An artificial intelligence-based radio frequency fingerprinting method and system

Abstract

This invention relates to the field of door lock control and management technology, specifically disclosing an artificial intelligence-based radio frequency fingerprint recognition method and system. The method acquires environmental humidity data and idle baseband signals from an environmental humidity sensor, calculates the theoretical phase shift and gain loss of the medium under the current environment, and simultaneously acquires the radio frequency reflection signal and fingerprint image data when the recognition area is triggered. Multimodal data is input into a preset residual compensation neural network to output the finger surface humidity level. Combined with physical parameters, an iterative refractive index correction operation is performed on the radio frequency signal until the corrected radio frequency feature vector falls into a preset feature convergence interval, resulting in a converged radio frequency feature vector. Based on the humidity level, corresponding image and radio frequency feature extraction weights are dynamically assigned, and the matching degree between the fused verification vector and the pre-stored template is calculated to control the door lock. This effectively eliminates environmental medium interference, solves the problem of fingerprint recognition distortion in humid environments, and significantly improves the recognition accuracy and security of smart door locks.

Description

Technical Field

[0001] This invention belongs to the field of door lock control technology and relates to an artificial intelligence-based radio frequency fingerprint recognition method and system. Background Technology

[0002] As the core entry point of modern home security systems, smart locks directly impact users' life, property safety, and quality of life through the security and convenience of their authentication mechanisms. Fingerprint recognition technology, with its unique biometric characteristics and intuitive data acquisition process, has become the most widely used mainstream authentication method in the smart lock field. By extracting the unique ridge patterns and microstructure of the dermis from human fingerprints, it constructs a robust biometric defense, significantly enhancing the intelligence level of access control systems and improving user experience.

[0003] It is worth noting that the moisture content of the human finger surface and the ambient humidity are key variables affecting signal acquisition quality, and have a significant potential to interfere with the stability of the recognition system. Water molecules, as a polar medium with a high dielectric constant, not only fill the gaps between fingerprint ridges, causing optical imaging blurring, but also produce strong absorption attenuation and phase lag effects on the radio frequency detection signal. This change in medium properties caused by humidity fluctuations makes the physical transmission process of fingerprint signals extremely nonlinear and uncertain, easily leading to distortion in feature extraction, thus introducing significant noise interference at the source of the recognition chain.

[0004] Existing smart locks still have significant shortcomings in their detection and handling mechanisms for humidity interference. Current technologies mostly employ single image enhancement algorithms or fixed-parameter radio frequency compensation strategies, failing to quantify the specific phase shift and gain loss caused by environmental humidity on radio frequency signal transmission from a physical perspective. This results in a lack of accurate data support for the compensation process. Consequently, existing recognition systems often fail to recognize wet fingers or in high-humidity environments due to incomplete feature extraction or feature matching deviations, leading to recognition failures or delayed responses. They are unable to provide proactive signal correction and adaptive adjustment of multimodal feature weights when environmental interference occurs, resulting in insufficient robustness of smart locks under extreme conditions. Summary of the Invention

[0005] In view of the problems existing in the prior art, the present invention provides an artificial intelligence-based radio frequency fingerprint recognition method and system to solve the above-mentioned technical problems.

[0006] To achieve the above and other objectives, the technical solution adopted by the present invention is as follows: The first aspect of this invention provides an artificial intelligence-based radio frequency fingerprint recognition method applied to a smart door lock, the method comprising: Acquire ambient humidity data collected by an ambient humidity sensor and the no-load baseband signal under no-load conditions; based on the ambient humidity data and the no-load baseband signal, calculate the theoretical phase shift and gain loss of the medium under the current environment; Acquire the radio frequency reflection signal and fingerprint image data synchronously collected when the recognition area is triggered; The radio frequency reflection signal and fingerprint image data are input into a preset residual compensation neural network, and the corresponding finger surface humidity level data is output. Based on the phase shift and gain loss in dielectric theory, a refractive index correction operation is performed on the RF reflected signal to obtain the corrected RF eigenvector. Determine whether the corrected RF feature vector falls within the preset feature convergence interval; if it does not fall within the feature convergence interval, iteratively update the refractive index correction parameter and return to perform the refractive index correction operation on the RF reflected signal until the corrected RF feature vector falls within the feature convergence interval, thus obtaining the converged RF feature vector. Based on the surface humidity level data of the fingers, corresponding image feature extraction weights and radio frequency feature extraction weights are assigned; Fingerprint image feature vectors are extracted from fingerprint image data according to image feature extraction weights, and radio frequency dermal layer feature vectors are extracted from converged radio frequency feature vectors according to radio frequency feature extraction weights. The fingerprint image feature vector is fused with the radio frequency dermal layer feature vector to obtain the fused verification vector; Calculate the matching degree value between the fused verification vector and the pre-stored feature template; if the matching degree value is greater than the preset matching threshold, generate an unlocking control command and output it to the actuator of the smart door lock.

[0007] A second aspect of the present invention provides an artificial intelligence-based radio frequency fingerprint recognition system, the system comprising: Environmental parameter calculation module: acquires environmental humidity data collected by the environmental humidity sensor and the no-load baseband signal under no-load conditions; based on the environmental humidity data and the no-load baseband signal, calculates the theoretical phase shift and gain loss of the medium under the current environment; Finger humidity output module: acquires the radio frequency reflection signal and fingerprint image data synchronously collected when the recognition area is triggered; inputs the radio frequency reflection signal and fingerprint image data into a preset residual compensation neural network, and outputs the corresponding finger surface humidity level data; RF Vector Correction Module: Based on the phase shift and gain loss of the medium theory, the module performs a refractive index correction operation on the RF reflected signal to obtain a corrected RF feature vector; it determines whether the corrected RF feature vector falls into the preset feature convergence interval; if it does not fall into the feature convergence interval, it iteratively updates the refractive index correction parameters and returns to perform a refractive index correction operation on the RF reflected signal until the corrected RF feature vector falls into the feature convergence interval, thus obtaining a converged RF feature vector. Finger weight calculation module: Based on the finger surface humidity level data, it assigns corresponding image feature extraction weights and radio frequency feature extraction weights; it extracts fingerprint image feature vectors from fingerprint image data according to the image feature extraction weights, and at the same time extracts radio frequency dermal layer feature vectors from converged radio frequency feature vectors according to the radio frequency feature extraction weights. Door lock unlocking judgment module: The fingerprint image feature vector is fused with the radio frequency dermal layer feature vector to obtain the fused verification vector; the matching degree value between the fused verification vector and the pre-stored feature template is calculated; if the matching degree value is greater than the preset matching threshold, an unlocking control command is generated and output to the actuator of the smart door lock.

[0008] As described above, the radio frequency fingerprint recognition method and system based on artificial intelligence provided by the present invention has at least the following beneficial effects: This invention quantifies the theoretical phase shift and gain loss of the medium by collecting environmental humidity data and unloaded baseband signals, and then performs an iterative refractive index correction operation on the radio frequency reflected signal based on this data until the corrected radio frequency feature vector falls into a preset feature convergence interval. This closed-loop correction mechanism based on physical mechanisms effectively eliminates the nonlinear interference of environmental medium disturbances on signal transmission, overcomes the shortcomings of traditional signal processing in lacking environmental adaptability, and ensures the purity and stability of the feature extraction source data.

[0009] This invention utilizes residual compensation neural network to output finger surface humidity level data, and dynamically allocates image and radio frequency feature extraction weights accordingly, realizing differentiated extraction and fusion of fingerprint image feature vectors and radio frequency dermal layer feature vectors. This multimodal adaptive strategy accurately balances the contribution of optical and radio frequency signals under different humidity conditions, avoids the risk of single-modality failure in extreme environments, and greatly improves the robustness and anti-interference ability of biometric recognition systems. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram showing the connections between the steps of the method of the present invention.

[0012] Figure 2 This is a schematic diagram of the preset logic of the residual compensation neural network in this invention.

[0013] Figure 3 This is a schematic diagram showing the connections of the various modules in the system of the present invention. Detailed Implementation

[0014] The following description, in conjunction with the implementation of this invention, is merely an example and illustration of the concept of this invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the inventive concept or exceed the scope defined in these claims, all of which should fall within the protection scope of this invention.

[0015] In traditional smart door lock fingerprint recognition systems, fixed image processing parameters and static radio frequency gain configurations cannot adapt to the complex and ever-changing environmental humidity and finger surface moisture conditions. When a finger touches the recognition area, the change in refractive index caused by the surface water film leads to nonlinear distortion of the radio frequency signal. The system cannot establish a dynamic correlation between environmental physical parameters and the signal distortion model, making it difficult for the feature extraction stage to distinguish between effective dermal layer signals and water film interference noise. This static configuration mechanism severely reduces the reliability of fingerprint image and radio frequency signal fusion, resulting in feature vectors containing a large number of redundant environmental interference components, ultimately affecting the accuracy and stability of matching decisions.

[0016] For example, during the rainy season or in high-humidity environments, a layer of water of uneven thickness adheres to the surface of a user's finger, causing the fingerprint image ridges to blur and the amplitude of the radio frequency reflection signal to attenuate by more than 30%, resulting in a significant phase shift. In this situation, traditional systems still use preset standard gain and phase compensation parameters to process the signal, leading to the epidermal signal, interfered with by the water film refraction, being misidentified as the primary feature, while the attenuated effective signal from the dermis is submerged. The feature vector output by the fusion model loses key texture details characterizing a real biological fingerprint, the matching algorithm cannot extract effective feature points, the system frequently misidentifies and fails to recognize the fingerprint, and may even misidentify a wet finger as belonging to another user.

[0017] If the aforementioned issues are not addressed, signal distortion caused by environmental interference will significantly reduce the robustness of fingerprint recognition systems, leading to frequent rejections of legitimate users and severely impacting user experience. The lack of a physical layer correction mechanism for environmental humidity will also result in incomplete extraction of dermal biometric features, increasing the security risk of forgery attacks using fake silicone fingerprints or two-dimensional fingerprint photographs. Asynchronous mismatch between radio frequency signals and image data at the feature level will cause an imbalance in weight distribution during multimodal fusion, reducing the system's security defense level and ultimately creating a negative effect of both recognition obstacles and security vulnerabilities.

[0018] To address the aforementioned issues, this application first considers establishing a dynamic mapping mechanism between environmental medium characteristics and signal correction parameters. Traditional systems neglect the fundamental impact of unloaded baseband signals and environmental humidity on signal transmission, resulting in a lack of benchmark references for correction parameters. To resolve this, this application attempts to construct a physical-level signal benchmark correction model by calculating the theoretical phase shift and gain loss of the medium using environmental humidity data and unloaded baseband signals. Further analysis reveals that a single correction operation is insufficient to handle complex and variable humid conditions, necessitating an iterative correction strategy. This strategy optimizes the refractive index correction parameters by determining whether the corrected RF feature vector falls within the convergence interval. Simultaneously, the humidity level output by the residual compensation neural network is used to dynamically adjust the multimodal feature extraction weights, allowing the feature extraction strategy to adapt to the actual medium state, thereby resolving feature extraction distortion and matching failure issues caused by water film interference.

[0019] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0020] Example 1: Please refer to Figure 1-2 As shown, an artificial intelligence-based radio frequency fingerprint recognition method, applied to a smart door lock, specifically includes the following steps: Acquire ambient humidity data collected by an ambient humidity sensor and the no-load baseband signal under no-load conditions; based on the ambient humidity data and the no-load baseband signal, calculate the theoretical phase shift and gain loss of the medium under the current environment.

[0021] Preferably, calculating the theoretical phase shift and gain loss of the medium under the current environment specifically includes the following steps: The unloaded baseband signal is demodulated to extract the reference phase feature value and reference amplitude feature value of the unloaded baseband signal; According to the preset dielectric constant mapping table, find and extract the environmental equivalent dielectric constant value corresponding to the environmental humidity data; based on the environmental equivalent dielectric constant value, determine the dielectric phase shift coefficient and dielectric signal attenuation coefficient of the current environment. The theoretical phase shift of the medium is obtained by multiplying the reference phase characteristic value with the medium phase offset coefficient; the gain loss is obtained by multiplying the reference amplitude characteristic value with the medium signal attenuation coefficient.

[0022] In one specific embodiment, ambient humidity data is acquired through a capacitive humidity sensor integrated on the identification terminal, while the idle baseband signal in the no-load state is collected by the radio frequency transceiver circuit inside the smart lock. The specific acquisition process is as follows: The smart lock uses its integrated capacitive sensing sensor to monitor the dielectric load of the fingerprint collection window in real time. When the capacitive sensing sensor does not detect any change in capacitance due to human finger contact within a preset monitoring period (specifically, the change in capacitance is less than a preset contact threshold), and the smart lock is in standby, sleep, or idle state, the main control chip determines that the current state is unloaded and sends a calibration trigger command to the radio frequency (RF) front-end. Upon receiving the calibration trigger command, the voltage-controlled oscillator in the RF front-end generates a radio frequency detection signal at a specific frequency. This signal is radiated into the dielectric space where the fingerprint collection window is located via the transmitting antenna. Since there is no finger obstructing the window surface, the emitted electromagnetic wave signal mainly interacts with the air medium and the window protective cover, producing reflection, refraction, and scattering phenomena. The receiving antenna captures these echo signals after they have been affected by the environmental medium, forming an unloaded RF echo analog signal. The RF processing chip mixes the received unloaded RF echo analog signal with the local oscillator signal generated by the local oscillator, filtering out high-frequency carrier components and retaining low-frequency components containing environmental feature information. Subsequently, the low-frequency analog signal is converted into a digital signal by an analog-to-digital converter, thereby generating an unloaded baseband signal. The unloaded baseband signal specifically includes in-phase and quadrature components, and its data form is a complex sequence, where the real part represents the signal amplitude and the imaginary part represents the signal phase information.

[0023] Further demodulation processing is performed on the unloaded baseband signal. Specifically, the original RF waveform is decomposed into in-phase and quadrature components using a quadrature demodulator. The reference phase characteristic value of the unloaded baseband signal is extracted by calculating the arctangent function. At the same time, the reference amplitude characteristic value is extracted by the square root operation. The dimension of the reference phase characteristic value is radians, and the dimension of the reference amplitude characteristic value is millivolts.

[0024] Next, based on a pre-defined dielectric constant mapping table, the equivalent environmental dielectric constant value corresponding to the ambient humidity data is found and extracted. This mapping table is pre-calibrated based on the Debye relaxation model and records the influence of air and water vapor mixtures on electromagnetic wave propagation at different humidity percentages. Based on the equivalent environmental dielectric constant value, the dielectric phase shift coefficient and dielectric signal attenuation coefficient of the current environment are determined. Both coefficients are dimensionless correction factors used to quantify the wavelength shortening caused by dielectric polarization and the energy absorption caused by dielectric loss.

[0025] Based on this, the theoretical phase shift of the medium under the current environment can be calculated using the following formula: ; In the above formula, This is the phase shift quantity in medium theory, with dimensions in radians; The reference phase characteristic value is expressed in radians. The phase shift coefficient of the medium is a dimensionless coefficient.

[0026] Simultaneously, the gain loss is calculated using the following formula: ; In the above formula, This is the gain loss, measured in millivolts. The reference amplitude characteristic value is expressed in millivolts. is the dielectric signal attenuation coefficient, and is a dimensionless coefficient.

[0027] In this embodiment of the invention, the dielectric phase shift coefficient The preset logic is based on the rate of change of refractive index; for every 10% increase in ambient humidity, The preset step size is adjusted accordingly, for example, within a fine-tuning range of 0.001 to 0.005, to compensate for the slowdown in wave velocity. Medium signal attenuation coefficient. The preset logic is related to the resonant absorption cross section of water molecules, and its value is usually set between 0.950 and 0.999. The higher the humidity, the smaller the coefficient (meaning the lower the proportion of signal retained).

[0028] Acquire the radio frequency reflection signal and fingerprint image data synchronously collected when the recognition area is triggered; The radio frequency reflection signal and fingerprint image data are input into a preset residual compensation neural network, which outputs the corresponding finger surface humidity level data.

[0029] Preferably, the corresponding finger surface moisture level data is output, including: Spatial coordinate mapping and time synchronization processing are performed on fingerprint image data and radio frequency reflection signals to construct a multimodal fusion data matrix; The multimodal fusion data matrix is ​​input into the residual feature extraction layer of the residual compensation neural network to extract image blur residual features and radio frequency moisture absorption features, respectively. By utilizing the attention mechanism layer of the residual compensation neural network, the image blur residual features and radio frequency moisture absorption features are weighted and fused according to the preset feature weights to generate a comprehensive moisture characterization vector. The comprehensive moisture characterization vector is input into the classification output layer of the residual compensation neural network and mapped and compared with the preset humidity level range to determine and output the humidity level data of the finger surface.

[0030] Preferably, the image blur residual features and radio frequency moisture absorption features are extracted, including: The multimodal fusion data matrix is ​​processed by channel separation to independently acquire the image channel data components and the radio frequency channel data components; The image channel data components are edge detection convolution operation is performed on the image channel data components by using the preset image convolution kernel in the residual feature extraction layer to calculate the high-frequency missing values ​​of image texture, and the high-frequency missing values ​​of image texture are marked as image blur residual features. The spectral response of the RF channel data components is analyzed by using the preset RF analysis filter in the residual feature extraction layer. The amplitude attenuation value of a specific frequency band is extracted and marked as the RF moisture absorption feature.

[0031] In one specific embodiment, fingerprint image data in the form of a two-dimensional pixel array acquired by a fingerprint sensor, and radio frequency reflection signals in the form of a one-dimensional time series obtained by a radio frequency transceiver front end, are input into a preset residual compensation neural network. During the feedforward process of this network, spatial coordinate mapping and time synchronization processing are first performed. The pixel coordinate system of the fingerprint image is registered with the spatial gain distribution of the radio frequency antenna array through an affine transformation. Then, the image shutter exposure time is aligned at the millisecond level based on the trigger start time of the radio frequency pulse, thereby constructing a multimodal fusion data matrix. Subsequently, this matrix enters the residual feature extraction layer for channel separation, independently acquiring the image channel data components and the radio frequency channel data components.

[0032] For each image channel data component, an edge detection convolution operation is performed using a preset image convolution kernel in the residual feature extraction layer to identify the sharpness of ridge and valley edges in the image, and to calculate the high-frequency missing values ​​of image texture. The calculation formula is as follows: ; In the above formula, This represents the high-frequency missing value of the image texture; the closer the value is to 1, the more blurred the image is. The gradient magnitude of the current fingerprint image at coordinates (x, y) is extracted through convolution operation to represent the edge feature strength of the current image; The baseline gradient magnitude, defined as a pre-defined standard dry fingerprint image at the same coordinates, is derived from a pre-stored dry sample library within the system. This formula quantifies the loss of visual features due to moisture filling fingerprint valleys by comparing the high-frequency energy differences between the current image and a standard clear image, and labels this loss as an image blur residual feature.

[0033] Simultaneously, for the data components of the radio frequency channel, the system uses a preset radio frequency analysis filter to perform spectral response analysis in the frequency domain, extracting the amplitude attenuation value of a specific frequency band. The calculation formula is as follows: ; In the above formula, This represents the amplitude attenuation value for a specific frequency band, measured in watts (power unit). For a specific frequency band range under no-load or dry conditions The reference power spectral density within; This represents the real-time power spectral density of the currently acquired radio frequency reflection signal within this frequency band. and These are the preset lower and upper limits of the radio frequency monitoring band. In this embodiment, the 2.4GHz to 2.5GHz band is preferred because this band is in the non-polar resonance absorption range of water molecules and is extremely sensitive to humidity changes.

[0034] Further into the attention mechanism layer, the two types of residual features are weighted and fused according to preset feature assignments to generate a comprehensive moisture characterization vector. The fusion calculation formula is as follows: ; In the above formula, The scalar modulus of the comprehensive moisture characterization vector is a dimensionless coefficient. and These are preset image feature weights and radio frequency feature weights, respectively. Their preset logic is based on the dynamic adjustment of the signal-to-noise ratio of the features: when the ambient light is lower than the preset lumen threshold, the system automatically lowers the weights. And compensate proportionally to To ensure the reliability of recognition in low-light environments, the weights of both are in the range of [0,1] and their sum is 1. The total transmit power of the radio frequency signal is used to... Normalization is performed.

[0035] Finally, the comprehensive moisture characterization vector is input into the classification output layer, and a Softmax classifier is used to map and compare it with a preset humidity level range (e.g., set to level 1 to 5, corresponding to dry, normal, slightly damp, humid, and wet, respectively). For example, when... When the value falls within the range of [0.6, 0.8], it is determined to be a level 4 humid condition.

[0036] Preferably, the construction steps of the pre-defined residual compensation neural network include: Acquire historical fingerprint image data and historical radio frequency reflection signals collected when historical identification is triggered, and combine the two to form a historical multimodal sample set; Obtain the actual surface moisture measurement values ​​corresponding to the historical multimodal sample set, divide the actual surface moisture measurement values ​​into multiple continuous moisture span intervals, and set each moisture span interval as a standard humidity level label. An initial neural network structure is constructed, which includes a residual feature extraction layer, an attention mechanism layer, and a classification output layer. The network iteration parameters are initialized, including the residual convolution weight matrix, the attention bias vector, and the feature fusion allocation coefficient. The historical multimodal sample set is input into the residual feature extraction layer of the initial neural network structure, and the historical image blur residual features and historical radio frequency moisture absorption features are extracted through the residual convolution weight matrix. The blurred residual features of historical images and the moisture absorption features of historical radio frequency images are input into the attention mechanism layer, and weighted calculation is performed using feature fusion allocation coefficients to generate a predicted moisture characterization vector. The predicted moisture characterization vector is input into the classification output layer and superimposed with the attention bias vector to map the output predicted humidity level value. Calculate the classification residual error between the predicted humidity level value and the corresponding standard humidity level label, and at the same time calculate the feature overlap and divergence between the blurred residual features of historical images and the historical radio frequency moisture absorption features. The classification residual error and feature overlap divergence are added together according to a preset penalty weight to generate a comprehensive training loss value; The gradient descent algorithm is used to calculate the adjustment gradient of each parameter based on the comprehensive training loss value. The residual convolution weight matrix, attention bias vector and feature fusion allocation coefficient are updated and iterated by backpropagation until the comprehensive training loss value is less than the preset loss convergence threshold. The network iteration parameters at this time are saved to generate a residual compensation neural network.

[0037] In this embodiment, a historical multimodal sample set is acquired through a preset experimental environment when historical recognition is triggered. This sample set includes multiple sets of historical fingerprint image data and synchronously acquired historical radio frequency reflection signals. A high-precision contact moisture meter is used to obtain the actual surface moisture measurement value corresponding to each set of data. Dimensions are The sample is divided into multiple continuous moisture ranges according to a preset moisture gradient, thereby assigning a discrete standard humidity level label to each sample. The value range is usually an integer in the range [1, 5].

[0038] Next, an initial neural network structure is constructed, including a residual feature extraction layer, an attention mechanism layer, and a classification output layer, and the network iteration parameters are initialized, where the residual convolution weight matrix... The attention bias vector is determined using the Xavier initialization algorithm to extract nonlinear spatial features. Initialized as a zero vector, used to adjust feature weight offset; feature fusion allocation coefficients. Initialized to 0.5 to balance the contributions of image and radio frequency modes. Subsequently, the sample set is input into the residual feature extraction layer, where blurred residual features of historical images are extracted by performing cross-layer convolution operations on the image channels. Simultaneously, complex-domain convolution is performed on the radio frequency channel to extract historical radio frequency moisture absorption features. .

[0039] In the attention mechanism layer, the system utilizes feature fusion to assign coefficients. Perform feature reconstruction to generate predicted moisture representation vectors. The calculation formula is as follows: ; In the above formula, The feature fusion allocation coefficient has a value range of [0,1]. Its preset logic is determined based on the correlation statistics between image sharpness and radio frequency signal-to-noise ratio in historical data. If the effective feature ratio of the image under low light is less than 40%, it will be automatically lowered through preset inverse proportional logic. Values ​​are added to increase the decision weight of radio frequency characteristics; The function is used to map eigenvalues ​​to the (0,1) interval to eliminate dimensional differences.

[0040] The subsequent classification output layer will... Mapped to predicted humidity level values The system then proceeds to the loss calculation stage. To ensure that the model not only classifies accurately but also extracts complementary, non-redundant features, the system calculates the comprehensive training loss value using the following composite formula. : ; The comprehensive training loss value is a dimensionless pure number; the left-hand side of the equation is the traditional cross-entropy loss, used to characterize the classification residual error; the right-hand side is the feature overlap divergence term, where... The preset penalty weight is used to adjust the constraint strength of feature independence. It is recommended to set its value to an adjustable range of 0.1 to 0.4, preferably 0.25, to ensure that the differential information of the two modalities is extracted to the maximum extent without sacrificing classification accuracy. The fractional part uses the cosine similarity principle to calculate the correlation between the two types of features. The larger this term is (i.e. the smaller the similarity), the higher the feature divergence, which is more conducive to the robustness of the model under complex conditions.

[0041] Finally, the Adam gradient descent algorithm is used based on... right as well as Perform backpropagation updates and iterations, calculating partial derivatives and adjusting parameter weights layer by layer using the chain rule, until... If the fluctuation amplitude of N consecutive training cycles is less than the preset loss convergence threshold, then the optimal network iteration parameters are saved, and the construction of the residual compensation neural network is completed.

[0042] Based on the phase shift and gain loss in dielectric theory, a refractive index correction operation is performed on the RF reflected signal to obtain the corrected RF eigenvector. The specific operation steps are as follows: The original phase data component and the original amplitude data component are separated from the radio frequency reflected signal; The phase shift quantity of the medium theory is used to perform phase reverse offset processing on the original phase data components to eliminate the phase shift deviation caused by environmental medium disturbance and generate reference corrected phase data. Linear gain compensation is performed on the original amplitude data components using the gain loss to restore the signal energy attenuated by the absorption of the environmental medium and generate reference correction amplitude data. The reference correction phase data and reference correction amplitude data are mapped to a preset signal feature space, and a corrected radio frequency feature vector is generated through feature recombination operation.

[0043] In this embodiment, by invoking an orthogonal demodulation algorithm, the in-phase baseband component and the quadrature baseband component are separated from the original RF reflected signal, and the original phase data component and the original amplitude data component are extracted through coordinate system transformation. Specifically, for the sampled data at discrete time point t, the arctangent function is used... The original phase data components were calculated. The dimension is radians, and the modulus is calculated using the formula. The original amplitude data components were calculated. Its dimension is millivolt.

[0044] Phase offset processing is performed on the original phase data components using the phase shift quantity of medium theory to eliminate the additional signal flight time delay caused by environmental medium disturbances (such as water film refraction), generating reference-corrected phase data. The formula for calculating the offset is: ; In the above formula, The phase data is used as a reference for correction, and the dimension is radians; These are the original phase data components, with the dimension in radians; This is the phase shift quantity in medium theory, with dimensions in radians; is the preset phase compensation matching coefficient, which is a dimensionless coefficient. The adjustable range of 0.90 to 1.10 is used to fine-tune the aging frequency offset of the RF antenna under different service years; to ensure compatibility with both new and old equipment, a value of 1.0 is recommended at the factory. This formula is an inverse variation of the classical wave equation's phase superposition principle. The necessity of this variation lies in the fact that traditional uniform phase filtering often mistakenly misses subtle ridge-valley depth differences in the dermis, while using a phase shift incorporating specific media... The subtractive hedging method can precisely strip away the additional phase shift of the water film covering the epidermis, preserving the original phase difference of the fingerprint's three-dimensional topology to the greatest extent.

[0045] Subsequently, linear gain compensation is performed on the original amplitude data components using the gain loss to restore the signal energy attenuated by dielectric absorption when the radio frequency electromagnetic wave penetrates the water molecule layer, generating reference-corrected amplitude data. The linear gain compensation formula is as follows: ; The reference correction amplitude data is in millivolts. These are the original amplitude data components, with units of millivolts. This is the gain loss, measured in millivolts. The preset interface acoustic / electromagnetic coupling efficiency coefficient is a dimensionless pure number. The preset logic is based on the impedance matching of the pressing contact surface. When the internal pressure sensor detects a pressing force greater than 2 Newtons (i.e., tight contact), a value of 0.95 is recommended. This formula ensures that while increasing the attenuation signal (especially the weak dermal reflection valley value), it does not cause an exponential expansion of the base noise.

[0046] The acquired reference correction phase data Reference correction amplitude data Mapped to a preset signal feature space, a modified radio frequency feature vector is generated through feature recombination operations. The feature recombination formula is as follows: ; The output corrected radio frequency feature vector is a multidimensional dimensionless column vector; This is the maximum reference rated amplitude of the radio frequency transmitter, with the dimension in mV. It is used to divide the amplitude data by this rated value to achieve dimension unification. and Using the concept of Euler's formula, the one-dimensional phase angle (which exists) is expanded. The jump cutoff problem is transformed into a two-dimensional continuous trigonometric function mapping, both of which are dimensionless values; The preset orthogonal projection matrix has a size of , where D is the dimension of the target feature.

[0047] Determine whether the corrected RF feature vector falls within the preset feature convergence interval; if it does not fall within the feature convergence interval, iteratively update the refractive index correction parameter and return to perform the refractive index correction operation on the RF reflected signal until the corrected RF feature vector falls within the feature convergence interval, thus obtaining the converged RF feature vector.

[0048] Preferably, determining whether the modified radio frequency feature vector falls within a preset feature convergence interval and obtaining the converged radio frequency feature vector includes: Calculate the spatial geometric distance between the corrected radio frequency eigenvector and the center point of the eigenconvergence interval to obtain the eigenbia value; The feature deviation value is compared with the preset deviation threshold. If the feature deviation value is greater than the preset deviation threshold, the offset polarity data of the feature deviation value relative to the feature convergence interval is extracted. Based on the offset polarity data and characteristic deviation values, the refractive index increment adjustment value is calculated and then superimposed on the current refractive index correction parameter to complete the iterative update of the refractive index correction parameter. The phase reverse offset and linear gain compensation operation of the RF reflected signal is re-executed using the iteratively updated refractive index correction parameters until the recalculated characteristic deviation value is within the preset deviation threshold. The corrected RF characteristic vector that meets the condition at this time is marked as the converged RF characteristic vector.

[0049] Preferably, the phase reversal offsetting and linear gain compensation operations for the RF reflected signal are re-executed using the iteratively updated refractive index correction parameters until the recalculated characteristic deviation value is within a preset deviation threshold. The corrected RF characteristic vector that meets the conditions at this point is marked as the converged RF characteristic vector, including: Obtain the refractive index correction parameters after iterative updates, and decompose them to obtain the iterative phase offset coefficient and iterative gain compensation coefficient in the current period. The original phase data components are subjected to difference cancellation processing using iterative phase offset coefficients, and the original amplitude data components are subjected to product amplification processing using iterative gain compensation coefficients to generate the current iteratively corrected phase component and iteratively corrected amplitude component. A three-dimensional feature mapping is performed on the iteratively corrected phase component and the iteratively corrected amplitude component to generate the corrected radio frequency feature vector for the current period. Calculate the Euclidean space distance of the current corrected radio frequency feature vector relative to the feature convergence interval to obtain the recalculated feature deviation value; Compare the recalculated feature deviation value with the preset deviation threshold. If the recalculated feature deviation value is within the acceptable range determined by the preset deviation threshold, then the current corrected RF feature vector is determined to meet the convergence condition, and its output is marked as the converged RF feature vector.

[0050] In this embodiment, the smart lock's main control chip first calculates the spatial geometric distance between the corrected radio frequency feature vector and the center point of the feature convergence interval to obtain the feature deviation value. The calculation formula is as follows: In the formula is the characteristic deviation value, and is the dimensionless scalar distance; To correct the radio frequency feature vector, it is a dimensionless multidimensional column vector; The center point vector of the preset feature convergence interval is derived from the statistical mean vector of the standard dry state preset at the factory of the system. It is also a dimensionless feature; T represents the transpose symbol; the distance calculation formula is derived from the Euclidean space distance metric. Its necessity lies in its ability to comprehensively evaluate the overall offset in the multi-dimensional feature space, rather than the linear error of a single dimension. The feature deviation value is then compared with a preset deviation threshold, which is set based on three standard deviations of the intraclass variance of historical normal fingerprint features. If the feature deviation value is greater than the preset deviation threshold, it indicates that the current moisture compensation has not completely offset environmental interference. At this point, the offset polarity data of the feature deviation value relative to the feature convergence interval is extracted, that is, the positive and negative offset directions in each dimension are extracted using a sign function. Based on the offset polarity data and the feature deviation value, the refractive index increment adjustment value is calculated. The calculation formula is as follows: In the formula This is the refractive index increment adjustment value, which is dimensionless. The preset learning step size coefficient is set to a fixed constant between 0.01 and 0.05 to prevent oscillation; sgn represents the sign function, and its specific mathematical definition is as follows: x represents an unknown; Vector addition is performed to directly superimpose the calculated refractive index increment adjustment value onto the current refractive index correction parameter matrix or vector, thereby generating updated refractive index correction parameters to correct deviations in previous iterations. Based on a preset parameter structure mapping relationship, the updated refractive index correction parameters are decomposed or their components extracted, separating the parameters containing comprehensive correction information into independently controlled iterative phase offset coefficients and iterative gain compensation coefficients. Then, the correction operation is re-executed using the iteratively updated parameters, specifically using the iterative phase offset coefficients to perform difference cancellation processing on the original phase data components, as shown in the formula: ,in To iteratively correct the phase component, the dimension is radians. These are the original phase data components, measured in radians. This is the phase shift quantity in medium theory, with dimensions in radians. The iterative phase offset coefficient is a dimensionless coefficient; the iterative gain compensation coefficient is used simultaneously to perform product amplification on the original amplitude data components, as shown in the formula. ,in The amplitude component is iteratively corrected, with dimensions in millivolts. The original amplitude data components are measured in millivolts. The dimensionless iterative gain compensation coefficient is used because the fine-tuning in the later stages of iteration needs to keep the envelope ratio of the original waveform unchanged.

[0051] Then, a three-dimensional feature mapping is performed on the iteratively corrected phase component and the iteratively corrected amplitude component to generate the corrected RF feature vector for the current period. The Euclidean space distance of the current corrected RF feature vector relative to the feature convergence interval is calculated again to obtain the recalculated feature deviation value. The recalculated feature deviation value is compared with the preset deviation threshold. If the recalculated feature deviation value is within the qualified range determined by the preset deviation threshold, it is determined that the current corrected RF feature vector has eliminated all media interference and meets the convergence condition. Its output is marked as the converged RF feature vector.

[0052] Based on the surface humidity level data of the finger, corresponding image feature extraction weights and radio frequency feature extraction weights are assigned; fingerprint image feature vectors are extracted from fingerprint image data according to the image feature extraction weights, and radio frequency dermal layer feature vectors are extracted from converged radio frequency feature vectors according to the radio frequency feature extraction weights.

[0053] Preferably, fingerprint image feature vectors are extracted from fingerprint image data according to image feature extraction weights, and radio frequency dermal layer feature vectors are extracted from converged radio frequency feature vectors according to radio frequency feature extraction weights, including: The fingerprint image data is used to extract feature points using a preset image texture parsing algorithm to obtain an initial surface ridge feature vector; the initial surface ridge feature vector is then multiplied with the image feature extraction weights to adjust the representation ratio of the image features and generate a fingerprint image feature vector. The converged radio frequency feature vector is subjected to feature filtering operation using a preset frequency band isolation matrix to remove residual signal data in the epidermal layer and separate the initial dermal layer reflection feature vector. The initial dermal reflection feature vector is multiplied with the radio frequency feature extraction weights to amplify the representation ratio of dermal biological features and generate a radio frequency dermal feature vector.

[0054] It should be added that the preset process for the preset frequency band isolation matrix includes: A preset number of historical radiofrequency reflection frequency sequences are obtained, the amplitude reflectivity and phase delay value corresponding to each frequency point are extracted, and combined with the synchronously acquired skin layer depth reference data, a radiofrequency tissue layer corresponding sample set is established. Based on the sample set corresponding to the radio frequency tissue hierarchy, an initial frequency band isolation matrix with a multi-layer filter structure is constructed, and the iteration parameters of the initial frequency band isolation matrix are set. The iteration parameters include: frequency band selection weight component, inter-layer penetration attenuation coefficient, phase difference correction operator, and nonlinear tissue scattering deviation. The sample set corresponding to the radio frequency tissue layer is input into the initial frequency band isolation matrix. The signal intensity of different frequency bands is weighted and aggregated by selecting the frequency band weight component. The attenuation suppression of the reflected energy of the epidermal layer is performed by using the interlayer penetration attenuation coefficient, thereby separating the predicted dermal layer feature vector. Calculate the feature reconstruction residual between the predicted dermal feature vector and the skin layer depth reference data, and at the same time calculate the spatial independence metric of signals from different tissue layers at the output of the initial frequency band isolation matrix; By introducing a preset regularization ratio, the feature reconstruction residuals and spatial independence measures are weighted and fused to generate a comprehensive matrix loss function value. The second-order gradient optimization algorithm is applied to iteratively correct the frequency band selection weight component, interlayer penetration attenuation coefficient, phase difference correction operator, and nonlinear tissue scattering deviation based on the comprehensive matrix loss function value. In each iteration, the step size weight of the second-order gradient optimization algorithm is adjusted by comparing the slope of the change in the comprehensive matrix loss function value before and after the update, until the comprehensive matrix loss function value reaches the preset model convergence accuracy. The iteration parameters at this time are then saved to generate the preset frequency band isolation matrix.

[0055] In this embodiment, the main control chip first acquires the humidity level data of the finger surface. Input a preset allocation regulator, and allocate corresponding image feature extraction weights and radio frequency feature extraction weights through a linear inverse proportional function. The weight allocation calculation formula is as follows: At the same time satisfy In the formula Extract weights for image features. The radio frequency feature extraction weights are both dimensionless proportionality coefficients with values ​​ranging from [0,1]. The preset humidity sensitivity attenuation step size is usually set to an adjustable range of 0.15 to 0.25, preferably 0.20. The logic behind this step size setting is to ensure that when the humidity reaches the highest level 5, the image feature weight automatically drops to the bottom line of 0.2 to preserve the basic macroscopic contour, while the weight of radio frequency penetration features is increased to 0.8 to dominate the recognition process.

[0056] Fingerprint image feature vectors are extracted from fingerprint image data according to image feature extraction weights. Specifically, a preset image texture parsing algorithm is used to extract feature points from fingerprint image data. An initial surface ridge feature vector is obtained by calculating the local texture energy density in multiple directions. The initial surface ridge feature vector is then multiplied by the image feature extraction weights to dynamically adjust the representation ratio of image features in the final fusion decision, thereby generating the fingerprint image feature vector.

[0057] Simultaneously, a preset frequency band isolation matrix is ​​used to perform feature filtering on the converged radio frequency feature vector to remove residual signal data in the epidermis and separate the initial dermal reflection feature vector.

[0058] The fitting steps for the preset frequency band isolation matrix are as follows: First, under a constant temperature and humidity environment in the laboratory, a preset number of historical radio frequency reflection frequency sequences are acquired. The amplitude reflectivity and phase delay values ​​corresponding to each frequency point are extracted, and combined with high-precision skin layer depth reference data synchronously acquired by an optical coherence tomography (OCT) device, a radio frequency tissue layer corresponding sample set is established. Next, based on this radio frequency tissue layer corresponding sample set, an initial frequency band isolation matrix with a multi-layer filtering structure is constructed, and the iteration parameters of the initial frequency band isolation matrix are set. The iteration parameters include: a frequency band selection weight component characterizing the sensitivity of different frequency bands to deep tissues. Interlayer penetration attenuation coefficient following Beer-Lambert absorption law Phase difference correction operator for aligning phase shifts under different dielectric constants and nonlinear tissue scattering bias to compensate for microscopic inhomogeneity of tissue. All of the above parameters are set to dimensionless random numbers of a standard normal distribution during initialization.

[0059] During the training and convergence process of the matrix, the sample set corresponding to the radio frequency organization level is input into the initial frequency band isolation matrix, and its forward separation calculation formula is: ; In the above formula, The separated predicted dermal feature vector is a dimensionless column vector. The input historical radio frequency sample vector is dimensionless. The relative penetration depth is normalized based on layered depth reference data and is dimensionless. This formula is derived from a variation of the exponential decay model of electromagnetic waves propagating in layered lossy media. The advantage of this variation lies in the introduction of a nonlinear exponential term. It can force the matrix to learn the high-frequency rapid attenuation characteristics of the superficial epidermal signal, thereby naturally filtering out the reflected energy of the epidermal layer at the output.

[0060] Subsequently, the system calculates the comprehensive matrix loss function value using the following composite formula: ; In the above formula, The loss function value is a dimensionless matrix; the left-hand side represents the predicted dermal feature vector and the baseline true dermal feature vector. The mean square reconstruction residuals between them; the right-hand side is the spatial independence measure (i.e., cosine similarity) of signals from different organizational levels at the output. This refers to residual feature data of the epidermis; The preset regularization ratio is recommended to be set in the range of 0.1 to 0.3, preferably 0.15. The necessity of this variation of the loss formula lies in the fact that relying solely on error approximation can easily lead to the introduction of extremely strong shallow noise into the deep signal. Adding a spatial independence penalty term can orthogonalize the dermal and epidermal features, ensuring the complete removal of residual signals from the epidermal layer.

[0061] During the parameter optimization phase, the system applies a second-order gradient optimization algorithm to calculate the inverse approximation of the Hessian matrix based on the comprehensive matrix loss function value. This approximation is then used to iteratively correct the frequency band selection weight components, interlayer penetration attenuation coefficient, phase difference correction operator, and nonlinear scattering bias. In each iteration, the slope of the comprehensive matrix loss function value before and after the update is compared in real time. If the ratio of the current slope to the previous slope is greater than a preset smoothing threshold, it indicates that the gradient descent is too slow, and the system dynamically increases the step size weight of the second-order gradient optimization algorithm; conversely, it decreases the step size weight to prevent exceeding the global minimum. This iteration is repeated until the comprehensive matrix loss function value is less than the preset model convergence accuracy. Saving the iteration parameters at this point generates the preset frequency band isolation matrix.

[0062] The fingerprint image feature vector is fused with the radio frequency dermal layer feature vector to obtain a fused verification vector; the matching degree value between the fused verification vector and the pre-stored feature template is calculated; if the matching degree value is greater than the preset matching threshold, an unlocking control command is generated and output to the actuator of the smart door lock.

[0063] In this embodiment, the fusion calculation formula is: ; In the above formula, To reconstruct the fused verification vector of the output; This is the fingerprint image feature vector output after weighted processing; The first part is the dermal feature vector of the radiofrequency layer after weighted processing in the previous step; both are dimensionless vectors. The concatenation operation represents the concatenation of matrix dimensions, which is to concatenate two vectors into a single-dimensional, high-dimensional long vector; The preset cross-modal alignment weight matrix contains the association probabilities accumulated from the comparison of massive historical user features, and is a dimensionless coefficient matrix. is the preset fitting bias vector, dimensionless; tanh is the hyperbolic tangent activation function. This formula is a variation of the spatial mapping formula for the fully connected layer of a classic multilayer perceptron. The necessity and benefit of this variation are: when the finger surface experiences extreme high humidity, the extreme residuals of a single mode may cause abrupt changes in the local values ​​of the spliced ​​vector, exceeding the limit. Introducing the hyperbolic tangent activation function can strictly converge all cross-fused features and map them to a dimensionless range of negative one to positive one, effectively suppressing the distortion spikes caused by water film reflection in extreme environments. Subsequently, the system calculates the matching degree value between the fused verification vector and the pre-stored feature template. The pre-stored feature template is a dimensionless reference vector collected in a dry standard environment during the user's initial registration and also distributed and solidified in the local security chip through the same channel. The matching degree calculation formula is: ; In the above formula, The output matching score represents the authentication confidence of the current verifier's identity, and is a dimensionless pure number ranging from zero to one. This refers to the pre-stored feature template that is invoked. is the preset absolute deviation penalty coefficient, which is a dimensionless constant. The symbol represents the L2 norm of the orientation quantity; Finally, the calculated matching degree value is logically compared with the preset matching threshold. The preset matching threshold reflects the balance between the false rejection rate and the false acceptance rate; a range of 0.80 to 0.95 is recommended. If the comparison finds that the current matching degree value is greater than the preset matching threshold, the system determines that the user's identity is legitimate and authentication is successful. The underlying microcontroller then generates an unlocking control command containing a specific drive duty cycle and outputs it to the smart lock's actuator via a pulse width modulation signal bus. This triggers the drive motor to rotate and engages the mechanical clutch, ultimately completing the physical unlocking operation.

[0064] Example 2: Figure 3 As shown, an artificial intelligence-based radio frequency fingerprint recognition system includes an environmental parameter calculation module, a finger humidity output module, a radio frequency vector correction module, a finger weight calculation module, and a door lock unlocking judgment module. The various modules are connected via wired and / or wireless connections to enable data transmission between them; Environmental parameter calculation module: acquires environmental humidity data collected by the environmental humidity sensor and the no-load baseband signal under no-load conditions; based on the environmental humidity data and the no-load baseband signal, calculates the theoretical phase shift and gain loss of the medium under the current environment; Finger humidity output module: acquires the radio frequency reflection signal and fingerprint image data synchronously collected when the recognition area is triggered; inputs the radio frequency reflection signal and fingerprint image data into a preset residual compensation neural network, and outputs the corresponding finger surface humidity level data; RF Vector Correction Module: Based on the phase shift and gain loss of the medium theory, the module performs a refractive index correction operation on the RF reflected signal to obtain a corrected RF feature vector; it determines whether the corrected RF feature vector falls into the preset feature convergence interval; if it does not fall into the feature convergence interval, it iteratively updates the refractive index correction parameters and returns to perform a refractive index correction operation on the RF reflected signal until the corrected RF feature vector falls into the feature convergence interval, thus obtaining a converged RF feature vector. Finger weight calculation module: Based on the finger surface humidity level data, it assigns corresponding image feature extraction weights and radio frequency feature extraction weights; it extracts fingerprint image feature vectors from fingerprint image data according to the image feature extraction weights, and at the same time extracts radio frequency dermal layer feature vectors from converged radio frequency feature vectors according to the radio frequency feature extraction weights. Door lock unlocking judgment module: The fingerprint image feature vector is fused with the radio frequency dermal layer feature vector to obtain the fused verification vector; the matching degree value between the fused verification vector and the pre-stored feature template is calculated; if the matching degree value is greater than the preset matching threshold, an unlocking control command is generated and output to the actuator of the smart door lock.

[0065] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0066] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0067] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0068] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An artificial intelligence-based radio frequency fingerprinting method, characterized by, Applications in smart door locks include: Acquire ambient humidity data collected by an ambient humidity sensor and the no-load baseband signal under no-load conditions; based on the ambient humidity data and the no-load baseband signal, calculate the theoretical phase shift and gain loss of the medium under the current environment; Acquire the radio frequency reflection signal and fingerprint image data synchronously collected when the recognition area is triggered; input the radio frequency reflection signal and fingerprint image data into a preset residual compensation neural network, and output the corresponding finger surface humidity level data; Based on the phase shift and gain loss in dielectric theory, a refractive index correction operation is performed on the RF reflected signal to obtain the corrected RF eigenvector. Determine whether the corrected RF feature vector falls within the preset feature convergence interval; if it does not fall within the feature convergence interval, iteratively update the refractive index correction parameter and return to perform the refractive index correction operation on the RF reflected signal until the corrected RF feature vector falls within the feature convergence interval, thus obtaining the converged RF feature vector. Based on the surface humidity level data of the finger, corresponding image feature extraction weights and radio frequency feature extraction weights are assigned; fingerprint image feature vectors are extracted from the fingerprint image data according to the image feature extraction weights, and radio frequency dermal layer feature vectors are extracted from the converged radio frequency feature vectors according to the radio frequency feature extraction weights. The fingerprint image feature vector is fused with the radio frequency dermal layer feature vector to obtain a fused verification vector; the matching degree value between the fused verification vector and the pre-stored feature template is calculated; if the matching degree value is greater than the preset matching threshold, an unlocking control command is generated and output to the actuator of the smart door lock. 2.The method of claim 1, wherein, The calculation of the theoretical phase shift and gain loss of the medium under the current environment includes the following steps: The unloaded baseband signal is demodulated to extract the reference phase feature value and reference amplitude feature value of the unloaded baseband signal; Based on the preset dielectric constant mapping table, find and extract the environmental equivalent dielectric constant value corresponding to the environmental humidity data; Based on the equivalent dielectric constant of the environment, the dielectric phase shift coefficient and dielectric signal attenuation coefficient of the current environment are determined. The theoretical phase shift of the medium is obtained by multiplying the reference phase eigenvalue with the medium phase shift coefficient. The gain loss is obtained by multiplying the reference amplitude characteristic value with the dielectric signal attenuation coefficient. 3.The method of claim 1, wherein, Output the corresponding finger surface moisture level data, including: Spatial coordinate mapping and time synchronization processing are performed on fingerprint image data and radio frequency reflection signals to construct a multimodal fusion data matrix; The multimodal fusion data matrix is ​​input into the residual feature extraction layer of the residual compensation neural network to extract image blur residual features and radio frequency moisture absorption features, respectively. By utilizing the attention mechanism layer of the residual compensation neural network, the image blur residual features and radio frequency moisture absorption features are weighted and fused according to the preset feature weights to generate a comprehensive moisture characterization vector. The comprehensive moisture characterization vector is input into the classification output layer of the residual compensation neural network and mapped and compared with the preset humidity level range to determine and output the humidity level data of the finger surface.

4. The radio frequency fingerprint recognition method based on artificial intelligence according to claim 3, characterized in that, The pre-defined steps for constructing a residual compensation neural network include: Acquire historical fingerprint image data and historical radio frequency reflection signals collected when historical identification is triggered, and combine the two to form a historical multimodal sample set; Obtain the actual surface moisture measurement values ​​corresponding to the historical multimodal sample set, divide the actual surface moisture measurement values ​​into multiple continuous moisture span intervals, and set each moisture span interval as a standard humidity level label. An initial neural network structure is constructed, which includes a residual feature extraction layer, an attention mechanism layer, and a classification output layer. The network iteration parameters are initialized, including the residual convolution weight matrix, the attention bias vector, and the feature fusion allocation coefficient. The historical multimodal sample set is input into the residual feature extraction layer of the initial neural network structure, and the historical image blur residual features and historical radio frequency moisture absorption features are extracted through the residual convolution weight matrix. The blurred residual features of historical images and the moisture absorption features of historical radio frequency images are input into the attention mechanism layer, and weighted calculation is performed using feature fusion allocation coefficients to generate a predicted moisture characterization vector. The predicted moisture characterization vector is input into the classification output layer and superimposed with the attention bias vector to map the output predicted humidity level value. Calculate the classification residual error between the predicted humidity level value and the corresponding standard humidity level label, and at the same time calculate the feature overlap and divergence between the blurred residual features of historical images and the historical radio frequency moisture absorption features. The classification residual error and feature overlap divergence are added together according to a preset penalty weight to generate a comprehensive training loss value; The gradient descent algorithm is used to calculate the adjustment gradient of each parameter based on the comprehensive training loss value. The residual convolution weight matrix, attention bias vector and feature fusion allocation coefficient are updated and iterated by backpropagation until the comprehensive training loss value is less than the preset loss convergence threshold. The network iteration parameters at this time are saved to generate a residual compensation neural network.

5. The radio frequency fingerprint recognition method based on artificial intelligence according to claim 3, characterized in that, Image blur residual features and radio frequency moisture absorption features were extracted, including: The multimodal fusion data matrix is ​​processed by channel separation to independently acquire the image channel data components and the radio frequency channel data components; The image channel data components are edge detection convolution operation is performed on the image channel data components by using the preset image convolution kernel in the residual feature extraction layer to calculate the high-frequency missing values ​​of image texture, and the high-frequency missing values ​​of image texture are marked as image blur residual features. The spectral response of the RF channel data components is analyzed by using the preset RF analysis filter in the residual feature extraction layer. The amplitude attenuation value of a specific frequency band is extracted and marked as the RF moisture absorption feature.

6. The method of claim 1, wherein the method is based on artificial intelligence. The convergent radio frequency feature vector is obtained, including: Calculate the spatial geometric distance between the corrected radio frequency eigenvector and the center point of the eigenconvergence interval to obtain the eigenbia value; The feature deviation value is compared with the preset deviation threshold. If the feature deviation value is greater than the preset deviation threshold, the offset polarity data of the feature deviation value relative to the feature convergence interval is extracted. Based on the offset polarity data and characteristic deviation values, the refractive index increment adjustment value is calculated and then superimposed on the current refractive index correction parameter to complete the iterative update of the refractive index correction parameter. The phase reverse offset and linear gain compensation operation of the RF reflected signal is re-executed using the iteratively updated refractive index correction parameters until the recalculated characteristic deviation value is within the preset deviation threshold. The corrected RF characteristic vector that meets the condition at this time is marked as the converged RF characteristic vector.

7. The method of claim 6, wherein the method further comprises: The phase reversal offset and linear gain compensation operations for the RF reflected signal are re-executed using the iteratively updated refractive index correction parameters until the recalculated characteristic deviation value is within a preset deviation threshold. The corrected RF characteristic vector that meets the conditions at this point is marked as the converged RF characteristic vector, including: Obtain the refractive index correction parameters after iterative updates, and decompose them to obtain the iterative phase offset coefficient and iterative gain compensation coefficient in the current period. The original phase data components are subjected to difference cancellation processing using the iterative phase offset coefficient, and the original amplitude data components are subjected to product amplification processing using the iterative gain compensation coefficient, thereby generating the current iteratively corrected phase component and iteratively corrected amplitude component. A three-dimensional feature mapping is performed on the iteratively corrected phase component and the iteratively corrected amplitude component to generate the corrected radio frequency feature vector for the current period. Calculate the Euclidean space distance of the current corrected radio frequency feature vector relative to the feature convergence interval to obtain the recalculated feature deviation value; Compare the recalculated feature deviation value with the preset deviation threshold. If the recalculated feature deviation value is within the acceptable range determined by the preset deviation threshold, then the current corrected RF feature vector is determined to meet the convergence condition, and its output is marked as the converged RF feature vector.

8. The method of claim 1, wherein the method is based on artificial intelligence. Fingerprint image feature vectors are extracted from fingerprint image data according to image feature extraction weights, and radiofrequency dermal layer feature vectors are extracted from converged radiofrequency feature vectors according to radiofrequency feature extraction weights, including: The fingerprint image data is used to extract feature points using a preset image texture parsing algorithm to obtain an initial surface ridge feature vector; the initial surface ridge feature vector is then multiplied with the image feature extraction weights to generate a fingerprint image feature vector. The initial dermal layer reflection feature vector is separated by performing feature filtering operation on the converged radio frequency feature vector using a preset frequency band isolation matrix. The initial dermal reflection feature vector is multiplied with the radio frequency feature extraction weights to generate the radio frequency dermal feature vector.

9. An artificial intelligence based radio frequency fingerprinting system, characterized in that, It is implemented based on any one of claims 1-8, using an artificial intelligence-based radio frequency fingerprint recognition method, comprising: Environmental parameter calculation module: acquires environmental humidity data collected by the environmental humidity sensor and the no-load baseband signal under no-load conditions; based on the environmental humidity data and the no-load baseband signal, calculates the theoretical phase shift and gain loss of the medium under the current environment; Finger humidity output module: acquires the radio frequency reflection signal and fingerprint image data synchronously collected when the recognition area is triggered; inputs the radio frequency reflection signal and fingerprint image data into a preset residual compensation neural network, and outputs the corresponding finger surface humidity level data; RF Vector Correction Module: Based on the phase shift and gain loss of the medium theory, the module performs a refractive index correction operation on the RF reflected signal to obtain a corrected RF feature vector; it determines whether the corrected RF feature vector falls into the preset feature convergence interval; if it does not fall into the feature convergence interval, it iteratively updates the refractive index correction parameters and returns to perform a refractive index correction operation on the RF reflected signal until the corrected RF feature vector falls into the feature convergence interval, thus obtaining a converged RF feature vector. Finger weight calculation module: Based on the finger surface humidity level data, it assigns corresponding image feature extraction weights and radio frequency feature extraction weights; it extracts fingerprint image feature vectors from fingerprint image data according to the image feature extraction weights, and at the same time extracts radio frequency dermal layer feature vectors from converged radio frequency feature vectors according to the radio frequency feature extraction weights. Door lock unlocking judgment module: The fingerprint image feature vector is fused with the radio frequency dermal layer feature vector to obtain the fused verification vector; the matching degree value between the fused verification vector and the pre-stored feature template is calculated; if the matching degree value is greater than the preset matching threshold, an unlocking control command is generated and output to the actuator of the smart door lock.

Images