PPG signal denoising system and method based on multi-task noise modeling
By employing a multi-task noise modeling approach, utilizing multi-layer convolutional networks and bidirectional gated recurrent units, combined with self-supervised training, the denoising problem of PPG signals under multimodal noise interference was solved, achieving high-precision, low-computation denoising results and improving the robustness and accuracy of the signal.
Patent Information
- Application Number
- CN202511828626.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-05
- Publication Date
- 2026-03-03
AI Technical Summary
Existing technologies struggle to effectively process PPG signals under multimodal and non-stationary noise interference, leading to reduced signal-to-noise ratio, waveform distortion, and physiological characteristic deviations, thus affecting the accuracy of clinical monitoring.
A multi-task noise modeling method is adopted, which extracts features through a multi-layer one-dimensional convolutional network, combines parameter prediction of white noise, ramp noise and saturation noise, uses causal dilated convolution and bidirectional gated recurrent units for denoising, and combines self-supervised training and dynamic environment verification to generate high-precision denoised signals.
Achieving high-precision, low-computation denoising in multimodal noise environments enhances the robustness and generalization ability of the model, reduces dependence on clean labeled data, and improves the structural fidelity and robustness of the signal.
Smart Images
Figure CN121597977A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of signal processing and intelligent monitoring technology, specifically to a PPG signal denoising system and method based on multi-task noise modeling, which is applicable to highly robust physiological signal acquisition and analysis in clinical monitoring and wearable medical devices. Background Technology
[0002] In clinical monitoring and wearable devices, photoplethysmography (PPG) signals are widely used to monitor physiological parameters such as heart rate, blood oxygen saturation, and blood flow. However, during actual acquisition, PPG signals are often interfered with by multi-source, multi-modal noise, mainly including thermal noise from photodetectors and white noise caused by ambient light disturbances, ramp noise caused by slow changes in the sensor's position relative to the skin (such as body movement and posture adjustments), and saturation noise caused by strong ambient light or sudden changes in skin transmittance. In clinical monitoring and dynamic motion environments, these noises often coexist in a multi-modal coupling form and exhibit non-stationary characteristics, severely reducing the signal-to-noise ratio (SNR), leading to waveform distortion and physiological feature shifts, and consequently affecting the accuracy of feature extraction, pattern recognition, and parameter estimation.
[0003] Traditional denoising methods (such as bandpass filtering, wavelet transform, empirical mode decomposition, and adaptive filtering) are typically based on a single noise model (e.g., the Gaussian noise assumption). While they perform stably when dealing with a single type of noise, their performance degrades significantly when faced with multimodal noise coupling and non-stationary interference. Furthermore, these methods rely on manually setting filter parameters or selecting basis functions, lacking the ability to explicitly model the time-series dependence of PPG signals and the noise generation mechanism. This makes it difficult to effectively suppress motion artifacts and saturation distortion, leading to waveform distortion and physiological characteristic shifts under motion artifacts and saturation noise interference.
[0004] In recent years, deep learning methods (such as denoising autoencoders and convolutional-recurrent networks) have improved generalization ability to some extent by implicitly modeling noise distribution, but most of them rely on large-scale "clean-noisy" paired training data. However, in real clinical scenarios, it is difficult to obtain a clean PPG reference waveform synchronized with the noisy signal because it is impossible to maintain the patient's static state for a long time and the measurement conditions are not fully controllable, thus limiting the practicality of these methods.
[0005] Some self-supervised learning methods (such as masked modeling with random masking) attempt to reduce reliance on labeled data by hiding parts of the signal and using contextual inference for reconstruction. However, these methods inherently rely on statistical context to fill in missing information and lack explicit constraints on the physical generation mechanism of noise and the characteristics of multimodal distribution. In noisy environments, this method is prone to generating spurious peaks or distorted waveforms that do not conform to physiological laws, affecting clinical usability. Some self-supervised learning schemes (such as masked modeling) reduce reliance on labels by randomly masking clean signal segments and training the network to fill in the missing parts. However, they are essentially context-based inference reconstruction and lack explicit constraints on the physical distribution of the original noise. In noisy multimodal environments, they are prone to generating spurious peaks or distorted waveforms that do not conform to physiological laws. Summary of the Invention
[0006] This application provides a PPG signal denoising system and method based on multi-task noise modeling, which can achieve high-precision, low-computation, and robust denoising under multimodal and non-stationary noise interference without relying on pairs of clean labeled data.
[0007] In a first aspect, the system includes: The shared feature extraction module is used to extract features from the input time-series signal through a multi-layer one-dimensional convolutional network and output a multi-channel feature vector F; The noise parameter prediction branch includes a white noise intensity prediction submodule, a ramp noise slope prediction submodule, and a saturation noise localization amplitude prediction submodule. These submodules receive the value of F and predict the white noise variance parameter respectively. , slope of the ramp noise α, saturation noise mask Mask(t) and its saturation amplitude γ; The dynamic noise synthesis module, based on the above α, Mask(t), and γ are used to generate two independent noisy signals using different random seeds. and ; Denoising network, receiving respectively and Multi-scale features are extracted using a causal dilated convolutional encoder, and these features are then input into a bidirectional gated recurrent unit (Bi-GRU) decoder to output the corresponding denoised signal. and ; The self-supervised training module, based on and Calculate the L2 loss between noise pairs ; Generate reference signal based on filtering algorithm Calculate separately and ,as well as and The reference consistency loss is obtained by weighted averaging. ;based on and The convolution kernel weights of the shared feature extraction module, the regression layer parameters of the noise parameter prediction branch, and the gating unit parameters of the Bi-GRU decoder are jointly optimized.
[0008] In conjunction with the first aspect, in one embodiment, the system further includes: The dynamic environment verification module verifies the denoised signal by simulating motion artifacts and ambient light fluctuations. and Robustness; the motion artifacts include ramp noise caused by body movement or posture changes. The ambient light fluctuations include white noise caused by changes in ambient illuminance or light source flicker. .
[0009] In conjunction with the first aspect, in one implementation, the 1D convolutional network of the shared feature extraction module comprises L layers, with each layer having a kernel size of... Step length and number of channels The values are incremented according to preset rules to simultaneously cover multi-scale features ranging from local peaks and troughs to global trends.
[0010] In conjunction with the first aspect, in one implementation, the white noise variance parameter Channel-level energy features are extracted using global average pooling, and the variance of the output is estimated by combining it with a fully connected regression layer. The slope noise slope α is extracted by extracting low-frequency trends through temporal convolution with a large receptive field, and then the linear slope is output by the regression layer. The saturation noise mask Mask(t) and its saturation amplitude γ are located using local convolution and Gumbel-Softmax, and the saturation amplitude is regressed by combining the amplitude features of the mask region.
[0011] In conjunction with the first aspect, in one implementation, the slope noise slope prediction submodule extracts low-frequency trends through causal dilation convolution with increasing void ratio, where the void ratio... satisfy ,in For network depth.
[0012] In conjunction with the first aspect, in one embodiment, the dynamic noise synthesis module includes: The noise superposition unit generates white noise. Slope noise and saturation noise Linear superposition generates a composite noise signal. Two sets of independent noisy signals are generated by combining different random seeds. and The calculation formula is as follows: ; in For the initial estimated signal, Mask(t) is a continuous mask for white noise components sampled from a standard normal distribution N(0,1), where Mask(t)∈{0,1} or Mask(t)∈[0,1]. The and As input to the denoising network, the corresponding denoised signal is output through the model. and .
[0013] In conjunction with the first aspect, the hidden state fusion formula of the Bi-GRU decoder is: ; in , For trainable parameters, This indicates a splicing operation. This indicates that the forward GRU is at time point t The hidden state, Indicates the reverse GRU at time point t The hidden state.
[0014] In conjunction with the first aspect, in one implementation, the self-supervised training module further includes a local statistical constraint unit, which divides the time-series signal into sliding windows of length T with a fixed sliding step size S, and calculates the variance of each window. and global variance Comparison, through local variance consistency constraints, prevents overfitting or distortion of generated noise in local regions; differences The sum of the squared errors of the window variance and the global variance is calculated using the following formula:
[0015] Where N is the number of windows, T is the window length, and S is the fixed step size, in units of sampling points or time units.
[0016] Secondly, embodiments of this application provide a PPG signal denoising method based on multi-task noise modeling, the method comprising: Feature extraction is performed on the input time-series signal through a multi-layer one-dimensional convolutional network, and a multi-channel feature vector F is output. Receive the F and predict the white noise variance parameter respectively. , slope of the ramp noise α, saturation noise mask Mask(t) and its saturation amplitude γ; Based on the above α, Mask(t), and γ are used to generate two independent noise signals using different random seeds. and ; Receive separately and Multi-scale features are extracted using a causal dilated convolutional encoder, and these features are input into a Bi-GRU decoder to output the corresponding denoised signal. and ; according to and Calculate the L2 loss between noise pairs ; Generate reference signal based on filtering algorithm Calculate separately and ,as well as and The reference consistency loss is obtained by weighted averaging. ;based on and The convolutional kernel weights of the shared feature extraction module, the regression layer parameters of the noise parameter prediction branch, and the gating unit parameters of the Bi-GRU decoder are jointly optimized.
[0017] In conjunction with the second aspect, in one embodiment, the method further includes: The denoised signal was verified by simulating motion artifacts and ambient light fluctuations. and Robustness; the motion artifacts include ramp noise caused by body movement or posture changes. The ambient light fluctuations include white noise caused by changes in ambient illuminance or light source flicker. .
[0018] The beneficial effects of the technical solutions provided in this application include: Multi-task collaborative modeling: Simultaneously predicting white noise variance parameters The slope of the ramp noise α, the saturation noise mask Mask(t) and its saturation amplitude γ are used to reduce redundant calculations while ensuring the accuracy of noise parameter estimation, thus significantly improving the denoising efficiency.
[0019] Dynamic noise synthesis enhances robustness: Training is performed using two sets of independent composite noise signals generated based on different random seeds, which effectively avoids the denoising network from overfitting to a single noise pattern and enhances the model's adaptability to various noise environments.
[0020] Bidirectional temporal feature fusion: The Bi-GRU decoder achieves bidirectional temporal modeling by splicing the hidden states of the forward and backward GRUs. Combined with the multi-scale feature extraction capability of causal dilated convolution, it takes into account both local waveform details and long-range periodic patterns, thereby significantly improving the structural fidelity of the denoised signal.
[0021] Self-supervised multi-loss collaborative constraints: applying noise to losses Loss of consistency with reference The joint optimization not only prevents the model from collapsing into an identity mapping, but also takes into account the consistency between global trends and local forms, effectively improving the model's generalization ability and robustness. Attached Figure Description
[0022] Figure 1 This is a functional module diagram of an embodiment of the PPG signal denoising system based on multi-task noise modeling of this application.
[0023] Figure 2 This is a flowchart illustrating an embodiment of the PPG signal denoising method based on multi-task noise modeling of this application. Detailed Implementation
[0024] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application. First, some of the technical terms used in this application will be explained to help those skilled in the art understand this application.
[0025] Time-series signals: encompass all types of time-domain signals (such as PPG, ECG, EDA, etc.). Time-series signals include the original noisy signal, which refers to the signal in a specific state. For example, a time-series signal is a PPG signal, which includes ramp noise caused by motion artifacts, baseline drift, and high-frequency white noise.
[0026] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.
[0027] In a first aspect, embodiments of this application provide a PPG signal denoising system based on multi-task noise modeling (see...). Figure 1 ).
[0028] The aforementioned denoising system includes: a shared feature extraction module 1, a noise parameter prediction branch 2, a dynamic noise synthesis module 3, a denoising network 4, and a self-supervised training module 5.
[0029] Shared feature extraction module 1: Extracts features from the input time-series signal y through a multi-layer 1D convolutional network and outputs a multi-channel feature vector F=[F1,F2,..,FC], where C is the number of channels.
[0030] Noise parameter prediction branch 2: Receive F, and predict the white noise variance parameter respectively. The slope of the ramp noise α, the saturation noise mask Mask(t), and its saturation amplitude γ.
[0031] Dynamic noise synthesis module 3: under the same distribution parameters { Under the conditions of α, Mask(t) and γ, two independent noisy signals are generated by combining different random seeds. and .
[0032] Denoising Network 4: Receives separately and Multi-scale features are extracted using a causal dilated convolutional encoder, and these features are then input into a bidirectional gated recurrent unit decoder to output a denoised signal. and The Bi-Gated Cyclic Unit Decoder (Bi-GRU) recovers the waveform structure by splicing forward and reverse hidden states and performing linear mapping.
[0033] Self-supervised training module 5: Calculate the noise-to-loss ratio (symmetric form): ; Reference signal generated based on filtering algorithm Calculate the reference consistency loss: .
[0034] Total loss function: ; The parameters of the shared feature extraction module 1, the noise parameter prediction branch 2, and the Bi-GRU decoder are jointly optimized according to preset weights.
[0035] In this embodiment, a general temporal feature F is extracted from the temporal signal (i.e., the original noisy signal), and noise parameters are predicted based on F to drive dynamic noise synthesis. The synthesized noise and the original signal jointly participate in the training of the denoising network 4 to optimize the model parameters through self-supervised loss (…). + Further reverse optimization of the parameters of shared feature extraction module 1.
[0036] During training, two independent noisy signals are generated from the predicted noise parameters. These signals, along with the original noisy signal, serve as input to the denoising network, thereby achieving multi-task collaborative modeling. Simultaneously predict white noise, ramp noise, and saturation noise parameters to reduce redundant calculations and improve denoising efficiency.
[0037] Dynamic noise synthesis: Two sets of independent noisy signals are used to enhance training robustness and avoid model overfitting to a single noise pattern.
[0038] Bidirectional temporal modeling: Combining causal dilated convolution with Bi-GRU takes into account both local morphology and long-range periodicity, significantly improving signal fidelity.
[0039] Self-supervised joint optimization: through and At the same time, it constrains the consistency between the model output and the reference signal, avoids collapse into an identity mapping, and improves generalization ability.
[0040] In one specific embodiment, firstly, a multi-branch noise parameter estimation network is used to regress the shared features to obtain the white noise variance. The parameters include the slope α of the ramp noise and the interval location and amplitude γ of the saturation noise. Based on these parameters and different random seeds, two sets of noisy signals with the same distribution but different random implementations are generated. and .
[0041] To ensure that the generated samples are consistent with the original noisy signal in terms of local statistical characteristics, a patch-wise variance constraint is introduced to limit the deviation of the global and local variances of the generated signal from each window of the original noise signal to not exceed a set threshold.
[0042] Next, respectively and The signal is input into denoising network 4 to obtain the denoised signal. and Since both correspond to the same potential clean signal (sharing the potential "gold standard"), a noise-to-noise loss is constructed: ; This loss constraint ensures the consistency of the mapping between different noisy instances, thereby enabling self-supervised training under conditions without clean labels.
[0043] To prevent the model from collapsing into an identity mapping (output equals input), a reference consistency loss is introduced in addition to the noise pair loss. Specifically, the reference signal is obtained by applying traditional methods (such as wavelet denoising or adaptive filtering) to the original noisy signal y. And calculate: ; in It is used only for loss calculation and is not used as further input to the network. This "double loss" ( + Collaborative constraints not only improve model robustness, but also maintain physiological consistency of the denoised waveform at both global and local scales.
[0044] Furthermore, in one embodiment, the system also includes a dynamic environment verification module 6, used to verify the robustness of the denoising results by simulating motion artifacts and ambient light fluctuations / breathing modulation. The motion artifacts employ slow-trend / slope noise. The respiratory effect was modeled as low-frequency periodic modulation. Random ambient light disturbances are modeled as white noise. .
[0045] Motion artifacts are signal distortions caused by changes in body movement or posture during the testing process.
[0046] Slow drift: refers to the trend of signal intensity changing slowly over time (e.g., the gradual change in signal caused by a patient moving slowly or breathing slowly).
[0047] Slope noise refers to the linear intensity variation pattern in an image (i.e., the artifact where the signal slopes from one side to the other in the image).
[0048] The slow trend is the cause, and the slope noise is the result; they describe different aspects of the same type of artifact.
[0049] This embodiment compares the denoised signals using a sliding window. and The difference between local and global variance is used to assess stability.
[0050] Furthermore, in one embodiment, the 1D convolutional network of the shared feature extraction module 1 comprises L layers, with each layer having a convolutional kernel size of... Step length and number of channels The values increment according to preset rules to cover multi-scale features ranging from local peaks and troughs to global trends. An example setting is acceptable. Or increase appropriately with the number of floors. void ratio Set separately when modeling slope trends.
[0051] Shared convolutional encoders extract redundant low-level features (periodicity, amplitude trends, edge changes, etc.) from the same PPG segment in a unified manner, which can significantly reduce parameters and computational load, and improve stability and generalization through implicit regularization of multi-task learning; at the same time, it helps to model the coupling relationship of different noise components (such as saturation + drift caused by motion artifacts).
[0052] In this embodiment, the white noise intensity The estimates include: The original noisy PPG is processed by a shared feature extraction module to obtain global temporal features, which are then transformed into a statistical vector F by global average pooling. This vector F is then input into a fully connected regression layer to obtain a nonnegative scalar. This parameter is used to generate white noise components. The dynamic noise synthesis module 3 is used to construct composite noisy signals.
[0053] The slope noise slope α estimation includes: Similarly, after extracting time-series features F using a shared encoder, a set of causal convolutions or dilated convolutions is introduced to enhance the modeling ability for low-frequency trends. The extracted trend features undergo a series of temporal compression operations (such as 1D average pooling or stride convolution) to suppress high-frequency components and preserve the smooth trend of the signal. Finally, a real scalar α is output through the regression branch (FC layer) as the slope simulating linear drift or baseline drift trends. Slope noise can be expressed by the formula... Generate and simulate low-frequency disturbances such as changes in skin temperature and light shifts.
[0054] The estimation of saturation noise amplitude and interval location γ includes: First, a probability sequence P(t) aligned with the time axis is generated based on shared convolutional features, representing the probability of being in a saturated state at each time point. To achieve differentiable modeling of the saturation interval boundaries during training, Gumbel-Softmax reparameterization is used to discretize P(t), thereby obtaining a learnable saturation interval mask while preserving gradient propagation. Simultaneously, another network branch predicts the saturation amplitude parameter γ corresponding to these saturation intervals, and ensures its non-negativity and numerical stability through ReLU or Sigmoid activation functions. Finally, the saturation noise is generated by the following equation: ,in The saturation mask obtained by Gumbel–Softmax sampling.
[0055] Furthermore, in one embodiment, the slope noise slope prediction submodule extracts low-frequency trends through causal dilated convolution with increasing void ratio, where void ratio... satisfy ,in The number of network layers is represented by the exponentially increasing hole rate, which significantly improves the ability to capture low-frequency trends while avoiding the computational overhead of simply stacking more convolutional layers. This reduces complexity while improving the accuracy of slope noise estimation.
[0056] Furthermore, in one embodiment, the dynamic noise synthesis module 3 includes a noise superposition unit for superimposing the predicted white noise. Slope noise and saturation noise Linear superposition generates a composite noise signal. The calculation formula is as follows: ; in For the initial estimated signal, Mask(t) is a white noise component sampled from a standard normal distribution N(0,1), where Mask(t) ∈ {0,1} (or a continuous mask of [0,1]). The As input to denoising network 4, the corresponding denoised signal is output by the model. By superimposing white noise, ramp noise, and saturation noise, the mixed noise distribution in real PPG signals can be realistically simulated, making the training data closer to actual application scenarios. This noise synthesis process is differentiable, allowing end-to-end optimization. The gradient can be directly backpropagated to the noise parameter prediction branch 2, thereby jointly optimizing the noise prediction and denoising tasks, avoiding parameter fragmentation between modules, and significantly improving training efficiency.
[0057] In one specific embodiment, two sets of independent noise signals are generated. and These methods rely on the same set of noise parameters but use different random seeds for sampling. This introduces statistical differences while maintaining the consistency of noise patterns, forcing the model to focus on the commonalities of noise features rather than randomness, thus reducing the risk of overfitting. During training, a contrastive loss is incorporated. (based on and and reference signals generated based on traditional filtering .
[0058] In one specific embodiment, As input to denoising network 4, the model outputs the denoising result. and .according to and Calculate the L2 loss between noise pairs ; Generate reference signal based on filtering algorithm Calculate separately and ,as well as and The reference consistency loss is obtained by weighted averaging. ;based on and The convolution kernel weights of the shared feature extraction module, the regression layer parameters of the noise parameter prediction branch, and the gating unit parameters of the Bi-GRU decoder are jointly optimized according to preset weights.
[0059] Furthermore, in one embodiment, the dynamic environment verification module 6 simulates the interference of mixed noise (white noise, ramp noise, saturation noise) on the PPG signal using composite noise signals to verify the denoised signal. Robustness in dynamic environments. The verification method is to calculate the local variance of the denoised signal within a sliding window and compare it with the global variance to ensure that there is no overfitting or structural distortion in key waveform regions such as peaks and troughs.
[0060] Furthermore, in one embodiment, the hidden state fusion formula of the Bi-GRU decoder is: , in , For trainable parameters, This represents a vector concatenation operation. and These represent the hidden states of the forward and backward GRUs at time step t, respectively.
[0061] In this embodiment, the Bi-GRU simultaneously captures both forward and backward timing dependencies, enhancing its ability to recover the periodic and symmetrical characteristics of the PPG signal. By splicing together to preserve bidirectional timing information, the risk of waveform distortion caused by unidirectional modeling can be effectively reduced.
[0062] Furthermore, in one embodiment, the self-supervised training module 5 further includes: Local statistical constraint unit: Divide the time series signal y into sliding windows of length T with a fixed step size S, and calculate the variance of each window. and the global variance of the time series signal y By comparing and using local variance consistency constraints, overfitting or distortion of generated noise in local regions (such as peaks and troughs) is prevented. Differences The sum of the squared errors of the window variance and the global variance is calculated using the following formula: ; Where N is the number of windows, T is the window length, and S is the fixed step size, in units of sampling points or time units, such as milliseconds.
[0063] In this embodiment, in order to ensure that the local statistical characteristics of the generated noise signal are consistent with those of the original noise signal, we adopt a patch-wise variance constraint: comparing the variance within the window with the global variance to ensure that the local fluctuation intensity of the generated sample matches the original data, which serves as an auxiliary loss term for training, improving the realism of the synthesized sample and suppressing the distortion of the denoising result in key areas.
[0064] Local statistical characteristics refer to the low-order statistical features of the noise signal in each segment after the noisy PPG signal is divided into several small segments (Patch) in the time dimension. They are used to describe the degree of noise disturbance and structural characteristics in the local area.
[0065] Local variance describes the amplitude fluctuation of a signal within a certain time window. It is a direct measure of noise intensity and reflects the local noise energy.
[0066] Local mean drift describes whether the local signal baseline has shifted, and is mainly used to limit slope noise.
[0067] Local autocorrelation describes whether noise has a certain sequence dependency within a local window.
[0068] A local region refers to a patch in which the PPG signal is divided into fixed-length time windows (such as 64 or 128 sampling points) to measure local disturbance characteristics. This is to ensure the generation of noise signals. and Not only does it approximate the original noise signal in terms of global distribution, but it also maintains consistency in local perturbation intensity. This invention designs a patch-wise variance constraint mechanism. This mechanism ensures that the statistical characteristics (such as fluctuation intensity) of the generated noise in local regions are consistent with the original noise by comparing the variance within a window with the global variance. This prevents the denoised signal from becoming distorted due to overfitting or underfitting in key areas such as peaks / troughs. It serves as an auxiliary loss term to guide the generator in generating more realistic and structurally consistent noise samples during the training phase. By constraining global and local variance, the model's adaptability to different noise intensities is enhanced.
[0069] Secondly, embodiments of this application also provide a PPG signal denoising method based on multi-task noise modeling.
[0070] In one embodiment, reference is made to Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the PPG signal denoising method based on multi-task noise modeling according to this application. Figure 2 As shown, the PPG signal denoising method based on multi-task noise modeling includes: Step S1: Extract features from the input time-series signal using a multi-layer one-dimensional convolutional network, and output a multi-channel feature vector F=[F1,F2,..,FC], where C is the number of channels.
[0071] Step S2: Receive F and predict the white noise variance parameter respectively. The slope of the ramp noise α, the saturation noise mask Mask(t), and its saturation amplitude γ.
[0072] Step S3, with the same distribution parameter { α, Mask(t), and γ are used to generate two independent noisy signals using different random seeds. and .
[0073] Step S4: Separately and The input to the denoising network is processed by a causal dilated convolutional encoder to extract multi-scale features, which are then input into a Bi-GRU decoder to output the corresponding denoised signal. and Bi-GRU recovers the waveform structure by splicing forward and reverse hidden states.
[0074] Step S5: Calculate the self-supervised loss and jointly optimize the parameters: Noise on loss (Symmetrical form) , Reference consistency loss : Generate reference signal from y based on filtering algorithm ,calculate: , (Optional) Add local statistical constraints , , The convolutional kernel weights for shared feature extraction, the regression layer parameters for the noise parameter prediction branch, and the gating unit parameters of the Bi-GRU decoder are jointly optimized.
[0075] Step S6: Verify the denoised signal by simulating motion artifacts and ambient light fluctuations. and Robustness; the motion artifacts include ramp noise caused by body movement or posture changes. The ambient light fluctuations include white noise caused by changes in ambient illuminance or light source flicker. .
[0076] It should be noted that the sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.
[0077] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.
[0078] In the description of the embodiments of this application, terms such as "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a concrete manner.
[0079] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.
[0080] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.
[0081] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of this application.
[0082] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.
Claims
1. A PPG signal denoising system based on multi-task noise modeling, characterized in that, The system includes: The shared feature extraction module is used to extract features from the input time-series signal through a multi-layer one-dimensional convolutional network and output a multi-channel feature vector F; The noise parameter prediction branch includes a white noise intensity prediction submodule, a ramp noise slope prediction submodule, and a saturation noise localization amplitude prediction submodule. These submodules receive the value of F and predict the white noise variance parameter respectively. , slope of the ramp noise α, saturation noise mask Mask(t) and its saturation amplitude γ; The dynamic noise synthesis module, based on the above α, Mask(t), and γ are used to generate two independent noisy signals using different random seeds. and ; Denoising network, receiving respectively and Multi-scale features are extracted using a causal dilated convolutional encoder, and these features are then input into a bidirectional gated recurrent unit (Bi-GRU) decoder to output the corresponding denoised signal. and ; The self-supervised training module, based on and Calculate the L2 loss between noise pairs ; Generate reference signal based on filtering algorithm Calculate separately and ,as well as and The reference consistency loss is obtained by weighted averaging. ;based on and The convolution kernel weights of the shared feature extraction module, the regression layer parameters of the noise parameter prediction branch, and the gating unit parameters of the Bi-GRU decoder are jointly optimized.
2. The PPG signal denoising system based on multi-task noise modeling as described in claim 1, characterized in that, The system also includes: The dynamic environment verification module verifies the denoised signal by simulating motion artifacts and ambient light fluctuations. and Robustness; the motion artifacts include ramp noise caused by body movement or posture changes. The ambient light fluctuations include white noise caused by changes in ambient illuminance or light source flicker. .
3. The PPG signal denoising system based on multi-task noise modeling as described in claim 1, characterized in that, The 1D convolutional network of the shared feature extraction module contains L layers, with each layer having a kernel size of [missing information]. Step length and number of channels The values are incremented according to preset rules to simultaneously cover multi-scale features ranging from local peaks and troughs to global trends.
4. The PPG signal denoising system based on multi-task noise modeling as described in claim 1, characterized in that, The white noise variance parameter Channel-level energy features are extracted using global average pooling, and the variance of the output is estimated by combining it with a fully connected regression layer. The slope noise slope α is extracted by extracting low-frequency trends through temporal convolution with a large receptive field, and then the linear slope is output by the regression layer. The saturation noise mask Mask(t) and its saturation amplitude γ are located using local convolution and Gumbel-Softmax, and the saturation amplitude is regressed by combining the amplitude features of the mask region.
5. The PPG signal denoising system based on multi-task noise modeling as described in claim 1, characterized in that, The slope noise slope prediction submodule extracts low-frequency trends through causal dilation convolution with increasing void ratio. satisfy ,in For network depth.
6. The PPG signal denoising system based on multi-task noise modeling as described in claim 1, characterized in that, The dynamic noise synthesis module includes: The noise superposition unit generates white noise. Slope noise and saturation noise Linear superposition generates a composite noise signal. Two sets of independent noisy signals are generated by combining different random seeds. and The calculation formula is as follows: ； in For the initial estimated signal, Mask(t) is a continuous mask for white noise components sampled from a standard normal distribution N(0,1), where Mask(t)∈{0,1} or Mask(t)∈[0,1]. The and As input to the denoising network, the corresponding denoised signal is output through the model. and .
7. The PPG signal denoising system based on multi-task noise modeling as described in claim 1, characterized in that, The hidden state fusion formula of the Bi-GRU decoder is: ； in , For trainable parameters, This indicates a splicing operation. This indicates that the forward GRU is at time point t The hidden state, Indicates the reverse GRU at time point t The hidden state.
8. The PPG signal denoising system based on multi-task noise modeling as described in claim 1, characterized in that, The self-supervised training module also includes a local statistical constraint unit, which divides the time-series signal into sliding windows of length T with a fixed sliding step size S, and calculates the variance of each window. and global variance Comparison, through local variance consistency constraints, prevents overfitting or distortion of generated noise in local regions; differences The sum of the squared errors of the window variance and the global variance is calculated using the following formula: ； Where N is the number of windows, T is the window length, and S is the fixed step size, in units of sampling points or time units.
9. A PPG signal denoising method based on multi-task noise modeling, characterized in that, The method includes: Feature extraction is performed on the input time-series signal through a multi-layer one-dimensional convolutional network, and a multi-channel feature vector F is output. Receive the F and predict the white noise variance parameter respectively. , slope of the ramp noise α, saturation noise mask Mask(t) and its saturation amplitude γ; Based on the above α, Mask(t), and γ are used to generate two independent noise signals using different random seeds. and ; Receive separately and Multi-scale features are extracted using a causal dilated convolutional encoder, and these features are input into a Bi-GRU decoder to output the corresponding denoised signal. and ; according to and Calculate the L2 loss between noise pairs ; Generate reference signal based on filtering algorithm Calculate separately and ,as well as and The reference consistency loss is obtained by weighted averaging. ;based on and The convolutional kernel weights of the shared feature extraction module, the regression layer parameters of the noise parameter prediction branch, and the gating unit parameters of the Bi-GRU decoder are jointly optimized.
10. The PPG signal denoising method based on multi-task noise modeling as described in claim 9, characterized in that, The method further includes: The denoised signal was verified by simulating motion artifacts and ambient light fluctuations. and Robustness; the motion artifacts include ramp noise caused by body movement or posture changes. The ambient light fluctuations include white noise caused by changes in ambient illuminance or light source flicker. .