Reliable knowledge routing and adaptive memory method for PPG identification

Abstract

The present application relates to the PPG identity recognition technical field, especially provide a kind of reliable knowledge routing and adaptive memory method for PPG identity recognition.The method includes obtaining uniform KRAM format sample, and constructs physiological multi-view;Physiological multi-view is input into multi-scale pulse state encoder, and current timing state is obtained;Five categories of pulse knowledge state are constructed;Reliability vector is constructed, and sample reliability score is obtained by KAN reliability mapper;Coding feature is divided into local morphological segment, and morphological dictionary is matched and quantitatively reconstructed;The updating amplitude of exponential moving average is adjusted by sample reliability score;Token matrix is organized, and routing weight is generated, and output fusion identity representation;Fusion identity representation is input into identity discrimination head, and output identity recognition result, the method improves the time stability of continuous identity authentication result, reduces the false update of low-quality sample to morphological knowledge base and identity memory bank.

Description

Technical Field

[0001] This invention relates to the field of PPG identity recognition technology, and in particular to a reliable knowledge routing and adaptive memory method for PPG identity recognition. Background Technology

[0002] Photoplethysmography (PPG)-based biometric identification technology can perform continuous, contactless identity authentication using devices such as smartwatches, wristbands, and finger clip sensors, and has significant application value in wearable payments, mobile medical security, remote health monitoring, and human-computer interaction. PPG signals contain identity differences related to vascular elasticity, cardiac cycle, local peak-valley morphology, and individual physiological structure. However, PPG acquisition conditions in real wearable environments are complex, with signal quality, activity state, and sensor contact state constantly changing, leading to insufficient stability of existing identity recognition methods in low-quality segments and motion scenarios.

[0003] Traditional single-embedded representations struggle to distinguish stable identity knowledge from state perturbations. Existing deep learning methods typically map a PPG window directly to a single identity embedding before inputting it into a classifier or metric learning head for identity determination. This static paradigm of single window—single embedding—single classification can achieve some success on high-quality resting data, but in real-world wearable scenarios, the pulse amplitude, rise slope, dichroic waveform morphology, and rhythm stability all change when the same user is at rest, walking, running, or experiencing variations in wearing tightness or contact pressure. If the model only learns a single embedding representation, it easily couples stable identity morphology, changes in activity state, contact artifacts, and device noise into the same representation space, making it difficult to determine which information belongs to identity knowledge and which belongs to unreliable perturbations.

[0004] Quality-aware methods only suppress noise at the input or loss level, lacking knowledge base update control. Existing signal quality assessment or robust characterization methods typically screen, weight, or enhance training samples based on factors such as spectral concentration, rhythmic regularity, peak reliability, and motion intensity. These methods can reduce the impact of low-quality segments on training loss, but rarely address the contamination of the model's internal knowledge construction process by low-reliability samples. When low-quality segments are used for morphological dictionary learning or identity prototype updating, motion artifacts and state drift are written into the model's morphological dictionary or identity memory, leading to repeated referencing of contaminated knowledge during subsequent retrieval and fusion.

[0005] Fixed fusion structures struggle to adaptively select reliable knowledge sources based on different activity states. Real wearable PPG samples exhibit varying dependencies on knowledge sources: high-quality resting segments typically possess clear current temporal and pulse morphology features, while walking, running, or low-quality segments often require greater reliance on cross-window identity memory, physiological prior statistics, and perturbation state modeling for compensation. Traditional fixed-feature fusion structures primarily rely on hidden features or task loss, lacking explicit modeling of PPG sample quality reliability, activity state changes, and motion perturbation intensity. Therefore, they cannot dynamically adjust dependencies on temporal, morphological, memory, prior, and perturbation states based on sample status.

[0006] The temporal stability of identity representation is insufficient under continuous PPG windows. PPG signals have continuous temporal properties, and adjacent windows usually come from the same subject and maintain the same identity semantics. However, due to sliding window segmentation, local artifacts, or changes in motion state, the model output of adjacent windows may change irrationally. Existing methods mostly focus on the classification accuracy of a single window and lack constraints on the consistency between changes in the representation of adjacent windows and changes in identity prediction, resulting in insufficient temporal stability of identity representation in continuous identity authentication scenarios. Summary of the Invention

[0007] In view of this, the present invention provides a reliable knowledge routing and adaptive memory method for PPG identity recognition, which improves the temporal stability of continuous identity authentication results and reduces the erroneous updates of the morphological knowledge base and identity memory base by low-quality samples.

[0008] In a first aspect, the present invention provides a reliable knowledge routing and adaptive memory method for PPG identity recognition, the method comprising: Step 1: Unified Sample Construction and Physiological Multiview Construction: The original photoplethysmography (PPG) signal is processed to obtain a unified KRAM format sample; and a physiological multiview is constructed. Step 2, Pulse state encoding and current temporal state extraction: Input the physiological multi-view into the multi-scale pulse state encoder to extract local peaks and valleys, complete pulse cycles and long-term dependencies, and obtain the current temporal state; Step 3: Construction of five types of pulse knowledge states: Based on the current temporal state, construct the pulse morphology state, identity memory state, physiological prior state, and potential perturbation state to form a set of five types of pulse knowledge states; Step 4, Reliability Mapping: Construct a reliability vector and obtain the sample reliability score through the KAN reliability mapper; Step 5, Reliability-weighted morphological dictionary learning mechanism: The encoded features are divided into local morphological segments, matched and quantized with the morphological dictionary for reconstruction, and a reliability-weighted dictionary loss is adopted. Step 6, Reliability-Weighted Identity Memory Update Mechanism: Adjust the update magnitude of the exponential moving average by the sample reliability score; and mask uninitialized prototypes during the retrieval phase; Step 7, Reliability-Aware Knowledge-State Routing Mechanism: Based on the Token matrix, and combined with quality, reliability, activity, residual and disturbance information, routing weights are generated, and a fused identity representation is output. Step 8, Identity Discrimination and Joint Training: Input the fused identity representation into the identity discrimination head, output the identity recognition result, and constrain the multi-source knowledge state learning through joint optimization.

[0009] Optionally, the unified sample construction in step 1 includes: The input for PPG identification is a time window segmented from continuous physiological recordings. The raw photoplethysmography (PPG) signal undergoes missing value interpolation, bandpass filtering, outlier processing, sliding window segmentation, and length resampling to obtain a unified KRAM format sample set containing the PPG window, identification tag, activity status, quality surrogate score, time sequence number, and preceding window index. ; in, Let C be the number of channels and T be the number of uniform sampling points for the i-th PPG window. For identity tags; In active state; For quality agency scores; Use the window's start time or sequence number; The index of the preceding window that is located within the same data partition as the current window; In step 1, a physiological multiview is constructed from the original PPG signal, the first-order PPG signal, the second-order PPG signal, the smoothing trend, and the residual signal. This multiview includes: First, the input window is standardized at the fragment level, and then physiological multi-view is constructed: ; in, Indicates the first Physiological multi-view of a PPG window; This indicates splicing along the channel dimension or feature dimension; This represents the standardized PPG window; This represents the first-order discrete difference operator. Used to characterize the rising and falling edge variations of the PPG waveform; This represents a second-order discrete difference operator. Used to characterize waveform acceleration changes, curvature changes, and local peak-to-valley transitions; This represents the moving average smoothing operator, used to extract low-frequency trends; This represents the residual signal after detrending, used to preserve high-frequency disturbances, local artifacts, and fine-grained morphological changes; , , , and The five views correspond to the original PPG signal, the first-order PPG signal, the second-order PPG signal, the smoothing trend, and the residual signal, respectively.

[0010] Optionally, step 2 includes: The physiological multi-view is input into a multi-scale pulse state encoder composed of small kernel convolution, large kernel convolution and dilated convolution to extract local peaks and valleys, complete pulse cycles and long-term time dependencies, and the current time state is obtained through average pooling, max pooling and linear projection. The multi-scale pulse state encoder consists of small-kernel convolution, large-kernel convolution, dilated convolution, normalization layers, and nonlinear activation functions, used to capture local peak-valley morphology, complete pulse cycle structure, and long-term dependence, respectively; the output of the multi-scale pulse state encoder is denoted as... The current time-series state is obtained by average pooling, max pooling, and linear projection: ; ; in, Indicates the first Physiological multi-view of a PPG window; This represents a multi-scale pulse state encoder. Indicates model parameters; This represents the temporal feature map or hidden feature sequence output by the encoder, i.e., the encoded features; and These represent average pooling and max pooling operations, respectively, used to extract global statistical information and significant peak responses from time-series features; Indicates the projection layer. This indicates vector concatenation.

[0011] Optionally, step 3 includes: For the i-th sample, construct five types of pulse knowledge states. : ; in, This represents the current time series state, used to characterize the local peaks and valleys, periodic structure, and long-range time series dependencies within the current window; This refers to the pulse morphology, used to summarize relatively stable waveform structures. This serves as an identity memory state, used to maintain identity prototypes across windows, activities, and quality conditions. Physiological prior states are used to explicitly encode quality, rhythm, residuals, slope, and activity state; The potential perturbation state is used to represent the identity representation shift caused by motion artifacts and state drift.

[0012] Optionally, step 4 includes: Construct a reliability vector based on quality agent score, morphological dictionary matching confidence, morphological stability, identity memory consistency, potential perturbation energy, activity state, residual energy, and slope statistics; In a forward computation process, firstly, based on the existing pulse morphology dictionary, identity memory, and latent perturbation encoder, reliability evidence of morphology dictionary matching confidence, morphological stability, identity memory consistency, and latent perturbation energy is extracted from the current PPG window; subsequently, the KAN reliability mapper obtains the... Reliability score of each sample The reliability score is used to control the loss weight of the current morphology dictionary, the update magnitude of the identity memory EMA, and the routing weight of the knowledge state. Construct a reliability vector consisting of low-dimensional interpretable factors: ; in, Indicates the quality agent score; This indicates the confidence level of the morphological dictionary match. Indicates morphological stability; This indicates consistency with identity memory; Indicates potential disturbance energy; Indicates the activity status; Represents residual energy; Represents slope statistics; the reliability vector, after normalization, is input into the lightweight KAN reliability mapper to obtain the first... Reliability score of each sample : ; in, ∈[0.05,1], The larger the value, the more reliable the sample and the more suitable it is for participating in morphological dictionary learning and strong identity memory updates; The smaller the value, the more likely the sample is affected by motion artifacts, low-quality acquisition, or state drift, and its contribution to knowledge base writing and loss should be weakened. Construct a soft reliability proxy objective and constrain it with consistency loss: ; ; in, Transparent monitoring signals are jointly provided by factors such as quality, morphology, memory, and perturbation. This indicates that the gradient operation is stopped.

[0013] Optionally, step 5 includes: Pulse morphology fragments first match the morphology dictionary Matching and retrieval are performed, where K represents the number of dictionary prototypes and d represents the feature dimension after projection of the local morphological fragment, to obtain the corresponding dictionary prototypes. These prototypes are then quantized and reconstructed to form reconstructed fragments, and the sample-level dictionary loss is calculated. ; Given a temporal feature map or hidden feature sequence output by the encoder. The fragments are divided into local morphological segments and matched and quantized to reconstruct them using a learnable morphological dictionary. ; ; ; in, Indicates from The j-th local morphological fragment extracted from it; ( ) represents the fragment projection function, which maps local morphological fragments to the same feature dimension as the morphological dictionary; Let j be the projection representation of the j-th local morphological segment. It is the k-th prototype in the learnable morphology dictionary; Indicates the cosine similarity between the two; Represents the dictionary prototype index that best matches the current local morphological fragment; This represents the quantized local morphological fragment representation; This indicates that the gradient operation has been stopped. The quantized local morphological fragments are aggregated and residual compensated to obtain the pulse morphology state. ; A reliability-weighted dictionary loss is employed to ensure that high-reliability samples dominate the learning of stable morphological prototypes and suppress the dictionary update contribution of low-reliability samples. ; in, For the vector quantization and commitment loss of the i-th sample, This is a separation constraint between dictionary prototypes. These are the weighting coefficients.

[0014] Optionally, step 6 includes: Current sample information includes identity tags. Basic identity embedding Activity status and quality agent score For identity category y, maintain global identity memory. It also memorizes information based on activity status and quality bin maintenance conditions; during the training phase, the reliability-weighted average of samples with the same identity is first calculated: ; in, Indicates identity category Reliability-weighted identity prototype; Indicates the first The reliability score of each sample; To prevent stable terms with a denominator of zero; ( The ) indicates a normalization operation; subsequently, the identity memory is updated using the reliability-adjusted exponential moving average (EMA). ; in, For identity categories Global identity memory; Based on the update rate, The average reliability of the identity samples; During forward propagation, similar prototypes are retrieved from the initialized identity memory based on the current embedding, and the identity memory state is obtained by weighting the results. The memory retrieval weights and the identity memory state of the current sample are expressed as follows: ; ; in, This indicates the embedding of the basic identity of the current sample. Highest similarity A set of initialized identity memory prototypes; Indicates identity category The global identity memory prototype; Indicates the current sample pair of the th The retrieval weight of an individual identity memory prototype; ( This is used to normalize similarity into weights; This represents the identity memory state obtained by weighted fusion of similar identity memory prototypes.

[0015] Optionally, in step 7, the current temporal state, pulse morphology state, identity memory state, physiological prior state, and potential disturbance state are organized into a Token matrix; Physiological prior states are used to explicitly preserve low-dimensional, interpretable quality and physiological statistics, directly describing the amplitude, slope, residuals, quality, activity, and rhythm changes of the current PPG window: ; ; in, Indicates the first Low-dimensional physiological prior feature vectors of PPG windows; This indicates the statistics of window amplitude; Indicates the window mean or baseline level; Indicates the intensity of the first-order difference; Represents residual energy; Indicates the quality agent score; Indicates the normalized activity state; This represents the zero-crossing ratio of the differential signal; the constant 1 is used to provide the bias term. ( ) represents the physiological prior projection network; This represents the physiological prior state obtained by mapping low-dimensional physiological statistical information; In step 7, the potential perturbation state is used to describe motion artifacts, wearing variations, low-quality acquisition, and identity representation shifts caused by noise. The potential perturbation direction is constructed from various perturbation evidence, including the direction of random noise, the direction of perturbation samples obtained by motion enhancement or noise enhancement, the embedding difference between high-quality and low-quality windows of the same identity, the embedding difference under different activity states, and the perturbation direction inferred from acceleration, activity state, residual energy, or quality proxy score. Constructing from basic identity embeddings in the identity embedding space to the perturbation target vector The interpolated state is used to train the conditional perturbation encoder to predict the current potential perturbation direction. : ; ; in, The target vector for perturbation. These are the interpolation coefficients. For a conditional perturbation encoder, the output energy of the conditional perturbation encoder near s=0 is used as a perturbation intensity estimate and as input for reliability mapping and knowledge state routing.

[0016] Optionally, step 7 further includes: The pulse knowledge state router receives temporal state, morphological state, identity memory state, physiological prior state, and potential disturbance state. The pulse knowledge state router includes a KAN gated network, a core token, and a state routing weight α. Its goal is to adaptively allocate the fusion ratio of each knowledge source based on sample reliability and state information. Five categories of pulse knowledge states In the reliability-aware knowledge state routing phase, the tokens are organized into a fixed order based on the current temporal state, pulse morphology state, identity memory state, physiological prior state, and potential disturbance state. : The five types of knowledge states are ultimately determined by the reliability-aware router based on the token matrix. Perform fusion to obtain a fused identity representation and input it into the identity discrimination header: ; ; in, Indicates the first A set of five types of pulse knowledge states for each sample; Indicates by A token matrix organized in a fixed order; This represents the routing weight of the i-th sample to the k-th knowledge state; This represents the representation of the k-th knowledge state after interaction with the core token. ( ) indicates the normalization operation; This represents a fusion of identity representations; Indicates identity identification head; This represents the identity classification output logits or identity score vector corresponding to the i-th PPG window; Calculate the probability of each identity category based on the identity score vector, and select the category with the highest probability as the predicted identity result: ; in, Represents the predicted identity label for the i-th PPG window; k represents the identity category index. This represents the predicted probability of the k-th identity category; Construct routing metadata that includes quality, reliability, activity, and disturbance information: ; in, This represents routing metadata, used to provide routers with auxiliary information on quality, reliability, activity, residuals, identity memory consistency, and disturbance strength. During the training phase, non-basic knowledge sources are randomly deactivated before routing. Then, the router first extracts the core knowledge token, and then combines linear gating and KAN gating to obtain the fusion weights of the five types of states: ; in, Represents the linear gating parameters; ( ) represents the KAN gating function used for generating route weights; This represents the fusion coefficient of the KAN-gated branch.

[0017] Optionally, the pairwise window change consistency in step 8 is used to enhance the temporal stability of continuous PPG identity representation; for the i-th PPG window and its preceding window within the same data partition The models respectively obtain fused identity representations , and identity classification output , First, construct the change features between the i-th PPG window and the preceding window, and then use the change prediction head. Predict changes in logits to ensure that changes in the representation align with predicted changes: ; ; in, Represents the i-th PPG window The corresponding preceding window index; and Let i and n represent the fused identity representations of the i-th sample window and the preceding window, respectively. and These represent the identity classification output logits or identity score vectors of the i-th sample window and the preceding window, respectively. This represents the change characteristics consisting of the current window, the previous window, and the differences in their representations; This indicates vector concatenation; Represents the absolute difference element by element; This indicates a change prediction head, used to predict changes in identity scores from changes in characterization; Indicates the target of changes in the true identity score; This indicates that the gradient operation has been stopped. Indicates smoothness Loss is used to reduce the impact of abnormal windows on consistency constraints; Identity discrimination employs cross-entropy and ArcFace boundary enhancement objectives to improve inter-class separation capability. Simultaneously, it utilizes reliability-weighted morphological dictionary loss, identity memory consistency loss, latent perturbation modeling loss, reliability estimation loss, and pairwise window consistency loss to jointly constrain multi-source knowledge state learning. The final training objective is: ; in, The identity discrimination loss includes cross-entropy and ArcFace boundary enhancement targets; and λ represents the morphological dictionary loss, identity memory consistency loss, potential perturbation modeling loss, reliability estimation loss, and pairwise window change consistency loss, respectively; each λ term is the corresponding loss weight; through joint optimization, identity discrimination ability, stable morphological knowledge, reliable identity memory, perturbation perception ability, and cross-window time stability are learned.

[0018] The technical solution provided by this invention includes a method comprising: unified sample construction and physiological multi-view construction: processing the original photoplethysmography (PPG) signal to obtain a unified KRAM format sample; and constructing a physiological multi-view; pulse state encoding and current temporal state extraction: inputting the physiological multi-view into a multi-scale pulse state encoder to extract local peaks and valleys, the complete pulse cycle, and long-term dependence, and obtaining the current temporal state; construction of five types of pulse knowledge states: based on the current temporal state, constructing pulse morphology state, identity memory state, physiological prior state, and potential perturbation state, forming a set of five types of pulse knowledge states; reliability mapping: constructing a reliability vector, and obtaining the sample reliability score through a KAN reliability mapper; and a reliability-weighted morphological dictionary learning mechanism: encoding features... The feature is divided into local morphological segments, matched and quantized with the morphological dictionary for reconstruction, and a reliability-weighted dictionary loss is adopted. The reliability-weighted identity memory update mechanism adjusts the update magnitude of the exponential moving average by the sample reliability score and masks uninitialized prototypes during the retrieval stage. The reliability-aware knowledge state routing mechanism organizes the token matrix and generates routing weights by combining quality, reliability, activity, residual and perturbation information, and outputs the fused identity representation. Identity discrimination and joint training: the fused identity representation is input into the identity discrimination head, the identity recognition result is output, and the multi-source knowledge state learning is constrained by joint optimization. This method improves the temporal stability of continuous identity authentication results and reduces the erroneous updates of the morphological knowledge base and identity memory base by low-quality samples.

[0019] Compared with the prior art, the present invention has the following advantages: ① Reduce confusion between identity information and state perturbations. This invention extends the single PPG identity embedding to the current temporal state, pulse morphology state, identity memory state, physiological prior state, and potential perturbation state, enabling the model to distinguish stable identity knowledge, collect state changes and motion perturbation information, thereby improving the stability of identity discrimination in complex wearable scenarios.

[0020] ② Reduce erroneous updates of the morphological knowledge base and identity memory base by low-quality samples. The sample reliability score is used as a unified control variable to participate in the weighting of the morphological dictionary loss and the adjustment of the identity memory update amplitude. This allows high-reliability samples to dominate the formation of stable morphological prototypes and identity prototypes, while weakening the contribution of low-reliability samples to the writing of the knowledge base. This reduces the impact of low-quality fragments, motion artifacts, and state drift on the knowledge base.

[0021] ③ Improve the temporal stability of continuous identity authentication results. By using pairwise window change consistency constraints, the representation changes of adjacent PPG windows are kept consistent with the identity prediction changes, reducing abnormal jumps in identity output during continuous authentication.

[0022] ④ Improve the adaptability of knowledge source selection. The reliability-aware knowledge state router can dynamically adjust its dependence on temporal sequence, morphology, memory, prior and perturbation states based on sample quality, activity state, residual energy and perturbation intensity, thereby selecting more reliable identity discrimination criteria under different quality and activity conditions. Attached Figure Description

[0023] 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.

[0024] Figure 1 A flowchart illustrating a reliable knowledge routing and adaptive memory method for PPG identity recognition provided in an embodiment of the present invention; Figure 2 The flowchart below shows another reliable knowledge routing and adaptive memory method for PPG identity recognition provided in this embodiment of the invention, wherein (a) is the KAN reliability mapping mechanism, (b) is the reliability weighted morphological dictionary learning mechanism, (c) is the reliability weighted identity memory update mechanism, and (d) is the reliability-aware knowledge state routing mechanism. Detailed Implementation

[0025] 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.

[0026] 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.

[0027] 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.

[0028] 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)."

[0029] This invention provides a reliable knowledge routing and adaptive memory method for PPG identity recognition, such as... Figure 1 and Figure 2 As shown in (a) to (d) above, the method includes: The method of this invention includes a training phase and a deployment inference phase. In the training phase, identity tags, quality agent scores, activity status, time sequence numbers, and preceding window indices can be used to construct supervision signals and pairwise window change consistency constraints. In the deployment inference phase, only the current PPG window and optional quality agent scores and activity status information are input. After reliability estimation, knowledge state routing, and identity discrimination, the identity recognition result is output.

[0030] Step 1: Unified Sample Construction and Physiological Multiview Construction: Process the original photoplethysmography (PPG) signal to obtain a unified KRAM format sample; and construct a physiological multiview.

[0031] In this embodiment of the invention, the unified sample construction in step 1 includes: The input for PPG identification is not an isolated signal point, but a time window segmented from continuous physiological recordings. The raw photoplethysmography (PPG) signal undergoes missing value interpolation, bandpass filtering, abnormal segment processing, sliding window segmentation, and length resampling to obtain a unified KRAM format sample set containing the PPG window, identification tag, activity status, quality surrogate score, time sequence number, and preceding window index. ; in, Let C be the number of channels and T be the number of uniform sampling points. In the preferred embodiment, T is 256. For identity tags; In active state; For quality agency scores; Use the window's start time or sequence number; This is the index of the preceding window located within the same data partition as the current window. Preceding windows are preferentially selected from adjacent windows of the same subject and in the same activity state; if none exist, they degenerate into preceding windows of the same subject. This definition ensures that subsequent pairwise window consistency constraints are constructed only within the training, validation, or testing partitions, avoiding cross-partition time leakage.

[0032] Quality Agent Score Factors such as spectral concentration, rhythm regularity, heart rate reasonableness, artifact intensity, and exercise intensity are estimated to provide initial quality cues for the reliability mapper. It is important to emphasize that... It is not the final reliability score, nor does it directly determine identity prediction; the final reliability is determined by a combination of observational evidence and internal model evidence, including morphological dictionary matching confidence, morphological stability, identity memory consistency, potential perturbation energy, residual energy, and slope statistics. Precedence Window The pairwise window change consistency constraint is only used during the training phase and may not be required as input during the deployment and inference phase.

[0033] In this embodiment of the invention, step 1 involves constructing a physiological multiview from the original PPG signal, the first-order PPG signal, the second-order PPG signal, the smoothing trend, and the residual signal. This multiview includes: PPG identity information is not only present in the original waveform, but also reflected in local slope, second-order changes, low-frequency trends, and residual signals. To reduce the sensitivity of a single view to noise, changes in wearing status, and local artifacts, the input window is first normalized at the segment level, and then a physiological multi-view is constructed: ; in, This represents the physiological multiview of the i-th PPG window; This indicates splicing along the channel dimension or feature dimension; This represents the standardized PPG window; This represents the first-order discrete difference operator. Used to characterize the rising and falling edge variations of the PPG waveform; This represents a second-order discrete difference operator. Used to characterize waveform acceleration changes, curvature changes, and local peak-to-valley transitions; This represents the moving average smoothing operator, used to extract low-frequency trends; This represents the residual signal after detrending, used to preserve high-frequency disturbances, local artifacts, and fine-grained morphological changes; , , , and The five views correspond to the original PPG signal, the first-order PPG signal, the second-order PPG signal, the smoothing trend, and the residual signal, respectively.

[0034] Step 2, Pulse state encoding and current temporal state extraction: Input the physiological multi-view into the multi-scale pulse state encoder to extract local peaks and valleys, complete pulse cycles and long-term dependencies, and obtain the current temporal state.

[0035] In this embodiment of the invention, step 2 includes: The physiological multi-view is input into a multi-scale pulse state encoder composed of small kernel convolution, large kernel convolution and dilated convolution to extract local peaks and valleys, complete pulse cycles and long-term time dependencies, and the current time state is obtained through average pooling, max pooling and linear projection. The multi-scale pulse state encoder consists of small-kernel convolution, large-kernel convolution, dilated convolution, normalization layers, and nonlinear activation functions, used to capture local peak-valley morphology, complete pulse cycle structure, and long-term dependence, respectively; the output of the multi-scale pulse state encoder is denoted as... The current time-series state is obtained by average pooling, max pooling, and linear projection: ; ;in, Indicates the first Physiological multi-view in a PPG window; This represents a multi-scale pulse state encoder. Indicates model parameters; This represents the temporal feature map or hidden feature sequence output by the encoder, i.e., the encoded features; and These represent average pooling and max pooling operations, respectively, used to extract global statistical information and significant peak responses from time-series features; Indicates the projection layer. Represents vector concatenation. Current time-series state. It provides the most direct evidence for identity determination, but it is easily affected by short-term artifacts during motion or in low-quality windows. Therefore, this invention does not use it as the sole identity representation, but rather as one of the five types of knowledge states to participate in subsequent routing.

[0036] Step 3: Construction of five types of pulse knowledge states: Based on the current temporal state, construct pulse morphology state, identity memory state, physiological prior state, and potential disturbance state to form a set of five types of pulse knowledge states.

[0037] In this embodiment of the invention, step 3 includes: For the i-th sample, construct five types of pulse knowledge states. : ; in, This represents the current time series state, used to characterize the local peaks and valleys, periodic structure, and long-range time series dependencies within the current window; This refers to the pulse morphology, used to summarize relatively stable waveform structures. This serves as an identity memory state, used to maintain identity prototypes across windows, activities, and quality conditions. Physiological prior states are used to explicitly encode quality, rhythm, residuals, slope, and activity state; The potential perturbation states are used to represent the identity representation shift caused by motion artifacts and state drift, as shown in Table 1.

[0038] Table 1 Knowledge Status .

[0039] Step 4, Reliability Mapping: Construct a reliability vector and obtain the sample reliability score through the KAN reliability mapper.

[0040] In this embodiment of the invention, step 4 includes: Construct a reliability vector based on quality agent score, morphological dictionary matching confidence, morphological stability, identity memory consistency, potential perturbation energy, activity state, residual energy, and slope statistics; In a forward computation process, firstly, based on the existing morphology dictionary, identity memory, and latent perturbation encoder, reliability evidence of morphology dictionary matching confidence, morphological stability, identity memory consistency, and latent perturbation energy is extracted from the current PPG window; subsequently, the KAN reliability mapper obtains the... Reliability score of each sample The reliability score is used to control the weight of the current morphological dictionary loss, the magnitude of the identity memory EMA update, and the weight of the knowledge state routing, rather than relying on the updated dictionary or memory results of the current round.

[0041] The pulse morphology dictionary is a learnable dictionary used to store stable pulse morphology prototypes, capable of summarizing morphological knowledge such as rising edge, falling edge, peak-valley relationship, and local periodic patterns; the identity memory is a memory bank used to store identity prototypes across windows, activities, and quality conditions, including global identity memory, activity condition memory, and quality condition memory; the knowledge state routing mechanism adaptively selects and integrates multiple types of pulse knowledge states based on sample reliability, quality, activity, residual, and perturbation information.

[0042] This invention constructs a reliability vector from eight categories of explainable factors: observation quality, morphological matching, morphological stability, identity memory consistency, potential perturbation, activity state, residual signal, and slope variation. Specifically, the reliability vector includes the quality surrogate score, morphological dictionary matching confidence, morphological stability, identity memory consistency, potential perturbation energy, activity state, residual energy, and slope statistics.

[0043] Construct a reliability vector consisting of low-dimensional interpretable factors: ; in, Indicates the quality agent score; This indicates the confidence level of the morphological dictionary match. Indicates morphological stability; This indicates consistency with identity memory; Indicates potential disturbance energy; Indicates the activity status; Represents residual energy; Represents slope statistics; the reliability vector, after normalization, is input into the lightweight KAN reliability mapper to obtain the first... Reliability score of each sample : ; in, ∈[0.05,1], The larger the value, the more reliable the sample and the more suitable it is for participating in morphological dictionary learning and strong identity memory updates; The smaller the value, the more likely the sample is affected by motion artifacts, low-quality acquisition, or state drift, and its contribution to knowledge base writing and loss should be weakened. To prevent the reliability score from degenerating into meaningless small values, a soft reliability surrogate objective is constructed and constrained by consistency loss: ; ; in, Transparent monitoring signals are jointly provided by factors such as quality, morphology, memory, and perturbation. This indicates that the gradient operation is stopped.

[0044] Step 5, Reliability-weighted morphological dictionary learning mechanism: The encoded features are divided into local morphological segments, matched and quantized with the morphological dictionary, and reconstructed using reliability-weighted dictionary loss.

[0045] In this embodiment of the invention, step 5 includes: Pulse morphology fragments first match the morphology dictionary Matching and retrieval are performed, where K represents the number of dictionary prototypes and d represents the feature dimension after projection of the local morphological fragment, to obtain the corresponding dictionary prototypes. These prototypes are then quantized and reconstructed to form reconstructed fragments, and the sample-level dictionary loss is calculated. ; Given a temporal feature map or hidden feature sequence output by the encoder. The fragments are divided into local morphological segments and matched and quantized to reconstruct them using a learnable morphological dictionary. ; ; ; in, Indicates from The j-th local morphological fragment extracted from it; ( ) represents the fragment projection function, which maps local morphological fragments to the same feature dimension as the morphological dictionary; Let j be the projection representation of the j-th local morphological segment. It is the k-th prototype in the learnable morphology dictionary; Indicates the cosine similarity between the two; Represents the dictionary prototype index that best matches the current local morphological fragment; This represents the quantized local morphological fragment representation; This indicates that the gradient operation has been stopped. The quantized local morphological fragments are aggregated and residual compensated to obtain the pulse morphology state. ; A reliability-weighted dictionary loss is employed to ensure that high-reliability samples dominate the learning of stable morphological prototypes and suppress the dictionary update contribution of low-reliability samples. ;in, For the vector quantization and commitment loss of the i-th sample, This is a separation constraint between dictionary prototypes. These are the weighting coefficients.

[0046] Step 6, Reliability-Weighted Identity Memory Update Mechanism: Maintain global identity memory, activity condition memory, and quality condition memory for each identity category, and adjust the update magnitude of the exponential moving average by adjusting the sample reliability score; and mask uninitialized prototypes during the retrieval phase to avoid invalid memories in the early training stage affecting identity state construction.

[0047] In this embodiment of the invention, step 6 includes: Current sample information includes identity tags. Basic identity embedding Activity status and quality agent score For identity category y, maintain global identity memory. It also memorizes information based on activity status and quality bin maintenance conditions; during the training phase, the reliability-weighted average of samples with the same identity is first calculated: ; in, Indicates identity category Reliability-weighted identity prototype; Indicates the first The reliability score of each sample; To prevent stable terms with a denominator of zero; ( The ) indicates a normalization operation; subsequently, the identity memory is updated using the reliability-adjusted exponential moving average (EMA). ; in, For identity categories Global identity memory; Based on the update rate, The average reliability of the identity samples; In embodiments of the present invention, such as Figure 2 As shown in (c), M y ← (1-ηr i M y + ηr i e i It is a sample-level update form, representing r i When the value is large, perform a strong update, r i Weak updates are performed every hour. Activity-conditional memory and quality-conditional memory use the same reliability-weighted update strategy. When a prototype is not yet initialized, the model initializes it using the current reliability-weighted mean; uninitialized prototypes are masked during the retrieval phase to avoid invalid memories in the early training stages affecting identity state construction.

[0048] During forward propagation, similar prototypes are retrieved from the initialized identity memory based on the current embedding, and the identity memory state is obtained by weighting the results. The memory retrieval weights and the identity memory state of the current sample are expressed as follows: ; ; in, This indicates the embedding of the basic identity of the current sample. Highest similarity A set of initialized identity memory prototypes; Indicates identity category The global identity memory prototype; Indicates the current sample pair of the th The retrieval weight of an individual identity memory prototype; ( This is used to normalize similarity into weights; This represents the identity memory state obtained by weighted fusion of similar identity memory prototypes.

[0049] This identity memory state provides cross-window identity references for low-quality or moving windows. Because the memory update strength is determined by r... i Controlled, reliable samples can more strongly reinforce identity prototypes, while low-reliability samples only produce weaker updates, thereby improving the stability of the identity memory knowledge base in an open state.

[0050] Step 7, Reliability-Aware Knowledge-State Routing Mechanism: Based on the Token matrix, and combined with quality, reliability, activity, residual and disturbance information, routing weights are generated, and a fused identity representation is output.

[0051] In this embodiment of the invention, step 7 organizes the current temporal state, pulse morphology state, identity memory state, physiological prior state, and potential perturbation state into a Token matrix; Physiological prior states are used to explicitly preserve low-dimensional, interpretable quality and physiological statistics. Unlike deep hidden features, these physiological prior states directly describe the amplitude, slope, residuals, quality, activity, and rhythm changes of the current PPG window. ; ; in, Indicates the first Low-dimensional physiological prior feature vectors of PPG windows; This indicates the statistics of window amplitude; Indicates the window mean or baseline level; Indicates the intensity of the first-order difference; Represents residual energy; Indicates the quality agent score; Indicates the normalized activity state; This represents the zero-crossing ratio of the differential signal; the constant 1 is used to provide the bias term. ( ) represents the physiological prior projection network; This represents the physiological prior state obtained by mapping low-dimensional physiological statistical information; In this embodiment of the invention, the potential perturbation state in step 7 is used to describe motion artifacts, wearing variations, low-quality acquisition, and identity representation shifts caused by noise; the potential perturbation direction is constructed from various perturbation evidence, including the direction of random noise, the direction of perturbation samples obtained by motion enhancement or noise enhancement, the embedding difference between high-quality and low-quality windows of the same identity, the embedding difference under different activity states, and the perturbation direction inferred from acceleration, activity state, residual energy, or quality proxy score; Constructing from basic identity embeddings in the identity embedding space to the perturbation target vector The interpolated state is used to train the conditional perturbation encoder to predict the current potential perturbation direction. : ; ; in, The perturbation target vector can be a random noise vector in the preferred embodiment, or it can be constructed from motion-enhanced samples, noise-enhanced samples, low-quality windows, high-low quality embedding differences, or activity / acceleration information. These are the interpolation coefficients. For a conditional perturbation encoder, the output energy of the conditional perturbation encoder near s=0 is used as a perturbation intensity estimate and as input for reliability mapping and knowledge state routing.

[0052] The Pulse Knowledge State Routing and Adaptive Memory (PKS-KRAM) framework is a reliability-driven knowledge enhancement model for PPG identity recognition proposed in this invention. Pulse Knowledge State refers to the multi-source knowledge representation explicitly organized from the PPG window, including the current temporal state, pulse morphology state, identity memory state, physiological prior state, and potential perturbation state. Kolmogorov-Arnold Networks (KANs) are nonlinear mapping networks with learnable one-dimensional functions / spline functions set on the network edges, suitable for handling complex nonlinear relationships between low-dimensional interpretable factors.

[0053] In this embodiment of the invention, step 7 further includes: The pulse knowledge state router receives temporal state, morphological state, identity memory state, physiological prior state, and potential disturbance state. The pulse knowledge state router includes a KAN gated network, a core token, and a state routing weight α. Its goal is to adaptively allocate the fusion ratio of each knowledge source based on sample reliability and state information. Five categories of pulse knowledge states In the reliability-aware knowledge state routing phase, the tokens are organized into a fixed order based on the current temporal state, pulse morphology state, identity memory state, physiological prior state, and potential disturbance state. : The five types of knowledge states are ultimately determined by the reliability-aware router based on the token matrix. Perform fusion to obtain a fused identity representation and input it into the identity discrimination header: ; ; in, Indicates the first A set of five types of pulse knowledge states for each sample; Indicates by A token matrix organized in a fixed order; This represents the routing weight of the i-th sample to the k-th knowledge state; This represents the representation of the k-th knowledge state after interaction with the core token. ( ) indicates the normalization operation; This represents a fusion of identity representations; Indicates identity identification head; This represents the identity classification output logits or identity score vector corresponding to the i-th PPG window; Calculate the probability of each identity category based on the identity score vector, and select the category with the highest probability as the predicted identity result: ; in, Represents the predicted identity label for the i-th PPG window; k represents the identity category index. This represents the predicted probability of the k-th identity category; Construct routing metadata that includes quality, reliability, activity, and disturbance information: ; in, This represents routing metadata, used to provide routers with auxiliary information on quality, reliability, activity, residuals, identity memory consistency, and disturbance strength. During the training phase, before routing, random deactivation of knowledge states is introduced for non-basic knowledge sources to prevent the model from over-relying on a single fixed knowledge source. Subsequently, the router first extracts the core knowledge token, and then combines linear gating and KAN gating to obtain the fusion weights of the five types of states: ; in, Represents the linear gating parameters; ( ) represents the KAN gating function used for generating route weights; This represents the fusion coefficient of the KAN-gated branch.

[0054] Step 8, Identity Discrimination and Joint Training: Input the fused identity representation into the identity discrimination head, output the identity recognition result, and constrain the multi-source knowledge state learning through joint optimization.

[0055] In this embodiment of the invention, the consistency of paired window changes in step 8 is used to enhance the temporal stability of continuous PPG identity representation; for the i-th PPG window and its preceding window within the same data partition The models respectively obtain fused identity representations , and identity classification output , First, construct the change features between the i-th PPG window and the preceding window, and then use the change prediction head. Predict changes in logits to ensure that changes in the representation align with predicted changes: ; ; in, Represents the i-th PPG window The corresponding preceding window index; and Let i and n represent the fused identity representations of the i-th PPG window and the preceding window, respectively. and These represent the identity classification output logits or identity score vectors of the i-th sample window and the preceding window, respectively. This represents the change characteristics consisting of the current window, the previous window, and the differences in their representations; This indicates vector concatenation; Represents the absolute difference element by element; This indicates a change prediction head, used to predict changes in identity scores from changes in characterization; Indicates the target of changes in the true identity score; This indicates that the gradient operation has been stopped. Indicates smoothness The loss is used to reduce the impact of outlier windows on the consistency constraint, which requires that changes in the representation of adjacent windows can explain changes in identity prediction, thereby enhancing the temporal stability of identity representation under consecutive PPG windows. Since preceding windows are only selected within the same data partition, this constraint does not introduce information leakage between training, validation, and testing.

[0056] Paired-window Change Consistency: This constraint ensures consistency between the representational and predicted changes of the current PPG window and its preceding windows within the same partition, enhancing the temporal stability of continuous identity representations. ArcFace: An identity classification loss function based on angular margins, used to increase inter-class margins and improve identity discrimination in the hyperspherical feature space. Identity Discrimination Head: A discrimination module that outputs identity category, identity probability, or identity verification score based on the fused identity representation. In a preferred embodiment, the identity discrimination head is implemented using the ArcFace boundary enhancement classification head.

[0057] Identity discrimination employs cross-entropy and ArcFace boundary enhancement objectives to improve inter-class separation capability. Simultaneously, it utilizes reliability-weighted morphological dictionary loss, identity memory consistency loss, latent perturbation modeling loss, reliability estimation loss, and pairwise window consistency loss to jointly constrain multi-source knowledge state learning. The final training objective is: ; in, The identity discrimination loss includes cross-entropy and ArcFace boundary enhancement targets; and λ represents the morphological dictionary loss, identity memory consistency loss, potential perturbation modeling loss, reliability estimation loss, and pairwise window change consistency loss, respectively; each λ term is the corresponding loss weight; through joint optimization, identity discrimination ability, stable morphological knowledge, reliable identity memory, perturbation perception ability, and cross-window time stability are learned.

[0058] Identity identification and output: After reliability-aware knowledge state routing, the fused identity representation z is obtained. i This representation is input into the ArcFace identity discrimination head, outputs the classification probability of each identity category, and yields the final identity recognition result. ArcFace improves inter-class separability and open-set identity verification stability by introducing an angular margin m at the angle between the feature vector and the class weight vector, making samples of the same identity more compact on the hypersphere and increasing the angular distance between different identity samples.

[0059] During deployment, the model can perform forward inference using only the current PPG window, activity state, and quality proxy score. Identity memory can be frozen after training for prototype retrieval, or updated online with reliability threshold control in continuous authentication scenarios with user authorization and secure update mechanisms to avoid low-reliability samples contaminating user identity memory.

[0060] Alternative solution to the invention: 1. Alternatives to reliability factors and reliability mapping methods; In preferred embodiments of this invention, a reliability vector is constructed using quality proxy scores, morphological dictionary matching confidence, morphological stability, identity memory consistency, potential disturbance energy, activity state, residual energy, and slope statistics. In other embodiments, factors such as heart rate variability, peak-to-valley interval stability, contact pressure, acceleration information, or device wearing status may also be introduced; the reliability mapper may also be implemented using spline function networks, radial basis function networks, lightweight MLPs, gated expert networks, or other nonlinear mapping modules.

[0061] 2. Alternative methods for constructing pulse knowledge states; In preferred embodiments of this invention, five types of pulse knowledge states are constructed: current temporal state, pulse morphology state, identity memory state, physiological prior state, and potential disturbance state. In other embodiments, the above knowledge states can be merged, split, or expanded, for example, by adding device state, sensor contact state, heart rate dynamic state, or cross-day stable state.

[0062] 3. Alternatives to morphological dictionaries, identity memory, and state routing methods; morphological knowledge can be implemented through prototype networks, clustering dictionaries, learnable codebooks, memory queues, or comparative morphological libraries; identity memory can be implemented through queued updates, gating updates, momentum encoder updates, Bayesian prototype updates, or secure online updates; knowledge state routing can be implemented using attention routing, hybrid expert routing, Transformer interactive routing, graph attention routing, or selection routing based on confidence thresholds.

[0063] 4. Alternatives to training objectives and application scenarios; In addition to the loss described in this paper, training objectives can also incorporate contrastive learning loss, center loss, triplet loss, mutual information constraint, or domain adaptation loss. This invention is not only applicable to wristbands, smartwatches, and finger-clip PPG identity recognition, but can also be extended to continuous identity authentication, wearable payment, remote health and safety login, in-vehicle driver status authentication, and other biometric recognition scenarios that use PPG or similar pulse waves as input.

[0064] The core of this invention is not a simple network structure replacement, but rather a remodeling of PPG identity recognition as a reliability-driven knowledge enhancement process: a reliability mapper outputs a sample reliability score, which is then used as a unified control variable throughout pulse morphology dictionary learning, identity memory update, and routing of five types of pulse knowledge states; in a preferred embodiment, the reliability mapper is implemented using a KAN nonlinear mapper.

[0065] Compared with the prior art, the present invention has the following advantages: ① Reduce confusion between identity information and state perturbations. This invention extends the single PPG identity embedding to the current temporal state, pulse morphology state, identity memory state, physiological prior state, and potential perturbation state, enabling the model to distinguish stable identity knowledge, collect state changes and motion perturbation information, thereby improving the stability of identity discrimination in complex wearable scenarios.

[0066] ② Reduce erroneous updates of the morphological knowledge base and identity memory base by low-quality samples. The sample reliability score is used as a unified control variable to participate in the weighting of the morphological dictionary loss and the adjustment of the identity memory update amplitude. This allows high-reliability samples to dominate the formation of stable morphological prototypes and identity prototypes, while weakening the contribution of low-reliability samples to the writing of the knowledge base. This reduces the impact of low-quality fragments, motion artifacts, and state drift on the knowledge base.

[0067] ③ Improve the temporal stability of continuous identity authentication results. By using pairwise window change consistency constraints, the representation changes of adjacent PPG windows are kept consistent with the identity prediction changes, reducing abnormal jumps in identity output during continuous authentication.

[0068] ④ Improve the adaptability of knowledge source selection. The reliability-aware knowledge state router can dynamically adjust its dependence on temporal sequence, morphology, memory, prior and perturbation states based on sample quality, activity state, residual energy and perturbation intensity, thereby selecting more reliable identity discrimination criteria under different quality and activity conditions.

[0069] The technical solution provided by this invention includes a method comprising: unified sample construction and physiological multi-view construction: processing the original photoplethysmography (PPG) signal to obtain a unified KRAM format sample; and constructing a physiological multi-view; pulse state encoding and current temporal state extraction: inputting the physiological multi-view into a multi-scale pulse state encoder to extract local peaks and valleys, the complete pulse cycle, and long-term dependence, and obtaining the current temporal state; construction of five types of pulse knowledge states: based on the current temporal state, constructing pulse morphology state, identity memory state, physiological prior state, and potential perturbation state, forming a set of five types of pulse knowledge states; reliability mapping: constructing a reliability vector, and obtaining the sample reliability score through a KAN reliability mapper; and a reliability-weighted morphological dictionary learning mechanism: encoding features... The feature is divided into local morphological segments, matched and quantized with the morphological dictionary for reconstruction, and a reliability-weighted dictionary loss is adopted. The reliability-weighted identity memory update mechanism adjusts the update magnitude of the exponential moving average by the sample reliability score and masks uninitialized prototypes during the retrieval stage. The reliability-aware knowledge state routing mechanism organizes the token matrix and generates routing weights by combining quality, reliability, activity, residual and perturbation information, and outputs the fused identity representation. Identity discrimination and joint training: the fused identity representation is input into the identity discrimination head, the identity recognition result is output, and the multi-source knowledge state learning is constrained by joint optimization. This method improves the temporal stability of continuous identity authentication results and reduces the erroneous updates of the morphological knowledge base and identity memory base by low-quality samples.

[0070] 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 reliable knowledge routing and adaptive memory method for PPG identity recognition, characterized in that, The method includes: Step 1: Unified Sample Construction and Physiological Multiview Construction: The original photoplethysmography (PPG) signal is processed to obtain a unified KRAM format sample; and a physiological multiview is constructed. Step 2, Pulse state encoding and current temporal state extraction: Input the physiological multi-view into the multi-scale pulse state encoder to extract local peaks and valleys, complete pulse cycles and long-term dependencies, and obtain the current temporal state; Step 3: Construction of five types of pulse knowledge states: Based on the current temporal state, construct the pulse morphology state, identity memory state, physiological prior state, and potential perturbation state to form a set of five types of pulse knowledge states; Step 4, Reliability Mapping: Construct a reliability vector and obtain the sample reliability score through the KAN reliability mapper; Step 5, Reliability-weighted morphological dictionary learning mechanism: The encoded features are divided into local morphological segments, matched and quantized with the morphological dictionary for reconstruction, and a reliability-weighted dictionary loss is adopted. Step 6, Reliability-Weighted Identity Memory Update Mechanism: Adjust the update magnitude of the exponential moving average by the sample reliability score; and mask uninitialized prototypes during the retrieval phase; Step 7, Reliability-Aware Knowledge-State Routing Mechanism: Based on the Token matrix, and combined with quality, reliability, activity, residual and disturbance information, routing weights are generated, and a fused identity representation is output. Step 8, Identity Discrimination and Joint Training: Input the fused identity representation into the identity discrimination head, output the identity recognition result, and constrain the multi-source knowledge state learning through joint optimization; Step 4 includes: Construct a reliability vector based on quality agent score, morphological dictionary matching confidence, morphological stability, identity memory consistency, potential perturbation energy, activity state, residual energy, and slope statistics; In a forward computation process, firstly, based on the existing pulse morphology dictionary, identity memory, and latent perturbation encoder, reliability evidence of morphology dictionary matching confidence, morphological stability, identity memory consistency, and latent perturbation energy is extracted from the current PPG window; subsequently, the KAN reliability mapper obtains the... Reliability score of each sample The reliability score is used to control the loss weight of the current morphology dictionary, the update magnitude of the identity memory EMA, and the routing weight of the knowledge state. Construct a reliability vector consisting of low-dimensional interpretable factors: ； in, Indicates the quality agent score; This indicates the confidence level of the morphological dictionary match. Indicates morphological stability; This indicates consistency with identity memory; Indicates potential disturbance energy; Indicates the activity status; Represents residual energy; Represents slope statistics; the reliability vector, after normalization, is input into the lightweight KAN reliability mapper to obtain the first... Reliability score of each sample : ； in, ∈[0.05,1], The larger the value, the more reliable the sample and the more suitable it is for participating in morphological dictionary learning and strong identity memory updates; The smaller the value, the more likely the sample is affected by motion artifacts, low-quality acquisition, or state drift, and its contribution to knowledge base writing and loss should be weakened. Construct a soft reliability proxy objective and constrain it with consistency loss: ； ； in, Transparent monitoring signals are jointly provided by factors such as quality, morphology, memory, and perturbation. This indicates that the gradient operation is stopped.

2. The method according to claim 1, characterized in that, The unified sample construction in step 1 includes: The input for PPG identification is a time window segmented from continuous physiological recordings. The raw photoplethysmography (PPG) signal undergoes missing value interpolation, bandpass filtering, outlier processing, sliding window segmentation, and length resampling to obtain a unified KRAM format sample set containing the PPG window, identification tag, activity status, quality surrogate score, time sequence number, and preceding window index. ； in, Let C be the number of channels and T be the number of uniform sampling points for the i-th PPG window. For identity tags; In active state; For quality agency scores; Use the window's start time or sequence number; The index of the preceding window that is located within the same data partition as the current window; In step 1, a physiological multiview is constructed from the original PPG signal, the first-order PPG signal, the second-order PPG signal, the smoothing trend, and the residual signal. This multiview includes: First, the input window is standardized at the fragment level, and then physiological multi-view is constructed: ； in, Indicates the first Physiological multi-view of a PPG window; This indicates splicing along the channel dimension or feature dimension; This represents the standardized PPG window; This represents the first-order discrete difference operator. Used to characterize the rising and falling edge variations of the PPG waveform; This represents a second-order discrete difference operator. Used to characterize waveform acceleration changes, curvature changes, and local peak-to-valley transitions; This represents the moving average smoothing operator, used to extract low-frequency trends; This represents the residual signal after detrending, used to preserve high-frequency disturbances, local artifacts, and fine-grained morphological changes; , , , and The five views correspond to the original PPG signal, the first-order PPG signal, the second-order PPG signal, the smoothing trend, and the residual signal, respectively.

3. The method according to claim 2, characterized in that, Step 2 includes: The physiological multi-view is input into a multi-scale pulse state encoder composed of small kernel convolution, large kernel convolution and dilated convolution to extract local peaks and valleys, complete pulse cycles and long-term time dependencies, and the current time state is obtained through average pooling, max pooling and linear projection. The multi-scale pulse state encoder consists of small-kernel convolution, large-kernel convolution, dilated convolution, normalization layers, and nonlinear activation functions, used to capture local peak-valley morphology, complete pulse cycle structure, and long-term dependence, respectively; the output of the multi-scale pulse state encoder is denoted as... The current time-series state is obtained by average pooling, max pooling, and linear projection: ； ； in, Indicates the first Physiological multi-view of a PPG window; This represents a multi-scale pulse state encoder. Indicates model parameters; This represents the temporal feature map or hidden feature sequence output by the encoder, i.e., the encoded features; and These represent average pooling and max pooling operations, respectively, used to extract global statistical information and significant peak responses from time-series features; Indicates the projection layer. This indicates vector concatenation.

4. The method according to claim 3, characterized in that, Step 3 includes: For the i-th sample, construct five types of pulse knowledge states. : ； in, This represents the current time series state, used to characterize the local peaks and valleys, periodic structure, and long-range time series dependencies within the current window; This refers to the pulse morphology, used to summarize relatively stable waveform structures. This serves as an identity memory state, used to maintain identity prototypes across windows, activities, and quality conditions. Physiological prior states are used to explicitly encode quality, rhythm, residuals, slope, and activity state; The potential perturbation state is used to represent the identity representation shift caused by motion artifacts and state drift.

5. The method according to claim 1, characterized in that, Step 5 includes: Pulse morphology fragments first match the morphology dictionary Matching and retrieval are performed, where K represents the number of dictionary prototypes and d represents the feature dimension after projection of the local morphological fragment, to obtain the corresponding dictionary prototypes. These prototypes are then quantized and reconstructed to form reconstructed fragments, and the sample-level dictionary loss is calculated. ; Given a temporal feature map or hidden feature sequence output by the encoder. The fragments are divided into local morphological segments and matched and quantized with a learnable morphological dictionary for reconstruction. ； ； ； in, Indicates from The j-th local morphological fragment extracted from it; ( ) represents the fragment projection function, which maps local morphological fragments to the same feature dimension as the morphological dictionary; Let j be the projection representation of the j-th local morphological segment. It is the k-th prototype in the learnable morphology dictionary; Indicates the cosine similarity between the two; Represents the dictionary prototype index that best matches the current local morphological fragment; This represents the quantized local morphological fragment representation; This indicates that the gradient operation has been stopped. The quantized local morphological fragments are aggregated and residual compensated to obtain the pulse morphology state. ; A reliability-weighted dictionary loss is employed to ensure that high-reliability samples dominate the learning of stable morphological prototypes and suppress the dictionary update contribution of low-reliability samples. ； in, For the vector quantization and commitment loss of the i-th sample, This is a separation constraint between dictionary prototypes. These are the weighting coefficients.

6. The method according to claim 5, characterized in that, Step 6 includes: Current sample information includes identity tags. Basic identity embedding Activity status and quality agent score For identity category y, maintain global identity memory. It also memorizes information based on activity status and quality bin maintenance conditions; during the training phase, the reliability-weighted average of samples with the same identity is first calculated: ； in, Indicates identity category Reliability-weighted identity prototype; Indicates the first The reliability score of each sample; To prevent stable terms with a denominator of zero; ( The ) indicates a normalization operation; subsequently, the identity memory is updated using the reliability-adjusted exponential moving average (EMA). ； in, For identity categories Global identity memory; Based on the update rate, The average reliability of the identity samples; During forward propagation, similar prototypes are retrieved from the initialized identity memory based on the current embedding, and the identity memory state is obtained by weighting the results. The memory retrieval weights and the identity memory state of the current sample are expressed as follows: ； ； in, This indicates the embedding of the basic identity of the current sample. Highest similarity A set of initialized identity memory prototypes; Indicates identity category The global identity memory prototype; Indicates the current sample pair of the th The retrieval weight of an individual identity memory prototype; ( This is used to normalize similarity into weights; This represents the identity memory state obtained by weighted fusion of similar identity memory prototypes.

7. The method according to claim 6, characterized in that, In step 7, the current temporal state, pulse morphology state, identity memory state, physiological prior state, and potential disturbance state are organized into a Token matrix. Physiological prior states are used to explicitly preserve low-dimensional, interpretable quality and physiological statistics, directly describing the amplitude, slope, residuals, quality, activity, and rhythm changes of the current PPG window: ； ； in, Indicates the first Low-dimensional physiological prior feature vectors of PPG windows; This indicates the statistics of window amplitude; Indicates the window mean or baseline level; Indicates the intensity of the first-order difference; Represents residual energy; Indicates the quality agent score; Indicates the normalized activity state; This represents the zero-crossing ratio of the differential signal; the constant 1 is used to provide the bias term. ( ) represents the physiological prior projection network; This represents the physiological prior state obtained by mapping low-dimensional physiological statistical information; In step 7, the potential perturbation state is used to describe motion artifacts, wearing variations, low-quality acquisition, and identity representation shifts caused by noise. The potential perturbation direction is constructed from various perturbation evidence, including the direction of random noise, the direction of perturbation samples obtained by motion enhancement or noise enhancement, the embedding difference between high-quality and low-quality windows of the same identity, the embedding difference under different activity states, and the perturbation direction inferred from acceleration, activity state, residual energy, or quality proxy score. Constructing from basic identity embeddings in the identity embedding space to the perturbation target vector The interpolated state is used to train the conditional perturbation encoder to predict the current potential perturbation direction. : ； ； in, The target vector for perturbation. These are the interpolation coefficients. For a conditional perturbation encoder, the output energy of the conditional perturbation encoder near s=0 is used as a perturbation intensity estimate and as input for reliability mapping and knowledge state routing.

8. The method according to claim 7, characterized in that, Step 7 also includes: The pulse knowledge state router receives temporal state, morphological state, identity memory state, physiological prior state, and potential disturbance state. The pulse knowledge state router includes a KAN gated network, a core token, and a state routing weight α. Its goal is to adaptively allocate the fusion ratio of each knowledge source based on sample reliability and state information. Five categories of pulse knowledge states In the reliability-aware knowledge state routing phase, the tokens are organized into a fixed order based on the current temporal state, pulse morphology state, identity memory state, physiological prior state, and potential disturbance state. : The five types of knowledge states are ultimately determined by the reliability-aware router based on the token matrix. Perform fusion to obtain a fused identity representation and input it into the identity discrimination header: ； ； in, Indicates the first A set of five types of pulse knowledge states for each sample; Indicates by A token matrix organized in a fixed order; This represents the routing weight of the i-th sample to the k-th knowledge state; This represents the representation of the k-th knowledge state after interaction with the core token. ( ) indicates the normalization operation; This represents a fusion of identity representations; Indicates identity identification head; This represents the identity classification output logits or identity score vector corresponding to the i-th PPG window; Calculate the probability of each identity category based on the identity score vector, and select the category with the highest probability as the predicted identity result: ； in, Represents the predicted identity label for the i-th PPG window; k represents the identity category index. This represents the predicted probability of the k-th identity category; Construct routing metadata that includes quality, reliability, activity, and disturbance information: ； in, This represents routing metadata, used to provide routers with auxiliary information on quality, reliability, activity, residuals, identity memory consistency, and disturbance strength. During the training phase, non-basic knowledge sources are randomly deactivated before routing. Then, the router first extracts the core knowledge token, and then combines linear gating and KAN gating to obtain the fusion weights of the five types of states: ； in, Represents the linear gating parameters; ( ) represents the KAN gating function used for generating route weights; This represents the fusion coefficient of the KAN-gated branch.

9. The method according to claim 8, characterized in that, In step 8, the consistency of paired window changes is used to enhance the temporal stability of continuous PPG identity representation; for the i-th PPG window and its preceding window within the same data partition The models respectively obtain fused identity representations , and identity classification output , First, construct the change features between the i-th PPG window and the preceding window, and then use the change prediction head. Predict changes in logits to ensure that changes in the representation align with predicted changes: ； ； in, Represents the i-th PPG window The corresponding preceding window index; and Let i and j represent the fused identity representations of the i-th sample window and the preceding window, respectively. and These represent the identity classification output logits or identity score vectors of the i-th sample window and the preceding window, respectively. This represents the change characteristics consisting of the current window, the previous window, and the differences in their representations; This indicates vector concatenation; Represents the absolute difference element by element; This indicates a change prediction head, used to predict changes in identity scores from changes in characterization; Indicates the target of changes in the true identity score; This indicates that the gradient operation has been stopped. Indicates smoothness Loss is used to reduce the impact of abnormal windows on consistency constraints; Identity discrimination employs cross-entropy and ArcFace boundary enhancement objectives to improve inter-class separation capability. Simultaneously, it utilizes reliability-weighted morphological dictionary loss, identity memory consistency loss, latent perturbation modeling loss, reliability estimation loss, and pairwise window consistency loss to jointly constrain multi-source knowledge state learning. The final training objective is: ； in, The identity discrimination loss includes cross-entropy and ArcFace boundary enhancement targets; and λ represents the morphological dictionary loss, identity memory consistency loss, potential perturbation modeling loss, reliability estimation loss, and pairwise window change consistency loss, respectively; each λ term is the corresponding loss weight; through joint optimization, identity discrimination ability, stable morphological knowledge, reliable identity memory, perturbation perception ability, and cross-window time stability are learned.

Images