Physical information driven dual-flow network ppg signal denoising and reconstruction method and system

Abstract

The application provides a physical information driven double-flow network PPG signal denoising and reconstruction method and system, and belongs to the technical field of computer data processing and digital signal processing. The method extracts deep shared features from an original signal; oscillation components and slow trend components are decoupled through Hilbert analytic signal processing and anisotropic diffusion processing respectively; and finally, the two components are superimposed after being calibrated with a benchmark to obtain a reconstructed PPG time sequence signal. The application can suppress signal noise and extract slow change trends while retaining main waveform features of the PPG, and improve the stability of the reconstruction result and the interpretability of the model processing process.

Description

Technical Field

[0001] This invention belongs to the field of computer data processing and digital signal processing technology, specifically relating to a method and system for denoising and reconstructing PPG signals in a two-stream network driven by physical information. Background Technology

[0002] PPG (photoplethysmography) signal is a one-dimensional physiological time series signal reflecting changes in human blood volume, widely used in the analysis of heart rate, blood oxygen saturation, and related physiological states. During actual acquisition, PPG signals are easily affected by factors such as motion, ambient light, changes in sensor contact, and equipment noise, resulting in localized rapid fluctuations and slow trend drifts, which degrade signal quality and affect subsequent feature extraction and analysis results.

[0003] Existing PPG signal denoising methods mainly include digital filtering, wavelet transform, and deep learning methods. Traditional methods typically rely on fixed parameters, thresholds, or basis functions, making it difficult to simultaneously achieve noise suppression and effective waveform preservation when dealing with non-stationary and morphologically variable PPG signals. While existing deep learning methods have strong nonlinear mapping capabilities, they usually directly output the complete reconstructed signal, lacking a clear distinction between oscillating and slow-trending components, resulting in insufficient stability and interpretability in the processing.

[0004] Furthermore, PPG signals from different individuals vary in oscillation rhythm and rate of trend change. Existing methods typically employ uniform constraint parameters, making it difficult to adaptively adjust based on conditional features such as blood oxygenation and age. Therefore, it is necessary to propose a dual-stream PPG signal denoising and reconstruction method that extracts oscillatory and slow-trend components separately, generates corresponding parameters based on conditional features, and trains the method using reconstruction supervision, component supervision, and discrete ordinary differential equation constraints to improve denoising performance, reconstruction stability, and model interpretability. Summary of the Invention

[0005] In view of the above, the main objective of this invention is to propose a physical information-driven method and system for PPG signal denoising and reconstruction in a two-stream network, in order to solve the problem that existing PPG signal denoising methods are unable to simultaneously achieve noise suppression, waveform preservation, and separation of signal components at different time scales, and to improve the stability of the reconstruction results and the interpretability of the model processing.

[0006] This invention proposes a method for denoising and reconstructing PPG signals in a two-stream network driven by physical information, the method comprising the following steps: The original PPG time series signal and the corresponding prior condition features are acquired and preprocessed to obtain the condition input; the prior condition features include blood oxygenation features and age features. The original PPG time series signal is subjected to stepwise feature extraction by a shared encoder, and the encoded features are upsampled to the same time resolution as the original PPG time series signal to obtain deep shared features. The deep shared features are input into the oscillating flow branch and the slow trend flow branch respectively; amplitude features and phase increment features are extracted by Hilbert analytical signal processing, and oscillating components are generated; slow trend components are generated by anisotropic diffusion processing. Based on the input conditions, an oscillation frequency and trend relaxation coefficient corresponding to the current PPG signal are generated through a parameter adaptive mapping network. The oscillation component is zero-mean calibrated in the time dimension, the slow trend component is mean-aligned with the original PPG time series signal, and the calibrated oscillation component is superimposed with the mean-aligned slow trend component to obtain the reconstructed PPG time series signal. During model training, a discrete oscillation ordinary differential equation constraint loss is constructed based on the oscillation frequency and the oscillation component; a slow trend ordinary differential equation constraint loss is constructed based on the trend relaxation coefficient and the slow trend component; and a reconstruction loss and a component supervision loss are constructed based on the reconstructed PPG time series signal and the oscillation component, respectively. The reconstruction loss, component supervision loss, discrete oscillatory ordinary differential equation constraint loss, and slow trend ordinary differential equation constraint loss are weighted and combined. The two-stream reconstruction network is trained based on the combined total loss, and the trained two-stream reconstruction network is used for signal reconstruction.

[0007] This invention also proposes a physical information-driven dual-stream network PPG signal denoising and reconstruction system, wherein the system applies the physical information-driven dual-stream network PPG signal denoising and reconstruction method described above, and the system includes: The data feature preprocessing module is used for: The original PPG time series signal and the corresponding prior condition features are acquired and preprocessed to obtain the condition input; the prior condition features include blood oxygenation features and age features. The feature processing module is used for: The original PPG time series signal is subjected to stepwise feature extraction by a shared encoder, and the extracted encoded features are restored to the same time resolution as the original PPG time series signal to obtain deep shared features. The deep shared features are input into the oscillating flow branch and the slow trend flow branch, respectively; the oscillating flow branch extracts amplitude features and phase increment features through Hilbert analytic signal processing, and generates oscillating components based on the deep shared features, amplitude features, and phase increment features; the slow trend flow branch generates slow trend components through anisotropic diffusion processing. The parameter adaptive mapping network module is used for: Based on the input conditions, the oscillation frequency and trend relaxation coefficient are generated through a parameter adaptive mapping network. Signal synthesis module, used for: Zero-mean calibration is performed on the oscillation component in the time dimension, mean alignment is performed on the slow trend component to the original PPG time series signal, and the calibrated oscillation component and the mean-aligned slow trend component are superimposed to obtain the reconstructed PPG time series signal. The model training module is used for: During model training, a discrete oscillation ordinary differential equation constraint loss is constructed based on the oscillation frequency and the oscillation component; a slow trend ordinary differential equation constraint loss is constructed based on the trend relaxation coefficient and the slow trend component; and a reconstruction loss and a component supervision loss are constructed based on the reconstructed PPG time series signal and the oscillation component, respectively. The reconstruction loss, component supervision loss, discrete oscillatory ordinary differential equation constraint loss, and slow trend ordinary differential equation constraint loss are weighted and combined. The two-stream reconstruction network is trained based on the combined total loss, and the trained two-stream reconstruction network is used for signal reconstruction.

[0008] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By setting up a shared encoder, an oscillating flow branch, and a slow trend flow branch, the PPG signal reconstruction process is divided into two parallel processing processes: oscillating component extraction and slow trend component extraction. This allows signal components at different time scales to have a clearer functional division of labor and reduces mutual interference between different signal components.

[0009] 2. The oscillating flow branch obtains the amplitude and phase increment features of the depth-shared feature through Hilbert analytical signal processing, and fuses the depth-shared feature, amplitude feature and phase increment feature, which is beneficial to retain the oscillation changes and local waveform features in the PPG signal.

[0010] 3. The slow trend flow branch performs multiple iterations of smoothing on the deep shared features through anisotropic diffusion processing, and adjusts the diffusion propagation degree according to the local differential gradient, which is beneficial for extracting the slowly changing trend components in the PPG signal, while reducing the possibility of local abrupt changes being over-smoothed.

[0011] 4. By setting up a parameter adaptive mapping network, oscillation frequency and trend relaxation coefficient are generated based on blood oxygenation characteristics and age characteristics, so that the constraint parameters of discrete ordinary differential equations can be adjusted according to different input samples, thereby enhancing the model's adaptability to the signal change characteristics of different individuals.

[0012] 5. By performing zero-mean calibration on the oscillation component and aligning the slow trend component with the mean of the original PPG time series signal, the oscillation component mainly represents the signal components that change around the reference level, while the slow trend component retains the overall amplitude reference of the original signal, thereby improving the stability when the two components are superimposed and reconstructed.

[0013] 6. During model training, reconstruction loss, component supervision loss, oscillating component discrete ordinary differential equation constraint loss, and slow trend component discrete ordinary differential equation constraint loss are used in combination. The weights of the two types of discrete ordinary differential equation constraint losses are gradually increased. This allows the model to approximate the wavelet smoothed reference signal while maintaining the time variation characteristics of the oscillating component and the slow trend component, thereby improving the stability and interpretability of PPG signal denoising and reconstruction results.

[0014] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by means of embodiments of the invention. Attached Figure Description

[0015] Figure 1 This is a flowchart of the physical information-driven dual-stream network PPG signal denoising and reconstruction method proposed in this invention; Figure 2 This is a schematic diagram of the internal structure of the dual-stream reconstruction network of the present invention and the generation process of oscillating and slow-trend components. Figure 3 The computational flow and hierarchical structure diagram for generating oscillation frequency and trend relaxation coefficient for the parameter adaptive mapping network of this invention are shown below. Figure 4 This presents the overall framework of the physical information-driven dual-stream network PPG signal denoising and reconstruction system proposed in this invention. Detailed Implementation

[0016] The following describes specific embodiments of the present invention in detail. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements, or elements having the same or similar functions, throughout. The specific embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the scope of protection of the present invention.

[0017] Other aspects of the invention will become clear from the following description and accompanying drawings. These descriptions and drawings specifically disclose some embodiments of the invention to illustrate its basic principles and processes; however, it should be understood that the scope of protection of the invention is not limited to the specific embodiments described below.

[0018] Please see Figure 1This embodiment discloses a physical information-driven dual-stream network PPG signal denoising and reconstruction method. The method is implemented through a dual-stream reconstruction network, which includes a shared encoder, an oscillating stream branch, a slow trend stream branch, and a parameter adaptive mapping network. During training, it relies on the parameter adaptive mapping network and prior condition features. After training is completed, the parameter adaptive mapping network is deactivated and no longer accepts prior condition features. The original PPG time series signal is directly used for signal reconstruction.

[0019] The method includes the following steps: Step 1: Obtain the original PPG time series signal and the corresponding blood oxygen and age characteristics, and process missing values ​​to obtain the conditional input; The original PPG time series signal, the blood oxygenation features and age features corresponding to the original PPG time series signal are obtained, and a wavelet smoothing reference signal is generated for model training.

[0020] Data feature preprocessing module: Suppose the original PPG time series signal input to the dual-stream reconstruction network is: ; in, This represents the original PPG time series signal; Represents the set of real numbers; Indicates the number of signal samples contained in a batch; This indicates the number of PPG signal channels contained in each signal sample; This indicates the number of time sampling points contained in each PPG signal channel.

[0021] For the i-th signal sample in a batch, extract its corresponding blood oxygenation feature and age feature, and form a conditional feature vector, as shown in the following formula: ; Where i represents the signal sample index within the batch, and satisfies ; This represents the conditional feature vector corresponding to the i-th signal sample; This represents the blood oxygenation characteristic corresponding to the i-th signal sample; This represents the age characteristic corresponding to the i-th signal sample.

[0022] The blood oxygenation and age features are numerically processed, and missing values ​​are detected. When any conditional feature has a missing value, a preset default value corresponding to that conditional feature is used to fill it. The processed blood oxygenation and age features are combined into a conditional feature vector without altering the original PPG time series signal.

[0023] The conditional input is obtained by sequentially combining the conditional feature vectors corresponding to all signal samples within the same batch, as shown in the following formula: ; in, Represents conditional input in matrix form; This represents the vector transpose operation; This represents the conditional feature vector corresponding to the Bth signal sample; the number 2 indicates that the conditional input for each signal sample includes two feature dimensions: blood oxygenation feature and age feature.

[0024] Conditional input The i-th row in the input corresponds one-to-one with the i-th signal sample in the original PPG time series signal X. The original PPG time series signal X is used as input to the two-stream reconstruction network, and the conditional input is used as input to the parameter adaptive mapping network.

[0025] During model training, the original PPG time series signal is subjected to wavelet decomposition, thresholding, and inverse wavelet reconstruction to obtain a wavelet-smoothed reference signal, as shown in the following equation: ; in, Indicates the wavelet-smoothed reference signal; This represents the wavelet decomposition operation performed along the time dimension of the original PPG time series signal; This indicates a soft thresholding operation performed on the detail coefficients obtained from wavelet decomposition with a threshold τ; τ represents the threshold determined based on the wavelet detail coefficients. This indicates the inverse wavelet reconstruction operation.

[0026] The wavelet-smoothed reference signal Y has the same dimension as the original PPG time series signal X, and the relationship is as follows: ; in, , , , and The meaning is consistent with the aforementioned definition.

[0027] The wavelet-smoothed reference signal Y is used only as a supervisory reference target during model training and is not used as input to the forward processing of the two-stream reconstruction network. The input to the two-stream reconstruction network is the original PPG time series signal X, and the input to the parameter adaptive mapping network is the conditional input. .

[0028] Step 2: Extract deep shared features through a shared encoder, and obtain oscillating and slow trend components through Hilbert analytic signal processing and anisotropic diffusion processing, respectively. Please see Figure 2 The original PPG time series signal X is input into a dual-stream reconstruction network. A shared encoder is used to extract deep shared features from the original PPG time series signal; an oscillating flow branch is used to generate oscillating components based on the deep shared features; and a slow trend flow branch is used to generate slow trend components based on the deep shared features.

[0029] Deep shared feature extraction: The shared encoder comprises multiple one-dimensional convolutional layers and a bottleneck convolutional layer connected in sequence; the original PPG time series signal X is sequentially input into the one-dimensional convolutional layers to obtain the encoded features at different levels, as shown in the following formula: ; ; ; in, , and These represent the encoded features obtained after one-dimensional convolution processing of the first, second, and third layers, respectively. , and These represent the first, second, and third one-dimensional convolution operations, respectively; This represents the linear rectification activation function.

[0030] The third layer encoded feature E3 is input into the bottleneck convolutional layer to obtain the bottleneck feature, as shown in the following formula: ; in, Indicates bottleneck characteristics; the subscript b indicates the bottleneck layer; This indicates the bottleneck convolution operation.

[0031] During the encoding process, partial one-dimensional convolutional layers are used to reduce the temporal resolution of features, thereby expanding the network's receptive range in the temporal dimension. This addresses bottleneck features. Linear interpolation upsampling is performed to restore it to the same time length as the original PPG time series signal X, yielding the deep shared features, as shown in the following formula: ; in, Indicates deep shared features; This indicates a linear interpolation upsampling operation; T represents the number of time sampling points of the aforementioned original PPG time series signal.

[0032] The deep shared feature F is input to the oscillating flow branch and the slow trend flow branch, respectively; Oscillating flow branch: The oscillating flow branch performs Hilbert analytical signal processing on the depth-shared feature F to obtain amplitude and phase increment features.

[0033] First, perform a Fourier transform on the deep shared feature F along the time dimension: ; in, Frequency domain representation of deep shared features; This indicates a Fourier transform performed along the time dimension; the subscript t indicates that the transform is performed along the time dimension.

[0034] An analytical signal kernel is constructed based on the number of time sampling points T. When the number of time sampling points T is even, the analytical signal kernel has the following relationship: ; When the number of time sampling points T is odd, the analytical signal kernel relationship is as follows: ; in, This represents the value of the analytic signal kernel at the k-th frequency position; the subscript H indicates Hilbert analytic signal processing; k represents the frequency position index, and satisfies... .

[0035] Frequency domain representation of deep shared features Element-wise multiplication with the analytic signal kernel followed by inverse Fourier transform yields a complex analytic signal, as shown in the following equation: ; Where Z represents a complex analytic signal; This represents the inverse Fourier transform performed along the time dimension; Indicates that it is composed of all Composed of analytical signal kernels; This indicates element-wise multiplication.

[0036] The amplitude characteristics are obtained from the modulus of the complex analytic signal Z, and the relationship is as follows: ; Where A represents the amplitude characteristic; This indicates taking the modulus of a complex analytic signal element by element.

[0037] The phase characteristics are obtained from the argument of the complex analytic signal Z, and the relationship is as follows: ; in, Indicates phase characteristics; This represents the operation of calculating the argument of a complex number element by element.

[0038] The initial phase difference between adjacent time positions is calculated using the following formula: ; in, This represents the initial phase difference between time position t and time position t+1; This represents the value of the phase characteristic at time position t; This represents the value of the phase feature at time position t+1; t represents the time position index, and satisfies... The subscript 0 indicates the initial phase difference before periodic wrapping.

[0039] Perform periodic wrapping on the initial phase difference to obtain the phase increment feature, as shown in the following formula: ; in, Indicates phase increment characteristics; The phase increment characteristic representing time position t; Represents the two-parameter arctangent function; Represents the sine function; The cosine function is represented in this specific embodiment. The preset phase interval is (-π, π], and the phase difference is automatically limited to this interval by the two-parameter arctangent function.

[0040] Boundary copy padding is performed at the end position of the phase increment feature to make its time length consistent with that of the depth-shared feature F.

[0041] The depth-sharing feature F, amplitude feature A, and phase increment feature are combined. When splicing along the channel dimension, the relationship is as follows: ; in, Indicates the oscillation characteristics after splicing; subscript Indicates a branch of the oscillating flow; This indicates a feature splicing operation performed along the channel dimension.

[0042] The spliced ​​oscillation characteristics The input oscillatory projection layer is used, and the initial oscillatory components are obtained through the hyperbolic tangent activation function, as shown in the following equation: ; in, Indicates the initial oscillation component; The convolution projection operation represents the branch of the oscillating flow; This represents the hyperbolic tangent activation function.

[0043] The mean of the initial oscillation component along the time dimension is calculated using the following formula: ; in, This represents the mean of the initial oscillation component along the time dimension; This indicates the operation of calculating the mean along the time dimension.

[0044] Subtracting the mean from the initial oscillation component yields the oscillation component, as shown in the following formula: ; in, This represents the oscillation component after zero-mean calibration.

[0045] By using zero-mean calibration, the oscillation component is made to mainly represent the time-series components that vary around the signal reference level, reducing the impact of fixed bias on subsequent signal reconstruction.

[0046] Slow trend flow branch: The slow trend flow branch performs anisotropic diffusion processing on the deep shared feature F to extract slowly changing trend features.

[0047] The deep shared features are initialized as diffusion iteration variables, as shown in the following formula: ; in, This represents the initial iteration variable before the diffusion process begins; the superscript (0) indicates the initial iteration state.

[0048] In the nth diffusion iteration, the forward difference gradient is calculated, and the relationship is as follows: ; And backward differential gradient: ; Where n represents the diffusion iteration index; This represents the value of the diffusion variable at time position t during the nth iteration; This indicates a forward difference operation performed along the positive time direction; This indicates a backward difference operation performed in the reverse direction of time.

[0049] For positions at both ends of the time series that lack adjacent values, a boundary copying method is used to fill them, ensuring that the forward and backward difference gradients have the same time length as the diffusion iteration variables.

[0050] The forward diffusion conduction coefficient is calculated based on the forward differential gradient, and the relationship is as follows: ; The backward diffusion conduction coefficient is calculated based on the backward differential gradient, and the relationship is as follows: ; in, This represents the forward diffusion conduction coefficient at time position t during the nth iteration; This represents the backward diffusion conduction coefficient at time position t during the nth iteration; The diffusion threshold parameter represents the relationship between the local differential amplitude and the degree of diffusion conduction.

[0051] Based on the diffusion conduction coefficient and the corresponding difference gradient, the diffusion iteration variables are updated, as shown in the following formula: ; in, This represents the diffusion variable at time position t after the nth iteration update; This represents the evolution step size of a single diffusion iteration.

[0052] When the absolute value of the local differential gradient is small, the corresponding diffusion conduction coefficient is large to enhance the smoothing of slowly changing regions; when the absolute value of the local differential gradient is large, the corresponding diffusion conduction coefficient is small to reduce over-smoothing of local change boundaries.

[0053] After completing a preset number of diffusion iterations, the final diffusion variable is input into the slow trend projection layer, and the initial slow trend component is obtained through the Softplus activation function, as shown in the following formula: ; in, Indicates the initial slow trend component; This indicates the Softplus activation function; Indicates the convolution projection operation of the slow-trending flow branch; subscript Indicates a slow trend flow branch; This indicates the preset number of diffusion iterations; This represents the diffusion variable obtained after completing all diffusion iterations.

[0054] The mean of the initial slow trend component along the time dimension is calculated using the following formula: ; in, This represents the mean of the initial slow trend component along the time dimension.

[0055] The mean of the original PPG time series signal X along the time dimension is calculated using the following formula: ; in, This represents the mean of the original PPG time series signal along the time dimension.

[0056] The initial slow trend components are mean-aligned to obtain the slow trend components, as shown in the following formula: ; in, This represents the slow trend component after mean alignment.

[0057] By aligning the mean as described above, the slow trend component retains the overall amplitude reference of the original PPG time series signal, reducing the overall offset between the oscillation component and the slow trend component during subsequent superposition.

[0058] Step 3: Based on the input conditions, generate the oscillation frequency and trend relaxation coefficient corresponding to the current signal through a parameter adaptive mapping network; Please see Figure 3 Input the conditions obtained in step 1 An adaptive input parameter mapping network is used. This network comprises a first fully connected layer, a linear rectified activation function, and a second fully connected layer, connected sequentially. These layers are used to generate oscillation frequencies and trend relaxation coefficients corresponding to each PPG signal sample based on blood oxygenation and age characteristics.

[0059] First, input the conditions. The first fully connected layer is input, and the hidden features are obtained by passing a linear rectified activation function, as shown in the following formula: ; in, This represents the hidden feature matrix generated by the parameter adaptive mapping network; This represents the weight matrix of the first fully connected layer; This represents the bias vector of the first fully connected layer.

[0060] Hidden feature matrix Inputting the second fully connected layer yields an unrestricted parameter output matrix, as shown in the following formula: ; in, This represents the undefined parameter output matrix of the parameter adaptive mapping network; This represents the weight matrix of the second fully connected layer; This represents the bias vector of the second fully connected layer.

[0061] The undefined parameter output matrix P is divided into an oscillation parameter vector and a trend parameter vector according to the output dimension, as shown in the following formula: ; in, This represents an undefined parameter vector used to generate the oscillation frequency; Represents the undefined parameter vector used to generate the trend relaxation coefficient; subscript Indicates the oscillation frequency; subscript This represents the trend relaxation coefficient.

[0062] To ensure that the generated parameters are within a preset range, the oscillation parameter vector is adjusted accordingly. and trend parameter vector Perform Sigmoid range mapping.

[0063] The relationship between the oscillation frequency vectors is as follows: ; in, This represents a vector consisting of the oscillation frequencies of each PPG signal sample in a batch. Indicates the preset lower limit of the oscillation frequency; Indicates the preset upper limit of the oscillation frequency; This represents a sigmoid activation function that maps input values ​​to a range between 0 and 1.

[0064] Among them, the preset lower limit of the oscillation frequency 0.5Hz, preset upper limit Since the frequency is 3.5Hz, therefore: ; The number 3.0 represents the difference between the preset upper limit and the preset lower limit of the oscillation frequency.

[0065] The relationship between the trend relaxation coefficient vector is as follows: ; in, This represents a vector consisting of the trend relaxation coefficients corresponding to each PPG signal sample in a batch. This represents the preset lower limit of the trend relaxation coefficient; This represents the preset upper limit of the trend relaxation coefficient.

[0066] Among them, the preset lower limit of the trend relaxation coefficient The value is 0.001, with a preset upper limit. Since it is 0.05, therefore: ; The number 0.049 represents the difference between the preset upper limit and the preset lower limit of the trend relaxation coefficient.

[0067] For the i-th PPG signal sample in the batch, the oscillation frequency and trend relaxation coefficient generated by the parameter adaptive mapping network are denoted as follows: and .in, Represents the oscillation frequency vector The i-th element in Represents the trend relaxation coefficient vector The i-th element in the original PPG time series signal X, and both of them correspond to the i-th signal sample in the original PPG time series signal X.

[0068] The oscillation frequency vector and trend relaxation coefficient vector The discrete ordinary differential equation constraints used to construct oscillatory and slow-trend components during model training do not directly participate in the forward superposition of oscillatory and slow-trend components.

[0069] Step 4: Perform zero-mean calibration on the oscillating components, mean alignment on the slow trend components, and superimpose the two to obtain the reconstructed PPG time series signal; Signal synthesis module: The oscillation component obtained in step 2 With slow trend component By performing element-wise superposition, the reconstructed PPG time series signal is obtained, and the relationship is as follows: ; in, This indicates the reconstruction of the PPG time series signal; This represents the oscillation component after zero-mean calibration; This represents the slow trend component after mean alignment; the symbol "+" indicates element-wise addition of values ​​at the same sample, same channel, and same time position in the oscillating component and the slow trend component; the subscript rec indicates the reconstruction result.

[0070] Oscillating component O, slow trend component S, and reconstructed PPG time series signal Having the same tensor dimension, the relationship is as follows: ; in, , , and The meaning is consistent with the aforementioned definition.

[0071] The oscillating component O, calibrated to zero mean over time, characterizes the oscillating components varying around a signal baseline. The slow-trend component S, aligned to the mean of the original PPG time series signal, characterizes the overall amplitude baseline and the slow-trend variation of the signal. By element-wise superposition of the two components, a reconstructed PPG time series signal exhibiting both oscillating and slow-trend variations is obtained. .

[0072] The oscillation frequency vector f and the trend relaxation coefficient vector ρ generated in step 3 do not directly participate in the above signal superposition operation. Instead, they are used to construct the discrete ordinary differential equation constraints for the oscillation component and the slow trend component, respectively, during model training. Therefore, the forward signal synthesis process only corresponds to the element-wise superposition of the oscillation component and the slow trend component.

[0073] Step 5: Based on the reconstructed PPG time series signal, oscillation frequency, trend relaxation coefficient, oscillation component and slow trend component, construct a composite total loss to train the two-stream reconstruction network, and use the trained two-stream reconstruction network to reconstruct the signal.

[0074] After completing the PPG signal reconstruction in step 4, the model training module uses the reconstructed PPG time series signal... Wavelet smoothed reference signal Y, oscillation component O, slow trend component S, oscillation frequency vector and trend relaxation coefficient vector The composite total loss is calculated, and the learnable parameters in the two-stream reconstruction network and the parameter adaptive mapping network are updated based on the composite total loss.

[0075] Reconstruction loss calculation: Calculate and reconstruct the PPG time series signal The mean square error between the wavelet-smooth reference signal Y and the reconstruction loss is obtained from the following formula: ; in, Indicates the reconstruction loss; Represents the reconstruction of the PPG time series signal The i-th signal sample, the i-th The values ​​of each PPG signal channel at the t-th time position; This represents the value at the corresponding position in the wavelet-smoothed reference signal Y; Represents the PPG signal channel index, and satisfies The meanings of i, t, B, C, and T are consistent with the aforementioned definitions. This represents the loss function, and different loss functions are distinguished by their corresponding subscripts.

[0076] Reference component construction: To separately monitor the oscillating flow branch and the slow-trend flow branch, a moving average is performed on the wavelet-smoothed reference signal Y along the time dimension to obtain the slow-trend reference component, as shown in the following equation: ; in, Indicates the slow trend reference component; superscript This represents the reference component used for model training; This indicates a moving average processing operation performed along the time dimension.

[0077] Based on the wavelet smoothing reference signal Y and the slow trend reference component The difference between them yields the oscillation reference component, and the relationship is as follows: ; in, Indicates the oscillation reference component; symbol " "" indicates element-wise subtraction of values ​​at the same sample, same channel, and same time position.

[0078] Through the above processing, the wavelet-smoothed reference signal is divided into oscillatory reference components for monitoring the oscillatory flow branch. and the slow trend reference component used to supervise the slow trend flow branches .

[0079] Component monitoring loss calculation: Calculate the oscillation component O and the oscillation reference component respectively. The mean square error between them, and the slow trend component S and the slow trend reference component The mean square error between them is used to obtain the component supervision loss, and the relationship is as follows: ; in, Indicates component supervision loss; the subscript comp indicates component supervision. Represents the i-th signal sample in the oscillation component O, the i-th... The values ​​of each PPG signal channel at the t-th time position; Represents the oscillation reference component The value at the corresponding position in the text; This represents the value at the corresponding position in the slow trend component S; Indicates the slow trend reference component The value at the corresponding position in the text.

[0080] The component supervision loss guides the oscillating flow branch and the slow trend flow branch to learn the corresponding reference components, thereby reducing functional confusion between the two branches during model training.

[0081] Constraints of discrete ordinary differential equations for oscillatory components: The discrete oscillation coefficients are calculated based on the oscillation frequency corresponding to the i-th signal sample, using the following formula: ; in, This represents the discrete oscillation coefficient corresponding to the i-th signal sample; Represents pi; This represents the i-th oscillation frequency in the oscillation frequency vector f; This indicates the sampling frequency of the original PPG time series signal; the subscript s indicates sampling.

[0082] For the time positions of the oscillation component other than the beginning and end, construct the discrete oscillation residuals, with the following relationship: ; in, Represents the i-th signal sample in the oscillation component O, the i-th... The values ​​of each PPG signal channel at the (t+1)th time position; Represents the i-th signal sample in the oscillation component O, the i-th... The values ​​of each PPG signal channel at the (t-1)th time position; Represents the i-th signal sample, the i-th... The discrete oscillation residuals of each PPG signal channel at time position t; the superscript osc indicates the oscillation constraint; the time position index t satisfies the following in this formula. .

[0083] To reduce the impact of amplitude differences between different signal samples and different PPG signal channels on the constraint loss, the normalized scale of the oscillation residual is calculated, and the relationship is as follows: ; in, Represents the i-th signal sample, the i-th... Normalized scale of oscillation residuals corresponding to each PPG signal channel; ε represents the absolute value; ε represents the numerical stability term used to prevent the normalization scale from being zero.

[0084] The constraint loss of the discrete ordinary differential equation for the oscillation component is calculated based on the normalized discrete oscillation residual, and the relationship is as follows: ; in, This represents the constraint loss of the discrete ordinary differential equation for the oscillating component.

[0085] During the calculation of model parameter gradients, the normalization scale of oscillation residuals is used. It is used only for numerical normalization and does not participate in gradient backpropagation.

[0086] Constraints on discrete ordinary differential equations for slow-trend components: First, calculate the i-th signal sample... The slow trend time averages corresponding to each PPG signal channel are expressed by the following formula: ; in, Represents the i-th signal sample in the slow trend component, the i-th signal sample ... The mean of each PPG signal channel along the time dimension; the horizontal line above indicates the result after averaging along the time dimension.

[0087] Based on the trend relaxation coefficient corresponding to the i-th signal sample The discrete relaxation residuals of the slow-trend components are constructed as follows: ; in, Represents the i-th signal sample, the i-th... Discrete relaxation residuals of each PPG signal channel at time t; the superscript tr indicates slow trend constraint. This represents the i-th trend relaxation coefficient in the trend relaxation coefficient vector ρ; the time position index t satisfies the following in this formula. .

[0088] The normalized scale for slow-trending residuals is calculated using the following formula: ; in, Represents the i-th signal sample, the i-th... Slow-trend residual normalization scale corresponding to each PPG signal channel.

[0089] The constraint loss of the discrete ordinary differential equation for the slow trend component is calculated based on the normalized discrete relaxation residual, and the relationship is as follows: ; in, This represents the constraint loss of the discrete ordinary differential equation for the slow trend component.

[0090] In the process of calculating the gradient of model parameters, the normalization scale of slow-trend residuals is used. It is used only for numerical normalization and does not participate in gradient backpropagation.

[0091] Constraint weight warm-up: In the early stages of model training, the weights of the constraint losses for the oscillating component discrete ordinary differential equations and the slow-trend component discrete ordinary differential equations in the total composite loss are gradually increased. The relationship between the constraint weight warm-up coefficients is as follows: ; in, This represents the constraint weight warm-up coefficient corresponding to the e-th training round; This indicates the current training round, counting from 0. Indicates the preset number of preheating cycles. This indicates taking the minimum value among the values ​​within the parentheses.

[0092] When the number of training rounds has not reached the preset number of constraint weight warm-up rounds, the constraint weight warm-up coefficient gradually increases with the number of training rounds; when the preset number of constraint weight warm-up rounds has been reached, the constraint weight warm-up coefficient remains at 1.

[0093] In this embodiment, the number of preheating rounds is set to 5 rounds, the base value of the preset weight of the oscillating ordinary differential equation constraint loss is set to 0.001, the base value of the preset weight of the slow trend ordinary differential equation constraint loss is set to 0.05, and the base value of the preset weight of the component supervision loss is set to 1.0.

[0094] Calculation of total compound loss: The reconstruction loss, component supervision loss, oscillating component discrete ordinary differential equation constraint loss, and slow trend component discrete ordinary differential equation constraint loss are weighted and combined to obtain the total composite loss, as shown in the following formula: ; in, Indicates the total compound loss; Indicates the reconstruction loss; This represents the component monitoring loss, which is obtained by adding the mean square error between the oscillating component and the oscillating reference component and the mean square error between the slow trend component and the slow trend reference component. The constraint loss of the discrete oscillatory ordinary differential equation is represented by the mean squared value of the normalized discrete oscillation residuals. The slow-trend ordinary differential equation constraint loss is represented by the mean squared value of the normalized discrete relaxation residuals. , , These represent the preset weights of the component supervision loss, the discrete oscillatory ordinary differential equation constraint loss, and the slow trend ordinary differential equation constraint loss, respectively.

[0095] By constraining the preheating coefficient In the early stages of model training, the impact of the constraint loss of the two types of discrete ordinary differential equations on the update of model parameters is reduced, and the corresponding constraints are gradually strengthened as the model is trained.

[0096] Please see Figure 4 This embodiment also discloses a physical information-driven dual-stream network PPG signal denoising and reconstruction system. The system applies the aforementioned physical information-driven dual-stream network PPG signal denoising and reconstruction method and includes a data feature preprocessing module, a feature processing module, a parameter adaptive mapping network module, a signal synthesis module, and a model training module.

[0097] Data feature preprocessing module: The data feature preprocessing module is used to acquire the original PPG time series signal and the blood oxygenation feature and age feature corresponding to the original PPG time series signal, to perform numerical conversion and missing value processing on the blood oxygenation feature and age feature, and to combine the processed blood oxygenation feature and age feature into a conditional input Q.

[0098] During the model training phase, the data feature preprocessing module is also used to perform wavelet smoothing on the original PPG time series signal to obtain a wavelet smoothed reference signal Y. This wavelet smoothed reference signal Y serves as the supervisory reference target for model training and is not used as input during the forward processing of the two-stream reconstruction network.

[0099] Feature processing module: The feature processing module includes a shared encoder, an oscillating flow branch, and a slow trend flow branch.

[0100] The shared encoder is used to perform multi-layer one-dimensional convolution processing on the original PPG time series signal X, and to perform linear interpolation upsampling on the bottleneck feature to obtain a deep shared feature F with the same time length as the original PPG time series signal.

[0101] The oscillating flow branch is used to perform Hilbert analytical signal processing on the depth shared feature F to obtain the amplitude feature A and the phase increment feature. And the depth-sharing feature F, amplitude feature A, and phase increment feature are combined. By performing stitching and convolution projection, the initial oscillation component is generated. ; by starting from the initial oscillation component Subtracting its time mean from the mean yields the oscillation component O.

[0102] The slow trend flow branch is used to perform anisotropic diffusion iteration on the deep shared feature F, calculate the diffusion propagation coefficient based on the local difference gradient, and update the diffusion variables; the features after the diffusion iteration are convolved and projected to generate the initial slow trend component. ; By analyzing the initial slow trend components The slow trend component S is obtained by performing mean alignment with the original PPG time series signal.

[0103] Parameter adaptive mapping network module: The parameter adaptive mapping network module is used to receive the conditional input Q and generate the oscillation frequency vector f and the trend relaxation coefficient vector ρ through a fully connected layer and a nonlinear activation function connected in sequence.

[0104] In this context, each element in the oscillation frequency vector f corresponds to a PPG signal sample and is used to construct the discrete ordinary differential equation constraints for the corresponding oscillation component; each element in the trend relaxation coefficient vector ρ corresponds to a PPG signal sample and is used to construct the discrete ordinary differential equation constraints for the corresponding slow trend component.

[0105] The oscillation frequency and trend relaxation coefficient are limited to a preset range through range mapping to avoid the parameter adaptive mapping network generating parameter values ​​that are unsuitable for model training.

[0106] Signal synthesis module: The signal synthesis module is used to superimpose the oscillating component O and the slow trend component S element by element to obtain the reconstructed PPG time series signal. .

[0107] The oscillation component O is zero-mean calibrated and is mainly used to characterize the oscillation component that changes around the signal reference level; the slow trend component S is aligned with the mean of the original PPG time series signal and is mainly used to characterize the overall amplitude reference and slow trend of the signal.

[0108] The oscillation frequency vector f and the trend relaxation coefficient vector ρ generated by the parameter adaptive mapping network do not directly participate in the superposition operation of the oscillation component O and the slow trend component S.

[0109] Model training module: The model training module is used to train the model based on the reconstructed PPG time series signal. Calculate the reconstruction loss using the wavelet-smoothed reference signal Y .

[0110] The model training module is also used to perform a moving average process on the wavelet-smoothed reference signal Y to obtain the slow-trend reference component. , And based on the wavelet smoothed reference signal Y and the slow trend reference component The difference value is used to obtain the oscillation reference component. Based on the oscillation component O and the oscillation reference component Slow trend component S and slow trend reference component Calculate component supervision loss .

[0111] The model training module constructs the constraint loss of the discrete ordinary differential equations of the oscillation components based on the oscillation frequency vector f. And based on the trend relaxation coefficient vector Constructing Constraint Loss of Discrete Ordinary Differential Equations for Slow Trend Components .

[0112] The model training module will reconstruct the loss. , component monitoring loss Constraint loss of discrete ordinary differential equations for oscillatory components and the constraint loss of the discrete ordinary differential equation with slow trend components By performing a weighted combination, the total composite loss is obtained. And based on the total composite loss The learnable parameters in the dual-stream reconstruction network and the parameter adaptive mapping network are updated synchronously.

[0113] In the early stages of model training, the model training module gradually increases the weight of the constraint loss of the two types of discrete ordinary differential equations by using the constraint weight warm-up coefficient w(e) to reduce the impact of the constraint terms on the data fitting process when the model output components are not yet stable.

[0114] The process of using the trained model: After model training is completed, the original PPG time series signal to be processed is input into the dual-stream reconstruction network. The deep shared features are obtained through the shared encoder, and the oscillating components and slow-trend components are obtained through the oscillating flow branch and the slow-trend flow branch, respectively.

[0115] The oscillation component and the slow trend component are superimposed to obtain the denoised and reconstructed PPG time series signal.

[0116] During model usage, reconstruction loss, component supervision loss, and discrete ordinary differential equation constraint loss are no longer calculated, nor are gradient backpropagation and model parameter updates performed. The parameter adaptive mapping network and corresponding conditional features are mainly used to form constraint parameters associated with sample conditions during the model training phase.

[0117] In summary, this invention performs deep feature extraction on PPG time series signals using a shared encoder, generates oscillating and slow-trend components using oscillating flow branches and slow-trend flow branches respectively, and clarifies the representation benchmarks of the two components through zero-mean calibration and mean alignment. Simultaneously, it generates oscillation frequency and trend relaxation coefficients using blood oxygenation and age features, and jointly trains the model using reconstruction supervision, component supervision, and discrete ordinary differential equation constraints, thereby achieving noise reduction and reconstruction of PPG signals.

[0118] In the description of this specification, references to terms such as "one specific embodiment," "some specific embodiments," "example," or "specific example" refer to specific features, structures, materials, or characteristics described in connection with that specific embodiment or example, which are included in at least one specific embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same specific embodiment or example, and the specific features, structures, materials, or characteristics described may be combined in a suitable manner in one or more specific embodiments or examples.

[0119] The specific embodiments described above are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, various modifications, substitutions, or improvements can be made to the above specific embodiments without departing from the concept of the present invention, and all such modifications, substitutions, or improvements should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for denoising and reconstructing PPG signals in a two-stream network driven by physical information, characterized in that, The method is implemented using a two-stream reconstruction network and includes the following steps: The original PPG time series signal and the corresponding prior condition features are acquired and preprocessed to obtain the condition input; the prior condition features include blood oxygenation features and age features. The original PPG time series signal is subjected to stepwise feature extraction by a shared encoder, and the extracted encoded features are restored to the same time resolution as the original PPG time series signal to obtain deep shared features. The deep shared features are respectively input into the oscillating flow branch and the slow trend flow branch; The oscillating flow branch extracts amplitude and phase increment features through Hilbert analytic signal processing, and generates oscillating components based on the depth-shared features, amplitude features, and phase increment features; The slow trend flow branch generates slow trend components through anisotropic diffusion processing; Based on the input conditions, the oscillation frequency and trend relaxation coefficient are generated through a parameter adaptive mapping network. Zero-mean calibration is performed on the oscillation component in the time dimension, mean alignment is performed on the slow trend component to the original PPG time series signal, and the calibrated oscillation component and the mean-aligned slow trend component are superimposed to obtain the reconstructed PPG time series signal. During model training, a discrete oscillation ordinary differential equation constraint loss is constructed based on the oscillation frequency and the oscillation component; a slow trend ordinary differential equation constraint loss is constructed based on the trend relaxation coefficient and the slow trend component; and a reconstruction loss and a component supervision loss are constructed based on the reconstructed PPG time series signal and the oscillation component, respectively. The reconstruction loss, component supervision loss, discrete oscillatory ordinary differential equation constraint loss, and slow trend ordinary differential equation constraint loss are weighted and combined. The two-stream reconstruction network is trained based on the combined total loss, and the trained two-stream reconstruction network is used for signal reconstruction.

2. The method for denoising and reconstructing PPG signals in a two-stream network driven by physical information according to claim 1, characterized in that, The prior condition features are preprocessed to obtain the condition input, specifically including: Blood oxygenation and age features were extracted from the feature data associated with the original PPG time-series signal; The blood oxygenation and age features are converted into floating-point values, and the missing blood oxygenation and age features are filled with preset normalized default values. The blood oxygenation features and age features, after processing for missing values, are combined into a two-dimensional conditional input for use by the parameter adaptive mapping network, thus obtaining the conditional input.

3. The method for denoising and reconstructing PPG signals in a two-stream network driven by physical information according to claim 2, characterized in that, The original PPG time series signal is subjected to stepwise feature extraction using a shared encoder, specifically including: The original PPG time series signal is sequentially channel-mapped and time-dimension-downsampled by multiple one-dimensional convolutional layers, and the encoded features are obtained by a non-linear activation function. The encoded features are further mapped using a bottleneck convolutional layer; The encoded features output by the bottleneck convolutional layer are upsampled to the same time length as the original PPG time series signal using linear interpolation to obtain the depth-shared features.

4. The method for denoising and reconstructing PPG signals in a two-stream network driven by physical information according to claim 3, characterized in that, The oscillating flow branch generates oscillating components, specifically including: Perform a Fast Fourier Transform on the deep shared features to obtain frequency domain features; Based on the sequence length of the deep shared features, a Hilbert analytic signal operator kernel is constructed. The kernel value corresponding to the DC component is set to 1, the kernel value corresponding to the positive frequency component is set to 2, and the kernel value corresponding to the remaining frequency components is set to 0. When the sequence length is even, the kernel value corresponding to the Nyquist frequency is set to 1. Multiply the frequency domain features by the Hilbert analytic signal operator kernel and perform an inverse fast Fourier transform to obtain a complex analytic signal; Amplitude features are extracted based on the modulus of the complex analytic signal, phase features are extracted based on the argument of the complex analytic signal, and the phase difference between adjacent time positions is calculated along the time dimension. The phase difference is limited to a preset phase interval by periodic wrapping operation to obtain phase increment features, and the phase increment features and depth-shared features are made to have the same time length by boundary copying and filling. The depth-shared features, amplitude features, and phase increment features are concatenated along the channel dimension, and the initial oscillation components are obtained through convolutional projection and hyperbolic tangent activation. Calculate the mean of the initial oscillation component along the time dimension, and subtract the mean from the initial oscillation component to obtain the oscillation component with zero mean.

5. The method for denoising and reconstructing PPG signals in a two-stream network driven by physical information according to claim 4, characterized in that, The slow trend flow branch generates a slow trend component, specifically including: The deep shared features are initialized as diffusion iteration variables; Within a preset number of iterations, the forward and backward difference gradients of the diffusion iteration variable are calculated along the time dimension, and the forward and backward difference gradients are kept consistent with the time length of the diffusion iteration variable by boundary copying and filling. The forward diffusion conduction coefficient and the backward diffusion conduction coefficient are calculated based on the forward differential gradient and the backward differential gradient, respectively, wherein the diffusion conduction coefficient decreases as the absolute value of the corresponding differential gradient increases; The diffusion iteration variables are updated based on the forward diffusion conduction coefficient, the backward diffusion conduction coefficient, the forward difference gradient, and the backward difference gradient; After updating the diffusion iteration variable according to the preset evolution step size, the diffusion iteration variable after completing a preset number of iterations is passed through convolution projection and the Softplus activation function to obtain the initial slow trend component; Calculate the mean of the initial slow trend component along the time dimension and the mean of the original PPG time series signal along the time dimension. Subtract the mean of the initial slow trend component from the initial slow trend component and add the mean of the original PPG time series signal to obtain the slow trend component aligned with the mean of the original PPG time series signal.

6. The method for denoising and reconstructing PPG signals in a two-stream network driven by physical information according to claim 5, characterized in that, The parameter adaptive mapping network includes an input layer, at least one hidden layer, and an output layer; The parameter adaptive mapping network receives a two-dimensional conditional input consisting of blood oxygenation features and age features, obtains hidden features through linear mapping and nonlinear activation of the hidden layer, and generates two unconstrained parameters through the output layer. Perform Sigmoid range mapping on the two unconstrained parameters respectively to generate the oscillation frequency and trend relaxation coefficient, where: The relationship corresponding to the oscillation frequency is as follows: ； The relationship corresponding to the trend relaxation coefficient is as follows: ； in, and These represent two unconstrained parameters generated by the output layer, where f represents the oscillation frequency, ranging from 0.5Hz to 3.5Hz. The slack coefficient represents the trend relaxation coefficient, which ranges from 0.001 to 0.05; the sigmoid function represents the S-shaped mapping function, whose output value is between 0 and 1.

7. The method for denoising and reconstructing PPG signals in a two-stream network driven by physical information according to claim 6, characterized in that, The constraints of the discrete oscillation ordinary differential equation are constructed based on the second-order time-domain difference of the oscillation components and the oscillation frequency; For the component value at time position t in the oscillation component The corresponding discrete oscillation residual relationship is as follows: ； in, This represents the discrete oscillation residual at time position t. This represents the component value of the oscillation component at time position t, where t represents the time position index. Indicates the oscillation frequency, This indicates the sampling frequency of the original PPG time series signal. Represents pi; The slow-trend ordinary differential equation constraint loss is constructed based on the first-order time-domain difference of the slow-trend component, the time mean of the slow-trend component, and the trend relaxation coefficient. For the time position in the slow trend component Component values ​​at the location The corresponding discrete relaxation residual relationship is as follows: ； in, This represents the discrete relaxation residual at time position t. This represents the component value of the slow trend component at time position t. This represents the trend relaxation coefficient. This represents the mean of the slow trend component along the time dimension.

8. The method for denoising and reconstructing PPG signals in a two-stream network driven by physical information according to claim 7, characterized in that, The model training process includes: Wavelet smoothing is performed on the original PPG time series signal to obtain a wavelet smoothed reference signal. The oscillating component and the slow trend component are superimposed to obtain the reconstructed PPG time series signal, and the reconstruction loss between the reconstructed PPG time series signal and the wavelet smoothing reference signal is calculated. The wavelet smoothed reference signal is subjected to moving average processing to obtain a slow trend reference component, and the oscillation reference component is obtained by subtracting the slow trend reference component from the wavelet smoothed reference signal. The component supervision loss is obtained by calculating the error between the oscillation component and the oscillation reference component, and the error between the slow trend component and the slow trend reference component, respectively. The constraint loss of the discrete oscillation ordinary differential equation is calculated based on the discrete oscillation residual, and the constraint loss of the slow trend ordinary differential equation is calculated based on the discrete relaxation residual. The composite total loss is obtained by weighting and combining the reconstruction loss, component supervision loss, discrete oscillatory ordinary differential equation constraint loss, and slow trend ordinary differential equation constraint loss. The gradient of the composite total loss with respect to the learnable parameters of the shared encoder, oscillating flow branch, slow trend flow branch, and parameter adaptive mapping network is calculated, and the learnable parameters are updated synchronously through backpropagation and the Adam optimizer.

9. The method for denoising and reconstructing PPG signals in a two-stream network driven by physical information according to claim 8, characterized in that, The relationship corresponding to the total composite loss is as follows: ； in, Indicates the total compound loss; Indicates the reconstruction loss; This represents the component monitoring loss, which is obtained by adding the mean square error between the oscillating component and the oscillating reference component and the mean square error between the slow trend component and the slow trend reference component. The constraint loss of the discrete oscillatory ordinary differential equation is represented by the mean squared value of the normalized discrete oscillation residuals. The slow-trend ordinary differential equation constraint loss is represented by the mean squared value of the normalized discrete relaxation residuals. , , These represent the preset weights of the component supervision loss, the discrete oscillatory ordinary differential equation constraint loss, and the slow trend ordinary differential equation constraint loss, respectively. The constraint warm-up coefficient represents the coefficient that increases progressively with each training round, and the corresponding relationship is as follows: ； in, This represents the constraint weight warm-up coefficient corresponding to the e-th training round; This indicates the current training round, counting from 0. Indicates the preset number of preheating cycles. This indicates taking the minimum value among the values ​​within the parentheses.

10. A physical information-driven dual-stream network PPG signal denoising and reconstruction system, characterized in that, The system employs the physical information-driven dual-stream network PPG signal denoising and reconstruction method as described in any one of claims 1 to 9, and the system comprises: The data feature preprocessing module is used for: The original PPG time series signal and the corresponding prior condition features are acquired and preprocessed to obtain the condition input; the prior condition features include blood oxygenation features and age features. The feature processing module is used for: The original PPG time series signal is subjected to stepwise feature extraction by a shared encoder, and the extracted encoded features are restored to the same time resolution as the original PPG time series signal to obtain deep shared features. The deep shared features are input into the oscillating flow branch and the slow trend flow branch, respectively; the oscillating flow branch extracts amplitude features and phase increment features through Hilbert analytic signal processing, and generates oscillating components based on the deep shared features, amplitude features, and phase increment features; the slow trend flow branch generates slow trend components through anisotropic diffusion processing. The parameter adaptive mapping network module is used for: Based on the input conditions, the oscillation frequency and trend relaxation coefficient are generated through a parameter adaptive mapping network. Signal synthesis module, used for: Zero-mean calibration is performed on the oscillation component in the time dimension, mean alignment is performed on the slow trend component to the original PPG time series signal, and the calibrated oscillation component and the mean-aligned slow trend component are superimposed to obtain the reconstructed PPG time series signal. The model training module is used for: During model training, a discrete oscillation ordinary differential equation constraint loss is constructed based on the oscillation frequency and the oscillation component; a slow trend ordinary differential equation constraint loss is constructed based on the trend relaxation coefficient and the slow trend component; and a reconstruction loss and a component supervision loss are constructed based on the reconstructed PPG time series signal and the oscillation component, respectively. The reconstruction loss, component supervision loss, discrete oscillatory ordinary differential equation constraint loss, and slow trend ordinary differential equation constraint loss are weighted and combined. The two-stream reconstruction network is trained based on the combined total loss, and the trained two-stream reconstruction network is used for signal reconstruction.

Images