Anti-noise ppg biometric identification method based on spectral-kan discretization and multi-scale convolution cooperation
By employing a noise-resistant PPG biometric method that combines spectral-KAN discretization with multi-scale convolution, the computational efficiency and recognition accuracy issues of PPG signals in dynamic scenarios are resolved, enabling real-time high-precision identity authentication on wearable devices.
Patent Information
- Application Number
- CN202610360995.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-24
- Publication Date
- 2026-05-19
- Estimated Expiration
- 2046-03-24
AI Technical Summary
Existing PPG biometric technologies face challenges in dynamic application scenarios, including low computational efficiency, difficulty in modeling long-range dependencies, and non-stationary distribution shifts, making real-time identification particularly difficult to achieve on wearable devices.
A noise-resistant PPG biometric method combining spectral-KAN discretization and multi-scale convolution is adopted. High-dimensional feature tensors are extracted through physiological feature front-end, combined with a dual-stream collaborative feature model and a state-space model, and the sampling step size is dynamically adjusted using Haar wavelet transform and KAN network. Combined with multi-scale convolution and channel attention, adaptive recognition of heart rate changes is achieved.
It significantly improves noise resistance and recognition accuracy, reduces computational latency and resource consumption, and meets the real-time deployment requirements of wearable devices.
Smart Images
Figure CN121905542B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a noise-resistant PPG biometric method based on spectral-KAN discretization and multi-scale convolution collaboration, belonging to the field of biometric recognition technology. Background Technology
[0002] Photoplethysmography (PPG)-based biometric identification technology holds great promise for wearable device authentication due to its non-invasive nature and liveness detection capabilities. However, in real-world dynamic applications, existing identification technologies face three major challenges: computational efficiency, long-range dependency modeling, and motion artifacts.
[0003] Traditional PPG biometric identification methods primarily rely on manually designed features, such as extracting morphological features of PPG signals like peaks, troughs, and time intervals, and then combining these with machine learning classifiers for identification. These methods have the following significant drawbacks:
[0004] ① Limitations of traditional CNN architecture: Although convolutional neural networks (CNNs) are lightweight, they are limited by local receptive fields and have difficulty capturing long-term hemodynamic dependencies in PPG signals that span multiple heartbeat cycles.
[0005] ② Computational bottleneck of Transformer architecture: Although Transformer can capture global dependencies, the computational complexity of its self-attention mechanism increases quadratically with the sequence length (O(L²)), resulting in high latency and high power consumption, making it difficult to deploy on edge devices (such as wristbands) with millisecond-level response requirements.
[0006] ③ The inherent defects of existing SSMs (such as Mamba) in physiological signal processing:
[0007] The drawback of linear discretization: Standard Mamba models typically calculate the discretization step size Δ based on the current input using linear projection. This linear approach assumes that the signal changes are smooth and regular.
[0008] Unable to handle non-stationarity: Real-world PPG signals exhibit significant non-stationarity due to user movement or emotional changes. Heart rate (frequency) fluctuates dramatically, causing the signal to "stretch and warp" over time. Simple linear discretization mechanisms cannot detect these complex physiological rhythm changes, leading to the model losing synchronization with the heartbeat rhythm and a sharp decline in recognition accuracy.
[0009] Incomplete feature extraction: Relying solely on SSM can easily focus on global trends and smooth out key morphological details (such as diphtheria notch), resulting in large intraclass variance and insufficient discriminative power.
[0010] Therefore, there is an urgent need for a lightweight PPG recognition method that can adapt to non-stationary heart rate changes, resist motion interference, and has extremely low computational latency. Summary of the Invention
[0011] To address the shortcomings of existing technologies, this invention provides a noise-resistant PPG biometric identification method based on spectral-KAN discretization and multi-scale convolution collaboration. This method can effectively adapt to non-stationary heart rate changes in PPG signals, overcome distribution offset problems, significantly improve noise robustness, and has extremely low computational latency and resource consumption, meeting the real-time deployment requirements of wearable devices.
[0012] The present invention adopts the following technical solution:
[0013] A noise-resistant PPG biometric identification method based on spectral-KAN discretization and multi-scale convolution collaboration includes the following steps:
[0014] S1 extracts physiological features from the single-channel PPG raw sequence and maps them to a high-dimensional feature tensor. ;
[0015] S2, Constructing a dual-stream collaborative feature model, high-dimensional feature tensor The final embedding vector is obtained through a two-stream collaborative feature model and global average pooling.
[0016] S3 involves inputting the embedded vector into the classification head and using a joint loss function for training and optimization.
[0017] Preferably, in step S1, the input data is a normalized single-channel PPG signal sequence. ,in, Represent real numbers, This indicates the number of channels, specifically the PPG signal per channel. The length of the time series is represented by the number of sampling points contained in the signal segment (e.g., L=256).
[0018] The physiological feature front-end performs dimensionality expansion and shallow feature extraction. The physiological feature front-end extraction process is implemented through one or more stacked convolutional blocks, each including a one-dimensional convolutional layer (Conv1d), batch normalization (BatchNorm), and a ReLU activation function. The physiological feature front-end extraction maps the single-channel PPG raw sequence into a high-dimensional feature tensor. The temporal resolution is gradually reduced to provide input rich in local information for subsequent deep modeling.
[0019] The number of convolutional blocks is not limited to two. It can be set to other numbers (such as one, three or more) according to actual needs. In practical applications, the number of convolutional blocks can be determined as needed. If the input signal is very long, more convolutional blocks may be needed for downsampling; if the signal is short, only one convolutional block may be needed.
[0020] Each convolutional block in this invention contains a convolution operation with a stride of 2. With each convolutional block, the temporal length (temporal resolution) of the signal is halved. The more convolutional blocks are stacked, the greater the compression factor of the signal in the temporal dimension, and the greater the reduction in temporal resolution (fewer remaining time points), but at the same time, the extracted features become more abstract and higher-level.
[0021] Preferably, in step S2, the implementation process of the dual-stream collaborative feature model includes three stages, each stage consisting of one or more stacked SK-Mamba modules;
[0022] Each SK-Mamba module consists of a first sub-layer of a two-stream mixer and a second sub-layer of a feedforward neural network. The first and second sub-layers are connected in series via residuals to form a complete SK-Mamba module.
[0023] In this invention, the number of SK-Mamba modules (i.e., the network depth) in each stage can be adjusted. It can be one or more, for example, two SK-Mamba modules can be stacked in stage 1, two SK-Mamba modules can be stacked in stage 2, three SK-Mamba modules can be stacked in stage 3, etc., depending on the requirements for model capacity.
[0024] Preferably, layer normalization is provided in both the first and second sub-layers. In the dual-stream mixer, after layer normalization (LayerNorm), the features are split into two parts in the channel dimension, denoted as... , They are then fed into the global branch and the local branch for parallel processing, respectively.
[0025] The outputs of the global and local branches are concatenated and then subjected to a channel shuffle operation. The channel shuffle disrupts the channel order, allowing global rhythmic information and local morphological information to fully interact in subsequent layers, and finally output through the projection layer.
[0026] Preferably, the projection layer is a 1×1 convolutional layer or a fully connected layer.
[0027] Preferably, the feedforward neural network consists of two fully connected layers (or convolutional layers) and an activation function for feature transformation.
[0028] Preferably, the global branch aims to solve the non-stationary distribution migration problem by utilizing frequency domain priors, and the process is as follows:
[0029] To capture physiological fundamental frequencies unaffected by motion, this invention introduces Haar wavelet transform. First, the characteristics... After passing through linear layers and convolutional layers, the model enters the Haar wavelet transform module. The linear layers are used for feature dimension mapping, projecting the dimensions of the input features onto the dimensions of the internal hidden layers, increasing model capacity. The convolutional layers capture local contextual information. Before entering the state-space model (SSM, such as Mamba) to handle long-range dependencies, a small convolutional kernel aggregates local neighborhood information, improving the modeling stability of the SSM.
[0030] The SiLU activation function is used for nonlinear transformations and gating. It is the standard activation function in the Mamba architecture, controlling the strength of information flow into the SSM.
[0031] In the Haar wavelet transform module, define a low-frequency filter. and high frequency filters , for features Perform convolution operations to separate low-frequency approximations. High-frequency details :
[0032]
[0033] Approximating low frequencies High-frequency details splicing together yields spectral prior features ;
[0034] Haar wavelet transform is sensitive to signal abrupt changes (motion artifacts) and can effectively extract time-frequency features.
[0035] In order to fuse high and low frequency information and restore the channel to its original size, Perform a 1×1 convolution operation to obtain fused features. ;
[0036] Then, a Kolmogorov-Arnold network (KAN) is used to replace the traditional linear layer to predict the discretized step size Δ. KAN uses learnable B-splines as activation functions, which can accurately fit the complex nonlinear relationship between heart rate changes and the optimal sampling step size;
[0037] With fusion features As input, the discretization step size Δ is predicted based on the KAN layer, and the expression is:
[0038]
[0039] Here, Softplus() represents a smooth nonlinear activation function, whose mathematical expression is: The purpose is to ensure that the output step size Δ is always positive (because the physical time step cannot be negative); KAN() means that the operation is performed through the KAN layer (Kolmogorov-Arnold Network), which passes the input data into a KAN layer and uses the learnable B-spline function in the layer to perform nonlinear transformation on the input;
[0040] Finally, based on the discretization step size Δ, the state-space model first utilizes the zero-order hold (ZOH) principle to divide the parameters of the continuous system ( , Discretize into ( , ):
[0041]
[0042]
[0043] in, Represents the continuous-time state transition matrix; These represent continuous-time input projection matrices; they are learnable parameter matrices in neural networks. They are generated during model initialization (typically using the HiPPO initialization strategy) and updated and optimized during training using the gradient descent algorithm.
[0044] This represents the discretized state transition matrix; These represent the discretized input projection matrices; they are the actual parameters used to perform the recursive scan at discrete time steps.
[0045] express and Multiply; Represents the matrix exponential function; express and Multiply; This represents the identity matrix, with the same dimensions as matrix A;
[0046] Then a recursive scan operation is performed. , This represents the input feature vector at the current time t. It represents the hidden state vector of the previous time step t-1, and outputs global features containing long-range physiological rhythms; the recursive scan operation is performed within the state-space model.
[0047] Preferably, in the local branch, two parallel one-dimensional convolutional layers are used, namely large kernel convolution and small kernel convolution, and the output features are concatenated and then subjected to batch normalization and ReLU activation function;
[0048] To suppress motion noise, KAN-gated attention weights for the generation channels are introduced. :
[0049]
[0050]
[0051] in, for function; This represents the global average pooling operation, which averages the feature maps over time, compressing the time series into channel descriptors. This represents the multi-scale features after splicing; This indicates element-wise multiplication; KAN gating can non-linearly filter out channels containing effective physiological information and suppress noise channels.
[0052] Preferably, the large kernel convolution has a kernel value of K=15 to capture peak contours, while the small kernel convolution has a kernel value of K=5 to capture notch details. The kernel values of the large and small kernel convolutions can be adaptively adjusted according to the sampling rate of the input PPG signal, such as to 3 and 11, to adapt to the acquisition frequency of different devices.
[0053] Preferably, in step S3, the joint loss function is:
[0054]
[0055] in, Cross-entropy loss is used to determine classification accuracy. Angular margin loss is used to enhance the intra-class compactness and inter-class separability of features in the embedding space; The classification probability distribution predicted by the model is represented by the embedding vector. Obtained through mapping by a linear classifier; Indicates the true identity label of the sample;
[0056] This is a preset balancing coefficient used to balance the weights of the two loss functions. It is usually set based on experience or experiments, for example... Take 0.1.
[0057] In this invention, during the training phase, the classification prediction results and loss values are output.
[0058] During the inference / application phase: Output the identity recognition result (i.e., the predicted user ID) or the identity embedding vector (for comparison).
[0059] For any details not covered in this invention, please refer to the prior art.
[0060] The beneficial effects of this invention are as follows:
[0061] 1. This invention employs a spectral-KAN discretization (SKD) mechanism, which differs from the fixed or linear parameter mapping in existing technologies. This invention introduces Haar wavelet transform to extract frequency domain prior features unaffected by motion, and utilizes the powerful nonlinear fitting capability of KAN (Kolmogorov-Arnold Network) to dynamically predict the discretization step size Δ of the state-space model. This allows the model to adjust the sampling step size in real time according to the heart rate, achieving "locking" and synchronization with the heart rhythm. Even when vigorous exercise causes non-stationary fluctuations in heart rate (i.e., distribution shift), the model maintains stable feature extraction capabilities, solving the problem of traditional methods failing in dynamic scenarios.
[0062] 2. This invention designs a dual-stream collaborative feature modeling architecture. The global branch utilizes a state-space model (SSM) to capture long-range dependencies (physiological rhythms) spanning multiple heartbeat cycles; the local branch utilizes parallel big and small kernel multi-scale convolutions to capture microscopic morphological details such as peaks and notches; and deep fusion of the two features is achieved through channel shuffling. This overcomes the shortcomings of single models (such as pure CNN with limited field of view or pure SSM with loss of detail), enabling the model to obtain a more comprehensive and discriminative representation of identity features, thereby significantly improving recognition accuracy.
[0063] 3. This invention utilizes the linear computational complexity of the state-space model to replace the Transformer architecture, whose computational complexity grows quadratically. It also employs computationally efficient Haar wavelets and lightweight one-dimensional convolutions. While maintaining high accuracy, it significantly reduces the number of model parameters and computational load, achieving millisecond-level inference latency and possessing extremely low computational latency and resource consumption, meeting the real-time deployment requirements of wearable devices. Attached Figure Description
[0064] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.
[0065] Figure 1This is a schematic diagram of the overall process of the noise-resistant PPG biometric identification method based on spectral-KAN discretization and multi-scale convolution collaboration of the present invention.
[0066] Figure 2 A schematic diagram of the structure of a single SK-Mamba module;
[0067] Figure 3 This is a schematic diagram of the processing flow of the dual-stream mixer in the SK-Mamba module. Detailed Implementation
[0068] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. However, this is not the only description; all aspects not described in detail herein are based on conventional techniques in the art.
[0069] Terminology Explanation:
[0070] PPG (Photoplethysmography): Photoplethysmography is an optical measurement method that uses optical technology to detect changes in blood volume in living tissue.
[0071] SSM: A deep learning architecture based on a state space model (SSM), such as Mamba, with linear computational complexity and good at handling long sequences.
[0072] KAN (Kolmogorov-Arnold Networks): Kolmogorov-Arnold networks are a new type of neural network that places learnable activation functions (such as B-splines) on the edges (weights), and has a stronger nonlinear fitting ability compared to traditional MLPs.
[0073] Spectral-KAN Discretization (SKD): The core mechanism proposed in this invention refers to the technique of extracting frequency priors using wavelet transform and nonlinearly predicting the SSM discretization step size using a KAN network.
[0074] Example 1
[0075] A noise-resistant PPG biometric method based on spectral-KAN discretization and multi-scale convolution collaboration, such as... Figure 1 As shown, it includes the following steps:
[0076] S1 extracts physiological features from the single-channel PPG raw sequence and maps them to a high-dimensional feature tensor. ;
[0077] S2, Constructing a dual-stream collaborative feature model, high-dimensional feature tensor The final embedding vector is obtained through a two-stream collaborative feature model and global average pooling.
[0078] S3 involves inputting the embedded vector into the classification head and using a joint loss function for training and optimization.
[0079] Example 2
[0080] A noise-resistant PPG biometric identification method based on spectral-KAN discretization and multi-scale convolution is proposed, as described in Example 1, except that in step S1, the input data is a normalized single-channel PPG signal sequence. ,in, Represent real numbers, This indicates the number of channels, specifically the PPG signal per channel. The length of the time series is represented by the number of sampling points contained in the signal segment (e.g., L=256).
[0081] The physiological feature front-end performs dimensionality expansion and shallow feature extraction. The physiological feature front-end extraction process is implemented through one or more stacked convolutional blocks, each including a one-dimensional convolutional layer (Conv1d), batch normalization (BatchNorm), and a ReLU activation function. The physiological feature front-end extraction maps the single-channel PPG raw sequence into a high-dimensional feature tensor. The temporal resolution is gradually reduced to provide input rich in local information for subsequent deep modeling.
[0082] In this embodiment, there are two convolutional blocks. Each convolutional block contains a convolution operation with a stride of 2. With each convolutional block, the temporal length (temporal resolution) of the signal is halved. The more convolutional blocks are stacked, the greater the compression factor of the signal in the temporal dimension, and the greater the reduction in temporal resolution (fewer remaining time points). However, the extracted features become more abstract and higher-level.
[0083] Example 3
[0084] A noise-resistant PPG biometric identification method based on spectral-KAN discretization and multi-scale convolution collaboration, as described in Example 2, differs in that the implementation process of the dual-stream collaborative feature model in step S2 includes three stages, each stage consisting of one or more stacked SK-Mamba modules.
[0085] like Figure 2 As shown, each SK-Mamba module includes a first sub-layer of a two-stream mixer and a second sub-layer of a feedforward neural network. The first and second sub-layers are connected in series via residuals to form a complete SK-Mamba module.
[0086] In this embodiment, stage 1 stacks 2 SK-Mamba modules, stage 2 stacks 2 SK-Mamba modules, and stage 3 stacks 3 SK-Mamba modules.
[0087] Layer normalization is implemented in both the first and second sub-layers. In the two-stream mixer, after layer normalization (LayerNorm), the features are split into two parts along the channel dimension, denoted as... , They are then fed into the global branch and the local branch for parallel processing, respectively.
[0088] The outputs of the global and local branches are concatenated and then subjected to a channel shuffle operation. The channel shuffle disrupts the channel order, allowing global rhythmic information and local morphological information to fully interact in subsequent layers, and finally output through the projection layer.
[0089] The projection layer is a 1×1 convolutional layer or a fully connected layer.
[0090] A feedforward neural network consists of two fully connected layers (or convolutional layers) and an activation function for feature transformation.
[0091] Example 4
[0092] A noise-resistant PPG biometric identification method based on spectral-KAN discretization and multi-scale convolution collaboration, as described in Example 3, differs in that the global branch aims to solve the non-stationary distribution shift problem using frequency domain priors. The process is as follows:
[0093] To capture physiological fundamental frequencies unaffected by motion, this invention introduces Haar wavelet transform. First, the characteristics... After passing through linear layers and convolutional layers, the model enters the Haar wavelet transform module. The linear layers are used for feature dimension mapping, projecting the dimensions of the input features onto the dimensions of the internal hidden layers, increasing model capacity. The convolutional layers capture local contextual information. Before entering the state-space model (SSM, such as Mamba) to handle long-range dependencies, a small convolutional kernel aggregates local neighborhood information, improving the modeling stability of the SSM.
[0094] The SiLU activation function is used for nonlinear transformations and gating. It is the standard activation function in the Mamba architecture, controlling the strength of information flow into the SSM.
[0095] In the Haar wavelet transform module, define a low-frequency filter. and high frequency filters , for features Perform convolution operations to separate low-frequency approximations. High-frequency details :
[0096]
[0097] Approximating low frequencies High-frequency details splicing together yields spectral prior features ;
[0098] Haar wavelet transform is sensitive to signal abrupt changes (motion artifacts) and can effectively extract time-frequency features.
[0099] In order to fuse high and low frequency information and restore the channel to its original size, Perform a 1×1 convolution operation to obtain fused features. ;
[0100] Then, a Kolmogorov-Arnold network (KAN) is used to replace the traditional linear layer to predict the discretized step size Δ. KAN uses learnable B-splines as activation functions, which can accurately fit the complex nonlinear relationship between heart rate changes and the optimal sampling step size;
[0101] With fusion features As input, the discretization step size Δ is predicted based on the KAN layer, and the expression is:
[0102]
[0103] Here, Softplus() represents a smooth nonlinear activation function, whose mathematical expression is: The purpose is to ensure that the output step size Δ is always positive (because the physical time step cannot be negative); KAN() means that the operation is performed through the KAN layer (Kolmogorov-Arnold Network), which passes the input data into a KAN layer and uses the learnable B-spline function in the layer to perform nonlinear transformation on the input;
[0104] Finally, based on the discretization step size Δ, the state-space model first utilizes the zero-order hold (ZOH) principle to divide the parameters of the continuous system ( , Discretize into ( , ):
[0105]
[0106]
[0107] in, Represents the continuous-time state transition matrix; These represent continuous-time input projection matrices; they are learnable parameter matrices in neural networks. They are generated during model initialization (typically using the HiPPO initialization strategy) and updated and optimized during training using the gradient descent algorithm.
[0108] This represents the discretized state transition matrix; These represent the discretized input projection matrices; they are the actual parameters used to perform the recursive scan at discrete time steps.
[0109] express and Multiply; Represents the matrix exponential function; express and Multiply; This represents the identity matrix, with the same dimensions as matrix A;
[0110] Then a recursive scan operation is performed. , This represents the input feature vector at the current time t. It represents the hidden state vector of the previous time step t-1, and outputs global features containing long-range physiological rhythms; the recursive scan operation is performed within the state-space model.
[0111] Example 5
[0112] A noise-resistant PPG biometric identification method based on spectral-KAN discretization and multi-scale convolution collaboration, as described in Example 4, differs in that two parallel one-dimensional convolutional layers are used in the local branch, namely large kernel convolution and small kernel convolution, and the output features are concatenated and then subjected to batch normalization and ReLU activation function;
[0113] To suppress motion noise, KAN-gated attention weights for the generation channels are introduced. :
[0114]
[0115]
[0116] in, for function; This represents the global average pooling operation, which averages the feature maps over time, compressing the time series into channel descriptors. This represents the multi-scale features after splicing; This indicates element-wise multiplication; KAN gating can non-linearly filter out channels containing effective physiological information and suppress noise channels.
[0117] Preferably, in this embodiment, the kernel K=15 of the large kernel convolution is used to capture the peak contour, and the kernel K=5 of the small kernel convolution is used to capture the notch details.
[0118] like Figure 3 The diagram shows a two-stream mixer. The left side is the global branch, which demonstrates the process of extracting frequency domain features using Haar wavelet transform and generating a discretized step size Δ through 1×1 convolution, a simplified KAN layer, and a Softplus activation function, thereby controlling the state-space model (SSM). The right side is the local branch, which demonstrates parallel convolution with large and small kernels (K=15, K=5) and a KAN-based gating mechanism. The top shows the feature concatenation, channel shuffling, and projection output process of the two branches.
[0119] Example 6
[0120] A noise-resistant PPG biometric identification method based on spectral-KAN discretization and multi-scale convolution collaboration, as described in Example 5, differs in that the joint loss function in step S3 is:
[0121]
[0122] in, Cross-entropy loss is used to determine classification accuracy. Angular margin loss is used to enhance the intra-class compactness and inter-class separability of features in the embedding space; The classification probability distribution predicted by the model is represented by the embedding vector. Obtained through mapping by a linear classifier; Indicates the true identity label of the sample;
[0123] This is a preset balancing coefficient used to balance the weights of the two loss functions. It is usually set based on experience or experiments. In this embodiment... Take 0.1.
[0124] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A noise-resistant PPG biometric identification method based on spectral-KAN discretization and multi-scale convolution collaboration, characterized in that, Includes the following steps: S1 extracts physiological features from the single-channel PPG raw sequence and maps them to a high-dimensional feature tensor. ; S2, Constructing a dual-stream collaborative feature model, high-dimensional feature tensor The final embedding vector is obtained through a two-stream collaborative feature model and global average pooling. S3 involves inputting the embedded vector into the classification head and using a joint loss function for training and optimization. In step S2, the implementation process of the dual-stream collaborative feature model includes three stages, each of which consists of one or more stacked SK-Mamba modules. Each SK-Mamba module consists of a first sub-layer of a two-stream mixer and a second sub-layer of a feedforward neural network. The first and second sub-layers are concatenated together via residuals to form a complete SK-Mamba module. Layer normalization is implemented in both the first and second sub-layers. In the dual-stream mixer, after layer normalization, the features are divided into two parts in the channel dimension, denoted as... , They are then fed into the global and local branches for parallel processing, respectively. The outputs of the global branch and the local branch are concatenated, then shuffled through the channel, and finally output through the projection layer. The global branch aims to solve the non-stationary distribution migration problem by utilizing frequency domain priors. The process is as follows: First, features After passing through a linear layer and a convolutional layer, the material enters the Haar wavelet transform module, where a low-frequency filter is defined. and high frequency filters , for features Perform convolution operations to separate low-frequency approximations. High-frequency details : Approximating low frequencies High-frequency details splicing together yields spectral prior features ; In order to fuse high and low frequency information and restore the channel to its original size, Perform a 1×1 convolution operation to obtain fused features. ; Then, using fusion features As input, the discretization step size Δ is predicted based on the KAN layer, and the expression is: Where Softplus() represents a smooth nonlinear activation function that ensures a positive step size; KAN() represents the operation performed through the KAN layer; Finally, based on the discretization step size Δ, the state-space model first uses the zeroth-order preservation principle to divide the parameters of the continuous system ( , Discretize into ( , ): in, Represents the continuous-time state transition matrix; This represents the continuous-time input projection matrix; This represents the discretized state transition matrix; This represents the discretized input projection matrix; express and Multiply; Represents the matrix exponential function; express and Multiply; This represents the identity matrix, with the same dimensions as matrix A; Then a recursive scan operation is performed. , This represents the input feature vector at the current time t. This represents the hidden state vector at the previous time t-1, and the output contains global features that include long-range physiological rhythms. In the local branch, two parallel one-dimensional convolutional layers are used, namely large kernel convolution and small kernel convolution. The output features are concatenated and then subjected to batch normalization and ReLU activation function. To suppress motion noise, KAN-gated attention weights for the generation channels are introduced. : in, for function; This indicates a global average pooling operation; This represents the multi-scale features after splicing; This indicates element-wise multiplication.
2. The noise-resistant PPG biometric identification method based on spectral-KAN discretization and multi-scale convolution collaboration as described in claim 1, characterized in that, In step S1, The physiological feature extraction process is implemented through one or more stacked convolutional blocks, each of which includes a one-dimensional convolutional layer, batch normalization, and a ReLU activation function.
3. The noise-resistant PPG biometric identification method based on spectral-KAN discretization and multi-scale convolution collaboration according to claim 2, characterized in that, The projection layer is a 1×1 convolutional layer or a fully connected layer.
4. The noise-resistant PPG biometric identification method based on spectral-KAN discretization and multi-scale convolution collaboration according to claim 3, characterized in that, A feedforward neural network consists of two fully connected layers and an activation function for feature transformation.
5. The noise-resistant PPG biometric identification method based on spectral-KAN discretization and multi-scale convolution collaboration according to claim 4, characterized in that, Large kernel convolution with kernel K=15 is used to capture peak contours, while small kernel convolution with kernel K=5 is used to capture notch details.
6. The noise-resistant PPG biometric identification method based on spectral-KAN discretization and multi-scale convolution collaboration according to claim 5, characterized in that, In step S3, the joint loss function is: in, Cross-entropy loss is used to determine classification accuracy. Angular margin loss is used to enhance the intra-class compactness and inter-class separability of features in the embedding space; This is the preset balance coefficient; The classification probability distribution predicted by the model is represented by the embedding vector. Obtained through mapping by a linear classifier; This represents the true identity label of the sample.