset and the number of sampling points corresponding to a single chirp, combined with the channel bandwidth. Extract the complex sample vectors corresponding to adjacent rising chirps from the preamble sample fragment, obtain the average phase difference based on the inner product phase of the complex sample vectors corresponding to adjacent rising chirps, and obtain the carrier frequency offset correction component according to the correspondence between the average phase difference, the spreading factor and the channel bandwidth. The carrier frequency offset principal component and the carrier frequency offset correction component are combined to obtain the carrier frequency offset estimation result; The compressed time-frequency features are concatenated with the carrier frequency offset estimation results to form a fused feature vector; the fused feature vector is input into a trained convolutional neural network classification model to output the device category and category confidence of the narrowband IoT transceiver to be identified.

2. The narrowband IoT transceiver radio frequency fingerprint recognition method according to claim 1, characterized in that, The candidate packet start position is determined based on the received signal energy change of the in-phase / orthogonal discrete sample sequence; the start position of the preamble in the wireless message is located using the candidate packet start position as a reference. Extract a preamble sample segment containing the complete preamble structure according to the preset preamble length; calculate the number of sampling points corresponding to a single chirp based on narrowband communication parameters, and perform boundary correction and time alignment on the preamble sample segment based on the number of sampling points.

3. The narrowband IoT transceiver radio frequency fingerprint recognition method according to claim 2, characterized in that, The number of sampling points corresponding to a single chirp is determined according to the receiver sampling rate, spreading factor, and channel bandwidth; the duration of a single chirp is determined according to the spreading factor and channel bandwidth, and the number of sampling points corresponding to a single chirp is determined according to the duration of the single chirp and the receiver sampling rate; the number of sampling points corresponding to a single chirp is used for boundary correction, chirp alignment, cross-correlation calculation, and phase difference estimation of preamble sample segments.

4. The narrowband IoT transceiver radio frequency fingerprint recognition method according to claim 1, characterized in that, The generated spectrogram energy matrix includes: The preamble sample segment is framed using a window function, and a short-time Fourier transform is performed on each frame signal to obtain the time-frequency representation result. The time-frequency representation result is subjected to modulus squaring to obtain the spectrogram energy matrix, wherein the spectrogram energy matrix is ​​used to characterize the sweep trajectory, local texture, energy boundary and frequency drift features of the preamble chirp on the time axis and frequency axis.

5. The narrowband IoT transceiver radio frequency fingerprint recognition method according to claim 1, characterized in that, Extracting compressed time-frequency features includes: The spectrogram energy matrix is ​​decomposed into a left singular vector matrix, a singular value diagonal matrix, and the conjugate transpose of a right singular vector matrix; A preset number of principal singular values ​​are selected in order of their magnitude, or the principal components corresponding to the principal singular values ​​are selected. The selected principal singular values ​​or principal components are used as compressed time-frequency features to convert the spectrogram energy matrix into a low-dimensional feature sequence.

6. The narrowband IoT transceiver radio frequency fingerprint recognition method according to claim 1, characterized in that, Select rising chirp segments and falling chirp segments from the preamble sample segments; calculate the cross-correlation result between the rising chirp segments and the falling chirp segments; determine the sampling point offset based on the peak position in the cross-correlation result; Based on the proportional relationship between the sampling point offset and the number of sampling points corresponding to a single chirp, the proportional relationship is mapped to the frequency range corresponding to the channel bandwidth, and the carrier frequency offset principal component is obtained by binary division based on the opposite frequency offset directions of rising and falling chirps.

7. The narrowband IoT transceiver radio frequency fingerprint recognition method according to claim 6, characterized in that, Extract multiple complex sample vectors corresponding to adjacent rising chirps from the preamble sample fragment; perform complex vector inner product calculation on the complex sample vectors corresponding to adjacent rising chirps to obtain multiple complex vector inner product results; The phase of the multiple complex vector inner product results is calculated to obtain the average phase difference; according to the phase accumulation relationship of the average phase difference in a single chirp period, and combined with the spreading factor and channel bandwidth, the carrier frequency offset correction component is obtained; the carrier frequency offset principal component and the carrier frequency offset correction component are added to obtain the carrier frequency offset estimation result.

8. The narrowband IoT transceiver radio frequency fingerprint recognition method according to claim 1, characterized in that, The compressed time-frequency features are arranged into a one-dimensional feature sequence; The carrier frequency offset estimation result is appended as a hardware frequency offset feature to the end of the one-dimensional feature sequence; According to the preset vector length and preset feature order of the input layer of the convolutional neural network classification model, the appended feature sequence is arranged in a fixed order to form a fused feature vector that matches the input layer of the convolutional neural network classification model. Wherein: the fused feature vector simultaneously includes the preamble time-frequency structure difference feature and the transmitter hardware frequency offset difference feature.

9. The narrowband IoT transceiver radio frequency fingerprint recognition method according to claim 1, characterized in that, The convolutional neural network classification model is trained in the following way: Collect wireless packets sent by multiple registered narrowband IoT transceivers on different collection dates and at different deployment locations; For each wireless message, perform preamble sample fragment extraction, spectrogram energy matrix generation, singular value decomposition, carrier frequency offset principal component estimation, carrier frequency offset correction component estimation, and feature concatenation in sequence to obtain the corresponding fused feature vector. The fused feature vectors obtained from the same registered narrowband IoT transceiver under different collection dates and deployment locations are configured as the same device label to form a labeled training set. A convolutional neural network classification model is trained using the labeled training set, enabling the convolutional neural network classification model to establish a mapping relationship between fused feature vectors and device categories.

10. A narrowband IoT transceiver radio frequency fingerprint recognition system, used to execute the narrowband IoT transceiver radio frequency fingerprint recognition method according to any one of claims 1-9, characterized in that, It includes a receiving radio frequency front end, a processor, and a memory. The memory stores program instructions, and when the processor executes the program instructions, it forms the following functional modules. The signal acquisition module is used to receive wireless messages sent by the narrowband IoT transceiver to be identified, and to perform down-conversion, sampling and digitization processing on the wireless messages to generate a complex in-phase / orthogonal discrete sample sequence. The preamble processing module, connected to the signal acquisition module, is used to perform packet start detection and preamble localization on the in-phase / orthogonal discrete sample sequence, extract preamble sample segments containing rising chirps and falling chirps, and perform amplitude normalization and time alignment on the preamble sample segments. The spectrogram construction module, connected to the preamble processing module, is used to perform a short-time Fourier transform on the preamble sample fragment to generate a spectrogram energy matrix. A compressed feature extraction module, connected to the spectrogram construction module, is used to perform singular value decomposition on the spectrogram energy matrix and extract compressed time-frequency features from the singular value decomposition results. The frequency offset estimation module, connected to the preamble processing module, is used to obtain the main component of carrier frequency offset based on the cross-correlation peak position offset between rising chirps and falling chirps and the proportional relationship between it and the number of sampling points corresponding to a single chirp, and to obtain the carrier frequency offset correction component based on the inner product phase of the complex sample vectors corresponding to adjacent rising chirps, and to combine the main component of carrier frequency offset and the carrier frequency offset correction component to output the carrier frequency offset estimation result. The feature fusion module is connected to the compressed feature extraction module and the frequency offset estimation module respectively, and is used to concatenate the compressed time-frequency features with the carrier frequency offset estimation result into a fused feature vector. The model training module, connected to the feature fusion module, is used to train a convolutional neural network classification model based on the fused feature vectors and device labels corresponding to multiple registered narrowband IoT transceivers. The online identification module is connected to the feature fusion module and the model training module respectively. It is used to input the fused feature vector corresponding to the narrowband IoT transceiver to be identified into the trained convolutional neural network classification model and output the device category and category confidence.

Images