Identity Recognition Method and System Based on Invariant Representation Learning of Biological Signals

Abstract

The present disclosure provides an identity recognition method and system based on invariant representation learning of biological signals, which relates to the technical fields of artificial intelligence and pattern recognition. It includes converting waveform heartbeat data into a graph structure, then extracting the topological feature graph of the graph structure through a graph structure feature extraction unit, performing multi-granularity semantic correlation evaluation on the waveform heartbeat data through a multi-granularity semantic converter, establishing cross-dimensional dependence relationships using an attention mechanism to obtain a score feature graph, and then inputting the score feature graph and the topological feature graph into a feature decoupling and recombination module. First, it is decomposed into an identity-sensitive invariant feature representation and a physiological variable feature representation, and then the invariant feature representation and the shuffled physiological variable feature representation are recombined to generate a recombined feature representation. The recombined feature is input into a loss function to obtain a predicted value for identity matching verification. The present disclosure realizes efficient identity recognition through cloud collaboration and the model structure architecture.

Description

Technical Field

[0001] The present disclosure relates to the technical fields of artificial intelligence and pattern recognition, and particularly to an identity recognition method and system based on invariant representation learning of biological signals. Background Art

[0002] The statements in this part merely provide background technical information related to the present disclosure and do not necessarily constitute prior art.

[0003] With the rapid growth of mobile Internet users and mobile devices, new social interaction paradigms based on mobile terminals, such as digital payment, instant messaging, and remote office, are increasing day by day. While this digital transformation improves service efficiency, it also poses unprecedented security challenges to device identity authentication mechanisms.

[0004] The current mainstream biometric authentication systems (including fingerprint recognition, iris verification, and 3D face modeling technologies) and traditional password verification methods have exposed significant security vulnerabilities in practice. In recent years, with the popularization of smart wearable devices, smart bracelets have become important carriers for daily health monitoring, sports management, and security authentication. The photoplethysmogram (PPG) sensor built into the wearable bracelet can collect the user's PPG signal in real time for identity recognition. Compared with traditional biometrics (such as fingerprints and facial recognition), the PPG signal has significant technical advantages: on the one hand, it obtains vascular pulsation information through the principle of light reflection and has the characteristics of being difficult to forge; on the other hand, the acquisition of the PPG signal does not require complex hardware, has low cost and strong compatibility, and can be seamlessly integrated into various existing wearable devices. With the rapid development of deep learning methods, various deep learning methods have shown good performance in actual identity recognition.

[0005] However, the PPG signal identity recognition method still faces the following problems in practical applications:

[0006] 1) The PPG signal collected by the smart bracelet is easily affected by motion artifacts. The frequency of motion artifacts coincides with the frequency of the PPG signal itself, and it is difficult for existing noise cancellation methods to eliminate the influence of motion artifacts.

[0007] 2) The distribution of the PPG signal is easily affected by physiological state changes, environmental noise, and device differences, resulting in the distribution of the PPG signal drifting over time (such as user heart rate variability, blood vessel elasticity changes, or sensor aging). Such distribution drift will significantly reduce the model generalization ability. Especially in open dynamic scenarios, the identity recognition effect of the PPG signal is low due to the influence of motion artifacts and distribution drift, which limits the practical implementation of the PPG signal identity recognition technology. Summary of the Invention

[0008] To solve the above problems, the present disclosure proposes an identity recognition method and system based on invariant representation learning of biological signals. Through preprocessing based on adaptive motion noise separation, a pure PPG waveform synchronized with the cardiac cycle is extracted, and an identity recognition model for PPG signals based on invariant representation learning of graph neural networks is used. The PPG signal is transformed into a graph structure data model through a visible graph topology mapping mechanism, and a graph space representation matrix is constructed. Through an adaptive attention mechanism, the contribution degree of feature dimensions is quantified in the latent semantic space. The graph space representation is decomposed into an identity-sensitive invariant feature set and a physiological variable feature set, and the robustness of the model is improved through a feature recombination strategy.

[0009] According to some embodiments, the present disclosure adopts the following technical solutions:

[0010] An identity recognition method based on invariant representation learning of biological signals, comprising:

[0011] Obtain the user's pulse wave PPG signal, and perform preprocessing on it based on an adaptive motion noise separation algorithm, and segment and extract the waveform heartbeat data synchronized with the cardiac cycle;

[0012] Input the waveform heartbeat data into an identity recognition model for PPG signals based on invariant representation learning of graph neural networks, and output the user's identity recognition result;

[0013] Among them, after the waveform heartbeat data is input into an identity recognition model for PPG signals based on invariant representation learning of graph neural networks, it enters the graph structure feature extraction unit and the multi-granularity semantic converter unit respectively for feature extraction; after the waveform heartbeat data enters the graph structure feature extraction unit, the waveform heartbeat data is transformed into a graph structure, and then the topological feature graph of the graph structure is extracted through the graph structure feature extraction unit; the waveform heartbeat data is evaluated for multi-granularity semantic correlation through a multi-granularity semantic converter, and a cross-dimensional dependence relationship is established by using a time-sensitive attention mechanism to obtain a score feature graph. Then, the score feature graph and the topological feature graph are input into the feature decoupling and recombination module, which is first decomposed into an identity-sensitive invariant feature representation and a physiological variable feature representation, and then the invariant feature representation and the scrambled physiological variable feature representation are recombined to generate a recombined feature representation. The recombined feature is input into the loss function to obtain a predicted value for identity matching verification.

[0014] According to some embodiments, the present disclosure adopts the following technical solutions:

[0015] An identity recognition system based on invariant representation learning of biological signals, comprising:

[0016] A signal acquisition module, configured to obtain the user's pulse wave PPG signal, and perform preprocessing on it based on an adaptive motion noise separation algorithm, and segment and extract the waveform heartbeat data synchronized with the cardiac cycle;

[0017] An identity recognition module, configured to input waveform heart rate data into a PPG signal identity recognition model based on invariant representation learning of a graph neural network, and output the identity recognition result of the user;

[0018] Wherein, after the waveform heart rate data is input into the PPG signal identity recognition model based on invariant representation learning of a graph neural network, it respectively enters a graph structure feature extraction unit and a multi-granularity semantic converter unit for feature extraction; after the waveform heart rate data enters the graph structure feature extraction unit, the waveform heart rate data is transformed into a graph structure, and then the topological feature graph of the graph structure is extracted by the graph structure feature extraction unit; the waveform heart rate data is subjected to multi-granularity semantic correlation evaluation through a multi-granularity semantic converter, a cross-dimensional dependence relationship is established by using a time-series sensitive attention mechanism to obtain a score feature graph, and then the score feature graph and the topological feature graph are input into a feature decoupling and recombination module, first decomposed into identity-sensitive invariant feature representations and physiological variable feature representations, and then the invariant feature representations and the scrambled physiological variable feature representations are recombined to generate a recombined feature representation, and the recombined feature is input into a loss function to obtain a predicted value for identity matching verification.

[0019] According to some embodiments, the present disclosure adopts the following technical solutions:

[0020] A computer program product, including a computer program, which when executed by a processor implements the identity recognition method based on invariant representation learning of biological signals.

[0021] According to some embodiments, the present disclosure adopts the following technical solutions:

[0022] A non-transitory computer-readable storage medium, which is used to store computer instructions, and when the computer instructions are executed by a processor, the identity recognition method based on invariant representation learning of biological signals is implemented.

[0023] According to some embodiments, the present disclosure adopts the following technical solutions:

[0024] 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 runs, the processor executes the computer program stored in the memory so that the electronic device executes and implements the identity recognition method based on invariant representation learning of biological signals.

[0025] Compared with the prior art, the beneficial effects of the present disclosure are:

[0026] The identity recognition method based on invariant representation learning of bio-signals disclosed in the present disclosure uses the PPG sensor built in the smart bracelet to collect the user's pulse wave signal in real time, transmits the signal to the user's mobile device using the Bluetooth protocol, and the mobile device uploads the data to the remote server through an encrypted channel for identity recognition. The smart mobile device establishes a two-way communication link with the smart bracelet through the Low Energy Bluetooth (BLE) protocol to achieve low-latency and high-fidelity transmission of PPG signals. In the connection initialization stage, the mobile device completes the two-way authentication of device identities based on the dynamic key negotiation protocol and negotiates an adaptive clock synchronization mechanism to ensure that the sampling timing is strictly aligned with the bracelet sensor. The PPG signal can be collected by the smart bracelet in real time, providing a dynamic, continuous, and covert verification method for identity recognition, which is more suitable for identity recognition in the mobile device environment. It will also promote the in-depth application of mobile bracelets in fields such as financial payment and medical health, laying a foundation for the development of multi-modal biometric systems.

[0027] The identity recognition method based on invariant representation learning of bio-signals disclosed in the present disclosure designs a preprocessing algorithm based on adaptive motion noise separation. Through multi-channel signal fusion and time-frequency domain filtering, high-frequency and low-frequency noises, ambient light interference, and baseline drift are eliminated. The improved adaptive empirical mode decomposition algorithm is used to extract the pure PPG waveform synchronized with the cardiac cycle, eliminating motion artifacts and signal distortion caused by limb movement in dynamic scenarios.

[0028] The identity recognition method based on invariant representation learning of bio-signals disclosed in the present disclosure proposes an identity recognition model for PPG signals using invariant representation learning of graph neural networks, which includes three core components: a graph structure feature extraction unit (GFEM), a multi-granularity semantic converter (MTM), and a feature decoupling and recombination module (SMM). First, the PPG signal is transformed into a graph structure data model through a visible graph topology mapping mechanism, and the GFEM unit is used to capture its topological features and construct a graph space representation. At the same time, the MTM component performs multi-granularity semantic correlation evaluation on the graph space features, and realizes the quantification of the contribution degree of feature dimensions in the latent semantic space through an adaptive attention mechanism. Finally, the SMM module uses a feature decoupling algorithm to decompose the graph space representation into identity-sensitive invariant features and physiologically variable features, and improves the robustness of the model through a feature recombination strategy. This solution achieves efficient recognition through an end-cloud collaborative architecture and the recognition model structure, significantly improving the accuracy and stability of PPG identity recognition in complex scenarios while ensuring privacy and security, and can be extended to high-security demand scenarios such as mobile payment and medical data authorization. Description of the Drawings

[0029] The specification drawings forming a part of the present disclosure are used to provide a further understanding of the present disclosure. The illustrative embodiments and descriptions of the present disclosure are used to explain the present disclosure and do not constitute an improper limitation of the present disclosure.

[0030] Figure 1 Flowchart of the identity recognition method based on invariant representation learning of biological signals according to an embodiment of the present disclosure;

[0031] Figure 2 Schematic diagram of mobile identity recognition of PPG signals of the smart bracelet according to an embodiment of the present disclosure;

[0032] Figure 3 Structure diagram of the invariant representation learning PPG signal identity recognition model of the graph neural network according to an embodiment of the present disclosure;

[0033] Figure 4 Schematic diagram of the structure of the graph structure feature extraction unit according to an embodiment of the present disclosure;

[0034] Figure 5 Schematic diagram of the structure of the multi-granularity semantic converter according to an embodiment of the present disclosure;

[0035] Figure 6 Schematic diagram of the structure of the feature decoupling and recombination module according to an embodiment of the present disclosure. Detailed implementation manners

[0036] The present disclosure will be further described below in conjunction with the accompanying drawings and embodiments.

[0037] It should be noted that the following detailed descriptions are all illustrative and are intended to provide further descriptions of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present disclosure belongs.

[0038] It should be noted that the terms used herein are only for describing specific implementation manners and are not intended to limit the exemplary implementation manners according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0039] Embodiment 1

[0040] In an embodiment of the present disclosure, an identity recognition method based on invariant representation learning of biological signals is provided, and the method steps include:

[0041] Step 1: Obtain the user's pulse wave PPG signal, preprocess it based on the adaptive motion noise separation algorithm, and segment and extract the waveform heartbeat data synchronized with the cardiac cycle;

[0042] Step 2: Input the waveform heartbeat data into the invariant representation learning PPG signal identity recognition model based on the graph neural network, and output the user's identity recognition result;

[0043] Among them, after inputting the waveform heartbeat data into the PPG signal identity recognition model based on graph neural network for invariant representation learning, it enters the graph structure feature extraction unit and the multi-granularity semantic converter unit respectively for feature extraction; after the waveform heartbeat data enters the graph structure feature extraction unit, the waveform heartbeat data is converted into a graph structure, and then the topological feature graph of the graph structure is extracted through the graph structure feature extraction unit; the waveform heartbeat data is evaluated for multi-granularity semantic correlation through the multi-granularity semantic converter, and a cross-dimensional dependence relationship is established by using the time-series sensitive attention mechanism to obtain a score feature graph. Then, the score feature graph and the topological feature graph are input into the feature decoupling and recombination module, first decomposed into identity-sensitive invariant feature representations and physiological variable feature representations, and then the invariant feature representations and the scrambled physiological variable feature representations are recombined to generate a recombined feature representation. The recombined feature is input into the loss function to obtain a predicted value for identity matching verification.

[0044] As an embodiment, the specific implementation process of the identity recognition method based on invariant representation learning of biological signals of the present disclosure is as follows:

[0045] Step 1: Obtain the user's pulse wave PPG signal, and preprocess it based on the adaptive motion noise separation algorithm, and segment and extract the waveform heartbeat data synchronized with the cardiac cycle;

[0046] Specifically, first, a wearable device is used to collect the user's pulse wave PPG signal, which can be a mobile bracelet. The PPG sensor built in the mobile bracelet is used to collect the user's pulse wave signal in real time, and the signal is transmitted to the mobile terminal through the Bluetooth protocol. The mobile terminal can be a smart phone. The smart phone establishes a two-way communication link with the smart bracelet through the low-power Bluetooth (BLE) protocol. The mobile phone uploads the data to the remote server for identity recognition through an encrypted channel, and finally returns the recognition result to the mobile phone to complete the identity authentication. Realize the low-latency and high-fidelity transmission of the PPG signal. In the connection initialization stage, the mobile phone completes the two-way authentication of the device identity based on the dynamic key negotiation protocol, and negotiates the adaptive clock synchronization mechanism to ensure that the sampling timing is strictly aligned with the bracelet sensor.

[0047] Furthermore, the collected user's pulse wave PPG signal is segmented into a training set and a test set.

[0048] Then, the collected PPG signal is preprocessed. The preprocessing includes two stages. The first stage eliminates high-frequency and low-frequency noises, ambient light interference and baseline drift, and the second stage eliminates motion artifacts. The specific process is as follows:

[0049] The first stage: Eliminate high-frequency and low-frequency noises, ambient light interference and baseline drift.

[0050] 1) Eliminate baseline drift. Remove the low-frequency baseline offset through moving average or polynomial fitting, and perform detrending and normalization on the original PPG signal to eliminate baseline drift.

[0051] 2) Initialize the coefficients of the filter and the step size factor μ , set the cut-off frequency of the low-pass filter to 50 Hz, and the cut-off frequency of the high-pass filter to 0.5 Hz.

[0052] 3) Design a finite impulse response filter, and its output is expressed as:

[0053]

[0054] where y n is the output signal of the finite impulse response filter, x[n] is the input signal, w k [n] is the adaptive coefficient of the filter, M is the order of the filter. The filter coefficient will be adaptively adjusted with the change of the input signal to minimize the error.

[0055] 4) According to the principle of minimizing the sum of squared errors, gradually update the coefficients of the filter, and its update formula is:

[0056]

[0057] where e n = d n - y n is the error signal, d n is the desired signal (usually the reference signal), is the step size factor, which controls the convergence speed of the algorithm.

[0058] 5) Use the finite impulse response filter to design a low-pass filter and a high-pass filter. The low-pass filter is used to remove high-frequency noise (such as noise above 50 Hz), and the high-pass filter is used to remove low-frequency noise (such as baseline drift below 0.5 Hz).

[0059] 6) Combine the low-pass filter and the high-pass filter into a band-pass filter.

[0060] 7) Adjust the filter coefficients at each sampling point through the least mean square error algorithm.

[0061] The second stage: Eliminate motion artifacts.

[0062] ​​​​​1) Adaptive decomposition. Use the improved adaptive empirical mode decomposition algorithm to perform empirical mode decomposition on the signal y after denoising in the first stage to obtain a number of intrinsic mode functions (IMFs):

[0063]

[0064] where is the i th mode function, is the residual term.

[0065] 2) Time-frequency analysis. Perform time-frequency analysis on each mode function to calculate its spectral characteristics , calculate the integral of in the low-frequency band [0, 5]. The specific formula is:

[0066]

[0067] If is higher than a predetermined threshold , then it is determined that the component contains motion artifacts, otherwise it indicates that the is a valid PPG signal.

[0068] 3) Motion artifact removal and signal reconstruction. After removing the motion artifacts, the artifact-free PPG signal is obtained:

[0069]

[0070] where S is the set of valid PPG signals .

[0071] 4) Use the Pan-Tompkins algorithm to accurately detect the R-wave peaks in the PPG signal, segment the PPG signal into heartbeats, and the segmented heartbeats are used as the input of the PPG signal identity recognition model for invariant representation learning based on the graph neural network.

[0072] Specifically, the process of using the Pan-Tompkins algorithm to accurately detect the R-wave peaks in the PPG signal and segment the PPG signal into heartbeats is as follows:

[0073] S1: Preprocessing, including: calculating the signal slope using the first-order difference to enhance the fast rising edge feature. The calculation formula is: y[n] = y[n + 1] - y[n - 1]. Square each point of the differential signal to highlight the high-frequency components and eliminate the negative value interference.

[0074] Further, perform moving window integration, including integrating the squared signal using a 150 ms window (assuming a sampling rate of 100 Hz, corresponding to 15 sampling points) to generate a smoothed envelope and aggregate pulse energy.

[0075] S2: Perform feature enhancement.

[0076] S3: Adaptive threshold detection, including: setting the signal threshold = 0.75 × the noise threshold, and the noise threshold = 0.25 × the signal threshold. Calculate the average amplitude of the initial 2-second signal as the noise threshold, and force-set an absolute refractory period of 200 ms to avoid false peak interference. Update the threshold every 0.5 seconds subsequently.

[0077] S4: R-wave localization, including: searching for local maxima within a ±100 ms window of the original PPG signal, excluding abnormal peaks with adjacent intervals < 300 ms or > 1500 ms, and precisely locating the systolic peak.

[0078] S5: Heartbeat segmentation, including: taking 200 ms before and 400 ms after the R-wave peak as the heartbeat segment.

[0079] S6: Length normalization, including: unifying each heartbeat to a fixed length through cubic spline interpolation.

[0080] Step 2: Input the waveform heartbeat data into the PPG signal identity recognition model based on graph neural network invariant representation learning, and output the user's identity recognition prediction label;

[0081] Specifically, the PPG signal identity recognition model based on graph neural network invariant representation learning includes a graph structure feature extraction unit (GFEM), a multi-granularity semantic converter (MTM), and a feature decoupling and recombination module (SMM).

[0082] Among them, the graph structure feature extraction unit (Graph-based Feature Extraction Module, GFEM) consists of two parts: a visible graph construction module and a feature extraction module, as Figure 4 shown.

[0083] The visible graph construction module converts the ECG data into a visibility graph, and the feature extraction module can extract the feature representation of its topological features.

[0084] The Multi-scale Transformer Module (MTM) includes a Multi-scale Extraction Module (MEM), a Bidirectional Transformer (BiTransformer) module, a Multi-Feature Attention (MFA) module, and a linear layer, as Figure 5 shown.

[0085] The feature decoupling and recombination module includes a decoupling module and a recombination module. The feature decoupling and recombination module is responsible for separating and mixing invariant features and spurious features, and capturing the invariant relationship between features and labels through a self-supervised learning mechanism. As Figure 6 shown.

[0086] As an embodiment, the specific process of training the PPG signal identity recognition model based on invariant representation learning of graph neural networks is as follows:

[0087] Step 1: After inputting the waveform heartbeat data of the original PPG signal into the PPG signal identity recognition model based on invariant representation learning of graph neural networks, it enters the graph structure feature extraction unit and the multi-scale semantic transformer unit for processing respectively, and the topological feature map and the score feature map are extracted respectively. The specific process is as follows:

[0088] 1) After inputting the waveform heartbeat data into the PPG signal identity recognition model based on invariant representation learning of graph neural networks, the waveform heartbeat data enters the graph structure feature extraction unit (GFEM), and the graph structure transformation of the waveform heartbeat data is carried out by using the visible graph construction module of the graph structure feature extraction unit (GFEM). Given a waveform heartbeat data , denotes the visible graph transformed from , where denotes the node set, denotes the edge set. Specifically, the amplitude points in each heartbeat are regarded as nodes in the visible graph, . There is an edge between node and node if and only if the following conditions are met:

[0089]

[0090] where denotes the sampling point interval, denotes the maximum slope between the target node and its adjacent node on the right. is initially set to , in each iteration, the slope between the target node and the neighbor nodes is updated.

[0091] Furthermore, the visible graph structure data is input into the feature extraction module in the Graph Structure Feature Extraction Unit (GFEM), which includes four graph isomorphism layers and two linear layers. The graph isomorphism layers capture the structure and features of the visible graph by iteratively aggregating information from nodes and edges. The role of the linear layer is to perform a linear transformation on the input features to obtain the topological feature map of the graph structure. The specific representation is as follows:

[0092]

[0093]

[0094]

[0095]

[0096]

[0097]

[0098] Among them, represents the th graph isomorphism layer, which is used to splice the features, represents the linear transformation of the input feature , represents the topological feature map of the PPG signal.

[0099] 2) After the waveform heartbeat data of the original PPG signal is input into the Multi-scale Granularity Semantic Transformer (MTM), the Multi-scale Granularity Semantic Transformer (MTM) designed in this disclosure realizes the dynamic evaluation of the contribution degree of feature dimensions to identity discrimination in the latent feature space through a hierarchical semantic association analysis framework. The Multi-scale Granularity Semantic Transformer (MTM) is composed of a third-order processing unit: a multi-scale feature extraction module (Multi-scale Extraction Module, MEM), a bidirectional transformer (Bidirectional Transformer, BiTransformer) module, and a multi-feature attention (Multi-Feature Attention, MFA) module. The specific process is as follows:

[0100] (1) The waveform heartbeat data of the original PPG signal first enters the multi-scale feature extraction module and is respectively input into three convolutions with different kernel sizes, and then the feature maps of different scales are fused. The formal representation is:

[0101]

[0102]

[0103]

[0104]

[0105] Among them, are multi-granularity features.

[0106] (2) The bidirectional converter module performs inverse mapping on the obtained multi-granularity feature map to obtain an inverse-mapped feature map, and inputs the multi-granularity feature map and its inverse-mapped feature map into the BiTransformer model for encoding respectively. The outputs are respectively expressed as:

[0107]

[0108]

[0109] Among them, represents the Transformer encoding layer, represents the inverse mapping method, and respectively represent the outputs corresponding to the multi-granularity feature map and the inverse-mapped feature map.

[0110] (3) The multi-granularity attention (Multi-Feature Attention, MFA) module adds the outputs of the multi-scale feature map and its inverse-mapped feature map as follows:

[0111]

[0112] Input into the channel attention mechanism to emphasize key channel features. The definition of the output feature is as follows:

[0113]

[0114] Next, input into the linear layer to further process the feature and obtain a score feature map similar to the separation feature :

[0115]

[0116] Among them, represents the Sigmoid activation function, and the purpose is to constrain each value in to the range of (0, 1), represents the linear layer.

[0117] 3) Input the score feature map and the topological feature map into the feature decoupling and recombination module simultaneously, decompose them into identity-sensitive invariant feature representations and physiological variable feature representations, then recombine the invariant feature representations and the shuffled physiological variable feature representations to generate a recombined feature representation, and input the recombined feature into the loss function to obtain a predicted value for identity matching verification.

[0118] Specifically, after obtaining the score feature map and the topological feature map, based on separate graph-level representations , capture the invariant representation (identity-sensitive invariant feature representation) and the spurious representation (physiological variable feature representation):

[0119]

[0120]

[0121] Among them, represents the invariant representation, represents the spurious representation, represents the element-wise product.

[0122] Furthermore, connect the invariant representation and the shuffled spurious representation to generate another new recombined representation. Then, connect the invariant representation and the spurious representation to generate a new concatenated representation .

[0123] Furthermore, the learning objective loss function can be defined as:

[0124]

[0125]

[0126]

[0127]

[0128]

[0129]

[0130] Among them, represents the predicted label, represents the true label, represents the cross-entropy loss function, and respectively represent and The predicted value.

[0131] Among them, and are hyperparameters used to control the weights of the invariant prediction loss and the invariant mixture loss.

[0132] Step 3: Divide all heartbeat sample data into three parts: a training set, a template set, and a validation set. The template set generates registered matching templates, and the validation set is used to test the performance of the proposed method. The present disclosure first trains a PPG signal identity recognition model based on graph neural network invariant representation learning using the training set, and then uses the trained model to perform identity matching on the test set. The specific process is as follows:

[0133] (1) Model training. Train a PPG signal identity recognition model based on graph neural network invariant representation learning using the training PPG signal set.

[0134] (2) Construct the template set. Obtain the feature distribution encoding of each user for the template sample set through the trained feature extraction model of graph neural network invariant representation learning, and use it as the registered matching template set .

[0135] (3) For the PPG signal verification sample , obtain the feature distribution encoding through the trained PPG signal identity recognition model of graph neural network invariant representation learning .

[0136] (4) Calculate the Euclidean distance between and the registered matching template set . The identity label of the verification sample is obtained by the following formula:

[0137]

[0138]

[0139] Among them, represents the Euclidean distance between and the template set, represents the template closest to the distance from , which is the identity label of the sample .

[0140] Step 4: Matching result evaluation. Finally, calculate the false acceptance rate (FAR), false rejection rate (FRR), and equal error rate (EER) to measure the effectiveness of the proposed identity recognition method. The specific formulas are as follows:

[0141] False acceptance rate:

[0142]

[0143] Rejection rate:

[0144]

[0145] Equal error rate:

[0146]

[0147] Wherein, NGRA is the total number of within-class tests, NIRA is the total number of between-class tests; NFR and NFA are the numbers of false rejections and false acceptances.

[0148] Embodiment 2

[0149] In an embodiment of the present disclosure, an identity recognition system based on invariant representation learning of biological signals is provided, including:

[0150] A signal acquisition module, configured to acquire a user's pulse wave PPG signal, and preprocess it based on an adaptive motion noise separation algorithm, segment and extract waveform heartbeat data synchronized with the cardiac cycle;

[0151] An identity recognition module, configured to input the waveform heartbeat data into an invariant representation learning PPG signal identity recognition model based on a graph neural network, and output a user's identity recognition result;

[0152] Wherein, after the waveform heartbeat data is input into the invariant representation learning PPG signal identity recognition model based on a graph neural network, it enters a graph structure feature extraction unit and a multi-granularity semantic converter unit respectively for feature extraction; after the waveform heartbeat data enters the graph structure feature extraction unit, the waveform heartbeat data is converted into a graph structure, and then the topological feature graph of the graph structure is extracted by the graph structure feature extraction unit; the waveform heartbeat data is evaluated for multi-granularity semantic correlation through a multi-granularity semantic converter, a cross-dimensional dependence relationship is established by using a time-series sensitive attention mechanism to obtain a score feature graph, and then the score feature graph and the topological feature graph are input into a feature decoupling and recombination module, first decomposed into identity-sensitive invariant feature representations and physiologically variable feature representations, and then the invariant feature representations and the scrambled physiologically variable feature representations are recombined to generate a recombined feature representation, and the recombined feature is input into a loss function to obtain a predicted value for identity matching verification.

[0153] As an embodiment, the application process of the identity recognition method of the identity recognition system based on invariant representation learning of biological signals includes:

[0154] The first step: Use a smart bracelet to collect PPG signals, and a mobile phone obtains the collected PPG signals through Bluetooth or the like;

[0155] Step 2: Preprocess the PPG signal.

[0156] Step 3: Use the preprocessed PPG signal set to perform identity recognition using the PPG signal identity recognition model based on invariant representation learning of the graph neural network;

[0157] Step 4: Use the template to perform identity matching on the PPG signal recognition result.

[0158] Step 5: Evaluate the identity of the recognition result.

[0159] Embodiment 3

[0160] In an embodiment of the present disclosure, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the identity recognition method based on invariant representation learning of biological signals is implemented.

[0161] Embodiment 4

[0162] In an embodiment of the present disclosure, a non-transitory computer-readable storage medium is provided, and the non-transitory computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the identity recognition method based on invariant representation learning of biological signals is implemented.

[0163] Embodiment 5

[0164] In an embodiment of the present disclosure, an electronic device is provided, 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 runs, the processor executes the computer program stored in the memory so that the electronic device executes the identity recognition method based on invariant representation learning of biological signals.

[0165] The present disclosure is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present disclosure. It should be understood that each process and / or block in the flowcharts and / or block diagrams, and the combination of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in one Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0166] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, so that a series of operation steps are executed on the computer or other programmable apparatus to produce a computer-implemented process, thereby the instructions executed on the computer or other programmable apparatus provide steps for realizing the functions specified in one process or a plurality of processes and / or blocks Figure 1 one process or a plurality of processes and / or blocks Figure 1 steps for realizing the functions specified in one block or a plurality of blocks.

[0167] Although the specific embodiments of the present disclosure have been described above in conjunction with the accompanying drawings, they are not intended to limit the scope of protection of the present disclosure. Those skilled in the art should understand that, based on the technical solutions of the present disclosure, various modifications or variations that can be made by those skilled in the art without creative efforts are still within the scope of protection of the present disclosure.

Claims

1. An identity recognition method based on invariant representation learning of biological signals, characterized in that, Including: Obtain the user's photoplethysmogram (PPG) signal, and preprocess it based on an adaptive motion noise separation algorithm, segment and extract the waveform heartbeat data synchronized with the cardiac cycle; Input the waveform heartbeat data into the PPG signal identity recognition model based on invariant representation learning of graph neural network, and output the user's identity recognition result; Among them, after the waveform heartbeat data is input into the PPG signal identity recognition model based on invariant representation learning of graph neural network, it enters the graph structure feature extraction unit and the multi-granularity semantic converter unit respectively for feature extraction; after the waveform heartbeat data enters the graph structure feature extraction unit, the waveform heartbeat data is transformed into a graph structure, and then the topological feature graph of the graph structure is extracted by the graph structure feature extraction unit; The waveform heartbeat data is evaluated for multi-granularity semantic correlation through the multi-granularity semantic converter, a cross-dimensional dependence relationship is established using the time-series sensitive attention mechanism to obtain a score feature graph, and then the score feature graph and the topological feature graph are input into the feature decoupling and recombination module. First, it is decomposed into identity-sensitive invariant feature representations and physiologically variable feature representations, and then the invariant feature representations and the shuffled physiologically variable feature representations are recombined to generate a recombined feature representation. The recombined feature is input into the loss function to obtain a predicted value for identity matching verification.

2. The identity recognition method based on invariant representation learning of biological signals according to claim 1, wherein Preprocess it based on an adaptive motion noise separation algorithm. The preprocessing process includes two stages. In the first stage, low-frequency baseline drift is removed by moving average or polynomial fitting, detrending and normalizing the original PPG signal, removing high-frequency noise using a low-pass filter, and removing low-frequency noise using a high-pass filter.

3. The identity recognition method based on invariant representation learning of biological signals according to claim 2, wherein In the second stage, the improved adaptive empirical mode decomposition algorithm is used to perform empirical mode decomposition on the signal denoised in the first stage to obtain several intrinsic mode functions. Time-frequency analysis is performed on each intrinsic mode function, its spectral characteristics are calculated and artifacts are removed to obtain the artifact-removed PPG signal. The Pan-Tompkins algorithm is used to detect the R-wave peaks in the artifact-removed PPG signal, and segmentation is performed to extract the waveform heartbeat data synchronized with the cardiac cycle.

4. The identity recognition method based on invariant representation learning of biological signals according to claim 1, wherein After the waveform heartbeat data is input into the graph structure feature extraction unit, the visible graph construction module of the graph structure feature extraction unit regards the amplitude points in each waveform heartbeat as nodes in the visible graph, and transforms the waveform heartbeat data into a visible graph structure; then the topological feature graph is extracted by the feature extraction module of the graph structure feature extraction unit. The feature extraction module includes four graph isomorphism layers and two linear layers. The graph isomorphism layers iteratively aggregate the information of nodes and edges from the visible graph structure to obtain the structure and features of the visible graph. The role of the linear layer is to perform a linear transformation on the input features to obtain the topological feature graph of the graph structure.

5. The identity recognition method based on invariant representation learning of biological signals according to claim 1, wherein, The multi-granularity semantic converter includes a multi-granularity feature extraction module, a bidirectional converter module, and a multi-granularity attention module. The multi-granularity feature extraction module includes convolutions with three different kernels to extract feature maps of different scales, and then fuses the feature maps of different features to obtain multi-granularity features. The bidirectional converter module performs inverse mapping on the obtained multi-granularity feature maps to obtain inverse-mapped feature maps, and inputs the multi-granularity feature maps and their inverse-mapped feature maps into the BiTransformer model for encoding respectively. The encoding representations obtained respectively are weighted by the multi-granularity attention module to obtain a score feature map.

6. The identity recognition method based on invariant representation learning of biological signals according to claim 1, characterized in that After obtaining the score feature map and the topological feature map, based on the separated graph-level representation, the feature decoupling and recombination module is used to capture the invariant representation and the spurious representation. The invariant representation and the shuffled spurious representation are connected to generate another new recombined feature representation.

7. An identity recognition system based on invariant representation learning of biological signals, characterized in that, Including: A signal acquisition module, configured to acquire the user's pulse wave PPG signal, and preprocess it based on an adaptive motion noise separation algorithm, and segment and extract the waveform heartbeat data synchronized with the cardiac cycle. An identity recognition module, configured to input the waveform heartbeat data into an identity recognition model for PPG signal based on invariant representation learning of a graph neural network, and output the user's identity recognition result. Among them, after the waveform heartbeat data is input into the identity recognition model for PPG signal based on invariant representation learning of a graph neural network, it enters the graph structure feature extraction unit and the multi-granularity semantic converter unit respectively for feature extraction. After the waveform heartbeat data enters the graph structure feature extraction unit, the waveform heartbeat data is transformed into a graph structure, and then the topological feature map of the graph structure is extracted by the graph structure feature extraction unit. The waveform heartbeat data is evaluated for multi-granularity semantic correlation through the multi-granularity semantic converter, and a cross-dimensional dependence relationship is established by using a time-series sensitive attention mechanism to obtain a score feature map. Then, the score feature map and the topological feature map are input into the feature decoupling and recombination module, which is first decomposed into an identity-sensitive invariant feature representation and a physiological variable feature representation, and then the invariant feature representation and the shuffled physiological variable feature representation are recombined to generate a recombined feature representation. The recombined feature is input into the loss function to obtain a predicted value for identity matching verification.

8. A computer program product comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the identity recognition method based on invariant representation learning of biological signals 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, and when the computer instructions are executed by the processor, it implements the identity recognition method based on invariant representation learning of biological signals according to any one of claims 1-6.

10. An electronic device, characterized in that, 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 runs, the processor executes the computer program stored in the memory so that the electronic device executes the identity recognition method based on invariant representation learning of biological signals according to any one of claims 1-6.

Images