Adaptive consistency detection method and system for radio frequency fingerprint open set identification

Abstract

The present application relates to the technical field of radio frequency fingerprint identification, and particularly relates to an adaptive consistency detection method and system for radio frequency fingerprint open set identification. Radio frequency baseband signals of a wireless device to be tested are collected, I / Q signal segments are divided, signal quality vectors are extracted, corresponding conditional bins are determined, corresponding consistency thresholds, session unknown thresholds and expert reliability weights are obtained; an expert calling sequence is generated according to the expert reliability weights, the average reasoning delay of the expert model and the current computing resource state, and a plurality of radio frequency fingerprint expert models are called in turn; early stop judgment is performed, a plurality of segment-level identification results within a device session window are fused, and a device-level open set identification result is obtained; and the conditional threshold table and the expert reliability weight are updated. The present application can adaptively adjust the judgment conditions according to different wireless environments, reduce the number of expert model calls, and improve the stability and real-time performance of radio frequency fingerprint open set identification.

Description

Technical Field

[0001] This invention relates to the field of radio frequency fingerprint recognition technology, and more specifically to an adaptive consistency detection method and system for open set recognition of radio frequency fingerprints. Background Technology

[0002] Radio frequency (RF) fingerprinting is a technology that utilizes the non-ideal hardware characteristics formed in power amplifiers, oscillators, mixers, filters, antennas, and other RF link devices of wireless transmitting equipment to identify and authenticate different wireless devices. Compared with authentication methods based on protocol fields, address identifiers, or keys, RF fingerprinting originates from differences in the physical layer hardware of devices. It is difficult to forge, can be passively eavesdropped on, and is applicable to non-cooperative scenarios. Therefore, it has significant application value in areas such as spectrum regulation, IoT access authentication, wireless intrusion detection, and abnormal device discovery.

[0003] With the development of deep learning technology, models such as convolutional neural networks, recurrent neural networks, Transformer networks, and complex neural networks have been increasingly used in radio frequency fingerprinting tasks. These methods can directly learn discriminative representations from raw I / Q signals, time-frequency plots, spectral features, or statistical features, achieving high recognition accuracy under closed-set conditions where the training and testing device categories are completely identical. However, real-world wireless environments are open and dynamic. Spectrum monitoring systems or IoT gateways often receive signals from unknown devices that were not present during the training phase. If a known category is forcibly output using a closed-set classification method, unknown devices may be mistakenly identified as legitimate devices, leading to security risks.

[0004] Existing open-set RFID fingerprinting methods typically utilize model output confidence, feature distance, prototype distance, reconstruction error, or differences in outputs from multiple models to determine whether a signal belongs to an unknown device. Confidence-based methods rely on maximum classification probability or entropy; when an unknown device fragment is highly similar to the local hardware distortion features of a known device, the model may still output a high confidence level, leading to the unknown device fragment being incorrectly classified into a known category. Distance- or prototype-based methods rely on stable feature boundaries, but RFID signals are affected by signal-to-noise ratio, frequency offset, phase noise, multipath fading, and sampling synchronization errors, causing feature distributions to drift with environmental changes, making it difficult to maintain stable fixed boundaries over long periods.

[0005] Consistency determination using multiple expert models can identify unknown devices by leveraging the differences in decision boundaries between different models. However, if all expert models are invoked for each signal segment under test, it will result in high inference latency and energy consumption on edge gateways, embedded monitoring nodes, or low-power IoT access devices. Furthermore, existing methods often employ fixed consistency thresholds, lacking adaptability to different channel quality conditions; and most methods use a single signal segment as the decision object, making them susceptible to transient noise, synchronization deviations, or locally similar segments, leading to unstable device-level identification results. Therefore, an open-set RF fingerprinting method is needed that can adaptively determine the threshold in complex wireless environments, reduce the average number of expert calls, and fuse results from multiple segments. Summary of the Invention

[0006] To address the aforementioned shortcomings of existing technologies, this invention provides an adaptive consistency detection method and system for open set recognition of radio frequency fingerprints, which can effectively solve the problems of insufficient adaptability of fixed thresholds, high inference overhead of full-scale expert models, and insufficient stability of single-fragment decisions in existing technologies.

[0007] To achieve the above objectives, the present invention provides the following technical solution: This invention provides an adaptive consistency detection method and system for open-set recognition of radio frequency fingerprints, comprising at least: The radio frequency baseband signal of the wireless device under test is collected within a preset time window and divided into multiple I / Q signal segments according to a preset number of sampling points; The corresponding signal quality vector is calculated based on the I / Q signal segment, and the signal quality vector is normalized. The corresponding conditional binning is determined by the conditional mapping function. The signal quality vector is composed of signal-to-noise ratio, carrier frequency offset, error vector magnitude, and channel spread parameters. Retrieve the consistency threshold, session unknown threshold, and expert reliability weight from the condition threshold table; Based on the expert reliability weight, the average inference latency of the expert model and the current computing resource status, an expert call sequence is generated, and multiple radio frequency fingerprint expert models are called sequentially according to the expert call sequence. A category count vector is generated based on the output results of multiple radio frequency fingerprint expert models. A consistency early stop decision is then performed based on the highest count, the second highest count, and the number of remaining uncalled expert models to obtain fragment-level recognition results. Multiple segment-level identification results corresponding to the same wireless device under test are fused to obtain device-level open set identification results; The threshold table for updating conditions is based on manually confirmed labels or posterior labels, along with expert reliability weights.

[0008] The method for calculating the corresponding signal quality vector based on the I / Q signal segment is as follows: Calculate the received signal power; This indicates the carrier frequency offset; if there are intervals in the signal... If there is a repeating preamble or pilot, then calculate the carrier frequency offset; After synchronizing and equalizing the current segment, record For the received symbols after equalization, If the corresponding ideal constellation point or reference symbol is used, then:

[0009] Let the channel spreading parameter be denoted as , and let the . The discrete channel impulse response estimated from each segment is: , Indicates the delay tap number, then the channel energy center and channel spread parameters They are respectively: ; ; This forms the signal quality vector.

[0010] When normalizing the signal quality vector, the mean of the corresponding parameter is subtracted from each quality parameter in the signal quality vector, and then divided by the sum of the standard deviation of the corresponding parameter and a preset positive number to obtain the normalized signal quality parameter.

[0011] The specific process of determining the corresponding conditional bin by the conditional mapping function is as follows: obtain the normalized signal quality vector corresponding to the i-th I / Q signal segment, perform mapping calculation on the normalized signal quality vector through the conditional mapping function, and divide the signal segments with the same reception conditions into the corresponding conditional bins.

[0012] The technical solution provided by this invention has the following advantages compared with the known prior art: This invention has five key features: First, it configures different thresholds for different wireless environments by binning signal quality conditions, reducing misjudgments caused by fixed thresholds in low signal-to-noise ratio, strong frequency offset, or multipath channels. Second, it introduces early stopping decisions during expert model collaborative inference, avoiding the need to call all expert models for each signal segment, thereby reducing average inference latency and energy consumption. Third, it uses the number of remaining uncalled expert models to determine whether results can be output in advance, making the early stopping process clearly interpretable. Fourth, it reduces the impact of a single abnormal segment on the final device-level identification result through device session-level multi-segment fusion. Fifth, it improves the adaptive capability during long-term operation through an expert reliability update mechanism. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0014] Figure 1 This is a schematic diagram of the overall method of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0016] Example 1 (see Figure 1 An adaptive consistency detection method for open-set recognition of radio frequency fingerprints includes at least the following: The system acquires the radio frequency baseband signal emitted by the wireless device under test (DUT) using a receiver. This radio frequency baseband signal refers to the complex digital signal obtained after down-conversion by the receiver. The system divides this complex digital signal into multiple I / Q signal segments according to a fixed number of sampling points M, where I represents the in-phase component and Q represents the quadrature component. The i-th I / Q signal segment is represented as:

[0017] in, Indicates the i-th segment A complex number of sampling points, For in-phase components, Here, represents the orthogonal components, and j represents the imaginary unit. For the same wireless device under test, the system acquires N I / Q signal segments within a continuous time window. These N segments together constitute the device session window. Here, the device session window refers to a set of multiple signal segments continuously transmitted or received by the same device within a certain period of time, which is used to reduce the risk of misjudgment caused by noise affecting a single segment.

[0018] For the i-th I / Q signal segment, the system extracts a signal quality vector. The signal quality vector is a vector composed of multiple values ​​reflecting the current wireless channel state, used to indicate whether the segment is under high signal-to-noise ratio (SNR), low SNR, strong frequency offset, significant multipath propagation, or large synchronization deviation. The system first calculates the received signal power. It represents the average magnitude of the signal energy in the current segment, and is calculated as follows:

[0019] Noise power This represents the noise energy measured by the receiver within the current frequency band, and can be estimated from idle frequency bands, guard intervals, silent segments, or the residual sequence after synchronization. Signal-to-noise ratio (SNR) The formula representing the relationship between received signal power and noise power is:

[0020] in, It is a very small positive number, used to avoid the denominator being zero. If The larger the value, the stronger the relative noise of the signal segment, and the more stable the recognition result is usually; if... The smaller the value, the greater the impact of noise, and the more cautious the subsequent open set decision threshold should be.

[0021] Carrier frequency offset This represents the deviation between the actual carrier frequency of the received signal and the receiver's local carrier frequency. Carrier synchronization refers to the process by which the receiver estimates and corrects this frequency deviation. If there are intervals in the signal... If there is a repeating preamble or pilot, the carrier frequency offset can be estimated by the following formula:

[0022] in, Sampling rate, This indicates taking the complex phase angle. This represents a conjugate complex number. If the communication system has already output a carrier frequency offset estimate in the synchronization module, then the output value of that synchronization module is directly used.

[0023] Error vector magnitude Used to indicate the degree of deviation between the received symbol and the ideal symbol. After the system completes synchronization and equalization of the current segment, it records... For the received symbols after equalization, If the corresponding ideal constellation point or reference symbol is used, then:

[0024] in, The larger the value, the more dispersed the constellation points are, and the more the current segment may be affected by phase noise, nonlinear distortion, channel estimation error, or synchronization error.

[0025] Channel spread parameters Used to represent the degree of time delay spread caused by multipath channels. Channel estimation refers to the process by which a receiver estimates the wireless channel response based on pilot or preamble sequences. Let the first... The discrete channel impulse response estimated from each segment is: ,in Indicates the delay tap number, then the channel energy center and channel spread parameters They are respectively:

[0026]

[0027] in, The larger the value, the more pronounced the multipath delay diffusion, and the more easily the RF fingerprint characteristics of the current segment change.

[0028] The system combines the above parameters into a signal quality vector: .

[0029] In practical implementation, the signal quality vector can also incorporate received signal strength, phase offset, spectral occupancy, or sampling synchronization deviation. Since these parameters have different dimensions—for example, signal-to-noise ratio is in dB, carrier frequency offset is in Hz, and error vector amplitude is a proportional value—the system needs to be normalized first. Normalization refers to converting parameters with different dimensions to a comparable numerical scale; the calculation formula is:

[0030] in, Indicates the first The d-th quality parameter in the fragment, and These are the mean and standard deviation of the quality parameter in the training or validation set, respectively. The normalized signal quality vector is denoted as... .

[0031] The system determines conditional binning based on the normalized signal quality vector. Conditional binning refers to grouping signal segments with similar reception conditions into the same group; for example, high SNR low-frequency bias segments are grouped into one group, and low SNR high-frequency bias segments are grouped into another. The system uses a conditional mapping function. Will Mapped to conditional bin number:

[0032] in, Indicates the first Conditional binning is defined for each segment. Conditional binning can be determined in three ways: first, equal interval partitioning, which divides the range of a quality parameter into several equal intervals; second, quantile partitioning, which sets boundaries according to the sample distribution in the validation set, making the number of samples in each interval as close as possible; and third, cluster partitioning, which calculates several cluster centers using the quality vectors in the validation set. Then assign the current fragment to the nearest cluster center:

[0033] in, This represents the Euclidean distance. Through conditional binning, the system can set different thresholds for different reception conditions, instead of using the same fixed threshold for all signal segments.

[0034] During the training phase, the system trains multiple RFID expert models using a known device training set. The known device training set refers to a collection of I / Q signal samples with real device identification tags. An expert model is a classification model capable of determining the device category of I / Q signal segments, such as a one-dimensional convolutional network, residual network, recurrent neural network, Transformer network, or complex neural network. For the first... For each segment, the z-th expert model outputs the predicted probability for each known device category:

[0035] in, Given the number of known equipment categories, This indicates that the segment belongs to the first... The probability of each known device category. The predicted category output by the expert model is:

[0036] in, This indicates the category number with the highest probability. Multiple expert models can be trained using different random initializations, different training data orders, different training subsets, or different network structures. The goal is to allow different expert models to form different judgment boundaries, thereby judging unknown devices through their consistency.

[0037] During the verification phase, the system establishes a condition threshold table. This table stores thresholds according to condition bins, with each condition bin (b) corresponding to at least two thresholds: a fragment-level consistency threshold and a session-unknown threshold. Among them, the consistency threshold Used to determine if a segment has enough expert models supporting the same known category; Session unknown threshold This is used to determine if the proportion of unknown segments in a device session is too high. The system prepares verification samples for known devices in the verification set. and unknown device agent samples Unknown device proxy samples refer to samples that do not participate in the training of known device classifiers but are used to simulate unknown devices. They can come from reserved untrained device categories or be obtained by applying frequency shift, phase perturbation, multipath mixing, amplitude scaling or noise injection to known device segments.

[0038] For each conditional bin b, the system enumerates candidate threshold combinations. Three metrics were calculated on known device verification samples and unknown device proxy samples. The known recognition accuracy represents the proportion of known device samples that were correctly identified.

[0039] The unknown device detection rate represents the proportion of proxy samples that are identified as unknown devices.

[0040] Average number of expert calls indicates the average number of expert models the system needs to call to output a result:

[0041] in, Indicates the combination of thresholds The next system for the first The output of each sample UNK represents the true category of known device samples, while UNK represents the result for unknown devices. Indicates the first The number of expert models actually invoked per sample This is an indicator function; it takes the value 1 if the condition is true, and 0 otherwise. Among the candidate threshold combinations that meet the preset requirements for known recognition accuracy and unknown device detection rate, the system selects the combination with the minimum average number of expert calls as the threshold for that conditional binning.

[0042] And satisfy:

[0043] in, Given the lower limit of the known recognition accuracy, This represents the lower limit of the detection rate for unknown devices; both are determined by application safety requirements or validation set acceptance criteria.

[0044] The expert reliability parameter is used to represent the reliability of a certain expert model under a certain conditional binning. For the first... An expert model in conditional binning Record the number of correct judgments in the historical rulings. The number of incorrect judgments was The system uses a smoothing method to calculate expert reliability weights:

[0045] in, Indicates the first Reliability weights of an expert model under conditional binning b. and To smooth out parameters, the reliability weights are set to avoid being too high or too low when the sample size is small. The order in which expert models are invoked is determined by both the reliability weights and the average inference latency. Let the average inference latency of the z-th expert model be... The sorted score is:

[0046] in, The larger the value, the more suitable the expert model is to be prioritized under the current binning conditions. The system follows... Generate an expert call sequence from largest to smallest.

[0047] During the inference phase, the system calls expert models one by one according to the expert call sequence. The system initializes the category count vector. Each time an expert model is invoked, the predicted category output by that expert is added to the count. After the first expert model, the... The count of the known categories is:

[0048] If weighted counting is used, then the first The weighted counts of the known categories are:

[0049] The system calculates the highest count in real time. Second highest count and the number of remaining unused expert models :

[0050] in, The category corresponding to the current highest count. This represents the total number of expert models. This represents the number of expert models that have been invoked.

[0051] When the highest count has reached the consistency threshold, and the leading number of the highest count relative to the second highest count is greater than the number of remaining uncalled expert models, the system outputs the known device category corresponding to the highest count. The determination condition is:

[0052] The meaning of this condition is: even if the remaining All expert models support the second-highest category, and the second-highest category cannot exceed the current highest category, so the known device categories can be output in advance.

[0053] When the sum of the highest count and the number of remaining unused expert models is still less than the consistency threshold, the system outputs the result for unknown devices, with the following judgment condition:

[0054] This condition means that even if all remaining expert models support the current highest category, that category cannot reach the consistency threshold, therefore the current fragment should not be forced into any known device category. If neither of the above two conditions is met, the system continues to call the next expert model.

[0055] During device-level fusion, the system aggregates N segment-level decision results for the same wireless device under test within a preset time window. Let the decision result of the i-th segment be denoted as... If the fragment is determined to be from an unknown device, then Unknown segment ratio for:

[0056] like If the system outputs the result for an unknown device, then the result will be... The conditional binning corresponding to this session can be determined by the mode or average signal quality vector of multiple segment conditional bins. If The system then calculates the aggregation consistency score for each known device category:

[0057] in, For the first The consistency score for each segment can be taken as the proportion of the highest category count for that segment to the number of experts invoked. System output. Largest known device category:

[0058] By using the above multi-segment fusion, the system will not rely solely on a single segment for the final judgment, thereby reducing the impact of short-term noise, synchronization deviation, or individual abnormal segments on the device-level recognition results.

[0059] The conditional threshold table can be updated based on feedback samples. Feedback samples refer to manually confirmed results, posterior tags, or trusted service results obtained after system operation. If an unknown device is missed or a known device is falsely rejected under a certain conditional bin, the system will add the new feedback samples to the statistical set of that conditional bin, recalculate the known identification accuracy, unknown device detection rate, and average number of expert calls for that bin, and update the consistency threshold and session unknown threshold according to the aforementioned threshold selection rules. The expert reliability weight is also updated based on the number of new correct decisions and the number of incorrect decisions, thereby enabling the system to adapt to changes in receiver location, equipment aging, channel environment, and service scenarios.

[0060] This embodiment also provides an adaptive consistency detection system. The system includes a signal acquisition module, a quality binning module, a threshold query module, an expert scheduling module, an early stop decision module, a session fusion module, and an experience update module. The signal acquisition module acquires RF baseband signals and segments them into I / Q signal segments; the quality binning module calculates the signal quality vector and determines the conditional binning; the threshold query module reads the consistency threshold, session unknown threshold, and expert reliability weight corresponding to the current conditional binning; the expert scheduling module determines the expert call sequence based on the expert reliability weight and inference delay; the early stop decision module outputs segment-level results based on the highest count, the second highest count, and the number of remaining uncalled expert models; the session fusion module fuses multiple segment-level results and outputs device-level identification results; and the experience update module updates the conditional threshold table and expert reliability weights based on feedback samples.

[0061] This invention does not limit the specific neural network structure of the expert model, nor does it limit the signal segment length, the number of conditional bins, the threshold search index, or the feedback update cycle. Those skilled in the art can adjust these parameters according to specific communication standards, receiver bandwidth, sampling rate, hardware deployment, and security strategies. Any scheme that employs signal quality conditional binning, sequential expert invocation, early stop decision, and multi-segment device-level fusion to achieve open-set RF fingerprinting should fall within the protection scope of this invention.

[0062] Furthermore, if the aforementioned function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0063] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0064] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), fiber optic devices, and portable optical disc read-only memory (CDROM). Additionally, computer-readable media can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0065] Furthermore, in order to provide a concise description of exemplary embodiments, not all features of actual embodiments (i.e., those features that are not relevant to the best mode of carrying out the invention as currently considered, or those features that are not relevant to implementing the invention) may be omitted.

[0066] It should be understood that numerous specific implementation decisions can be made during the development of any practical implementation, such as in any engineering or design project. Such development efforts may be complex and time-consuming, but for those skilled in the art who benefit from this disclosure, the development effort will be a routine work of design, manufacturing, and production without requiring much experimentation.

[0067] The above embodiments are only used to illustrate the technical solutions of the present invention; those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions described in the foregoing embodiments without causing the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of the present invention.

Claims

1. An adaptive consistency detection method for open-set recognition of radio frequency fingerprints, characterized in that, include: The radio frequency baseband signal of the wireless device under test is collected within a preset time window and divided into multiple I / Q signal segments according to a preset number of sampling points; The corresponding signal quality vector is calculated based on the I / Q signal segment, and the signal quality vector is normalized. The corresponding conditional binning is determined by the conditional mapping function. The signal quality vector is composed of signal-to-noise ratio, carrier frequency offset, error vector magnitude, and channel spread parameters. Retrieve the consistency threshold, session unknown threshold, and expert reliability weight from the condition threshold table; Based on the expert reliability weight, the average inference latency of the expert model and the current computing resource status, an expert call sequence is generated, and multiple radio frequency fingerprint expert models are called sequentially according to the expert call sequence. A category count vector is generated based on the output results of multiple radio frequency fingerprint expert models. A consistency early stop decision is then performed based on the highest count, the second highest count, and the number of remaining uncalled expert models to obtain fragment-level recognition results. Multiple segment-level identification results corresponding to the same wireless device under test are fused to obtain device-level open set identification results; The threshold table for updating conditions is based on manually confirmed labels or posterior labels, along with expert reliability weights.

2. The method according to claim 1, characterized in that, The I / Q signal segment is obtained by combining the in-phase and quadrature components in a complex form, specifically as follows: Obtain the nth sampling point in the i-th I / Q signal segment, take the corresponding in-phase component as the real part and the corresponding quadrature component as the imaginary part, and perform complex number synthesis through the imaginary unit to obtain the complex sampling point in the i-th I / Q signal segment.

3. The method according to claim 1, characterized in that, The method for calculating the corresponding signal quality vector based on the I / Q signal segment is as follows: Calculate the received signal power : ； Indicates noise power. Indicates the signal-to-noise ratio. This represents the nth complex sampling point in the i-th segment; This indicates the carrier frequency offset; if there are intervals in the signal... If there is a repeating preamble or pilot, the carrier frequency will shift: ； in, Sampling rate, This indicates taking the complex phase angle. Let M represent the conjugate complex number, M represent the fixed sampling point, and n represent the index of the complex sampling point. The magnitude of the error vector is denoted as , which is the value after synchronization and equalization of the current segment. For the received symbols after equalization, If the corresponding ideal constellation point or reference symbol is used, then: ； Let represent the channel spreading parameters, and let the discrete channel impulse response estimated for the i-th segment be . , Indicates the delay tap number, then the channel energy center and channel spread parameters They are respectively: ； ； Composition of signal quality vector .

4. The method according to claim 1, characterized in that, The specific process of determining the corresponding conditional bin by the conditional mapping function is as follows: obtain the normalized signal quality vector corresponding to the i-th I / Q signal segment, perform mapping calculation on the normalized signal quality vector through the conditional mapping function, and divide the signal segments with the same reception conditions into the corresponding conditional bins.

5. The method according to claim 4, characterized in that, For each condition bin Enumerate candidate threshold combinations The known recognition accuracy is calculated on known device verification samples and unknown device proxy samples. ; Unknown device detection rate ; Average number of expert calls ; in, As a consistency threshold, For unknown thresholds in the session, Indicates the combination of thresholds The next system for the first The output results for each sample UNK represents the true category of known device samples, while UNK represents the result for unknown devices. Indicates the first The number of expert models actually invoked per sample This is an indicator function that takes the value 1 if the condition is true and 0 otherwise. Validate samples for known devices. Sample for unknown devices; The combination with the lowest average number of expert calls is selected as the threshold for this conditional binning. ; And satisfy: ； in, Given the lower limit of the known recognition accuracy, This represents the lower limit of the detection rate for unknown devices.

6. The method according to claim 1, characterized in that, The method for obtaining expert reliability weights from the conditional threshold table based on the conditional binning is as follows: For the first... An expert model in conditional binning Record the number of correct judgments in the historical rulings. The number of incorrect judgments was The expert reliability weights are calculated using a smoothing method: ； in, This represents the reliability weight of the z-th expert model under conditional binning b. and For smoothing parameters.

7. The method according to claim 1, characterized in that, The process of obtaining the corresponding consistency threshold and session unknown threshold is as follows: A condition threshold table is established. This table stores thresholds according to condition bins, with each condition bin containing the threshold. Corresponding to at least two thresholds: Fragment-level consistency threshold and session unknown threshold: The consistency threshold is used to determine whether a fragment has enough expert models to support the same known category; the session unknown threshold is used to determine whether the proportion of unknown fragments in a device session is too high.

8. The method according to claim 1, characterized in that, The method of sequentially calling multiple RF fingerprint expert models is as follows: Let the average inference latency of the z-th expert model be... The sorted score is: ； Among them, according to Generate expert call sequences from largest to smallest. To prevent positive numbers from being zero; During the inference phase, the category count vector is initialized. Each time an expert model is called, the predicted category output by that expert is added to the count. After the first expert model, the... The count of the known categories is: ； This indicates the predicted category of the output. If a weighted count is used, then the [number]th [class]... The weighted counts of the known categories are: ； Calculate the highest count Second highest count And the number of remaining unused expert models, R: ； in, The category corresponding to the current highest count. This represents the total number of expert models. This represents the number of expert models that have been invoked. When the highest count has reached the consistency threshold, and the leading number of the highest count relative to the second highest count is greater than the number of remaining uncalled expert models, output the known device category corresponding to the highest count. The judgment condition is: ； When the sum of the highest count and the number of remaining uncalled expert models is still less than the consistency threshold, output the result for unknown devices. The judgment condition is: 。 9. The method according to claim 1, characterized in that, The method for fusing multiple segment-level identification results corresponding to the same wireless device under test to obtain device-level open set identification results is as follows: Record No. The judgment result of each segment is If the fragment is determined to be from an unknown device, then Unknown segment ratio for: ； like Output results for unknown devices, where Binning the conditions corresponding to this session, if Calculate the aggregation consistency score for each known device category. ; For the first Output the consistency score of each segment. Largest known device category .

10. An adaptive consistency detection system, applied to the method described in any one of claims 1-9, characterized in that, It includes a signal acquisition module, a quality binning module, a threshold query module, an expert scheduling module, an early stop decision module, a session fusion module, and an experience update module; the signal acquisition module acquires the radio frequency baseband signal and segments it into I / Q signal segments; The quality binning module calculates the signal quality vector and determines the conditional binning; the threshold query module reads the consistency threshold, session unknown threshold and expert reliability weight corresponding to the current conditional binning. The expert scheduling module determines the expert call sequence based on expert reliability weights and inference latency; The early stop decision module is used to output fragment-level results based on the highest count, the second highest count, and the number of remaining uncalled expert models; The session fusion module merges multiple segment-level results and outputs device-level recognition results; The experience update module is used to update the conditional threshold table and expert reliability weights based on feedback samples.

Images