PPG Big Data-Driven Biometric Recognition Methods

Abstract

This invention provides a biometric identification method driven by PPG big data, comprising: acquiring PPG signal data of trainees; preprocessing the PPG signal data by high-pass filtering and low-pass filtering to obtain preprocessed PPG signals; segmenting the preprocessed PPG signals into frames, extracting Mel-frequency cepstral coefficient features and Gammatone-frequency cepstral coefficient features respectively; fusing the Mel-frequency cepstral coefficient features and Gammatone-frequency cepstral coefficient features and performing principal component analysis for dimensionality reduction to obtain fused features; constructing a general background model (UBM) of Gaussian mixture model based on the fused features, calculating zero-order, first-order, and second-order Baum-Welch statistics, estimating the global difference space matrix from the Baum-Welch statistics, and extracting an i-vector identity authentication vector; inputting the i-vector identity authentication vector into a long short-term memory network for training and classification to obtain the identity recognition result of the trainees. This invention can eliminate the influence of motion artifacts, capture the overall variability of PPG signals, and provide an identity authentication scheme for wearable devices.

Description

Technical Field

[0001] This invention relates to the field of biometric identification technology, and in particular to a biometric identification method driven by PPG big data. Background Technology

[0002] Optical volumetric pulse wave (PPG) data can effectively reflect vital signs and provide important reference for personnel identification during training. However, due to the diverse training scenarios, it is susceptible to motion artifacts (MA), which can distort the fidelity of the signal and thus hinder the reliable estimation of important parameters.

[0003] Currently, the main problems in the field of PPG signal processing are as follows: First, PPG signals are susceptible to motion artifacts, leading to a decrease in signal quality; second, traditional feature extraction methods are difficult to effectively capture individual differences in PPG signals; and third, existing recognition models lack robustness in complex training environments.

[0004] Therefore, there is an urgent need for a biometric identification method that can effectively eliminate motion artifacts, accurately capture individual characteristics of PPG signals, and has strong robustness, in order to meet the needs of identifying trainees during training. Summary of the Invention

[0005] The purpose of this invention is to provide a biometric identification method driven by PPG big data, aiming to solve the problem that PPG signals are easily affected by motion artifacts, which leads to a decrease in the reliability of important parameter estimation in the prior art.

[0006] To achieve the above objectives, this invention provides a biometric identification method driven by PPG big data, comprising: acquiring PPG signal data of trainees; preprocessing the PPG signal data by high-pass filtering and low-pass filtering to obtain preprocessed PPG signals; segmenting the preprocessed PPG signals into frames, extracting Mel-frequency cepstral coefficient features and Gammatone-frequency cepstral coefficient features respectively; fusing the Mel-frequency cepstral coefficient features and Gammatone-frequency cepstral coefficient features and performing principal component analysis for dimensionality reduction to obtain fused features; constructing a general background model (UBM) of Gaussian mixture model based on the fused features, calculating zero-order, first-order, and second-order Baum-Welch statistics, estimating the global difference space matrix from the Baum-Welch statistics, and extracting an i-vector identity authentication vector; inputting the i-vector identity authentication vector into a long short-term memory network for training and classification to obtain the identity recognition result of the trainees.

[0007] Preferably, the preprocessing of the PPG signal data by high-pass and low-pass filtering includes: configuring filter parameters for the PPG signal data, setting the cutoff frequency of the fourth-order Chebyshev Type I high-pass filter to 0.5Hz, setting the passband ripple to within ±0.5dB, and the stopband attenuation slope to -80dB per decade, to obtain high-pass filter parameters; constructing the fourth-order Chebyshev Type I high-pass filter based on the high-pass filter parameters, applying the fourth-order Chebyshev Type I high-pass filter to the PPG signal data to eliminate baseline drift, and obtaining a first filtered signal; evaluating the filtering effect based on the time-domain and frequency-domain characteristics of the first filtered signal, calculating the signal-to-noise ratio and the degree of baseline drift suppression, and ensuring that preset signal quality indicators are met.

[0008] Preferably, the preprocessing of the PPG signal data by high-pass filtering and low-pass filtering includes: configuring filter parameters for the first filtered signal, setting the cutoff frequency of the fourth-order elliptic low-pass filter to 18Hz, setting the passband to 0Hz to 18Hz with fluctuation less than or equal to 0.1 dB, and setting the stopband to start from 22Hz with attenuation greater than or equal to 40 dB at 22Hz and attenuation greater than or equal to 60 dB at 50Hz, thus obtaining the low-pass filter parameters; constructing the fourth-order elliptic low-pass filter based on the low-pass filter parameters, applying the fourth-order elliptic low-pass filter to the first filtered signal to eliminate motion artifacts, ambient light interference, and muscle noise, thus obtaining the preprocessed PPG signal; evaluating the noise suppression effect based on the spectral characteristics of the preprocessed PPG signal, calculating the high-frequency interference suppression degree, and ensuring that the preset signal quality requirements are met.

[0009] Preferably, the step of segmenting the preprocessed PPG signal into frames and extracting Mel frequency cepstral coefficient features and Gammatone frequency cepstral coefficient features respectively includes: performing pre-emphasis processing on the preprocessed PPG signal to enhance high-frequency components and obtain a pre-emphasis signal; segmenting the pre-emphasis signal into frames according to a preset frame length, and applying a Hamming window to each frame signal for smooth transition to obtain a first windowed signal; performing a fast Fourier transform on the first windowed signal to obtain a spectrum, and inputting the spectrum into a Mel filter bank for frequency envelope extraction to obtain a Mel spectrum; and performing a discrete cosine transform on the Mel spectrum to compress the feature dimension and obtain the Mel frequency cepstral coefficient features.

[0010] Preferably, the step of framing the preprocessed PPG signal and extracting Mel frequency cepstral coefficient features and Gammatone frequency cepstral coefficient features respectively includes: framing the preprocessed PPG signal and applying a Hamming window to each frame to obtain a second windowed signal; performing a fast Fourier transform on the second windowed signal to obtain a frequency domain signal; inputting the frequency domain signal into a Gammatone filter bank, dividing the frequency band according to the ERB frequency scale to obtain the filtered output of each channel; performing energy calculation and logarithmic compression on the filtered output of each channel, and performing a discrete cosine transform to obtain the Gammatone frequency cepstral coefficient features.

[0011] Preferably, the step of constructing a general background model (UBM) based on the fusion features, calculating zero-order, first-order, and second-order Baum-Welch statistics, estimating the global difference space matrix from the Baum-Welch statistics, and extracting the i-vector authentication vector includes: based on the fusion features, setting the number of Gaussian mixture components to 32, constructing a general background model (UBM) with its covariance matrix set as a diagonal matrix to obtain initial model parameters; according to the initial model parameters, using the EM algorithm to calculate the zero-order, first-order, and second-order Baum-Welch statistics, iteratively estimating the global difference space matrix until the change in the log-likelihood value is less than a preset threshold, thus obtaining the global difference space matrix; based on the global difference space matrix and the Baum-Welch statistics, extracting the i-vector authentication vector with a dimension of 200 using the maximum a posteriori probability estimation method.

[0012] Preferably, the method further includes: acquiring the high-pass filter parameters, the low-pass filter parameters, and extraction parameters of the Mel frequency cepstral coefficient features and the Gammatone frequency cepstral coefficient features; constructing a parameter optimization framework based on a genetic algorithm; encoding the high-pass filter parameters, the low-pass filter parameters, and the extraction parameters into chromosomes of an initial population; calculating the fitness function value of each chromosome based on the chromosomes of the initial population, wherein the fitness function value comprehensively evaluates three indicators: signal-to-noise ratio gain, signal quality index, and feature discrimination, to obtain the population fitness distribution; and, according to the population fitness distribution, retaining the 10% of individuals with the highest fitness to enter the next generation, and using brine extraction on the remaining 90% of individuals. The competition selects crossover and mutation, with the crossover rate initially set at 0.8 and dynamically adjusted according to population diversity, and the mutation rate initially set at 0.1 and gradually reduced to 0.01 with the number of iterations, resulting in a new generation population. Every 10 generations, the 10% of individuals with the highest fitness in the new generation population undergo simulated annealing for local optimization until a preset number of iterations is reached to obtain the optimal parameter configuration. Based on the optimal parameter configuration, personalized preprocessing and feature extraction are performed on the PPG signal of each trainee to generate optimized PPG signals and features. An incremental optimization mechanism is set up so that when the recognition accuracy drops beyond a preset threshold or reaches a fixed optimization cycle, parameter re-optimization is triggered, and the optimization process is accelerated through parallel computing.

[0013] Preferably, the method further includes: obtaining the initial fusion weights and dimensionality reduction matrix of the Mel frequency cepstral coefficient features and the Gammatone frequency cepstral coefficient features; constructing a bi-objective optimization problem for feature fusion, setting the first objective as maximizing the Fisher discrimination ratio and the second objective as minimizing the feature dimension; based on the bi-objective optimization problem, generating an initial population using the NSGA-II algorithm, where each individual contains two parts: feature fusion weights and dimensionality reduction, wherein the fusion weights range from [0,1] and sum to 1, and the dimensionality reduction ranges from [8,20]; performing non-dominated sorting and crowding calculation on the initial population; generating the parent population through a binary tournament selection; and using simulation... Binary crossover and polynomial mutation generate the offspring population; every 10 generations, a local search is performed on the solutions on the non-dominated front, with the search step size dynamically adjusted according to the number of iterations to accelerate algorithm convergence and obtain the Pareto optimal solution set; based on the Pareto optimal solution set, the optimal solution is selected using the knee point detection method. The knee point detection method first sorts the solution set in ascending order according to the first objective, then calculates the curvature of adjacent solutions, selects the point with the largest curvature as the knee point, and generates a personalized feature fusion model; an adaptive adjustment mechanism is set to monitor the recognition performance. When the cumulative error rate exceeds 5% or 30 days have passed, the feature fusion model is automatically updated to ensure performance stability.

[0014] Preferably, the method further includes: constructing an i-vector framework parameter space and a decision model parameter space, wherein the i-vector framework parameter space includes the number of Gaussian mixture components, the rank of the T matrix, the dimension of the feature vector, and the number of iterations, and the decision model parameter space includes the number of layers of the LSTM network, the number of hidden units, the dropout rate, and the learning rate; optimizing the i-vector framework parameter space and the decision model parameter space respectively using the covariance matrix adaptive evolution strategy CMA-ES, and realizing interactive optimization between subspaces through a three-layer collaborative mechanism of parameter cross-evaluation, fitness sharing, and search direction coordination; dynamically adjusting the computational resources allocated to each subspace according to the progress speed and performance improvement of the interactive optimization of the subspaces. To optimize computational resources, multi-threading is used to simultaneously evaluate different parameter combinations. For different training scenarios, typical scenario configurations are predefined and optimization processes are executed separately to generate scenario-specific parameter configuration schemes. Parameter configurations are automatically switched based on real-time detected training scenarios. A regularization control strategy is employed, adding L2 regularization constraints to the i-vector subspace to control the complexity of the T matrix. Batch normalization and early stopping strategies are introduced in the decision model subspace to improve model generalization ability. Incremental optimization is achieved using a two-level strategy: a lightweight update strategy and a full update strategy. The lightweight update strategy fine-tunes parameters within a local range, while the full update strategy re-executes a complete co-evolutionary optimization after accumulating new data exceeding a preset data value.

[0015] Preferably, the step of inputting the i-vector identity authentication vector into a Long Short-Term Memory (LSTM) network for training and classification to obtain the identity recognition result of the trainee includes: normalizing and temporally rearranging the i-vector identity authentication vector to generate a training sample sequence and corresponding identity labels; based on the training sample sequence and corresponding identity labels, constructing a network structure containing an input layer, two LSTM layers, and a fully connected layer, wherein the first LSTM layer has 128 units, the second LSTM layer has 64 units, and the dropout rate is set to 0.3 to obtain an initial network model; using cross-entropy as the loss function, iteratively training the initial network model using the Adam optimizer, setting the initial learning rate to 0.001 and dynamically adjusting it using a cosine annealing strategy to obtain a trained LSTM network model; inputting the i-vector identity authentication vector to be recognized into the trained LSTM network model, calculating the probability distribution of each identity category using a Softmax classifier, and selecting the category with the highest probability as the recognition result.

[0016] The beneficial effects of this invention are:

[0017] 1. This invention effectively eliminates interference such as baseline drift, motion artifacts, ambient light interference, and muscle noise by performing high-pass and low-pass filtering preprocessing on PPG signals, thereby improving signal quality;

[0018] 2. This invention employs a fusion of MFCC and GFCC features, fully utilizing the complementarity of the two features. MFCC can capture the overall spectral envelope of the signal, while GFCC emphasizes finer spectral details in the low-frequency part through the Gammatone filter bank, thereby improving the distinguishability of the features.

[0019] 3. This invention utilizes the i-vector framework to model PPG signals, compressing the feature vector sequence into a compact low-dimensional representation, effectively capturing the overall variability of PPG signals, and improving the accuracy and efficiency of recognition;

[0020] 4. This invention uses a Long Short-Term Memory (LSTM) network to make decisions on the i-vector, which can effectively capture the temporal characteristics of the PPG signal and further improve the accuracy of recognition;

[0021] 5. This invention provides a parameter optimization framework based on evolutionary algorithms, a feature fusion strategy based on multi-objective optimization, and an i-vector and model parameter optimization method based on co-evolution, enabling the system to adapt to the needs of different individuals, different motion states, and different scenarios, thereby improving the robustness and adaptability of the system. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the biometric identification method driven by PPG big data in an embodiment of the present invention;

[0024] Figure 2 This is a comparison diagram of the PPG signal before and after filtering in an embodiment of the present invention;

[0025] Figure 3 This is a flowchart of MFCC feature extraction in an embodiment of the present invention;

[0026] Figure 4 This is a block diagram of GFCC feature extraction in an embodiment of the present invention;

[0027] Figure 5 This is an average energy distribution diagram of MFCC and GFCC in an embodiment of the present invention;

[0028] Figure 6 This is a cumulative contribution rate diagram of each dimension of MFCC feature values ​​in an embodiment of the present invention;

[0029] Figure 7 This is a cumulative contribution rate diagram of each dimension of GFCC feature values ​​in an embodiment of the present invention;

[0030] Figure 8 This is a flowchart of the EM algorithm for estimating the T matrix in an embodiment of the present invention;

[0031] Figure 9 This is a schematic diagram of stratified data sampling in an embodiment of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0033] This invention provides a biometric identification method driven by PPG big data. This method is designed for wristband sensor acquisition scenarios and utilizes the i-vector framework to model the inherent time series of trainees' heartbeat signals to achieve efficient identification. The invention will be described in detail below with specific embodiments.

[0034] Example 1

[0035] like Figure 1 As shown, this embodiment provides a biometric identification method driven by PPG big data, including the following steps:

[0036] Step S1: Obtain the PPG signal data of the trainees, and perform high-pass filtering and low-pass filtering preprocessing on the PPG signal data to obtain the preprocessed PPG signal.

[0037] Specifically, 30 healthy participants aged 18-58 were selected to complete three types of activities in a simulated training scenario. For each participant, PPG signals were obtained from a pulse oximeter worn on the wrist with a green LED light (wavelength: 609 nm). Each recording lasted approximately 20 minutes and included physiological signals sampled at a frequency of 125 Hz. Real-world data from the simulated training scenario was obtained from the 30 participants using the following methods:

[0038] Type 1 (T1): Participants 1-10 will complete a certain duration and speed of walking or running on a treadmill in the following order: 0.5 minutes of walking at 1-2 km / h, 1 minute of jogging at 6-8 km / h, 12-15 minutes of running at 15 km / h, 1 minute of jogging at 6-8 km / h, 1 minute of running at 12-15 km / h, and 0.5 minutes of walking at 1-2 km / h.

[0039] Type 2 (T2): Participants aged 9 to 22 participate, mainly performing general upper limb exercises, such as shaking hands, stretching, pushing, running, jumping, and push-ups.

[0040] Type 3 (T3): Participants aged 21-30 participate, mainly engaging in upper limb violent exercises, such as boxing, knife fighting, and pull-ups.

[0041] Since trainees 9 and 10 participated in both T1 and T2, and trainees 21 and 22 participated in both T2 and T3, a total of 34 data records were obtained in this simulation test.

[0042] The PPG signal has a frequency band ranging from 0.5 to 18 Hz. The main noise sources in PPG recordings include baseline drift (respiration, temperature changes), motion artifacts (MA), ambient light interference (power frequency interference), and muscle noise (EMG). To suppress these interferences, multi-stage filtering was implemented: firstly, a Chebyshev Type I high-pass filter with a cutoff frequency of 0.5 Hz was used to eliminate baseline drift.

[0043] The implementation of a 4th-order Chebyshev Type I high-pass filter with a cutoff frequency of 0.5Hz is as follows:

[0044]

[0045] Where the coefficient These are the filter coefficients, and their specific values ​​are:

[0046] b=[0.9391,-3.7564,5.6346,-3.7564,0.9391]

[0047] a=[1.0000,-3.8746,5.6280,-3.6407,0.8873]

[0048] The corresponding difference equation is:

[0049]

[0050] The characteristics of this filter:

[0051] Passband (>0.5Hz): Fluctuations within ±0.5 dB.

[0052] Stopband (<0.5Hz): Attenuation slope ≈ -80 dB / decade (≥40 dB at 0.1Hz).

[0053] Subsequently, motion artifacts, ambient light interference (power frequency interference), and muscle noise were filtered through an elliptic low-pass filter with a cutoff frequency of 18 Hz.

[0054] The implementation of a fourth-order elliptic low-pass filter with a cutoff frequency of 18 Hz is as follows:

[0055]

[0056] Where the coefficient These are the filter coefficients, and their specific values ​​are:

[0057] b = [0.0126, 0.0344, 0.0422, 0.0344, 0.0126]

[0058] a = [1.0000, -2.6236, 2.7247, -1.3226, 0.2406]

[0059] The corresponding difference equation is:

[0060]

[0061] The characteristics of this filter:

[0062] Passband: 0-18Hz signal retention, fluctuation ≤0.1 dB.

[0063] Stopband: Signals ≥22Hz are strongly suppressed (≥40 dB at 22Hz, ≥60 dB at 50Hz).

[0064] Extremely narrow transition band: Attenuation drops rapidly from -3dB to -40dB at only 4Hz (18Hz→22Hz).

[0065] The final stage of preprocessing includes smoothing to further improve signal quality, such as... Figure 2 As shown.

[0066] Step S2: The preprocessed PPG signal is segmented into frames, and Mel frequency cepstral coefficient features and Gammatone frequency cepstral coefficient features are extracted respectively. After fusing the Mel frequency cepstral coefficient features and Gammatone frequency cepstral coefficient features, principal component analysis is performed to reduce the dimensionality and obtain the fused features.

[0067] Specifically, after noise reduction, MFCC and GFCC are used for feature extraction as the main analytical features. MFCC is a mature algorithm widely used in signal processing (especially speech recognition). Cepstral calculation, as a homomorphic transformation from convolution to addition, is scaled according to the Mel scale (an auditory scale simulating the human ear's perception of sound frequency response). This invention chooses the MFCC calculation method based on Discrete Cosine Transform (DCT) to reduce computational cost, and its calculation process is as follows: Figure 3 As shown.

[0068] The first step is to process the input signal. (indicating the sample point) The original PPG signal is pre-emphasized to enhance high-frequency components. The calculation formula is as follows:

[0069]

[0070] In the formula, 0.97 is the pre-emphasis coefficient, which is often used in signal processing to enhance high-frequency components and reduce low-frequency dominance.

[0071] Next, the pre-emphasis signal The signal is divided into equal-length short frames (e.g., 25 milliseconds), and each frame is windowed to reduce spectral leakage and ensure a smooth transition. Windowed signal. The calculation formula is:

[0072]

[0073] Where N represents the number of samples per frame, and the expression in parentheses defines the Hamming window applied to each frame.

[0074] Finally, the windowed signal h[n] of each frame is converted to the frequency domain using Fast Fourier Transform (FFT):

[0075]

[0076] In the formula Indicates the first The complex spectrum of each frequency bin.

[0077] This step calculates the Mel spectrum by inputting the Fourier transform signal amplitude into a Mel filter bank (a set of triangular bandpass filters), expressed as:

[0078]

[0079] It is the first A triangular filter, and They represent the first The lower and upper cutoff frequencies of each filter are determined. Finally, the... Generate Mel-frequency cepstral coefficients (MFCCs) using Discrete Cosine Transform (DCT):

[0080]

[0081] In the formula Indicates the first MFCC coefficients, The total number of coefficients selected for the eigenvector.

[0082] Traditional MFCC (Mel frequency cepstral coefficients) are designed based on the Mel scale of human hearing, but the frequency domain characteristics of PPG signals (mainly concentrated in the 0.5–5 Hz range) differ significantly from speech signals. This paper introduces GFCC (Gammatone frequency cepstral coefficients), the core of which lies in the simulation of the frequency selectivity of the cochlear basilar membrane by a Gammatone filter bank. Gammatone filters have higher frequency resolution in the low-frequency band, and their transfer function… The pole-zero distribution characteristics are better suited to the narrowband spectral structure of the PPG signal. By calculating the energy of each filter channel and performing logarithmic compression, the DCT transform further extracts a compact representation of the cepstral domain, and the final GFCC coefficients retain the key time-frequency domain information of the PPG signal.

[0083] The impulse response of a Gammatone filter is defined as:

[0084]

[0085] in For amplitude gain, It is the filter order. Indicates the filter bandwidth. It is the center frequency measured in Hertz (Hz). This represents the phase shift. The center frequency is equidistantly distributed within the filter bank boundaries according to the equivalent rectangular bandwidth (ERB) scale. The Gammatone filter bandwidth B is given by the following formula:

[0086]

[0087] Among them, arbitrary frequency (unit ) The calculation formula is:

[0088]

[0089] here This represents the quality factor of a high-frequency asymptotic filter. It is the minimum bandwidth at low frequencies. Typically, the value is 1 or 2. This invention uses the Glasberg-Moore parameter, i.e. Parameters; EarQ and This parameter can better improve the resolution of the spectrum.

[0090] like Figure 4As shown, GFCC can be considered an improved algorithm of MFCC: First, the audio signal is divided into frames and windowed, and the spectrum is obtained by performing Fast Fourier Transform (FFT) frame by frame. Then, the FFT signal is processed with a Gammatone filter to enhance the perception of significant frequency bands. Finally, logarithmic operations and discrete cosine transforms are applied. The GFCC calculation formula is:

[0091]

[0092] In the formula Indicates the first Each spectrum band has energy. It is the number of Gammatone filters. For the first GFCC coefficients, It is the total number of coefficients selected in the eigenvector.

[0093] In addition, such as Figure 5 The average energy distributions of MFCC and GFCC for the same PPG signal segment are shown, indicating that the Mel filter method produces a more uniform energy distribution, while the Gammatone filter framework generates significant energy peaks that coincide with key regions such as the low-frequency bands of the PPG. This characteristic enhances the advantage of cubic compression in preserving subtle amplitude variations. This difference is also reflected in histogram analysis; the output of the Mel filter exhibits a broader distribution, while the output of the Gammatone filter is concentrated in a specific (lower) frequency range. Therefore, fusing the two features ensures richer preservation of subject-specific characteristics. This approach, which maintains mid-to-high frequency resolution while optimizing low-frequency characterization, captures more variability in individual PPG signals, thereby enhancing biometric recognition capabilities.

[0094] Principal component analysis (PCA) is a variance-based dimensionality reduction technique used to transform potentially correlated variables into a set of linearly independent principal components. This method effectively reduces data dimensionality and redundancy while preserving key feature information.

[0095] The PCA dimensionality reduction process includes:

[0096] 1. Zero-mean normalization: Subtract the mean of each row of the data matrix to make the data distribution revolve around the origin and eliminate the offset.

[0097] 2. Calculate the covariance matrix: to measure the correlation and degree of change between different attributes.

[0098] 3. Eigenvalue decomposition and dimensionality reduction: Calculate the eigenvalues ​​and eigenvectors of the covariance matrix; sort the eigenvalues ​​from largest to smallest, and select the first k eigenvectors to form the dimensionality reduction matrix.

[0099] To reduce computational complexity while preserving key information, the cumulative contribution rate of the feature vectors needs to be calculated, as shown in the following formula:

[0100]

[0101] In the formula, These are the corresponding eigenvalues ​​in the matrix; The cumulative contribution rate is close to 100%, indicating that the selected principal components have covered most of the information.

[0102] For dimensionality reduction of PPG signal features:

[0103] 1. Data source: PPG data of 30 trainees, 15 groups were selected as test samples.

[0104] 2. Feature fusion: MFCC (Mel frequency cepstral coefficients) + GFCC (Gammatone frequency cepstral coefficients) → fused into MGCC features (higher dimensionality).

[0105] 3. Dimensionality Reduction Analysis: MFCC: The cumulative contribution rate of the first 8 dimensions reaches 100%, such as... Figure 6 The figure shows the cumulative contribution rate of each dimension of the MFCC eigenvalues, so 8 dimensions are retained; GFCC: the first 4 dimensions already contain the main information, such as Figure 7 The figure shows the cumulative contribution rate of each dimension of GFCC eigenvalues, so 4 dimensions are retained.

[0106] 4. Final features: After dimensionality reduction, 12-dimensional MFCC+GFCC fusion features are obtained, which effectively improves the model training speed and recognition accuracy.

[0107] Step S3: Construct a general background model (UBM) of Gaussian mixture model based on the fusion features, calculate the zero-order, first-order and second-order Baum-Welch statistics, estimate the global difference space matrix from the Baum-Welch statistics, and extract the i-vector identity authentication vector.

[0108] Specifically, in real-world PPG signal recognition scenarios, PPG signals not only contain target physiological characteristics (such as heart rate and blood pressure) but also various noise and motion artifact interferences. Different acquisition devices and environmental conditions (such as skin contact pressure and ambient lighting) introduce signal variations, which interfere with the collected PPG data and thus affect the accuracy of physiological parameter recognition. Traditional GMM-UBM methods may encounter challenges in handling such signal variations, leading to unstable recognition system performance. To address this issue, the i-vector method models global differences in PPG data, incorporating both physiological characteristic differences and acquisition interference differences into the model. This method reduces dependence on training data quality, simplifies the computation process, and improves the overall system performance. The i-vector method has high computational efficiency and is suitable for processing large-scale PPG datasets. Furthermore, thanks to i-vector's inherent cross-device stability and the integration of signal compensation techniques such as Linear Discriminant Analysis (LDA), i-vector exhibits stronger robustness in cross-device recognition tasks.

[0109] i-vector models the system in a global difference space, which includes differences in sound features and channel differences. The model can be represented as:

[0110] M = m + Tw

[0111] Where M represents the mean hypervector of the GMM generated from the current training samples; m represents the mean hypervector of the UBM that is unrelated to the sound feature information and channel information of the current training samples; T represents the global difference space matrix, which describes all variability in the sound data that is related to and unrelated to the current fault type. It is learned from a large amount of training data with different fault types; w represents a total variable difference factor that follows a standard Gaussian distribution, i.e., the i-vector. Of these four quantities, M and m are known and can be calculated using GMM and UBM. Therefore, to extract the i-vector, the global difference space matrix T must first be estimated. This estimation is done through statistical modeling, usually using the EM algorithm on a large amount of training data. The process is as follows: Figure 8 As shown, the detailed steps are as follows:

[0112] (1) First, the Baum-Welch statistic corresponding to each fault in the training data must be calculated. Its zeroth, first, and second-order statistics are as follows:

[0113]

[0114]

[0115]

[0116] in It is the mean hypervector in the UBM model. The first in The mean of the Gaussian components, It is the dimension of the input PPG signal features. That is, the first Frame feature vector.

[0117] (2) Next, the global difference space matrix T is estimated using the EM algorithm. First, the T matrix is ​​initialized and the current zero-order, first-order and second-order Baum-Welch statistics are calculated.

[0118] (3) Perform the E-step to calculate the first and second expectations of the total variable difference factor w:

[0119]

[0120]

[0121]

[0122] in It is the precision matrix, which is the inverse covariance matrix of the i-vector posterior distribution. It is the identity matrix. It is the diagonal covariance matrix of the variances of the mixture components containing UBM on the diagonal.

[0123] (4) Next, perform the M-step to update the global difference space matrix T and the covariance matrix ∑ respectively, using the following update formulas:

[0124]

[0125] (5) Repeat steps 3 and 4. After each iteration, calculate the log-likelihood value for the current parameters. If the change is less than the threshold, the convergence is considered complete. Finally, the global difference space matrix T is obtained.

[0126] Log-likelihood value The calculation formula is as follows:

[0127]

[0128] in , For second-order and first-order statistics, and Let i be the expectation of i-vector.

[0129] The convergence condition is that the log-likelihood difference between two consecutive iterations satisfies:

[0130]

[0131] After estimating the global difference space matrix T, the i-vector is the maximum a posteriori estimate of w. Therefore, the formula for extracting the i-vector is:

[0132]

[0133] Where I is the identity matrix, N is the zero-order statistic, i.e. the sum of frame weights of each Gaussian component, the supervector F is the first-order central statistic, ∑ represents the covariance matrix (diagonal matrix) of UBM, and T is the trained global difference space matrix.

[0134] (6) Selection of the number of Gaussian mixture components:

[0135] The Gaussian mixture factor (GMM) is the number of independent Gaussian distributions in a model. Theoretically, a higher GMM indicates a stronger fit to the class. However, as the mixture factor increases, the model becomes overly complex, significantly increasing computational costs and training / testing times, while also increasing the risk of overfitting. Therefore, the mixture factor directly impacts the model's recognition accuracy. Experiments were conducted with GMMs of 8, 16, 32, and 64, while keeping other parameters constant. The results, showing the fused features composed of 12 extracted feature vectors, are presented in Table 1.

[0136] Table 1. Effect of Gaussian mixture component number on model performance

[0137]

[0138] As can be seen from the table, the EER decreases as the number of Gaussian mixture components increases. When the number of Gaussian mixture components increases from 8 to 32, the EER decreases significantly. However, when the number of Gaussian mixture components continues to increase to 64, the EER also decreases, but the magnitude is very small. Considering that increasing the number of Gaussian mixture components will significantly increase the computational cost, reduce the efficiency of the model, and worsen the real-time performance, the final model uses a Gaussian mixture component number of 32.

[0139] (7) Choice of i-vector dimension:

[0140] EER is a commonly used performance evaluation metric in biometric (such as speaker recognition, fingerprint recognition, etc.) and classification systems. It represents the error rate when the false acceptance rate (FAR) and false rejection rate (FRR) are equal. Its core meaning is the overall error probability of the system when the two types of errors (false positives and false negatives) are balanced. The lower the value, the better the system performance.

[0141] In the i-vector method, the i-vector dimension has a significant impact on model accuracy. A low-dimensional i-vector may not be able to fully capture relevant information about the category, resulting in poor model recognition accuracy. On the other hand, an excessively high i-vector dimension may contain more information, but when the training samples are limited, it may lead to overfitting and increase computational costs. The optimal i-vector dimension for this model is determined through experiments. The i-vector dimension of the model output was set to 100, 200, 300, and 400, with other parameters remaining unchanged by default. The extracted 12-dimensional fused features were used for the experiments, and the results are shown in Table 2.

[0142] Table 2. Impact of i-vector dimension on model performance

[0143]

[0144] As can be seen from the table, the EER of the model gradually decreases as the dimension of the i-vector increases. When the dimension is 200, the model accuracy is improved significantly. Although the EER decreases at 300 and 400 dimensions, the magnitude is not large. Considering that i-vector feature vectors with more dimensions will increase the risk of overfitting in subsequent classification decisions, 200 dimensions are selected as the i-vector dimension of the final output of the model.

[0145] (8) Selection of input feature duration:

[0146] In the i-vector method, the input feature duration has a significant impact on the model's recognition accuracy. A shorter duration means less effective information for training and testing, which may lead to insufficient representational power of the i-vector and reduced model accuracy. Conversely, a longer duration helps generate a more stable and discriminative i-vector. In practical applications, it is necessary to balance model performance and computational cost to select an appropriate feature duration. The input model feature duration was set to 2s, 3s, 4s, and 5s respectively. The feature duration here depends on the number of frames in the fused features because the input model's features are frame-based. However, regardless of the input model's feature duration, the i-vector model can output an i-vector of uniform length, which is one of its advantages. Other parameters remained unchanged by default. Experiments were conducted using the extracted 12-dimensional fused features, and the results are shown in Table 3.

[0147] Table 3. Impact of feature duration on model performance

[0148]

[0149] As can be seen from the table, the model performance gradually improves and the EER gradually decreases as the feature duration of the input model increases. When the feature duration increases from 2s to 4s, the model accuracy improves significantly. When the feature duration continues to increase to 5s, the EER decreases, but not by much. Since the PPG signal is periodic, the impact of further increasing the feature duration of the input model on the model performance is not significant. Considering the issue of computational efficiency, this paper selects 4s as the final input feature duration.

[0150] The final model input MGCC fusion features are 12-dimensional, Gaussian mixture components are 32, and the output i-vector is 200-dimensional. The input feature duration is 4 seconds. At this point, the model balances the quality of the output i-vector and computational efficiency, preparing it for the next classification decision.

[0151] Based on the aforementioned fusion features, the number of Gaussian mixture components (Gaussian Mixture Components) is set to 32, and a Universal Background Model (UBM) is constructed with its covariance matrix set as a diagonal matrix to obtain the initial model parameters. The initialization process employs an improved K-means clustering method, clustering all training data into 32 clusters. The mean and variance of each cluster are used as the initial mean and covariance of the GMM, and the initial weights of each component are set to be equal. This clustering-based initialization method can accelerate the convergence of the EM algorithm and improve the quality of the final model.

[0152] UBM training employs the Expectation-Maximization (EM) algorithm to iteratively optimize the parameters of the Gaussian Mixture Model (GMM). Each iteration consists of two steps: the Expectation step (E-step), which calculates the posterior probability of each sample belonging to each Gaussian component; and the Maximization step (M-step), which updates the model parameters (weights, mean, and covariance) based on these posterior probabilities. Specifically, in the t-th iteration:

[0153] E-step: For each training sample x, calculate its posterior probability of belonging to the k-th Gaussian component.

[0154] M-step: Update the weights, mean, and covariance matrix of each Gaussian component.

[0155] To improve the robustness of UBM, a lower variance bound constraint is used during training to ensure that the diagonal elements of the covariance matrix are not less than a preset threshold (usually 0.01), preventing the occurrence of singular covariance matrices. Simultaneously, weight regularization is implemented to prevent some components from having excessively small or zero weights.

[0156] Based on the initial model parameters, the EM algorithm is used to calculate the zeroth, first, and second-order Baum-Welch statistics, iteratively estimating the global difference space matrix until the change in the log-likelihood value is less than a preset threshold, at which point the iteration stops, thus obtaining the global difference space matrix. The convergence criterion for the EM algorithm is: the change in the log-likelihood function between two consecutive iterations is less than a preset threshold (usually set to 1e-5) or the maximum number of iterations is reached (usually set to 100).

[0157] Estimating the global difference space matrix T is a crucial step in i-vector extraction. The initial T matrix is ​​obtained through principal component analysis (PCA) of the covariance matrix of the training data, followed by iterative optimization to obtain the final result. In each iteration, the posterior distribution parameters (mean and covariance) are first calculated based on the current T matrix, and then the T matrix is ​​updated based on these posterior statistics until the algorithm converges.

[0158] Based on the global difference space matrix and Baum-Welch statistics, a 200-dimensional i-vector authentication vector is extracted using the maximum a posteriori probability estimation method. The i-vector extraction employs a closed-form solution, resulting in an efficient and stable computation process. For each utterance, the i-vector is a point estimate (maximum a posteriori estimate) extracted from the posterior probability distribution, capturing the offset information of that utterance relative to the UBM.

[0159] Step S4: Input the i-vector identity authentication vector into the long short-term memory network for training and classification to obtain the identity recognition results of the trainees.

[0160] Specifically, in the overall variance space, the two most commonly used compensation techniques for i-vectors are within-class covariance normalization (WCCN) and linear discriminant analysis (LDA), which are typically used to reduce variability. WCCN normalizes the cosine kernel using the inverse of the within-class covariance matrix. LDA minimizes the within-class variance caused by variability effects through coordinate axis transformation, while maximizing the between-class variance. Experiments show that the cosine kernel function is an effective classifier for classifying i-vectors. The cosine kernel between two i-vectors, w1 and w2, is defined as follows:

[0161]

[0162] To further optimize the performance of this method, various commonly used machine learning and deep learning methods can be used for effective classifier selection in i-vector classification, including Support Vector Machine (SVM), Multilayer Perceptron (MLP), Random Forest (RF), Convolutional Neural Network (CNN), and Recurrent Neural Network (RNN). These methods can be used to identify PPG signals by capturing their time / frequency domain features. Support Vector Machine (SVM) performs well in nonlinear classification problems by determining the optimal separating hyperplane between data points and combining kernel functions, especially known for its high accuracy on finite datasets and robustness in noisy environments. Multilayer Perceptron (MLP) models nonlinear relationships in the data through one or more hidden layers, revealing signal features through sequence weighting and activation functions. Random Forest (RF) captures different sub-features of the data through a multi-decision tree ensemble structure, achieving robust classification. Convolutional Neural Network (CNN) automatically extracts PPG signal features hierarchically through convolutional filters and pooling layers, capturing local patterns and building a noise-resistant model. Recurrent Neural Networks (RNNs) model temporal dependencies by updating hidden states. In particular, the addition of gating structures such as LSTM and GRU helps to alleviate long-range dependency problems and effectively captures continuous heartbeat patterns in PPG signals.

[0163] The i-vector identity authentication vector is normalized and temporally rearranged to generate a training sample sequence and corresponding identity labels. Normalization is achieved using the Z-score method, ensuring the data mean is 0 and the standard deviation is 1. Temporal rearrangement organizes the training samples in chronological order so that the LSTM network can capture temporal features. The training samples and labels are divided into training and validation sets in an 80%:20% ratio.

[0164] Based on the training sample sequences and corresponding identity labels, a network structure was constructed consisting of an input layer, two LSTM layers, and a fully connected layer. The first LSTM layer has 128 units, the second LSTM layer has 64 units, and the dropout rate was set to 0.3 to obtain the initial network model. The LSTM layers used the tanh activation function, and the fully connected layer had the same number of output nodes as the number of trainees, using the Softmax activation function. The detailed network structure is as follows:

[0165] Input layer: shape is (batch) size sequence length , feature dim ), where feature dim For the dimensions of the i-vector;

[0166] First LSTM layer: 128 units, return sequences =True, using the tanh activation function;

[0167] Dropout layer: ratio of 0.3 to prevent overfitting;

[0168] The second LSTM layer has 64 units. sequences =False, use the tanh activation function;

[0169] Dropout layer: ratio of 0.3 to prevent overfitting;

[0170] Fully connected layer: The number of output nodes equals the number of trainees;

[0171] Softmax layer: Converts the output into a probability distribution.

[0172] Cross-entropy was used as the loss function, and the initial network model was iteratively trained using the Adam optimizer. The initial learning rate was set to 0.001 and dynamically adjusted using a cosine annealing strategy to obtain the trained LSTM network model. Training employed a mini-batch method with a batch size of 32 and a maximum number of iterations of 100. To prevent overfitting, an early stopping strategy was implemented, stopping training when the validation set loss showed no improvement for five consecutive epochs. Simultaneously, the validation set accuracy was calculated after each epoch, and the optimal model parameters were recorded.

[0173] The i-vector identity authentication vector to be identified is input into the trained LSTM network model. The Softmax classifier calculates the probability distribution of each identity category, and the category with the highest probability is selected as the recognition result. In practical applications, a confidence threshold (usually 0.7) is set. When the highest probability is lower than the threshold, the system will mark it as "unknown identity" to improve the reliability of recognition.

[0174] Example 2

[0175] This embodiment provides a biometric identification method driven by PPG big data, including the following steps:

[0176] Step S1: Obtain the PPG signal data of the trainees, and perform high-pass filtering and low-pass filtering preprocessing on the PPG signal data to obtain the preprocessed PPG signal.

[0177] Specifically, filter parameters are configured for the PPG signal data. The cutoff frequency of the fourth-order Chebyshev Type I high-pass filter is set to 0.5Hz, the passband ripple is set to within ±0.5 dB, and the stopband attenuation slope is set to -80 dB per decade. Based on these parameters, the fourth-order Chebyshev Type I high-pass filter is constructed and applied to the PPG signal data to eliminate baseline drift, resulting in a first filtered signal. The filtering effect is evaluated based on the time-domain and frequency-domain characteristics of the first filtered signal, and the signal-to-noise ratio and baseline drift suppression level are calculated to ensure that preset signal quality indicators are met.

[0178] The first filtered signal is configured with filter parameters. The cutoff frequency of the fourth-order elliptic low-pass filter is set to 18Hz, the passband is set to 0Hz to 18Hz with a fluctuation of less than or equal to 0.1 dB, and the stopband starts from 22Hz with an attenuation of greater than or equal to 40 dB at 22Hz and greater than or equal to 60 dB at 50Hz. Based on the low-pass filter parameters, the fourth-order elliptic low-pass filter is constructed and applied to the first filtered signal to eliminate motion artifacts, ambient light interference, and muscle noise, resulting in the preprocessed PPG signal. The noise suppression effect is evaluated based on the spectral characteristics of the preprocessed PPG signal, and the high-frequency interference suppression degree is calculated to ensure that the preset signal quality requirements are met.

[0179] Step S2: The preprocessed PPG signal is segmented into frames, and Mel frequency cepstral coefficient features and Gammatone frequency cepstral coefficient features are extracted respectively. After fusing the Mel frequency cepstral coefficient features and Gammatone frequency cepstral coefficient features, principal component analysis is performed to reduce the dimensionality and obtain the fused features.

[0180] Specifically, the preprocessed PPG signal is pre-emphasized to enhance high-frequency components, resulting in a pre-emphasized signal. The pre-emphasized signal is then divided into frames according to a preset frame length, and a Hamming window is applied to each frame for smooth transition, resulting in a first windowed signal. The first windowed signal is subjected to a Fast Fourier Transform to obtain a spectrum, which is then input into a Mel filter bank for frequency envelope extraction, resulting in a Mel spectrum. Finally, the Mel spectrum is subjected to a Discrete Cosine Transform to compress the feature dimension, resulting in the Mel frequency cepstral coefficient features.

[0181] The preprocessed PPG signal is framed, and a Hamming window is applied to each frame to obtain a second windowed signal. The second windowed signal is subjected to a Fast Fourier Transform to obtain a frequency domain signal. The frequency domain signal is input into a Gammatone filter bank, and the frequency band is divided according to the ERB frequency scale to obtain the filter output of each channel. The filter output of each channel is subjected to energy calculation and logarithmic compression, and a Discrete Cosine Transform is performed to obtain the Gammatone frequency cepstral coefficient characteristics.

[0182] Step S3: Construct a general background model (UBM) of Gaussian mixture model based on the fusion features, calculate the zero-order, first-order and second-order Baum-Welch statistics, estimate the global difference space matrix from the Baum-Welch statistics, and extract the i-vector identity authentication vector.

[0183] Specifically, based on the fusion features, the number of Gaussian mixture components is set to 32, and a general background model UBM is constructed, with its covariance matrix set as a diagonal matrix to obtain initial model parameters. According to the initial model parameters, the EM algorithm is used to calculate the zero-order, first-order, and second-order Baum-Welch statistics, and the global difference space matrix is ​​iteratively estimated until the change in the log-likelihood value is less than a preset threshold, at which point the iteration stops, and the global difference space matrix is ​​obtained. Based on the global difference space matrix and the Baum-Welch statistics, the i-vector authentication vector with a dimension of 200 is extracted using the maximum a posteriori probability estimation method.

[0184] Step S4: Input the i-vector identity authentication vector into the long short-term memory network for training and classification to obtain the identity recognition results of the trainees.

[0185] Specifically, the i-vector identity authentication vector is normalized and temporally rearranged to generate a training sample sequence and corresponding identity labels. Based on the training sample sequence and corresponding identity labels, a network structure is constructed, including an input layer, two LSTM layers, and a fully connected layer. The first LSTM layer has 128 units, the second LSTM layer has 64 units, and the dropout rate is set to 0.3 to obtain the initial network model. Cross-entropy is used as the loss function, and the Adam optimizer is used to iteratively train the initial network model. The initial learning rate is set to 0.001 and dynamically adjusted using a cosine annealing strategy to obtain the trained LSTM network model. The i-vector identity authentication vector to be identified is input into the trained LSTM network model, and the probability distribution of each identity category is calculated using a Softmax classifier. The category with the highest probability is selected as the recognition result.

[0186] Example 3

[0187] This embodiment provides a biometric identification method driven by PPG big data, including the following steps:

[0188] Step S1: Obtain the PPG signal data of the trainees, and perform high-pass filtering and low-pass filtering preprocessing on the PPG signal data to obtain the preprocessed PPG signal.

[0189] Step S2: The preprocessed PPG signal is segmented into frames, and Mel frequency cepstral coefficient features and Gammatone frequency cepstral coefficient features are extracted respectively. After fusing the Mel frequency cepstral coefficient features and Gammatone frequency cepstral coefficient features, principal component analysis is performed to reduce the dimensionality and obtain the fused features.

[0190] Step S3: Construct a general background model (UBM) of Gaussian mixture model based on the fusion features, calculate the zero-order, first-order and second-order Baum-Welch statistics, estimate the global difference space matrix from the Baum-Welch statistics, and extract the i-vector identity authentication vector.

[0191] Step S4: Input the i-vector identity authentication vector into the long short-term memory network for training and classification to obtain the identity recognition results of the trainees.

[0192] Step S5: Obtain the high-pass filter parameters, the low-pass filter parameters, and the extraction parameters of the Mel frequency cepstral coefficient features and the Gammatone frequency cepstral coefficient features; construct a parameter optimization framework based on a genetic algorithm; encode the high-pass filter parameters, the low-pass filter parameters, and the extraction parameters into chromosomes of the initial population; based on the chromosomes of the initial population, calculate the fitness function value of each chromosome, where the fitness function value comprehensively evaluates three indicators: signal-to-noise ratio gain, signal quality index, and feature discrimination, to obtain the population fitness distribution; according to the population fitness distribution, retain the 10% of individuals with the highest fitness to enter the next generation, and use a tournament for the remaining 90% of individuals. The competition selects to perform crossover and mutation, with the crossover rate initially set at 0.8 and dynamically adjusted according to population diversity, and the mutation rate initially set at 0.1 and gradually reduced to 0.01 with the number of iterations, to obtain a new generation of population. Every 10 generations, the 10% of individuals with the highest fitness in the new generation of population are subjected to simulated annealing for local optimization until a preset number of iterations is reached to obtain the optimal parameter configuration. Based on the optimal parameter configuration, the PPG signal of each trainee is preprocessed and feature extracted in a personalized manner to generate optimized PPG signals and features. An incremental optimization mechanism is set up so that when the recognition accuracy drops beyond a preset threshold or reaches a fixed optimization cycle, parameter re-optimization is triggered, and the optimization process is accelerated by parallel computing.

[0193] Step S6: Obtain the initial fusion weights and dimensionality reduction matrix of the Mel frequency cepstral coefficient features and the Gammatone frequency cepstral coefficient features, and construct a bi-objective optimization problem for feature fusion. The first objective is to maximize the Fisher discrimination ratio, and the second objective is to minimize the feature dimension. Based on the bi-objective optimization problem, the NSGA-II algorithm is used to generate an initial population. Each individual contains two parts: feature fusion weights and dimensionality reduction. The fusion weights range from [0,1] and sum to 1, while the dimensionality reduction ranges from [8,20]. The initial population is subjected to non-dominated sorting and crowding calculation. A binary tournament selection is used to generate the parent population, employing a simulated binary selection process. The offspring population is generated through number crossover and polynomial mutation. Every 10 generations, a local search is performed on the solutions on the non-dominated front, with the search step size dynamically adjusted according to the number of iterations to accelerate algorithm convergence and obtain the Pareto optimal solution set. Based on the Pareto optimal solution set, the optimal solution is selected using the knee detection method. The knee detection method first sorts the solution set in ascending order according to the first objective, then calculates the curvature of adjacent solutions, selects the point with the largest curvature as the knee point, and generates a personalized feature fusion model. An adaptive adjustment mechanism is set to monitor the recognition performance. When the cumulative error rate exceeds 5% or 30 days have passed, the feature fusion model is automatically updated to ensure performance stability.

[0194] Step S7: Construct the i-vector framework parameter space and the decision model parameter space. The i-vector framework parameter space includes the number of Gaussian mixture components, the rank of the T matrix, the dimension of the feature vectors, and the number of iterations. The decision model parameter space includes the number of layers of the LSTM network, the number of hidden units, the dropout rate, and the learning rate. The i-vector framework parameter space and the decision model parameter space are optimized using the Covariance Matrix Adaptive Evolutionary Strategy (CMA-ES). Interactive optimization between subspaces is achieved through a three-layer collaborative mechanism of parameter cross-evaluation, fitness sharing, and search direction coordination. The computational resources allocated to each subspace are dynamically adjusted based on the progress speed and performance improvement of the interactive optimization of the subspaces. Resources are utilized, and multi-threading technology is employed to simultaneously evaluate different parameter combinations. For different training scenarios, typical scenario configurations are predefined and optimization processes are executed separately to generate scenario-specific parameter configuration schemes. Parameter configurations are automatically switched based on real-time detected training scenarios. A regularization control strategy is adopted, adding L2 regularization constraints to the i-vector subspace to control the complexity of the T matrix. Batch normalization and early stopping strategies are introduced in the decision model subspace to improve the model's generalization ability. Incremental optimization is achieved using a two-level strategy: a lightweight update strategy and a full update strategy. The lightweight update strategy fine-tunes parameters within a local range, while the full update strategy re-executes a complete co-evolutionary optimization after accumulating new data exceeding a preset data value.

[0195] Example 4

[0196] This embodiment provides a parameter optimization framework based on evolutionary algorithms to enhance the effects of PPG signal preprocessing and feature extraction.

[0197] Specifically, after acquiring the PPG signal data of the trainees, the parameters of the high-pass filter, low-pass filter, and extraction parameters of Mel frequency cepstral coefficient features and Gammatone frequency cepstral coefficient features were first obtained. A parameter optimization framework based on a genetic algorithm was constructed, and the high-pass filter parameters, low-pass filter parameters, and extraction parameters were encoded into chromosomes of the initial population. For the high-pass filter, the encoded content included the cutoff frequency (0.3-0.7Hz), filter order (2-6), and filter type (Chebyshev, Butterworth, or elliptic); for the low-pass filter, the encoded content included the cutoff frequency (15-25Hz), filter order (3-5), and filter type. Simultaneously, the parameters in the MFCC and GFCC feature extraction process were also encoded, including the number of Mel filters (20-40), the number of cepstral coefficients (8-16), frame length (20-30ms), and frame shift (10-15ms). The initial population randomly generated 100 chromosomes, covering different regions of the parameter space to ensure diversity.

[0198] Based on the chromosomes of the initial population, the fitness function value of each chromosome is calculated. This fitness function value comprehensively evaluates three metrics: signal-to-noise ratio gain, signal quality index, and feature discrimination, resulting in the population fitness distribution. The signal-to-noise ratio gain is calculated by comparing the signal-to-noise power ratio before and after filtering. The signal quality index is evaluated based on the periodicity and morphological characteristics of the PPG waveform, including peak detection rate, waveform symmetry, and baseline stability. Feature discrimination is measured by calculating the ratio of intra-class distance to inter-class distance, using the Fisher discriminant ratio as the quantification metric. The overall fitness function value is defined as:

[0199]

[0200] Among them, the weighting coefficient , , Adjustments are made according to the application scenario; in this solution, the values ​​are set to 0.3, 0.3, and 0.4.

[0201] Based on the population fitness distribution, the top 10% of individuals with the highest fitness are retained for the next generation. The remaining 90% are subjected to tournament selection for crossover and mutation. The crossover rate is initially set to 0.8 and dynamically adjusted based on population diversity. The mutation rate is initially set to 0.1 and gradually decreases to 0.01 with each iteration, resulting in the next generation population. Adaptive arithmetic crossover is used, with an initial crossover rate of 0.8 that is dynamically adjusted based on population diversity. Gaussian mutation is used, with an initial mutation rate of 0.1 that is gradually decreased to 0.01 with each iteration to balance global exploration and local exploitation capabilities.

[0202] Every 10 generations, simulated annealing is performed on the top 10% of fittest individuals in the new generation population for local optimization until a preset number of iterations is reached to obtain the optimal parameter configuration. Local optimization helps to accelerate convergence, finely explore the neighborhood space of high-quality solutions, and improve the quality of the final solution.

[0203] Based on the optimal parameter configuration, personalized preprocessing and feature extraction are performed on the PPG signal of each trainee to generate optimized PPG signals and features. This personalized processing can adapt to the differences in physiological characteristics among different trainees, improving the targeting and effectiveness of feature extraction.

[0204] Finally, an incremental optimization mechanism is implemented. When the recognition accuracy drops beyond a preset threshold or reaches a fixed optimization cycle, parameter re-optimization is triggered, and the optimization process is accelerated through parallel computing. In this way, the system can continuously adapt to changes in the physiological state of trainees and environmental conditions, maintaining stable recognition performance over the long term.

[0205] Experimental results show that, compared with the fixed-parameter scheme, the PPG signal processing scheme optimized by this evolutionary algorithm improves noise resistance by 32%, feature discrimination by 28%, and overall recognition accuracy by 3.5 percentage points. Particularly in high-intensity motion scenarios (T3 type), the recognition accuracy increased from 92.7% to 97.8%, verifying the superior performance of this scheme in complex training environments.

[0206] Example 5

[0207] This embodiment provides a feature fusion strategy based on multi-objective optimization to enhance the fusion effect of MFCC and GFCC features.

[0208] Specifically, firstly, the initial fusion weights and dimensionality reduction matrices of the Mel frequency cepstral coefficient features and the Gammatone frequency cepstral coefficient features are obtained. A bi-objective optimization problem for feature fusion is then constructed, with the first objective being to maximize the Fisher discrimination ratio and the second objective being to minimize the feature dimension. Let the MFCC feature vector be X. MThe GFCC eigenvector is X G The fusion weight vector is W=[w M , w G The goal is to find the optimal weights. The dimensionality reduction matrix P is used to ensure that the fused feature Z maximizes both the inter-class distance and minimizes the intra-class distance. The Fisher discriminant ratio is used as a measure of the discriminative power of the feature. Considering the feature dimension as a complexity constraint, a bi-objective optimization problem is formed: maximizing the discriminative power while minimizing the feature dimension.

[0209] Before implementing multi-objective optimization, Z-score standardization was performed on MFCC and GFCC features to ensure comparability between different features during the fusion process. Simultaneously, outlier detection was performed on each feature class, and an improved Local Outlier Factor (LOF) algorithm was used to identify and process anomalous samples, enhancing the stability of feature distribution. Furthermore, a feature correlation matrix was calculated to identify redundant information between MFCC and GFCC features, providing a reference for subsequent fusion.

[0210] Based on the aforementioned bi-objective optimization problem, the NSGA-II algorithm is used to generate the initial population. Each individual comprises two parts: feature fusion weights and a reduced dimension. The fusion weights range from [0,1] and sum to 1, while the reduced dimension ranges from [8,20]. The algorithm initializes 100 solutions, each containing a weight vector W and a reduced dimension k. Individual encoding uses real-number encoding, where w... M ,w G ∈[0,1] and w M +w G =1, k∈{8,9,...,20}. The algorithm runs for 200 generations, with the crossover probability set to 0.9 and the mutation probability set to 1 / n (where n is the number of decision variables).

[0211] The initial population is subjected to non-dominated sorting and crowding calculation. Parent populations are generated through binary tournament selection, and offspring populations are generated using simulated binary crossover and polynomial mutation. Every 10 generations, a local search is performed on the solutions at the non-dominated front, with the search step size dynamically adjusted with the number of iterations to accelerate algorithm convergence and obtain the Pareto optimal solution set. This non-dominated sorting and crowding calculation mechanism ensures the diversity and convergence of multi-objective optimization, enabling the discovery of a series of high-quality solutions representing different trade-offs.

[0212] During the optimization process, the first objective function evaluates the feature discrimination capability using an improved Fisher discriminant ratio. This metric considers not only the global inter-class and intra-class scatter matrices but also incorporates local nearest neighbor structural information, ensuring that the fused features retain local manifold structure while maintaining global discriminative capability. The second objective function evaluates feature complexity, considering not only the dimension k but also feature information entropy as an auxiliary metric to comprehensively measure complexity. During the evaluation of each solution, five-fold cross-validation is used to calculate the recognition performance index, ensuring the reliability of the evaluation results.

[0213] Based on the Pareto optimal solution set, a knee point detection method is used to select the optimal solution. This method first sorts the solution set in ascending order according to the first objective, then calculates the curvature of adjacent solutions, and selects the point with the largest curvature as the knee point to generate a personalized feature fusion model. This method first sorts the solutions on the frontier in ascending order according to the first objective (discriminative ability), then calculates the curvature between adjacent solutions, and selects the point with the largest curvature as the knee point. This point represents the location where marginal returns change significantly, representing the best choice to balance the two objectives.

[0214] Based on the selected optimal solution, the system generates a personalized feature fusion model. This model consists of two parts: a feature weighting module and an adaptive dimensionality reduction module. The feature weighting module weights the MFCC and GFCC features according to the optimized weights; the adaptive dimensionality reduction module projects the weighted features into a lower-dimensional space using the optimized dimensionality reduction matrix. Unlike traditional PCA, the dimensionality reduction matrix is ​​constructed by maximizing the preservation of class information, considering both variance contribution rate and class discriminative ability, thus generating a more discriminative low-dimensional representation.

[0215] Finally, an adaptive adjustment mechanism is set up to monitor recognition performance. When the cumulative error rate exceeds 5% or 30 days have elapsed, the feature fusion model is automatically updated to ensure performance stability. During operation, the system continuously monitors recognition performance and automatically triggers model updates when the cumulative error rate exceeds a preset threshold or a fixed time interval has elapsed. During the update process, multi-objective optimization is re-executed based on newly collected data, adjusting feature fusion weights and dimensionality reduction parameters to ensure the system can adapt to long-term changes in the physiological state of trainees and environmental conditions.

[0216] Experimental results show that, compared with the original scheme, the feature fusion strategy based on multi-objective optimization improves the recognition accuracy by an average of 4.2 percentage points, with the most significant improvement in the T2 type (moderate intensity motion) scene, increasing from 94.5% to 99.1%. In addition, the feature dimension is reduced by an average of 35%, significantly reducing the computational burden of the subsequent i-vector model, verifying the effectiveness of the scheme in improving recognition performance while optimizing the efficiency of computational resource utilization.

[0217] Example 6

[0218] This embodiment provides a method for optimizing i-vector and model parameters based on co-evolution, which can enhance the performance of the i-vector framework and decision model.

[0219] Specifically, the i-vector framework parameter space and the decision model parameter space are constructed first. The i-vector framework parameter space includes the number of Gaussian mixture components, the rank of the T matrix, the dimension of the feature vector, and the number of iterations. The decision model parameter space includes the number of layers, the number of hidden units, the dropout rate, and the learning rate of the LSTM network. The first subspace is the i-vector framework parameter space, which includes parameters such as the number of Gaussian mixture components (range 16-128), the rank of the T matrix (range 100-400), the dimension of the feature vector (range 8-20), and the number of iterations (range 5-15) of the UBM. The second subspace is the decision model parameter space, which mainly covers parameters such as the number of layers (range 1-4), the number of hidden units (range 32-256), the dropout rate (range 0.1-0.5), and the learning rate (range 0.0001-0.01) of the LSTM network.

[0220] The i-vector framework parameter space and the decision model parameter space are optimized using the Covariance Matrix Adaptive Evolutionary Strategy (CMA-ES). Interactive optimization between subspaces is achieved through a three-layer collaborative mechanism: parameter cross-evaluation, fitness sharing, and search direction coordination. CMA-ES efficiently handles complex nonlinear and multimodal optimization problems by adaptively adjusting the covariance matrix of the search distribution. For the i-vector subspace, the initial population size is 16, and the initial step size is 0.3; for the decision model subspace, the initial population size is 20, and the initial step size is 0.25. To handle the heterogeneity of different parameters, each parameter is normalized to distribute it within the [0,1] interval, and then mapped back to its original value range in practical applications.

[0221] In the process of co-evolution, a three-layer cooperation mechanism enables effective interaction between subspaces. The first layer is parameter cross-evaluation, where elite individuals from each subspace combine with elite individuals from another subspace to jointly evaluate system performance and form a cross-fitness matrix. The second layer is fitness sharing, where the populations of the two subspaces update their individual fitness based on the cross-fitness matrix, allowing the optimization process to take into account the mutual influence between parameters. The third layer is search direction coordination, where the CMA-ES algorithms of the two subspaces adjust their own covariance matrices according to the evolutionary trend of the other subspace, achieving coordinated adjustment of the search direction.

[0222] Based on the progress and performance improvement of the interactive optimization of subspaces, the computing resources allocated to each subspace are dynamically adjusted, and multi-threading technology is used to evaluate different parameter combinations simultaneously. When the optimization of a subspace enters a plateau, its computing resources are temporarily reduced, and more resources are allocated to subspaces with greater improvement potential; when a subspace finds a significantly better solution, its resource allocation is increased to accelerate the exploration of the parameter space around that region.

[0223] For different training scenarios, typical scenario configurations are predefined and optimized separately to generate scenario-specific parameter configuration schemes. The system automatically switches parameter configurations based on real-time detected training scenarios. The system predefines multiple typical scenario configurations (such as high-intensity physical training, fine motor skills training, and long-duration endurance training), performs optimization processes separately for each scenario, and generates scenario-specific parameter configuration schemes. During deployment, the system can automatically switch parameter configurations based on real-time detected training scenarios, or gradually adjust parameters through online learning to adapt to new scenarios.

[0224] A regularization control strategy is employed. In the i-vector subspace, L2 regularization constraints are added to control the complexity of the T matrix. In the decision model subspace, batch normalization and early stopping strategies are introduced to improve the model's generalization ability. In the i-vector subspace, L2 regularization constraints are added to control the complexity of the T matrix, avoiding overfitting to the training data. In the decision model subspace, in addition to the conventional dropout mechanism, batch normalization and early stopping strategies are introduced to jointly enhance the model's generalization ability.

[0225] Incremental optimization is achieved using a two-tiered strategy: a lightweight update strategy and a full update strategy. The lightweight update strategy fine-tunes parameters within a local range, while the full update strategy re-executes a complete co-evolutionary optimization after accumulating more new data than a preset value. After the initial parameter optimization is completed, the system enters an online learning phase, periodically collecting new PPG data to incrementally update the model. The update process employs a two-tiered strategy: lightweight updates use the current parameters as a starting point, fine-tuning them within a small range to adapt to new data; full updates, after accumulating a sufficient amount of new data, re-execute a complete co-evolutionary optimization.

[0226] Experimental results show that, compared with the original scheme, the parameter optimization scheme based on co-evolution improves the overall recognition accuracy by 5.3 percentage points, reaching an average of 99.4%. The improvement is particularly significant when handling complex dynamic scenes (such as T3 type high-intensity motion), increasing the recognition accuracy from 92.7% to 98.9%. Simultaneously, the optimized system reduces training time by 47% and inference latency by 63%, comprehensively verifying the superior effectiveness of this scheme in improving system performance while optimizing computational resource utilization efficiency.

[0227] Example 7

[0228] This embodiment provides an experiment and analysis of a biometric identification method driven by PPG big data.

[0229] To evaluate performance, this invention used a dataset from 30 trainees to verify the generalization ability of the proposed model for PPG biometric recognition. This invention employs a sliding window processing strategy based on a 4-second analysis window (500 sampling points, 125Hz sampling rate), with a window step size of 25 sampling points (0.2-second offset), and implements an improved hierarchical 5-fold cross-validation scheme. The specific implementation process is as follows:

[0230] 1. Data preprocessing and window division:

[0231] For each 20-minute PPG signal (150,000 sampling points), a rectangular window was used to strictly extract 500 consecutive sampling points (4 seconds), with adjacent windows overlapping by 475 sampling points (3.8 seconds), resulting in approximately 2,370 analysis windows.

[0232] 2. Hierarchical grouping strategy:

[0233] First tier: Grouped by age (18-30 / 31-45 / 46-58 years old, 10 people in each tier);

[0234] Second level: Classified by activity intensity (T1 = low intensity, T2 = medium intensity, T3 = high intensity);

[0235] Third layer: subjects with special markings overlapping (9 / 10 / 21 / 22);

[0236] An improved K-means clustering algorithm was used, with feature space distance as the metric, to ensure that each fold contains: the same proportion of windows for each age group (±2%); the same proportion of windows for each activity type (T1:T2:T3≈1:1:1); and all windows of overlapping subjects are completely preserved in a single fold.

[0237] 3. Implementation of cross-validation:

[0238] a) All windows (71,100 in total) of the 30 participants were stratified and divided into 5 folds, such as... Figure 9 As shown;

[0239] b) Each fold contains: approximately 14,220 windows (20%), complete records from 6 subjects; number of windows for each activity type: T1≈4,740, T2≈4,740, T3≈4,740;

[0240] c) Training and validation datasets: Training set: 4 folds (56,880 windows); Test set: 1 fold (14,220 windows).

[0241] The metrics used for evaluation are:

[0242] (1) Accuracy. The ratio of correctly classified observations (to be identified) to the total number of observations.

[0243] (2) Precision. The ratio of correctly predicted positive observations to the total number of predicted positive observations.

[0244] (3) Recall. The ratio of correctly predicted positive observations to all observations in the actual class.

[0245] (4) F1 score. The weighted average of precision and recall.

[0246] The accuracy, precision, recall, and F1 score of the 5x cross-validation (CV) test on the dataset for 30 participants are shown in Table 4. Several commonly used machine learning and deep learning methods were employed for selecting effective classifiers for classifying the i-vector, including Support Vector Machine (SVM), Multilayer Perceptron (MLP), Random Forest (RF), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM).

[0247] Table 4 Evaluation Indicators of Various Models under Different Decision-Making Methods

[0248]

[0249] Table 4 clearly shows that the MFCC+GFCC fusion feature and LSTM classification model achieved the best recognition rate, approximately 3.1% higher than the second-ranked MFCC+GFCC fusion feature and Random Forest (RF) classification model. Compared to the other two decision models, Multilayer Perceptron (MLP) and Convolutional Neural Network (CNN), the rates were improved by 6.3% and 3.6% respectively, and significantly higher than the traditional SVM model by approximately 7.7%. Therefore, it can be seen that adopting a more reasonable decision mechanism further improves the overall classification ability of the model. By comparing the recognition success rates before and after model optimization with the same feature values, it is evident that the optimized model consistently outperforms the unoptimized model, further enhancing the classification ability of traditional cosine kernel and SVM models.

[0250] To verify the effectiveness of our proposed method, we selected three traditional neural network models that directly process one-dimensional time series data—1D-CNN, LSTM, and Transformer—and conducted qualitative and quantitative comparative experiments on the same dataset. 1D-CNN extracts local features from the original time series data through one-dimensional convolutional operations and completes the classification task through fully connected layers. Its model structure is relatively simple and suitable for processing time series data with significant local features, but it has limitations in capturing long-term dependencies. LSTM is a recurrent neural network model with unique gating mechanisms (input gate, forget gate, output gate) and memory units, effectively capturing long-term dependencies and dynamic changes in time series data, thus performing well in handling time-dependent tasks. In contrast, the Transformer model, based on a self-attention mechanism, can efficiently capture global dependencies in time series data. Unlike traditional recurrent neural networks, the Transformer model utilizes a multi-head self-attention mechanism to simultaneously model dependencies between different time steps, making it suitable for tasks involving long-term time series data. Table 5 shows the performance of these four networks in terms of accuracy, precision, recall, and F1 score.

[0251] Table 5 Performance comparison between the proposed model and traditional neural network models

[0252]

[0253] As shown in Table 5, the proposed method using MFCC+GFCC fusion features outperforms traditional one-dimensional neural network models (1D-CNN, LSTM, and Transformer) that directly input PPG signals in all performance metrics. The 1D-CNN model has relatively lower metrics, with accuracy and F1 score both around 92%. In contrast, the LSTM model shows improvement in accuracy (94.53%) and F1 score (94.39%), while the Transformer model shows further improvement compared to the LSTM model but is still significantly lower than the proposed method. This result not only verifies the superior performance of the proposed model but also demonstrates the effectiveness of the method for making decisions based on extracting i-vectors from MFCC+GFCC fusion feature input. It improves model performance in identity recognition tasks by more effectively extracting and utilizing key features in the data. This further proves the enormous potential and value of improving model performance by transforming data representation in the field of signal processing.

[0254] To verify the performance of the proposed method for decision-making based on MFCC+GFCC fusion feature input extraction of i-vectors, ablation experiments were conducted. The experiments aimed to verify whether the various improvements to the model resulted in enhanced accuracy (Acc), precision (Pre), recall, and F1 score for identity recognition based on PPG signals.

[0255] Table 6 compares the performance differences between using MFCC features alone, GFCC features alone, and MFCC+GFCC fusion features. Compared to using different feature values ​​with the same SVM model, the model using MFCC+GFCC fusion features as input showed an average 2.1% improvement in accuracy and an average 6.4% improvement in recall compared to using MFCC and GFCC features alone. This indicates that the fusion feature values ​​used in this paper retain more voiceprint information, which is more conducive to improving the accuracy of biometric recognition.

[0256] Table 6 Ablation Experiment Results

[0257]

[0258] This invention proposes a novel biometric authentication method based on photoplethysmography (PPG), which utilizes the i-vector framework for robust verification. The i-vector method compresses the feature vector sequence into a compact low-dimensional representation, thereby capturing the overall variability of the PPG signal. After reducing PPG noise through preprocessing, feature vectors are extracted using Mel-frequency cepstral coefficients (MFCC) and Gammatone-frequency cepstral coefficients (GFCC). Although these two techniques were originally designed for audio processing, MFCC captures the overall spectral envelope of the signal, while GFCC emphasizes finer spectral details in the low-frequency range through a Gammatone filter bank. These features form the basis of a Universal Background Model (UBM) based on a Gaussian mixture model, from which a low-dimensional i-vector authentication vector is derived. Finally, optimal authentication performance is achieved by inputting the i-vector into a Long Short-Term Memory (LSTM) network. The results show that, on a dataset collected from 30 trainees simulating three motion scenarios, the proposed method achieves average accuracy, precision, recall, and F1 score of 98.7%, 98.4%, 98.5%, and 98.4% respectively after 5x cross-validation. The proposed model significantly outperforms traditional pure neural network models in biometric identification. Experimental results demonstrate that this method exhibits excellent biometric recognition performance, possesses strong application value in training practice, and can provide a high-precision identity authentication solution for wearable devices.

[0259] 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, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A biometric identification method driven by PPG big data, characterized in that, include: Acquire PPG signal data from trainees, and perform high-pass and low-pass filtering preprocessing on the PPG signal data to obtain preprocessed PPG signals; The preprocessed PPG signal is framed, and Mel frequency cepstral coefficient features and Gammatone frequency cepstral coefficient features are extracted respectively. The Mel frequency cepstral coefficient features and Gammatone frequency cepstral coefficient features are fused and then principal component analysis is performed to reduce the dimensionality, resulting in fused features. A general background model (UBM) is constructed based on the fusion features. Zero-order, first-order, and second-order Baum-Welch statistics are calculated. The global difference space matrix is ​​estimated from the Baum-Welch statistics, and an i-vector identity authentication vector is extracted, including: Based on the aforementioned fusion features, the number of Gaussian mixture components is set to 32, and the covariance matrix Σ is a diagonal matrix. A general background model UBM is constructed, and the UBM mean supervector m is obtained. The global difference space matrix T is estimated iteratively using the EM algorithm, and the zeroth-order statistic is calculated in each iteration: , First-order statistics: , And second-order statistics: ,in It is the mean hypervector in the UBM model. The first in The mean of the Gaussian components, It is the dimension of the input PPG signal features. That is, the first Frame feature vector; In step E, calculate the first and second expectations of the total variable difference factor w: in It is the inverse covariance matrix of the i-vector posterior distribution. It is the identity matrix. It is the diagonal covariance matrix of the variances of the mixture components containing UBM on the diagonal; In step M, update the global difference space matrix T and calculate the log-likelihood value. ,when Stop iteration when the time is right; Based on the final global difference space matrix T, using the formula: Extract the 200-dimensional i-vector authentication vector w; The i-vector identity authentication vector is input into a long short-term memory network for training and classification to obtain the identity recognition results of the trainees.

2. The method according to claim 1, characterized in that, The preprocessing of the PPG signal data by high-pass filtering and low-pass filtering includes: The PPG signal data was configured with filter parameters. The cutoff frequency of the fourth-order Chebyshev Type I high-pass filter was set to 0.5Hz, the passband ripple was set to within ±0.5 dB, and the stopband attenuation slope was set to -80 dB per decade. The high-pass filter parameters were obtained. The fourth-order Chebyshev Type I high-pass filter is constructed based on the high-pass filter parameters. The fourth-order Chebyshev Type I high-pass filter is applied to the PPG signal data to eliminate baseline drift and obtain the first filtered signal. Based on the time-domain and frequency-domain characteristics of the first filtered signal, the filtering effect is evaluated, the signal-to-noise ratio and baseline drift suppression are calculated, and the preset signal quality indicators are ensured to be met.

3. The method according to claim 2, characterized in that, The preprocessing of the PPG signal data by high-pass filtering and low-pass filtering includes: Configure the filter parameters for the first filtered signal. Set the cutoff frequency of the fourth-order elliptic low-pass filter to 18Hz, set the passband to 0Hz to 18Hz with fluctuation less than or equal to 0.1 dB, and set the stopband to start from 22Hz with attenuation greater than or equal to 40 dB at 22Hz and attenuation greater than or equal to 60 dB at 50Hz to obtain the low-pass filter parameters. The fourth-order elliptic low-pass filter is constructed based on the low-pass filter parameters. The fourth-order elliptic low-pass filter is applied to the first filtered signal to eliminate motion artifacts, ambient light interference and muscle noise, and to obtain the preprocessed PPG signal. Based on the spectral characteristics of the preprocessed PPG signal, the noise suppression effect is evaluated, the high-frequency interference suppression level is calculated, and the preset signal quality requirements are ensured to be met.

4. The method according to claim 1, characterized in that, The step of segmenting the preprocessed PPG signal into frames and extracting Mel-frequency cepstral coefficient features and Gammatone-frequency cepstral coefficient features respectively includes: The preprocessed PPG signal is pre-emphasized to enhance the high-frequency components, resulting in a pre-emphasized signal. The pre-emphasized signal is divided into frames according to a preset frame length, and a Hamming window is added to each frame signal for smooth transition to obtain the first windowed signal. The first windowed signal is subjected to a fast Fourier transform to obtain the spectrum, and the spectrum is input into a Mel filter bank for frequency envelope extraction to obtain the Mel spectrum; The Mel spectrum is subjected to discrete cosine transform to compress the feature dimension, thereby obtaining the Mel frequency cepstral coefficient features.

5. The method according to claim 1, characterized in that, The step of segmenting the preprocessed PPG signal into frames and extracting Mel-frequency cepstral coefficient features and Gammatone-frequency cepstral coefficient features respectively includes: The preprocessed PPG signal is divided into frames, and a Hamming window is applied to each frame to obtain a second windowed signal. Perform a Fast Fourier Transform on the second windowed signal to obtain the frequency domain signal; The frequency domain signal is input into the Gammatone filter bank, and the frequency band is divided according to the ERB frequency scale to obtain the filtered output of each channel. Energy calculation and logarithmic compression are performed on the filtered output of each channel, and discrete cosine transform is performed to obtain the Gammatone frequency cepstral coefficient characteristics.

6. The method according to claim 3, characterized in that, Also includes: The high-pass filter parameters, low-pass filter parameters, and extraction parameters of the Mel frequency cepstral coefficient features and Gammatone frequency cepstral coefficient features are obtained. A parameter optimization framework based on a genetic algorithm is constructed, and the high-pass filter parameters, low-pass filter parameters, and extraction parameters are encoded as chromosomes of the initial population. Based on the chromosomes of the initial population, the fitness function value of each chromosome is calculated, wherein the fitness function value comprehensively evaluates three indicators: signal-to-noise ratio gain, signal quality index, and feature discrimination, to obtain the population fitness distribution; Based on the population fitness distribution, the top 10% of individuals with the highest fitness are retained for the next generation. The remaining 90% of individuals are subjected to tournament selection for crossover and mutation. The crossover rate is initially set to 0.8 and dynamically adjusted according to the population diversity. The mutation rate is initially set to 0.1 and gradually decreases to 0.01 with the number of iterations, thus obtaining the new generation population. Every 10 generations, simulated annealing is performed on the 10% of individuals with the highest fitness in the new generation population until a preset number of iterations is reached to obtain the optimal parameter configuration. Based on the optimal parameter configuration, personalized preprocessing and feature extraction are performed on the PPG signal of each trainee to generate optimized PPG signal and features. An incremental optimization mechanism is set up so that when the recognition accuracy drops below a preset threshold or reaches a fixed optimization cycle, the parameters are re-optimized, and the optimization process is accelerated through parallel computing.

7. The method according to claim 1, characterized in that, Also includes: Obtain the initial fusion weights and dimensionality reduction matrix of the Mel frequency cepstral coefficient features and the Gammatone frequency cepstral coefficient features, and construct a bi-objective optimization problem for feature fusion. Set the first objective to maximize the Fisher discrimination ratio and the second objective to minimize the feature dimension. Based on the aforementioned bi-objective optimization problem, the NSGA-II algorithm is used to generate the initial population. Each individual contains two parts: feature fusion weights and dimensionality reduction. The fusion weights range from [0,1] and sum to 1, while the dimensionality reduction ranges from [8,20]. The initial population is sorted by non-dominated order and crowding is calculated. Parent population is generated by binary tournament selection and offspring population is generated by simulated binary crossover and polynomial mutation. Every 10 generations, a local search is performed on the solution on the non-dominated front. The search step size is dynamically adjusted with the number of iterations to accelerate the convergence of the algorithm and obtain the Pareto optimal solution set. Based on the Pareto optimal solution set, the optimal solution is selected using the knee point detection method. The knee point detection method first sorts the solution set in ascending order according to the first objective, then calculates the curvature of adjacent solutions, selects the point with the largest curvature as the knee point, and generates a personalized feature fusion model. An adaptive adjustment mechanism is set up to monitor recognition performance. When the cumulative error rate exceeds 5% or 30 days have passed, the feature fusion model is automatically updated to ensure stable performance.

8. The method according to claim 1, characterized in that, Also includes: Construct the i-vector framework parameter space and the decision model parameter space. The i-vector framework parameter space includes the number of Gaussian mixture components, the rank of the T matrix, the dimension of the feature vector, and the number of iterations. The decision model parameter space includes the number of layers of the LSTM network, the number of hidden units, the dropout rate, and the learning rate. The parameter space of the i-vector framework and the parameter space of the decision model are optimized using the covariance matrix adaptive evolution strategy CMA-ES. Interactive optimization between subspaces is achieved through a three-layer collaborative mechanism of parameter cross-evaluation, fitness sharing and search direction coordination. Based on the progress and performance improvement of the interaction optimization of the subspaces, the computing resources allocated to each subspace are dynamically adjusted, and multi-threading technology is used to evaluate different parameter combinations simultaneously. For different training scenarios, predefine typical scenario configurations and perform optimization processes separately to generate scenario-specific parameter configuration schemes, and automatically switch parameter configurations based on the training scenarios detected in real time. A regularization control strategy is adopted, adding L2 regularization constraints to the i-vector subspace to control the complexity of the T matrix, and introducing batch normalization and early stopping strategies in the decision model subspace to improve the model's generalization ability. Incremental optimization is achieved through a two-tiered strategy: a lightweight update strategy and a full update strategy. The lightweight update strategy fine-tunes parameters within a local range, while the full update strategy re-executes a complete co-evolutionary optimization after the accumulated new data exceeds a preset data value.

9. The method according to claim 1, characterized in that, The step of inputting the i-vector identity authentication vector into a long short-term memory network for training and classification to obtain the identity recognition result of the trainee includes: The i-vector identity authentication vector is normalized and time-series rearranged to generate a training sample sequence and corresponding identity labels; Based on the training sample sequence and the corresponding identity label, a network structure containing an input layer, two LSTM layers and a fully connected layer is constructed, wherein the number of LSTM units in the first layer is 128 and the number of LSTM units in the second layer is 64, and the dropout rate is set to 0.3 to obtain the initial network model. Cross-entropy was used as the loss function, and the Adam optimizer was used to iteratively train the initial network model. The initial learning rate was set to 0.001 and dynamically adjusted using a cosine annealing strategy to obtain the trained LSTM network model. The i-vector identity authentication vector to be identified is input into the trained LSTM network model, and the probability distribution of each identity category is calculated by the Softmax classifier. The category with the highest probability is selected as the recognition result.

Images