CN-122004803-A - PPG identification method based on biomechanical derivative and KAN-xLSTM

CN122004803ACN 122004803 ACN122004803 ACN 122004803ACN-122004803-A

Abstract

The invention relates to the technical field of PPG biological feature recognition, and particularly provides a PPG recognition method based on biomechanical derivative and KAN-xLSTM. The method comprises the steps of constructing three-channel input tensors according to an acquired original PPG signal, distributing the three-channel input tensors to a parallel time domain flow processing branch and a frequency domain flow gating branch, respectively obtaining a time domain feature vector and a global gating weight based on frequency spectrum, weighting the time domain feature vector element by utilizing the global gating weight to obtain a fusion feature vector, projectively decomposing the fusion feature vector into an identity embedded vector and a noise embedded vector to purify the identity embedded vector, inputting the purified identity embedded vector into a classification head based on an angle margin, calculating classification probability, and outputting a final identity recognition result.

Inventors

WANG CHUNXIAO
XIAO YANCHAO
HUANG YUWEN
ZHENG YUE
LI WENHAO
WANG HAO
FU XI
ZHANG WENZHE
LIU ZIQIANG

Assignees

齐鲁工业大学(山东省科学院)
山东省计算中心（国家超级计算济南中心）

Dates

Publication Date: 20260512
Application Date: 20260414

Claims (9)

1. A PPG identification method based on biomechanical derivatives and KAN-xLSTM, the method comprising: step 1, constructing a three-channel input tensor containing a hemodynamic prior according to an acquired original PPG signal; Step 2, distributing three channel input tensors to a parallel time domain stream processing branch and a frequency domain stream gating branch, and respectively obtaining a time domain feature vector and a global gating weight based on a frequency spectrum; Step 3, weighting the time feature vector element by utilizing the global gating weight to obtain a fusion feature vector; Step 4, projectively decomposing the fusion feature vector into an identity embedded vector and a noise embedded vector so as to purify the identity embedded vector; And 5, inputting the purified identity embedded vector into a classification head based on the angle margin, calculating classification probability, and outputting a final identity recognition result.
2. The method of claim 1, wherein step 1 comprises constructing a biomechanical derivative embedding module and input: according to the obtained original PPG signal, calculating a first derivative as a speed characteristic and a second derivative as an acceleration characteristic, splicing the original PPG signal, the speed characteristic and the acceleration characteristic in channel dimensions, and constructing a three-channel input tensor containing hemodynamic priori; Based on the phase space reconstruction principle, the motion derivative of the PPG signal is explicitly calculated to construct biomechanical characteristics with translational invariance, an input PPG signal sequence is set as X, wherein T is a time step, an enhanced three-channel input tensor is constructed, and the construction process is as follows: Step 11, original PPG signal: The normalized time sequence of the input is recorded as ; Step 12, speed characterization : The first derivative of the original PPG signal is calculated, physically corresponding to the instantaneous velocity of the blood flow in the blood vessel, expressed as: ; Step 13, acceleration characteristics : The second derivative of the original PPG signal is calculated and is physically used for reflecting the expansion and contraction capacity of the vascular wall under pulse wave impact, and the expression is as follows: ; Step 14, obtaining a three-channel input tensor containing hemodynamic prior through splicing operation, wherein the expression is as follows: 。
3. The method according to claim 2, wherein the step 2 includes time-frequency dual-stream cooperative feature extraction: the method comprises the steps of distributing three-channel input tensors to two parallel branches, extracting front ends in a time domain flow processing branch by utilizing the multi-scale characteristics of a Kolmogorov-Arnod network KAN to capture waveform details of different scales, inputting a Bi-directional expansion long-short-period memory network Bi-xLSTM module, carrying out lossless coding on long-sequence characteristics by utilizing a matrix memory of the Bi-directional expansion long-period memory network Bi-xLSTM module, compressing time sequence characteristics into time domain characteristic vectors by global average pooling GAP, carrying out fast Fourier transform FFT on the three-channel input tensors in a frequency domain flow gating branch, and generating global gating weight based on frequency spectrum by a KAN projection layer and an activation function Sigmoid; The time domain flow processing branch is the core of feature extraction and is used for capturing the morphological features of fine granularity from the enhanced three-channel input tensor, and comprises a KAN multi-scale feature extraction front end and a two-way expansion long-short-term memory network module; Step 21, KAN multi-scale feature extraction front end; The KAN convolution layer is adopted to replace the linear convolution layer, the KAN convolution calculation can learn the nonlinear combination of the B-spline basis function B-Splines for the input vector u, and the KAN convolution with the convolution kernel size k has the output y expression as follows: ; Wherein, the ; Is an i-order B-spline basis function, Is a corresponding learnable control coefficient; Setting 4 parallel KAN convolution branches, wherein the convolution kernel sizes are k epsilon {1,3,7,11}, respectively, and are used for capturing multi-scale features from high-frequency noise filtering and sharp wave crest extraction to low-frequency waveform contour fitting, and forming deep feature sequences after output splicing of all scales ; Step 22, a Bi-xLSTM module of a bidirectional expansion long-term memory network; using an extended long and short term memory network xLSTM, which introduces matrix memory And covariance update rules, at time step t, for input features The extended long-short-term memory unit generates a query vector by linear projection Key vector Sum vector And calculates the input gate Forgetting door Exponential gating is adopted to support gradient long-range propagation; the updating rule of the matrix memory is as follows: ; Wherein, the Representing an element-by-element multiplication, Representing the outer product; adopting a bidirectional expansion long-short-term memory network module to respectively process a forward sequence and a reverse sequence, and obtaining a time domain feature vector through global average pooling layer GAP after output splicing 。
4. The method of claim 3, wherein step 2 further comprises a frequency domain flow gating branch that uses global spectrum priors to clean time domain features: Step 23, fast Fourier transform FFT, namely performing FFT transformation on the three-channel input tensor and taking a modulus to obtain a frequency domain amplitude spectrum ; Step 24, KAN projection and gating generation, wherein a KAN projection layer is utilized to map the frequency domain amplitude spectrum to the channel dimension identical to the time domain characteristic; step 25, generating global gating weight, namely generating spectrum gating weight g with the range between (0 and 1) through an activating function Sigmoid, wherein the spectrum gating weight g represents channel importance coefficients with different feature dimensions, and the expression is as follows: 。
5. The method of claim 4, wherein step 3 comprises feature fusion: Multiplying the global gating weight and the time domain feature vector element by element, and splicing the product result and the frequency domain feature vector projected by KAN to obtain a final fusion feature vector, wherein the expression is as follows: 。
6. The method of claim 5, wherein step 4 comprises quadrature decoupling and signal reconstruction: The method comprises the steps of integrating characteristic vector projection and decomposition into an identity embedded vector and a noise embedded vector, forcing the identity embedded vector and the noise embedded vector to be mutually perpendicular in a characteristic space through an orthogonal constraint loss function, adding the identity embedded vector and the noise embedded vector element by element, and inputting the added identity embedded vector and the added noise embedded vector into a signal reconstruction decoder to restore a reconstructed PPG signal, so as to purify the identity embedded vector; Designing a generated orthogonal decoupling mechanism, (1) feature projection and separation, namely fusing feature vectors After the splicing operation, the two low-dimensional embedded vectors are respectively mapped into two low-dimensional embedded vectors, namely identity embedded vectors, through two independent full-connection layer projection heads Which contains only user identity information, noise embedded vectors Non-identity variation for absorbing motion artifact and wearing looseness, (2) orthogonal constraint loss function During training, forced minimization And (3) with The absolute value of cosine similarity between the two is expressed as follows: (3) a signal reconstruction decoder that performs element-wise addition of the separated identity embedded vector and noise embedded vector: Will be Input signal reconstruction decoder to restore the reconstructed PPG signal, introducing reconstruction losses 。
7. The method according to claim 6, wherein said step 5 comprises output layer and joint optimization: Step 51, embedding the purified identity into the vector only by the classification head based on the angle margin Input classification head, employing ArcFace loss function The included angle between the feature vector and the weight vector is added with an angle margin m, so that the distance between the classes is maximized on the hypersphere; step 52, joint loss function, performing end-to-end training by the following multitasking loss function: ; Wherein, the , Is a balance coefficient.
8. A computer readable storage medium, characterized in that the computer readable storage medium comprises a stored program, wherein the program when run controls a device in which the computer readable storage medium is located to perform the PPG identification method based on biomechanical derivatives and KAN-xLSTM as claimed in any one of claims 1 to 7.
9. An electronic device comprising one or more processors, memory, and one or more computer programs, wherein the one or more computer programs are stored in the memory, the one or more computer programs comprising instructions that, when executed by the device, cause the device to perform the biomechanical derivative and KAN-xLSTM-based PPG identification method of any of claims 1-7.

Description

PPG identification method based on biomechanical derivative and KAN-xLSTM Technical Field The invention relates to the technical field of PPG biological feature recognition, in particular to a PPG recognition method based on biomechanical derivative and KAN-xLSTM. Background Biometric identification techniques based on photoplethysmographic pulse waves (Photoplethysmography, PPG) are of great value in wearable devices. However, in practical applications, especially in motion disturbances and complex physiological conditions, the prior art suffers from the following drawbacks: The vector state information bottleneck of the existing lightweight model, the current mainstream time sequence model comprising a traditional cyclic neural network (Recurrent Neural Network, RNN) and the latest state space model (such as Mamba, namely the technology mainly adopted by the preamble patent) has fundamental theoretical limitation in processing long sequences, namely vector state compression (Vector State Compression). They typically compress infinitely long historical information into a fixed-size vector state ) Is a kind of medium. This compression mechanism forces the model to tend to average out historical information, resulting in those transient but highly discriminative high frequency microcosmic details (e.g., precise slope of dicrotic notch, small variation of systolic peak) being smoothed out or lost. In biometric identification, it is these tiny high frequency details that determine the key to distinguishing heart rate similar users. The existing model is difficult to break through in the crowd with high similarity due to insufficient memory capacity. Second, the mismatch of the non-linear manifold and linear activation function of the blood vessel, and the PPG signal is essentially a hydrodynamic waveform generated by the interaction of the cardiac pumping and elastic recoil of the blood vessel. The compliance (Vascular Compliance) of the vessel wall follows a smooth, continuous nonlinear physical law (stress-strain curve). Existing backbone networks (e.g., convolutional neural network CNN, transformer) rely primarily on piecewise linear activation functions (e.g., reLU). The efficiency of approximating a smooth physiological Manifold (Manifold) by a piecewise linear function is extremely low, and extremely deep network layers are often required to be stacked to achieve high fidelity, so that the waste of computing resources is caused. And the collected PPG signal is a nonlinear mixture of identity essence and motion noise under strong motion interference. Most of the existing methods adopt end-to-end black box learning, and lack an explicit mechanism to distinguish physiological characteristics in signals from external interference. The high coupling of such features results in a model that is prone to misjudging noise as an identity feature in the face of unseen noise patterns, lacking robustness in extreme environments. Disclosure of Invention In view of the above, the invention provides a PPG identification method based on biomechanical derivative and KAN-xLSTM, which is used for improving the memory capacity of a model on fine-grained identity characteristics, improving the physical interpretability and accuracy of feature extraction and realizing the explicit stripping of noise. In a first aspect, the invention provides a method of PPG identification based on biomechanical derivatives and KAN-xLSTM, the method comprising: step 1, constructing a three-channel input tensor containing a hemodynamic prior according to an acquired original PPG signal; Step 2, distributing three channel input tensors to a parallel time domain stream processing branch and a frequency domain stream gating branch, and respectively obtaining a time domain feature vector and a global gating weight based on a frequency spectrum; Step 3, weighting the time feature vector element by utilizing the global gating weight to obtain a fusion feature vector; Step 4, projectively decomposing the fusion feature vector into an identity embedded vector and a noise embedded vector so as to purify the identity embedded vector; And 5, inputting the purified identity embedded vector into a classification head based on the angle margin, calculating classification probability, and outputting a final identity recognition result. Optionally, the step 1 includes constructing a biomechanical derivative embedding module and an input: according to the obtained original PPG signal, calculating a first derivative as a speed characteristic and a second derivative as an acceleration characteristic, splicing the original PPG signal, the speed characteristic and the acceleration characteristic in channel dimensions, and constructing a three-channel input tensor containing hemodynamic priori; Based on the phase space reconstruction principle, the motion derivative of the PPG signal is explicitly calculated to construct biomechanical characteristics with translation