Photoplethysmography identity recognition method and system

Abstract

The invention provides a photoelectric volume pulse wave identity recognition method and system, and relates to the technical field of identity recognition, and the method comprises the steps: obtaining a to-be-recognized PPG signal; and inputting the PPG signals into a trained identification model, firstly extracting linear features and nonlinear features, then projecting the features fused by the two features into a feature space by using a learned discriminant projection matrix to obtain multi-view features, and finally classifying the PPG signals by using the multi-view features to obtain an identity identification result. According to the method, manifold regularization and inter-class-error double sparse constraints are combined, the problems of intra-class discretization and inter-class overlapping of a linear model are solved, a graph structure learning method for adaptive local density adjustment is provided, and the manifold modeling precision of non-stationary PPG signals is improved.

Description

Technical Field

[0001] This invention relates to the field of identity recognition technology, specifically to a photoplethysmography (PPG) identity recognition method and system. Background Technology

[0002] Photoplethysmography (PPG) signal identification is a biometric technology based on human physiological characteristics. It distinguishes individual identities by analyzing the periodic changes in blood volume under the skin (i.e., PPG signals). Its core principle is that each person's cardiovascular system (such as vascular elasticity, microcirculation characteristics, etc.) is unique, resulting in individual differences in the waveform, frequency, and dynamic characteristics of the PPG signal.

[0003] Existing PPG identity verification technology has the following clear defects and the resulting technical problems: 1. Linear models fail to model nonlinear manifolds. Existing methods (such as least squares regression LSR) are based on the Euclidean linearity assumption. However, due to individual physiological differences (such as vascular elasticity and age-related pulsatility characteristics) and acquisition noise (motion artifacts and equipment baseline drift), PPG signals exhibit a local nonlinear manifold distribution in a high-dimensional feature space, which leads to two major technical problems: Inter-class overlap: Signals from different individuals intersect in the projection space, causing the false recognition rate of the traditional LSR method to be as high as 64.00%. Intra-class discrepancy: The manifold structure of samples from the same subject is broken, resulting in a loose distribution of similar samples in the feature space and poor recognition consistency.

[0004] 2. Model overfitting under small sample conditions Due to limitations in the sampling frequency and storage capacity of wearable devices, the actual training samples for each user may be less than 50, while the signal feature dimensions exceed 1000. Traditional deep learning models (such as CNN-LSTM) have a large number of parameters (over 100,000 parameters), which significantly increase the risk of overfitting with small samples. When the training samples are reduced to 10 per person, the recognition accuracy decreases.

[0005] 3. The contradiction between robustness and discriminativeness in a single regularization strategy Existing identification methods that employ norm smoothing constraints (such as the DLSR method) have fundamental flaws: Oversuppression problem: Individual-specific harmonic components (such as the difference in second harmonics of pulse waves from different people) are misclassified as noise, resulting in loss of inter-class discriminative features and a lower macro-average F1-score of the LSR method; Undersuppression problem: Motion artifacts cannot be effectively filtered when they overlap with the signal spectrum, resulting in a noise retention rate of over 40%, leading to an overall false recognition rate of 6.71% for the PCA+SVM method.

[0006] Therefore, existing technologies are insufficient in terms of recognition accuracy, discrimination ability, and anti-interference performance. Summary of the Invention

[0007] To address the aforementioned problems, this invention proposes a photoplethysmography (PPG) pulse wave identification method and system. It combines manifold regularization with intra-class error dual sparsity constraints to solve the intra-class discretization and inter-class overlap problems of linear models. Furthermore, it proposes an adaptive local density adjustment graph structure learning method to improve the manifold modeling accuracy of non-stationary PPG signals.

[0008] According to some embodiments, the present invention adopts the following technical solution: A photoplethysmography (PPG) identification method includes: Acquire the PPG signal to be identified; The PPG signal is input into the trained recognition model. First, linear and nonlinear features are extracted and concatenated into fusion features. Then, the learned discriminant projection matrix is ​​used to project the fusion features into a low-dimensional discriminant space. Finally, the projected features are used to classify the PPG signal to obtain the identity recognition result. The training of the recognition model involves introducing manifold regularization and graph Laplacian constraints to construct a graph Laplacian matrix for the training sample set. Based on the graph Laplacian matrix, the recognition model is jointly optimized using in-class group sparse constraints and error matrix row sparse constraints to obtain the optimal discriminant projection matrix.

[0009] According to some embodiments, the present invention adopts the following technical solution: A photoplethysmography (PPG) identification system includes: The acquisition module is configured to acquire the PPG signal to be identified. The recognition module is configured to: input the PPG signal into the trained recognition model, first extract linear and nonlinear features, concatenate them into fusion features, then use the learned discriminant projection matrix to project the fusion features into a low-dimensional discriminant space, and finally use the projected features to classify the PPG signal to obtain the identity recognition result. The training of the recognition model involves introducing manifold regularization and graph Laplacian constraints to construct a graph Laplacian matrix for the training sample set. Based on the graph Laplacian matrix, the recognition model is jointly optimized using in-class group sparse constraints and error matrix row sparse constraints to obtain the optimal discriminant projection matrix.

[0010] According to some embodiments, the present invention adopts the following technical solution: A computer program product includes a computer program that, when executed by a processor, implements the aforementioned photoplethysmography (PPG) identification method.

[0011] According to some embodiments, the present invention adopts the following technical solution: A non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the aforementioned photoplethysmography (PPG) identification method.

[0012] According to some embodiments, the present invention adopts the following technical solution: An electronic device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform a photoplethysmography (PPG) identification method.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: The proposed "Identification Method Based on Photoplethysmography Pulse Wave Signal with Manifold Regularization and Inter-class Sparse Regression" improves upon traditional methods such as Least Squares Regression (LSR), Discriminant Least Squares Regression (DLSR), and Principal Component Analysis combined with Support Vector Machine (PCA+SVM). It maintains the geometric structure of neighboring samples through manifold regularization, indirectly achieving inter-class separation; enhances intra-class consistency through intra-class group sparse constraints; and filters noise through error sparsity terms. These three methods synergistically address inter-class overlap and intra-class discrepancy issues. Furthermore, intra-class group sparse constraints force similar samples to activate on the same feature dimensions, while manifold regularization preserves the local geometric structure of samples, thus resolving intra-class discrepancy. This significantly improves recognition accuracy, discriminative ability, and anti-interference performance, offering the following advantages: (1) The recognition accuracy is significantly improved. On the Real-World PPG dataset, the method of this invention achieves a recognition accuracy of 96.14% under the support vector machine classifier, which is 60.14% and 60.71% higher than the traditional LSR method (36.00%) and DLSR method (35.43%), respectively. Even compared with the better-performing PCA+SVM method (93.29%), it is nearly 3 percentage points higher, showing the superiority of this method in feature extraction and classification.

[0014] (2) The method of this invention has stronger discrimination ability and better class separation effect. The macro average F1-score of 0.9585 is achieved in multi-class identity recognition tasks, which is significantly improved compared with LSR (0.3235) and DLSR (0.3203) and is better than PCA+SVM (0.9233). This shows that the method is better in maintaining intra-class compactness and inter-class separation, and has higher recognition consistency and generalization ability.

[0015] (3) The false recognition rate is significantly reduced and the robustness is stronger. The overall false recognition rate of this method is only 3.86%, which is much lower than LSR (64.00%), DLSR (64.57%) and PCA+SVM (6.71%). This shows that the present invention has better robustness and practicality when processing PPG signals with high-dimensionality, physiological noise interference and complex nonlinear changes.

[0016] (4) It is highly adaptable and suitable for small sample high-dimensional recognition tasks. Experimental results show that the performance of traditional linear methods drops sharply under high-dimensional small sample conditions. However, this invention improves the discrimination ability while maintaining the compactness of the model by introducing graph structure learning, manifold preservation regularization term and inter-class sparse structure joint optimization. It is particularly suitable for practical scenarios with limited samples and complex signal distribution, such as wearable devices, medical monitoring and biometric authentication.

[0017] (5) The experimental verification is sufficient and has good prospects for promotion. Under the same dataset and the same classifier parameters, the method of this invention has been verified by a variety of mainstream comparative models and has shown comprehensive leading performance in terms of accuracy, robustness and stability. It has significant practical value and broad industrial application potential. Attached Figure Description

[0018] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0019] Figure 1 This is a flowchart of a photoplethysmography (PPG) identification method according to Example 1. Detailed Implementation The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0020] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0022] Example 1 One embodiment of the present invention provides a photoplethysmography (PPG) identification method, such as... Figure 1 As shown, it includes: Step S1: Acquire the PPG signal to be identified; Step S2: Input the PPG signal into the trained recognition model. First, extract linear and nonlinear features and concatenate them into fusion features. Then, use the learned discriminant projection matrix to project the fusion features into a low-dimensional discriminant space. Finally, use the projected features to classify the PPG signal and obtain the identity recognition result. The training of the recognition model involves introducing manifold regularization and graph Laplacian constraints to construct a graph Laplacian matrix for the training sample set. Based on the graph Laplacian matrix, the recognition model is jointly optimized using in-class group sparse constraints and error matrix row sparse constraints to obtain the optimal discriminant projection matrix.

[0023] As one embodiment, the photoplethysmography (PPG) identification method of the present invention aims to solve the following core technical problems: (1) The problem of accurate modeling of nonlinear manifold structures By introducing manifold regularization and graph Laplace constraints, the problem that existing linear models cannot capture the local geometric laws of the nonlinear manifold of PPG signals is solved, enabling similar samples to form a compact cluster structure in the feature space and reducing intra-class dispersion.

[0024] (2) The problem of improving generalization of small sample high-dimensional data By jointly optimizing the inter-class and error dual sparsity constraints, the risk of model overfitting is reduced by 60% with a small number of training samples, and the recognition accuracy is maintained above 90%.

[0025] (3) The balance between noise robustness and discriminability By employing a dual sparsity mechanism (intra-class group sparsity constraint + error matrix row sparsity), the separation efficiency of noise such as motion artifacts is improved, while individual-specific harmonic components are preserved, thereby enhancing the discriminative feature retention rate.

[0026] The specific implementation process of this embodiment is as follows: 1. Extraction and fusion of linear and nonlinear features The PPG signal is a one-dimensional time series signal. It is transformed into a three-dimensional matrix through continuous wavelet transformation. Linear and nonlinear features are then extracted from this three-dimensional matrix, and the two features are spliced ​​together to form a fused feature.

[0027] Specifically, to comprehensively capture the statistical characteristics and energy distribution of PPG signals in the time and frequency domains, and to address the technical challenges of processing complex continuous wavelet transform (CWT) coefficients, the following three types of linear features are designed: 1) Time-domain statistical characteristics: The PPG signal is divided into 5-second windows, and the mean and standard deviation of each segment are calculated in the frequency dimension to reflect the baseline level and fluctuation amplitude of the waveform.

[0028] 2) Frequency domain features: The signal is converted from the time domain to the frequency domain by Fast Fourier Transform (FFT), and the mean value is calculated along the time axis after taking the modulus to characterize the energy distribution of different frequency components.

[0029] 3) The information entropy feature uses Shannon entropy to quantify the complexity and irregularity of the signal, and is sensitive to subtle physiological differences such as heart rate variability (HRV).

[0030] While capturing linear features, we also consider the manifold structure mining of nonlinear features. Since the dynamics of signals in different frequency bands may follow different nonlinear modes, it is necessary to analyze each frequency band independently. In this embodiment, the three-dimensional matrix obtained by continuous wavelet transformation is split into multiple views according to different frequency scales. PCA (principal component analysis) is performed on each view to extract its nonlinear features. Finally, the dimensionality reduction results of all views are merged to form nonlinear features.

[0031] The linear and nonlinear features are then concatenated, and the influence of dimensions is eliminated by Z-score standardization. This process is used to fuse the features and then standardize them to obtain the fused features.

[0032] 2. Graph Structure Learning and Manifold Construction The goal of graph structure learning is to learn a graph structure from the raw PPG data that reflects the local relationships between samples, providing a foundation for subsequent manifold learning and classification. To this end, an adaptive similarity calculation and matrix optimization are used to generate a sparse, robust similarity matrix S with clear physiological meaning, specifically including: 1) Global Distance Metric and Normalization: When calculating the initial distance metric, cosine distance is used to characterize the similarity between samples. Global normalization is then used to eliminate the influence of dimensions, resulting in a normalized cosine distance. This can be expressed as a formula:

[0033]

[0034] in, It is a sample and The cosine distance between them is scale invariant by dividing by the global average distance; , They are the first The and the first Feature vectors of each sample; It is the entire cosine distance matrix The average value of all elements in the dataset is a global scalar. It is a very small smoothing constant used to ensure that the denominator is not zero.

[0035] 2) Dynamic K-Nearest Neighbors (KNN) Adjustment: The aim is to adaptively adjust the number of nearest neighbors for each sample based on local density, avoiding over-connectivity or under-connectivity. The value determines the selection How many nearest neighbor samples are there? It is the set of nearest neighbor indices of sample i, specifically: Calculate local density: ; Determine Adaptive value: ,

[0036] in, The global baseline neighbor count is a preset hyperparameter that represents the baseline number of neighbors set for samples in the average density region. Let be the local density of sample i. The larger the value, the better the sample density. The more sparse the area; Local density for all samples The vector formed; Local density vector The median represents the average density of the entire dataset.

[0037] This strategy is based on local density. Dynamic adjustment Values ​​that allow for the use of larger values ​​in high-density areas. Values ​​are used to avoid redundant connections in order to capture fine structures, and smaller values ​​are employed in low-density areas. To prevent isolated points and maintain connectivity, only the top nearest neighbors for each sample are retained when updating the nearest neighbor count. The non-zero similarity of the nearest neighbors.

[0038] 3) Local Adaptive Kernel Function: Set a local adaptive kernel function, designing the kernel bandwidth based on local statistical properties for each sample pair (i,j): Calculate the median local distance of the sample: ; Calculate the median global distance of the samples: Here, "global" refers to the sample. The median distance between j and all other samples in the dataset; Geometric mean bandwidth: ; Calculate the Gaussian kernel similarity between samples i and j:

[0039] in, They represent the first and second elements in the dataset, respectively. The and the first One sample; The normalized distance matrix, Indicates sample and sample The distance between them; :sample The set of K-nearest neighbor indices, which is determined by the previous dynamic KNN adjustment process, and whose size is determined by an adaptive value. The decision included the sample. of The sample number of the nearest neighbor; Distance matrix The Rows represent samples Distance to all other samples in the dataset; It is a very small smoothing constant used to ensure that the denominator is not zero, thus ensuring the numerical stability of the formula.

[0040] 4) Matrix symmetry and sparsification: The purpose is to improve the mathematical properties of the matrix and enhance its noise resistance.

[0041] First, the symmetry process preserves the product relationship of the original similarities and avoids directional bias. The similarity after symmetry processing is expressed by the formula:

[0042] Subsequently, a sparsity strategy based on the median threshold was adopted:

[0043] in, Let be the initial similarity matrix, which is an asymmetric matrix, where the elements are... This represents the sample calculated using a local adaptive kernel function. to sample Similarity; For the initial similarity matrix, the samples to sample The similarity, since the initial calculation is directional, and They are not necessarily equal; This is the similarity matrix after symmetry adjustment; A mathematical operator that represents extracting a matrix. All non-zero elements in the vector are used to form a vector. To calculate the matrix The median of all non-zero elements in the sample is a robust statistic that is not sensitive to outliers. This value represents the overall average level of non-zero similarity values. The sparsity threshold coefficient is a preset hyperparameter used to control the strength of sparsity. The larger the value, the higher the threshold, the more edges are set to zero, and the sparser the matrix. The sparsity threshold is determined by the threshold coefficient. The baseline value is obtained by multiplying the median of all non-zero similarities and using it to determine whether an edge (similarity) should be retained. The final sparsified similarity matrix is ​​the final output of graph structure learning and will be used to construct the Laplacian matrix. The foundation.

[0044] Furthermore, to enhance the stability of the feature space and prevent overfitting, Tikhonov regularization based on the degree matrix is ​​introduced to obtain the final similarity matrix S:

[0045] in, For a degree vector, its elements It is a matrix No. The sum of all elements in a row, i.e. It represents that in the sparsified graph, each sample node (the first...) The connection strength or importance of each sample (how many edges and what weights are connected to it); Degree vector The arithmetic mean of all elements, which is a scalar representing the average connection strength of all nodes in the graph, is used to automatically adjust the overall strength of regularization; For a degree vector The diagonal matrix, whose diagonal elements are all zero, has zero elements on its off-diagonal sides. Tikhonov regularization coefficient is a preset hyperparameter used to control the strength of regularization. The larger the value, the more obvious the enhancement effect on the matrix diagonal.

[0046] Manifold construction utilizes graph structures to reveal the inherent low-dimensional manifold of data. It transforms the similarity matrix S into a mathematical representation (Laplace matrix L) and uses it for subsequent projection optimization. Specific operations include: 1) Construction of the Laplace matrix: First, based on the similarity matrix S, calculate the degree matrix. ; The degree matrix here is different from the distance matrix D(i,j) used in the adaptive similarity calculation above. The distance matrix calculates the initial distance between samples using cosine distance, while the degree matrix is ​​a diagonal matrix, with the diagonal elements being the row sums of the similarity matrix S.

[0047] Then, normalize the Laplace matrix: This matrix encodes the topological structure of the graph and is a core tool for manifold learning.

[0048] 2) Popular regularization construction: In the projection matrix Add manifold constraint terms to the optimization objective. Features after forced projection Preserve the local neighborhood relationships on the original manifold.

[0049] 3. Design of joint objective function We construct a joint optimization objective that includes five terms: label consistency, intra-class group sparsity, error sparsity, Q-regularization of the projection matrix, and manifold regularization. Through multi-task collaboration, we achieve feature selection, noise robustness, manifold structure preservation, and alignment of supervision information.

[0050] Where Y is a label matrix, which is the mathematical form of the supervision information (i.e. the class label of the sample), X is the concatenated feature matrix, E is the error matrix, Q is the projection matrix, QX is the projected feature, i.e. the predicted label; F is the sparse feature representation matrix generated by optimization; and L is the graph Laplacian matrix of the input sample.

[0051] 1) Label Consistency Term: Projected features QX of samples of the same class are pulled towards the same label vector in space, directly reducing the intra-class distance. The difference between the predicted label QX and the true label Y is measured by least squares loss, and a sparse error matrix E is introduced to model anomalous noise.

[0052] 2) Within-class sparse terms: Sparse constraints are applied to the feature representations of each class of samples. The norm forces similar samples to activate on the same feature dimension and suppresses irrelevant dimensions.

[0053] 3) Error sparsity term: Apply row sparsity to the error matrix E to allow a small number of samples to have large errors, filtering out abnormal samples or noise interference.

[0054] 4) Projection matrix Q regularization term: constrains the Frobenius norm of Q to prevent false reductions in regression error caused by overfitting and ensures the smoothness and generalization of feature transformation.

[0055] 5) Manifold regularization term: By performing trace operation on the graph Laplacian matrix L, it forces neighboring samples to maintain geometric proximity in the projection space, preserving the local geometric structure of the original data space, and making the regression error change smoothly in the local neighborhood.

[0056] High-performance feature extraction is achieved through a dual mechanism, specifically: First, there is the manifold structure guidance mechanism, which introduces a graph Laplace regularization term ( The similarity matrix S constructed using dynamic KNN encodes the local geometric relationships of samples. This mechanism forces samples of the same class to maintain their proximity in the original space within the projection space (QX), forming a compact manifold cluster structure. This effectively alleviates the intra-class discrepancy problem caused by linear projection, and is particularly suitable for extracting physiological signal features with nonlinear structures, such as heart rate variability (HRV).

[0057] Secondly, a dual sparse discrimination mechanism was designed. 1) In-class sparsity: through Constraints force the feature representations (F) of similar samples to be activated in the same dimension, explicitly enhancing intra-class consistency.

[0058] 2) Error sparsity: through Constraining the row sparsity of the error matrix filters out interference from motion artifacts or anomalous samples, improving the model's stability in noisy environments.

[0059] 4. ADMM Optimization Solution Strategy The Alternating Directional Multiplier Method (ADMM) is used to perform block optimization of the objective function. An auxiliary variable F and a dynamic penalty parameter μ are introduced to achieve efficient iterative solution. Here, F is both an auxiliary variable and a sparse feature representation matrix. This dual role is a clever combination designed under the ADMM framework to simplify problem decomposition.

[0060] The Lagrange function is:

[0061] in, The dynamic penalty parameter has an initial value of Each iteration increases according to increase; The penalty parameter is the growth rate; Penalty parameter upper limit; Lagrange multiplier terms It achieves consistency between the sparse representation F and the projected feature QX through dynamic accumulation and feedback constraint bias during iteration; the quadratic penalty term It is used for dynamic enhancement The larger the μ value, the stronger the consistency.

[0062] The subproblem is then solved iteratively. Initially, Q is a random value, E=0, F=QX, and C1=0. The solution is then iteratively updated according to the formula until the maximum number of iterations is reached or the function converges, yielding the projection matrix Q, specifically: 1) Update the projection matrix Q:

[0063] Analytical solution: , ,

[0064] 2) Update the intra-class sparse matrix F:

[0065] 3) Update the error matrix E:

[0066] 4) Update the Lagrange multiplier C 1：

[0067] 5. Feature Projection and Classification Evaluation To eliminate the influence of the projection matrix norm on the feature distribution, normalized projection is used:

[0068] Where Q is the learned projection matrix, used to project the data X onto a new feature space; and The feature representation is generated by fusing and standardizing the training and test set data after linear and nonlinear feature extraction, respectively. and These are the norms of the projected features, used for normalization.

[0069] The above equation projects the training and test sets onto a new feature space and performs normalization. This operation ensures the feature vectors lie on a unit sphere, enhancing the classifier's robustness to scale changes. The normalized result yields new training set features. and test set features These, along with the training set labels and test set labels, are input into the classifier for classification.

[0070] The classifier can be either a Nearest Neighbor (NN) classifier or a Support Vector Machine (SVM). The Nearest Neighbor (NN) classifier, based on Euclidean distance to measure feature similarity, is suitable for linearly separable scenarios and has low computational complexity. The Support Vector Machine (SVM) uses the RBF kernel function to handle non-linear classification boundaries and optimizes parameters through grid search, achieving optimal generalization performance on the dataset of this invention.

[0071] Example 2 One embodiment of the present invention provides a photoplethysmography (PPG) identification system, comprising: The acquisition module is configured to acquire the PPG signal to be identified. The recognition module is configured to: input the PPG signal into the trained recognition model, first extract linear and nonlinear features, concatenate them into fusion features, then use the learned discriminant projection matrix to project the fusion features into a low-dimensional discriminant space, and finally use the projected features to classify the PPG signal to obtain the identity recognition result. The training of the recognition model involves introducing manifold regularization and graph Laplacian constraints to construct a graph Laplacian matrix for the training sample set. Based on the graph Laplacian matrix, the recognition model is jointly optimized using in-class group sparse constraints and error matrix row sparse constraints to obtain the optimal discriminant projection matrix.

[0072] Example 3 One embodiment of the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned photoplethysmography (PPG) identification method.

[0073] Example 4 In one embodiment of the present invention, a non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the aforementioned photoplethysmography (PPG) identification method.

[0074] Example 5 One embodiment of the present invention provides an electronic device, including: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the photoplethysmography (PPG) identification method.

[0075] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0076] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0077] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A photoplethysmography (PPG) identification method, characterized in that, include: Acquire the PPG signal to be identified; The PPG signal is input into the trained recognition model. First, linear and nonlinear features are extracted and concatenated into fusion features. Then, the learned discriminant projection matrix is ​​used to project the fusion features into a low-dimensional discriminant space. Finally, the projected features are used to classify the PPG signal to obtain the identity recognition result. The training of the recognition model involves introducing manifold regularization and graph Laplacian constraints to construct a graph Laplacian matrix for the training sample set. Based on the graph Laplacian matrix, the recognition model is jointly optimized using in-class group sparse constraints and error matrix row sparse constraints to obtain the optimal discriminant projection matrix.

2. The photoplethysmography (PPG) identification method as described in claim 1, characterized in that, The linear features include time-domain statistical features, frequency-domain features, and information entropy features.

3. The photoplethysmography (PPG) identification method as described in claim 1, characterized in that, The method for extracting the nonlinear features is as follows: The three-dimensional matrix obtained by continuous wavelet transform of the PPG signal is split into multiple views according to different frequency scales. Nonlinear features are extracted for each view using principal component analysis. Finally, the dimensionality reduction results of all views are stitched together to form a nonlinear feature matrix.

4. The photoplethysmography (PPG) identification method as described in claim 1, characterized in that, The training process of the recognition model includes: Obtain the raw multimodal training data, including the real labels; Linear and nonlinear features are extracted, and multi-view features are obtained after feature fusion. Classify the multi-view features to obtain predicted labels; Based on the real labels and predicted labels, an objective function is constructed through label consistency, intra-class group sparsity, error sparsity, Q-regularization term of the projection matrix, and manifold regularization. The objective function is solved to obtain the optimal discriminant projection matrix.

5. The photoplethysmography (PPG) identification method as described in claim 1, characterized in that, The method introduces manifold regularization and graph Laplacian constraints to construct the graph Laplacian matrix of the training sample set. Based on the similarity between samples, the graph Laplacian matrix is ​​obtained through adaptive local scaling Gaussian kernel and sparse regularization method.

6. The photoplethysmography (PPG) identification method as described in claim 1, characterized in that, The objective function is expressed by the formula: Where Y is a label matrix, X is the fused features, QX is the predicted label, E is the error matrix, Q is the projection matrix, F is the sparse feature representation matrix generated through optimization, and L is the graph Laplacian matrix of the input samples.

7. A photoplethysmography (PPG) identification system, characterized in that, include: The acquisition module is configured to acquire the PPG signal to be identified. The recognition module is configured to: input the PPG signal into the trained recognition model, first extract linear and nonlinear features, concatenate them into fusion features, then use the learned discriminant projection matrix to project the fusion features into a low-dimensional discriminant space, and finally use the projected features to classify the PPG signal to obtain the identity recognition result. The training of the recognition model involves introducing manifold regularization and graph Laplacian constraints to construct a graph Laplacian matrix for the training sample set. Based on the graph Laplacian matrix, the recognition model is jointly optimized using in-class group sparse constraints and error matrix row sparse constraints to obtain the optimal discriminant projection matrix.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the photoplethysmography (PPG) identification method according to any one of claims 1-6.

9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement a photoplethysmography (PPG) identification method as described in any one of claims 1-6.

10. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform a photoplethysmography (PPG) identification method as described in any one of claims 1-6.

Images