Anti-drift ppg recognition method based on rate perception and state space model

Abstract

The present application relates to the technical field of biometric identification, and particularly provides a rate-aware and state-space model-based anti-offset PPG identification method. The method comprises physiological feature front-end extraction on an original single-channel PPG signal to obtain a high-dimensional shallow feature sequence; double-flow collaborative processing is performed on the high-dimensional shallow feature sequence, which is distributed to two parallel branches of a control flow branch and a data flow branch; in the control flow branch, an amplitude spectrum and an instantaneous physiological rate curve are obtained; in the data flow branch, a deep global feature sequence with rate invariance is obtained; based on the high-dimensional shallow feature sequence and the deep global feature sequence, multi-scale refined features are obtained; and according to the multi-scale refined features, a final biometric identification result is obtained. The method can actively sense physiological rate changes and achieve efficient nonlinear modeling with extremely low parameter quantity.

Description

A Misaligned PPG Recognition Method Based on Rate Sensing and State-Space Model Technical Field

[0001] This invention relates to the field of biometric recognition technology, and in particular to an anti-offset PPG recognition method based on rate sensing and state space model. Background Technology

[0002] Photoplethysmography (PPG)-based biometric identification technology holds significant value in wearable device authentication. However, existing technologies still face the following challenges in practical applications:

[0003] The shortcomings of traditional signal processing methods: Traditional PPG biometric methods mainly rely on manually designed features, such as extracting morphological features of PPG signals such as peaks, troughs, and time intervals, and combining them with machine learning classifiers for recognition. These methods have the following obvious shortcomings: (1) Poor adaptability to physiological rhythm changes. PPG signals are not only affected by motion artifacts, but also by individual heart rate variability. When users are in different motion states, the waveform will undergo significant temporal warping on the time axis. Manually designed fixed features (such as fixed time window statistics) are difficult to adapt to this nonlinear time domain change, resulting in a sharp decline in recognition performance; (2) Limited feature expression ability. Manually designed features can only capture the preset, shallow local morphological information in PPG signals, and cannot fully characterize the global periodicity and high-dimensional dynamic patterns contained in long sequences. These shallow features often ignore the inherent dynamic mechanism of biological signals, resulting in limited recognition accuracy; (3) They have weak resistance to distribution shift. Feature extraction algorithms designed for specific static states or specific devices often fail to extract consistent identity representations when faced with non-stationary real-world scenarios (such as distribution shift caused by moving from stillness to violent motion). This requires domain experts to make tedious parameter adjustments, resulting in poor universality.

[0004] Limitations of existing deep learning methods: Although models such as Transformer and Convolutional Neural Networks (CNNs) have made progress in feature extraction, they have inherent defects when dealing with non-stationary physiological signals: (1) Poor adaptability to temporal warping. Existing deep learning models (including conventional Mamba and Transformer) usually assume that the input signal is on a fixed sampling grid. However, PPG signals have significant quasi-periodicity, and this period (heart rate) changes drastically with the user's movement or emotional fluctuations. Existing models cannot perceive this change in physiological frequency, resulting in the waveform of the same user at different heart rates being identified as different features (distribution shift), which seriously affects the recognition accuracy; (2) The contradiction between computational complexity and parameter efficiency. On resource-constrained edge devices (such as wristbands), the O(N2) complexity of Transformer is too high, making it difficult to achieve continuous monitoring. Lightweight CNNs lack a global receptive field. In addition, traditional multilayer perceptrons (MLPs) have extremely low parameter efficiency when fitting complex nonlinear physiological features, requiring a large number of parameters to achieve a good fitting effect; (3) the discriminative boundary is blurred, and biometrics not only needs to extract features, but also needs to maximize inter-class differences in the feature space. It is difficult to cope with open set recognition scenarios of millions of people by relying solely on cross-entropy loss. Summary of the Invention

[0005] In view of this, the present invention provides an anti-migration PPG identification method based on rate sensing and state space model, which can actively sense changes in physiological rate and achieve efficient nonlinear modeling with extremely low parameter count.

[0006] In a first aspect, the present invention provides a method for resisting PPG identification based on rate sensing and a state-space model, the method comprising:

[0007] Step 1: Feature mapping. Physiological feature front-end extraction is performed on the original single-channel PPG signal to obtain a high-dimensional shallow feature sequence.

[0008] Step 2: Dual-stream collaborative modeling. The high-dimensional shallow feature sequence is processed by dual-stream collaborative processing and distributed to two parallel branches: the control flow branch and the data flow branch.

[0009] Step 3: In the control flow branch, obtain the amplitude spectrum and instantaneous physiological rate curve; in the data flow branch, obtain the deep global feature sequence with rate invariance.

[0010] Step 4: Feature fusion, based on high-dimensional shallow feature sequences and deep global feature sequences, to obtain multi-scale refined features;

[0011] Step 5: Output. Based on the multi-scale refined features, obtain the final biometric recognition result.

[0012] Optionally, step 1 includes:

[0013] Physiological features are extracted from the original single-channel PPG signal, local morphological features are enhanced by a parameterized Gabor filter bank, and high-dimensional shallow feature sequences are obtained through a convolutional projection layer.

[0014] Step 11: Parameterized Gabor filtering;

[0015] The sequence of the original single-channel PPG signal is as follows A parameterized Gabor filter bank is introduced for convolution processing; the Gabor kernel function is defined as:

[0016] ;

[0017] in, The time index variable representing the convolution kernel; center frequency. ,bandwidth and amplitude These are all parameters learned automatically by the network;

[0018] Step 12: Convolutional projection and downsampling;

[0019] The filtered signal is then passed through a one-dimensional convolutional layer with a kernel size of 7 and a stride of 2 to expand the number of channels. And halve the length of the time series. Generate high-dimensional shallow feature sequences .

[0020] Optionally, step 3 includes:

[0021] In the control flow branch, a global frequency domain transformation is performed on the high-dimensional shallow feature sequence to obtain the amplitude spectrum, and the instantaneous physiological rate curve is regressed using the Kolmogorov-Arnold network (KAN).

[0022] Step 31: Obtain the amplitude spectrum;

[0023] First, the high-dimensional shallow feature sequence is input into a normalization layer for standardization, and then a linear layer is used for initial feature dimension mapping. Next, global average pooling is performed on the features to compress the time dimension and obtain the global context vector. Then, a Fast Fourier Transform (FFT) is used to transform the signal from the time domain to the frequency domain, and the amplitude spectrum is obtained by taking the modulus. And a frequency domain masking noise enhancement mechanism is introduced;

[0024] Step 32: Rate regression based on KAN;

[0025] The amplitude spectrum is processed using the Kolmogorov-Arnold network (KAN). KAN replaces the original activation function with a learnable B-spline function (B-Spline) on the edges.

[0026] ;

[0027] in, This represents the input values ​​on the connection edges in the KAN layer of the network, i.e., the components from the spectral features; This indicates the number of B-spline basis functions, used to determine the fineness of the fit; Indicates the first The learnable coefficients corresponding to each basis; Indicates the first One B-spline basis function;

[0028] The final output of the KAN network is processed by the Softplus activation function to ensure that the generated rate factor is always positive; the scalar broadcast of the rate factor is expanded to be proportional to the sequence length. A consistent vector, denoted as the instantaneous physiological rate curve. .

[0029] Optionally, step 3 further includes:

[0030] In the data flow branch, the high-dimensional shallow feature sequence is input into the periodic sensing Mamba backbone network. The discretization step size of the state space model SSM is dynamically modulated using the instantaneous physiological rate curve, and rate-sensing selective scanning is performed to output a deep global feature sequence with rate invariance.

[0031] Step 33: Rate-sensing discretization;

[0032] Based on the continuity equation of the state-space model SSM, a physiological rate modulation mechanism is introduced to affect the time step. Make corrections:

[0033] ;

[0034] in, This represents the effective sampling step size after physiological rate correction; This represents the original step size inferred solely from the current input; This represents the instantaneous physiological rate output by the frequency estimator guided by the spectrum. Indicates the current time Input features; and These represent the activation function and the linear projection layer, respectively.

[0035] Step 34: Parameter transformation and scanning;

[0036] Using the corrected effective sampling step size Calculate the parameters of the discretized system And perform a recursive scan operation:

[0037] ;

[0038] in, Indicates the current time The hidden state vector is used to contain historical context information; This represents the hidden state at the previous moment; This represents the time-varying state transition matrix calculated based on the corrected effective sampling step size; This represents the time-varying input projection matrix calculated based on the corrected effective sampling step size; This represents the input features at the current moment;

[0039] The final output is a deep global feature sequence that contains global contextual information and is insensitive to heart rate changes. .

[0040] Optionally, step 4 includes:

[0041] The high-dimensional shallow feature sequence and the deep global feature sequence are merged by the full-frequency sensing fusion module. The fusion weight is calculated by the KAN gating mechanism and the key features are screened by the channel attention mechanism to obtain multi-scale refined features.

[0042] Step 41: Calculate the KAN gating weights;

[0043] High-dimensional shallow feature sequences With deep global feature sequences Concatenate the inputs to the KAN gated network to calculate the fusion gate weights. :

[0044] ;

[0045] in, This represents the generated fusion gating weights, with values ​​ranging from [0,1]. This represents the Sigmoid activation function, used to normalize the output; This represents the KAN network module used to generate the gate weights; This represents a deep global feature sequence from the Mamba backbone network; This represents a high-dimensional shallow feature sequence originating from the front end of physiological features; This indicates a splicing operation along the channel dimension;

[0046] Step 42, Shape Perception and Residual Fusion Module;

[0047] Weighted fusion is performed using the LayerScale strategy:

[0048] ;

[0049] Subsequently, the fused features Input the shape-aware module to capture waveform details at multiple scales.

[0050] Optionally, step 5 includes:

[0051] Multi-scale refined features are input into the morphologically perceptual hybrid head, mapped to the discriminative feature space, and the final biometric recognition result is output based on the hybrid loss function.

[0052] To adaptively select key feature channels, a channel attention mechanism is introduced, the calculation process of which is as follows:

[0053] First, the input features are compressed in space by global average pooling to generate channel descriptors. Then, a non-linear transformation is performed through two 1×1 convolutional layers: the first convolution reduces the channel dimension and passes it through the GELU activation function, and the second convolution restores the number of channels to the original dimension. The output is processed by the Sigmoid activation function to generate a channel weight vector in the range [0,1], and then multiplied element-wise with the original input features.

[0054] Finally, the features enhanced by channel attention are input into the morphological perception hybrid head. In order to utilize both label information and the intrinsic structural information of the data, the morphological perception hybrid head is designed as two parallel task branches: the first branch is a classifier based on angle margins, and the second branch is a contrastive learning projection head.

[0055] Optionally, it includes:

[0056] Using a hybrid loss function Perform end-to-end training;

[0057] in, This represents the total loss function for model training; This represents the angular margin classification loss, used to maximize the angular distance between different identity features on the hypersphere to improve inter-class separability; InfoNCE represents the self-supervised contrastive loss, used to narrow the feature distance of the same user under different data augmentation views, thereby improving robustness against interference. This represents the balance coefficient, used to adjust the weight ratio of the two loss terms.

[0058] The technical solution provided by this invention includes a method comprising: feature mapping, which extracts physiological features from the original single-channel PPG signal to obtain a high-dimensional shallow feature sequence; dual-stream collaborative modeling, which performs dual-stream collaborative processing on the high-dimensional shallow feature sequence and distributes it to two parallel branches: a control flow branch and a data flow branch; in the control flow branch, the amplitude spectrum and instantaneous physiological rate curve are obtained; in the data flow branch, a deep global feature sequence with rate invariance is obtained; feature fusion, which obtains multi-scale refined features based on the high-dimensional shallow feature sequence and the deep global feature sequence; and output, which obtains the final biometric recognition result based on the multi-scale refined features. This method can actively sense changes in physiological rate and achieve efficient nonlinear modeling with extremely low parameter count. Attached Figure Description

[0059] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 is a flowchart of the anti-offset PPG identification method based on rate sensing and state space model provided in an embodiment of the present invention;

[0061] Figure 2 is a schematic diagram of the morphological perception and residual fusion module provided in an embodiment of the present invention. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] It should be understood that the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0064] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” used in the embodiments of this invention are also intended to include the plural forms unless the context clearly indicates otherwise.

[0065] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0066] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0067] Figure 1 is a flowchart of the anti-offset PPG identification method based on rate sensing and state space model provided in an embodiment of the present invention. As shown in Figure 1, the method includes:

[0068] This invention presents a lightweight PPG biometric recognition method based on physiological rate sensing and a Mamba-KAN collaborative architecture. Addressing the challenge of existing PPG recognition technologies struggling to balance computational efficiency and resistance to heart rate variability interference on edge devices, this invention innovatively designs an end-to-end dual-stream synergy processing architecture. The core logic of this invention breaks away from the traditional deep learning model's fixed sampling and passive extraction paradigm, introducing active physiological dynamics control. The processing flow mainly comprises five cascaded stages, with the specific logic as follows:

[0069] Step 1: Feature mapping. Physiological feature front-end extraction is performed on the original single-channel PPG signal to obtain a high-dimensional shallow feature sequence.

[0070] In this embodiment of the invention, step 1 includes:

[0071] Physiological features are extracted from the original single-channel PPG signal, local morphological features are enhanced by a parameterized Gabor filter bank, and high-dimensional shallow feature sequences are obtained through a convolutional projection layer.

[0072] In this embodiment of the invention, the physiological feature front-end extraction serves as the system's perception portal, primarily functioning to condition signals and expand dimensionality. Addressing the issues of low signal-to-noise ratio and limited dimensionality in the original single-channel PPG signal, this module utilizes learnable filters to remove baseline drift and high-frequency electromyographic noise, while simultaneously elevating the one-dimensional time series to a high-dimensional feature space, providing subsequent modules with clean input rich in local morphological details.

[0073] Step 11: Parameterized Gabor filtering;

[0074] The sequence of the original single-channel PPG signal is as follows A parameterized Gabor filter bank is introduced for convolution processing; the Gabor kernel function is defined as:

[0075] ;

[0076] in, The time index variable representing the convolution kernel; center frequency. ,bandwidth and amplitude All parameters are automatically learned by the network; this filter bank can adaptively enhance key physiological features such as peaks and troughs.

[0077] Step 12: Convolutional projection and downsampling;

[0078] The filtered signal is then passed through a one-dimensional convolutional layer with a kernel size of 7 and a stride of 2 to expand the number of channels. (e.g., 32), and halve the length of the time series. Generate high-dimensional shallow feature sequences .

[0079] Step 2: Dual-stream collaborative modeling. The high-dimensional shallow feature sequence is processed by dual-stream collaborative processing and distributed to two parallel branches: the control flow branch and the data flow branch.

[0080] Step 3: In the control flow branch, obtain the amplitude spectrum and instantaneous physiological rate curve; in the data flow branch, obtain the deep global feature sequence with rate invariance.

[0081] In this embodiment of the invention, step 3 includes:

[0082] In the control flow branch, a global frequency domain transformation is performed on the high-dimensional shallow feature sequence to obtain the amplitude spectrum, and the instantaneous physiological rate curve is regressed using the Kolmogorov-Arnold network (KAN).

[0083] In this embodiment of the invention, the spectrum-guided frequency estimator acts as the system's physiological pacemaker. Its core function is not to extract identity features, but rather to calculate physiological dynamic parameters. Addressing the non-stationarity (heart rate fluctuations) of PPG signals caused by exercise or emotional changes, it utilizes frequency domain stability to accurately regress the instantaneous physiological rate curve from the noisy signal, providing crucial dynamic control signals for the backbone network.

[0084] Step 31: Obtain the amplitude spectrum;

[0085] First, the high-dimensional shallow feature sequence is input into a normalization layer for standardization, and then a linear layer is used for initial feature dimension mapping. Next, global average pooling is performed on the features to compress the time dimension and obtain the global context vector. Then, a Fast Fourier Transform (FFT) is used to transform the signal from the time domain to the frequency domain, and the amplitude spectrum is obtained by taking the modulus. To improve the robustness of the model in complex noise environments, a frequency domain masking noise enhancement mechanism is introduced. Specifically, the amplitude spectrum is multiplied element-wise with a learnable frequency domain weight matrix and random masking noise is superimposed. This is equivalent to adaptively filtering and enhancing the signal in the frequency domain, forcing the model to focus on more physiologically meaningful frequency components rather than overfitting specific noise patterns.

[0086] Step 32: Rate regression based on KAN;

[0087] The amplitude spectrum is processed using Kolmogorov-Arnold Networks (KANs). KANs replace the original activation function with a learnable B-spline function on the edges.

[0088] ;

[0089] in, This represents the input values ​​on the connection edges in the KAN layer of the network, i.e., the components from the spectral features; The grid size represents the number of B-spline basis functions and is used to determine the fineness of the fit. Indicates the first The learnable coefficients corresponding to each basis; Indicates the first B-spline basis functions; Kolmogorov-Arnold network is a novel neural network structure that places learnable activation functions on edges (weights).

[0090] The final output of the KAN network is activated by the Softplus activation function. Processing is performed to ensure that the generated rate factor is always positive (consistent with the physical property that physiological frequencies are non-negative); the scalar broadcast of the rate factor is extended to be proportional to the sequence length. A consistent vector, denoted as the instantaneous physiological rate curve. .

[0091] In this embodiment of the invention, step 3 further includes:

[0092] In the data flow branch, high-dimensional shallow feature sequences are input into the periodic-aware Mamba backbone network. The discrete step size of the state-space model SSM is dynamically modulated using the instantaneous physiological rate curve to perform rate-aware selective scanning and output a deep global feature sequence with rate invariance. The Mamba backbone network is a linear complexity deep learning architecture based on the selective state-space sequence model.

[0093] In this embodiment of the invention, the periodic sensing Mamba backbone network consists of multiple stacked periodic sensing Mamba modules, as shown in Figure 1. The internal processing flow of each module is as follows: the input features are first preprocessed through a normalization layer and a linear layer, then enter a convolutional layer (e.g., a one-dimensional 1D convolution) to capture local contextual features, and undergo nonlinear transformation using the SiLU activation function. The processed features enter the core state-space model (SSM), where the step size of the SSM is dynamically scaled using the aforementioned instantaneous physiological rate curve (i.e., rate-aware selective scanning). Finally, the output of the SSM is reconstructed through the terminal linear layer, serving as the output of this module. This modular stacking design enables the network to extract deep global feature sequences from local waveforms to global rhythms layer by layer.

[0094] Step 33: Rate-sensing discretization;

[0095] Based on the continuity equation of the state-space model SSM, a physiological rate modulation mechanism is introduced to affect the time step. Make corrections:

[0096] ;

[0097] in, This represents the effective sampling step size after physiological rate correction; This represents the original step size inferred solely from the current input; This represents the instantaneous physiological rate output by the frequency estimator guided by the spectrum. Indicates the current time Input features; and These represent the activation function and the linear projection layer, respectively.

[0098] When the user's heart rate increases ( When the sampling step size is increased, the PPG waveform narrows, and the system automatically increases the sampling step size. This allows the model to span a longer phase per unit time, thus offsetting the effect of waveform narrowing.

[0099] Step 34: Parameter transformation and scanning;

[0100] Using the corrected effective sampling step size Calculate the parameters of the discretized system And perform a recursive scan operation:

[0101] ;

[0102] in, Indicates the current time The hidden state vector is used to contain historical context information; This represents the hidden state at the previous moment; This represents the time-varying state transition matrix calculated based on the corrected effective sampling step size; This represents the time-varying input projection matrix calculated based on the corrected effective sampling step size; This represents the input features at the current moment;

[0103] The final output is a deep global feature sequence that contains global contextual information and is insensitive to heart rate changes. .

[0104] Step 4: Feature fusion. Based on the high-dimensional shallow feature sequence and the deep global feature sequence, multi-scale refined features are obtained.

[0105] In this embodiment of the invention, step 4 includes:

[0106] The high-dimensional shallow feature sequence and the deep global feature sequence are merged by the full-frequency sensing fusion module. The fusion weight is calculated by the KAN gating mechanism and the key features are screened by the channel attention mechanism to obtain multi-scale refined features.

[0107] In this embodiment of the invention, the function of full-frequency sensing fusion and refinement is multi-scale detail recovery. Because the Mamba backbone network tends to extract low-frequency global rhythms, it may lose high-frequency microscopic morphological details (such as diphtheria notches). It utilizes the KAN gating mechanism to intelligently re-inject the local details from the front end into the global features, ensuring the integrity of the features.

[0108] Step 41: Calculate the KAN gating weights;

[0109] High-dimensional shallow feature sequences With deep global feature sequences Concatenate the inputs to the KAN gated network to calculate the fusion gate weights. Specifically, this KAN gated network employs a bottleneck structure to compress parameters and extract key information, as shown in Figure 1. It consists of two KAN linear layers: the first KAN linear layer maps the concatenated features to the latent space, followed by processing with the GELU activation function; the second KAN linear layer restores the features to the original channel dimension; finally, the output is processed by the Sigmoid activation function to generate fusion gate weights with values ​​ranging from [0,1]. :

[0110] ;

[0111] in, This represents the generated fusion gating weights, with values ​​ranging from [0,1]. This represents the Sigmoid activation function, used to normalize the output; This represents the KAN network module used to generate the gate weights; This represents a deep global feature sequence from the Mamba backbone network; This represents a high-dimensional shallow feature sequence originating from the front end of physiological features; This indicates a splicing operation along the channel dimension;

[0112] Step 42, Shape Perception and Residual Fusion Module;

[0113] Weighted fusion is performed using the LayerScale strategy:

[0114] ;

[0115] Subsequently, the fused features The input shape-aware module captures waveform details at multiple scales, as shown in Figure 2. The specific processing flow of this module is as follows:

[0116] First, features The features are then subjected to dimensionality reduction and nonlinear mapping using a 1×1 convolutional layer and the GELU activation function. Subsequently, the features are split into two parallel branches, which pass through 3×3 and 7×7 depthwise convolutions respectively, to capture local morphological features in the receptive field at different times.

[0117] The outputs of the two branches are concatenated along the channel dimension, and then the input channel attention mechanism is used for feature calibration.

[0118] The calculation process of the channel attention mechanism is as follows: features First, the spatial dimension is compressed using a global average pooling layer. Then, the input features are sequentially processed through a first convolutional layer, a GELU activation function, a second convolutional layer, and a Sigmoid activation function to generate channel weight coefficients. These weight coefficients are multiplied element-wise with the input features to achieve feature recalibration. Finally, the features are processed through a 1×1 convolutional layer for the final fusion and adjustment of channel information.

[0119] Step 5: Output. Based on the multi-scale refined features, obtain the final biometric recognition result.

[0120] In this embodiment of the invention, step 5 includes:

[0121] Multi-scale refined features are input into the morphologically perceptual hybrid head, mapped to the discriminative feature space, and the final biometric recognition result is output based on the hybrid loss function.

[0122] To adaptively select key feature channels, a channel attention mechanism is introduced, the calculation process of which is as follows:

[0123] First, the input features are compressed in spatial dimension using global average pooling to generate channel descriptors. Then, a non-linear transformation is performed through two 1×1 convolutional layers: the first convolution reduces the channel dimension (the reduction ratio can be set to 4 or 8) and passes it through the GELU activation function, while the second convolution restores the number of channels to the original dimension. The output is then processed by the Sigmoid activation function to generate a channel weight vector in the range [0,1], and is multiplied element-wise with the original input features to suppress noisy channels and enhance the feature response containing physiological information.

[0124] Finally, the features enhanced by channel attention are input into the morphological perception hybrid head. To simultaneously utilize label information and the inherent structural information of the data, the morphological perception hybrid head is designed as two parallel task branches, as shown in Figure 1. The first branch is a classifier based on angular margins: this branch is used for identity recognition tasks. Features are first mapped to the class space through fully connected layers, and the classification probability is calculated by combining angular margin loss, thereby maximizing inter-class differences in the feature space and improving recognition accuracy. The second branch is a contrastive learning projection head: this branch is used for self-supervised contrastive learning tasks. Features are subjected to nonlinear projection (usually including fully connected layers, normalization layers, etc.), and the output embedding vector is used to calculate the contrastive loss, so as to narrow the feature distance of the same user under different data augmentation views.

[0125] In this embodiment of the invention, the role of hybrid loss-supervised learning is to shape the feature space and enhance discriminative power. Considering the characteristics of PPG signals—large intra-class differences (different motion states) and small inter-class differences (similar waveforms)—the hybrid loss function constrains the model to learn discriminative features that are compact within classes and separable between classes.

[0126] In this embodiment of the invention, it includes:

[0127] Using a hybrid loss function Perform end-to-end training;

[0128] in, This represents the total loss function for model training; This represents the angular margin classification loss, used to maximize the angular distance between different identity features on the hypersphere to improve inter-class separability; InfoNCE represents the self-supervised contrastive loss, used to narrow the feature distance of the same user under different data augmentation views, thereby improving robustness against interference. This represents the balance coefficient, used to adjust the weight ratio of the two loss terms.

[0129] The alternatives to the rate estimation method in this invention, besides using FFT+KAN, can also be wavelet transform or short-time Fourier transform (STFT) combined with other nonlinear regression networks (such as multinomial regression networks) to estimate the physiological rate r; the alternative to the backbone network, although the embodiments of this invention use Mamba, the idea of ​​rate-aware step-size modulation can also be applied to other neural networks based on continuous-time recursion (such as CTRNN, Liquid Neural Networks); the alternative to the application scenario, this method is not only applicable to PPG signals, but also to other physiological signal recognition tasks with quasi-periodic characteristics such as electrocardiogram (ECG) and respiration.

[0130] The processing flow of the method of this invention mainly includes five cascaded stages:

[0131] In the feature mapping stage, the original single-channel PPG signal is first mapped into a high-dimensional shallow feature sequence containing rich local morphological information by passing through the physiological feature front end.

[0132] The dual-stream collaborative modeling stage (core): The data stream is distributed to two parallel branches, working collaboratively through a control-execution mechanism: Control Stream (Auxiliary Branch): The signal enters a spectrum-guided frequency estimator. This branch does not extract identity features but focuses on solving the physiological rhythm of the signal, using a KAN network to regress the instantaneous physiological rate curve (Rate Curve), i.e., the trend of heart rate changes, from the frequency domain; Data Stream (Main Branch): The signal enters the periodic sensing Mamba backbone network. This branch is responsible for extracting long-range identity dependencies. Its key feature is that it receives the physiological rate curve output from the control stream as instructions and dynamically adjusts the scan step size of the internal state-space model (SSM). When the heart rate increases, the step size automatically increases; when the heart rate decreases, the step size decreases. This mechanism achieves real-time cancellation of temporal scaling during feature extraction.

[0133] Feature fusion and output stage: Global features processed by Mamba and local features from the front end are merged in the frequency-aware fusion module. The system uses a gating mechanism to recover smoothed high-frequency details (such as diphtheria notch). Finally, the features are mapped to the discrimination space through the morphological-aware hybrid head, and the final biometric recognition result is output.

[0134] The technical solution provided by this invention includes a method comprising: feature mapping, which extracts physiological features from the original single-channel PPG signal to obtain a high-dimensional shallow feature sequence; dual-stream collaborative modeling, which performs dual-stream collaborative processing on the high-dimensional shallow feature sequence and distributes it to two parallel branches: a control flow branch and a data flow branch; in the control flow branch, the amplitude spectrum and instantaneous physiological rate curve are obtained; in the data flow branch, a deep global feature sequence with rate invariance is obtained; feature fusion, which obtains multi-scale refined features based on the high-dimensional shallow feature sequence and the deep global feature sequence; and output, which obtains the final biometric recognition result based on the multi-scale refined features. This method can actively sense changes in physiological rate and achieve efficient nonlinear modeling with extremely low parameter count.

[0135] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying PPGs with resistance to migration based on rate sensing and state-space model, characterized in that, The method includes: Step 1, feature mapping, extracting physiological features from the original single-channel PPG signal to obtain a high-dimensional shallow feature sequence; Step 2, dual-stream collaborative modeling, processing the high-dimensional shallow feature sequence using a dual-stream collaborative method and distributing it to two parallel branches: a control flow branch and a data flow branch; Step 3, in the control flow branch, acquiring the amplitude spectrum and instantaneous physiological rate curve; in the data flow branch, acquiring a deep global feature sequence with rate invariance; in the control flow branch, performing a global frequency domain transformation on the high-dimensional shallow feature sequence to obtain the amplitude spectrum, and utilizing Kjeldahl... The KAN network (Kinetic Anomaly AN) regresses the instantaneous physiological rate curve. In the data flow branch, the high-dimensional shallow feature sequence is input into the periodic sensing Mamba backbone network. The discrete step size of the state-space model (SSM) is dynamically modulated using the instantaneous physiological rate curve to perform rate-aware selective scanning and output a deep global feature sequence with rate invariance. Step 4: Feature fusion. Based on the high-dimensional shallow feature sequence and the deep global feature sequence, multi-scale refined features are obtained. Step 5: Output. Based on the multi-scale refined features, the final biometric recognition result is obtained.

2. The method according to claim 1, characterized in that, Step 1 includes: extracting physiological features from the original single-channel PPG signal, enhancing local morphological features using a parameterized Gabor filter bank, and obtaining a high-dimensional shallow feature sequence through a convolutional projection layer; Step 11, parameterized Gabor filtering; the sequence of the original single-channel PPG signal is... A parameterized Gabor filter bank is introduced for convolution processing; the Gabor kernel function is defined as: ;in, The time index variable representing the convolution kernel; center frequency. ,bandwidth and amplitude All parameters are automatically learned by the network; Step 12, convolutional projection and downsampling; The filtered signal is passed through a one-dimensional convolutional layer with a kernel size of 7 and a stride of 2 to expand the number of channels to And halve the length of the time series. Generate high-dimensional shallow feature sequences 。 3. The method according to claim 2, characterized in that, Step 3 includes: Step 31, obtaining the amplitude spectrum; First, the high-dimensional shallow feature sequence is input into a normalization layer for standardization, and a linear layer is used for preliminary mapping of the feature dimensions; then, global average pooling is performed on the features to compress the time dimension and obtain the global context vector; next, Fast Fourier Transform (FFT) is used to transform the signal from the time domain to the frequency domain, and the amplitude spectrum is obtained by taking the modulus. ; and introduce a frequency domain masking noise enhancement mechanism; Step 32, KAN-based rate regression; use the Kolmogorov-Arnold network (KAN) to process the amplitude spectrum, and KAN replaces the original activation function with a learnable B-spline function (B-Spline) on the edges: ;in, This represents the input values ​​on the connection edges in the KAN layer of the network, i.e., the components from the spectral features; This indicates the number of B-spline basis functions, used to determine the fineness of the fit; Indicates the first The learnable coefficients corresponding to each basis; Indicates the first The network uses B-spline basis functions; the final output of the KAN network is processed by the Softplus activation function to ensure that the generated rate factor is always positive; the scalar broadcast of the rate factor is extended to be proportional to the sequence length. A consistent vector, denoted as the instantaneous physiological rate curve. 。 4. The method according to claim 3, characterized in that, Step 3 further includes: Step 33, rate sensing discretization; based on the continuous equation of the state-space model SSM, a physiological rate modulation mechanism is introduced to adjust the time step. Make corrections: ;in, This represents the effective sampling step size after physiological rate correction; This represents the original step size inferred solely from the current input; This represents the instantaneous physiological rate output by the frequency estimator guided by the spectrum. Indicates the current time Input features; and Representing the activation function and linear projection layer respectively; Step 34, parameter transformation and scanning; utilizing the corrected effective sampling step size. Calculate the parameters of the discretized system And perform a recursive scan operation: ;in, Indicates the current time The hidden state vector is used to contain historical context information; This represents the hidden state at the previous moment; This represents the time-varying state transition matrix calculated based on the corrected effective sampling step size; This represents the time-varying input projection matrix calculated based on the corrected effective sampling step size; The input features are represented at the current moment; the final output is a deep global feature sequence that has global context information and is insensitive to heart rate changes. 。 5. The method according to claim 4, characterized in that, Step 4 includes: merging the high-dimensional shallow feature sequence and the deep global feature sequence through a full-frequency sensing fusion module, calculating the fusion weights using the KAN gating mechanism, and filtering key features using a channel attention mechanism to obtain multi-scale refined features; Step 41: KAN gating weight calculation; merging the high-dimensional shallow feature sequence... With deep global feature sequences Concatenate the inputs to the KAN gated network to calculate the fusion gate weights. ： ;in, This represents the generated fusion gating weights, with values ​​ranging from [0,1]. This represents the Sigmoid activation function, used to normalize the output; This represents the KAN network module used to generate the gate weights; This represents a deep global feature sequence from the Mamba backbone network; This represents a high-dimensional shallow feature sequence originating from the front end of physiological features; This indicates a stitching operation along the channel dimension; Step 42, Shape Awareness and Residual Fusion Module; Weighted fusion is performed using the LayerScale strategy: Subsequently, the fused features Input the shape-aware module to capture waveform details at multiple scales.

6. The method according to claim 5, characterized in that, Step 5 includes: inputting multi-scale refined features into the morphologically-aware hybrid head, mapping them to the discriminative feature space, and outputting the final biometric recognition result based on the hybrid loss function; in order to adaptively select key feature channels, a channel attention mechanism is introduced, the calculation process of which is as follows: first, the input features are compressed in space dimension by global average pooling to generate channel descriptors; then, a nonlinear transformation is performed through two 1×1 convolutional layers: the first convolution reduces the channel dimension and passes it through the GELU activation function, and the second convolution restores the number of channels to the original dimension; the output is processed by the Sigmoid activation function to generate a channel weight vector in the range [0,1], and multiplied element-wise with the original input features; finally, the features enhanced by channel attention are input into the morphologically-aware hybrid head; in order to utilize both label information and the inherent structural information of the data, the morphologically-aware hybrid head is designed as two parallel task branches, the first branch is a classifier based on angle margins, and the second branch is a contrastive learning projection head.

7. The method according to claim 6, characterized in that, include: Using a hybrid loss function Perform end-to-end training; among which, This represents the total loss function for model training; This represents the angular margin classification loss, used to maximize the angular distance between different identity features on the hypersphere to improve inter-class separability; InfoNCE represents the self-supervised contrastive loss, used to narrow the feature distance of the same user under different data augmentation views, thereby improving robustness against interference. This represents the balance coefficient, used to adjust the weight ratio of the two loss terms.

Images