PPG denoising method and system based on sensor topology and noise structure

Abstract

The application provides a PPG denoising method and system based on a sensor topology and a noise structure, the method comprising: collecting a PPG signal, and calculating the PPG signal by using an additive noise model; constructing an undirected graph according to a physical topology structure of a sensor array by using a signal collection and graph construction module, and processing node features on the undirected graph by using a cascaded graph convolution network; performing feature aggregation and time sequence modeling by using a stacked GRU network in a space-time feature extraction module; decoding and predicting the GRU network output by using a noise prediction module; inputting a noisy PPG signal into an optimized denoising model, predicting and removing noise, and obtaining a final reconstructed PPG signal. The application can effectively recover a high-quality PPG signal from strong noise, and provides a reliable guarantee for accurate physiological parameter monitoring in a dynamic scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biomedical signal processing and artificial intelligence, and particularly relates to a PPG denoising method and system based on sensor topology and noise structure. BACKGROUND

[0002] The photoelectric plethysmogram (PPG) technology has been widely used in continuous monitoring of physiological indicators such as heart rate and blood oxygen saturation due to its non-invasive and low-cost characteristics, and is the core technology of wearable health devices. However, the PPG signal is extremely weak and is easily severely disturbed by motion artifacts (MA) in the user's daily activities. The MA and the real heart rate signal are highly overlapped in the frequency domain, which makes it a challenging task to effectively separate the two, and is also the main technical bottleneck restricting the accurate application of PPG technology in dynamic scenes.

[0003] Existing PPG denoising methods can be mainly divided into the following two types:

[0004] (1) Traditional methods (such as adaptive filtering, wavelet transform, etc.): usually rely on strong prior assumptions and are not robust enough when dealing with complex and variable motion scenes.

[0005] (2) Deep learning-based models: although they can automatically learn noise patterns, most existing models have a common technical defect when applied to multi-channel sensor arrays: they tend to treat each sensor channel as an independent signal sequence, completely ignoring the inherent spatial correlation of noise generated by the same motion source on different sensors. This spatial structure information is strong prior knowledge for distinguishing motion artifacts from physiological signals, and its neglect limits the upper limit of the denoising performance of existing models.

[0006] Therefore, the prior art has the following technical defects:

[0007] 1. Existing deep learning methods fail to effectively utilize the physical topology of the multi-channel sensor array and the statistical characteristics of noise in the spatial dimension, resulting in insufficient recognition ability for motion artifacts with strong spatial correlation.

[0008] 2. Most use point-by-point error loss functions, which only focus on the similarity in the time domain amplitude and cannot effectively constrain the internal structure of the signal, resulting in the model generating noise with similar amplitudes but incorrect spatial structures, ultimately affecting the fidelity of signal reconstruction. SUMMARY

[0009] In view of the above situation, the main purpose of the present application is to provide a PPG denoising method and system based on sensor topology and noise structure to solve the above technical problems.

[0010] The application provides a PPG denoising method and system based on sensor topology and noise structure.

[0011] Step 1, a denoising model is constituted based on a signal acquisition and graph construction module, a space-time feature extraction module, a noise prediction module, and a training optimization module.

[0012] Step 2, PPG signals are acquired, an additive noise model is used to calculate the PPG signals to generate real noise signals, a signal acquisition and graph construction module is used to construct an undirected graph according to the physical topology structure of a sensor array, and a noisy PPG signal in the PPG signals is expressed as a node feature on the undirected graph, and a cascaded graph convolution network is used to process the node feature on the undirected graph to obtain a space-time feature of noise.

[0013] Step 3, the space-time feature of noise is subjected to feature aggregation and time sequence modeling by a stacked GRU network in the space-time feature extraction module to obtain a GRU network output.

[0014] Step 4, the GRU network output is subjected to decoding prediction by a noise prediction module to obtain a predicted noise signal.

[0015] Step 5, a difference between a covariance matrix of the predicted noise signal and a covariance matrix of the real noise signal is calculated and compared by a training optimization module to construct a space structure loss, a point-by-point difference between the predicted noise signal and the real noise signal is calculated to construct a point-by-point accuracy loss, the space structure loss and the point-by-point accuracy loss are combined by weighting to form a composite loss function, the denoising model is guided and optimized by using the composite loss function, and an optimized denoising model is obtained, the noisy PPG signal is input into the optimized denoising model, the noise is predicted and removed to obtain a final reconstructed PPG signal.

[0016] The application further provides a PPG denoising system based on sensor topology and noise structure, and the system comprises:

[0017] The construction module is used for:

[0018] The denoising model is constituted based on the signal acquisition and graph construction module, the space-time feature extraction module, the noise prediction module, and the training optimization module.

[0019] The feature extraction module is used for:

[0020] PPG signals are acquired, an additive noise model is used to calculate the PPG signals to generate real noise signals, a signal acquisition and graph construction module is used to construct an undirected graph according to the physical topology structure of a sensor array, and a noisy PPG signal in the PPG signals is expressed as a node feature on the undirected graph, and a cascaded graph convolution network is used to process the node feature on the undirected graph to obtain a space-time feature of noise.

[0021] The spatial features of the noise are aggregated and time-series modeled by a stacked GRU network in the space-time feature extraction module to obtain a GRU network output;

[0022] The prediction module is used for:

[0023] The GRU network output is input into the noise prediction module for decoding prediction to obtain a predicted noise signal;

[0024] The reconstruction module is used for:

[0025] The training optimization module is used for calculating the difference between the covariance matrix of the predicted noise signal and the covariance matrix of the real noise signal to construct a spatial structure loss; calculating the point-by-point difference between the predicted noise signal and the real noise signal to construct a point-by-point accuracy loss; and combining the spatial structure loss and the point-by-point accuracy loss by weighting to form a composite loss function to guide the optimization of the noise reduction model to obtain an optimized noise reduction model; inputting the noisy PPG signal into the optimized noise reduction model to predict and remove the noise to obtain a final reconstructed PPG signal.

[0026] Compared with the prior art, the present application has the following advantages:

[0027] 1. The present application encodes the sensor physical topology into a graph structure, and effectively extracts the space-time features of the motion artifacts by using the space-time feature extraction module framework, solving the defect that the existing method ignores the spatial information;

[0028] 2. The present application can constrain the learning process of the model from a higher dimension (covariance) by proposing a "noise spatial structure consistency loss function", significantly improving the reconstruction ability of the noise spatial structure and the performance of the traditional point-by-point error measurement.

[0029] 3. The present application can effectively recover high-quality PPG signals from strong noise, providing reliable protection for accurate physiological parameter monitoring in dynamic scenes.

[0030] Additional aspects and advantages of the present application will be partially given in the following description, partially become obvious from the following description, or be understood by the embodiments of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 The step flow chart of the PPG noise reduction method based on sensor topology and noise structure proposed by the present application.

[0032] Figure 2 The overall architecture schematic diagram of the cascaded space-time graph convolution noise reduction model of the PPG noise reduction method based on sensor topology and noise structure proposed by the present application.

[0033] Figure 3 The data acquisition to the intelligent noise reduction flowchart of the PPG noise reduction method based on sensor topology and noise structure proposed in the application.

[0034] Figure 4 The training sample construction process schematic diagram of the PPG noise reduction method based on sensor topology and noise structure proposed in the application.

[0035] Figure 5 The PPG noise reduction system architecture diagram based on sensor topology and noise structure proposed in the application. DETAILED DESCRIPTION

[0036] Embodiments of the application are described below in detail, examples of which are shown in the drawings, wherein the same or similar reference numbers are the same or similar elements or elements with the same or similar functions throughout. The embodiments described below by referring to the drawings are exemplary, only for explaining the application, and cannot be understood as limiting the application.

[0037] These and other aspects of embodiments of the application will be apparent from the following description and drawings. In these descriptions and drawings, some specific embodiments of the embodiments of the application are specifically disclosed to illustrate the principles of implementing the embodiments of the application, but it should be understood that the scope of the embodiments of the application is not limited thereto.

[0038] Referring to Figure 1 , the application proposes a PPG noise reduction method based on sensor topology and noise structure, which comprises the following steps:

[0039] Step 1, based on the signal acquisition and graph construction module, the space-time feature extraction module, the noise prediction module, and the training optimization module to constitute the noise reduction model.

[0040] Step 2, collect PPG signals, calculate PPG signals using additive noise model to generate real noise signals; construct an undirected graph according to the physical topology structure of the sensor array through the signal acquisition and graph construction module, and express the noisy PPG signal in the PPG signal as a node feature on the undirected graph; use cascaded graph convolution network to process the node feature on the undirected graph to obtain the spatial feature of the noise.

[0041] Referring to Figure 2 , Figure 3 and Figure 4In step 2, the PPG signal is collected, the PPG signal is calculated using an additive noise model to generate a real noise signal; through the signal acquisition and graph construction module, the physical topology structure of the sensor array is constructed into an undirected graph, and the noisy PPG signal in the PPG signal is represented as a node feature on the undirected graph; the node feature on the undirected graph is processed by using a cascaded graph convolution network to obtain the spatial feature of the noise, which specifically includes the following steps:

[0042] Collect the pure PPG signal of the user in the static state and the noisy PPG signal in the motion state;

[0043] Calculate using an additive noise model, subtract the pure PPG signal from the noisy PPG signal to generate a real noise signal;

[0044] The noisy PPG signal and the real noise signal are used as a training sample pair, and the sliding window technology is used to divide into fixed-length signal segments;

[0045] Through the signal acquisition and graph construction module, the physical adjacency relationship of the sensor array is abstracted into an undirected graph, and an original adjacency matrix is constructed according to the undirected graph; self-loop is added to the original adjacency matrix and symmetric normalization processing is performed to obtain a normalized adjacency matrix;

[0046] The fixed-length signal segment and the normalized adjacency matrix are associated, so that the signal slice of each time step in the signal segment is defined as a node feature on the normalized adjacency matrix;

[0047] The node feature on the normalized adjacency matrix is taken as an initial node feature, which is sent into a stacked two-layer graph convolution network for spatial feature extraction to obtain the spatial feature of the noise.

[0048] Calculate using an additive noise model, subtract the pure PPG signal from the noisy PPG signal to generate a real noise signal, and the relationship formula corresponding to the process is as follows:

[0049] ;

[0050] Wherein, The real noise signal is represented by n, The noisy PPG signal is represented by n, The pure PPG signal is represented by n;

[0051] In the step of abstracting the physical adjacency relationship of the sensor array into an undirected graph, the relationship formula corresponding to the process is as follows:

[0052] ;

[0053] Wherein, The undirected graph is represented by G, denotes a set of nodes, denotes a set of edges;

[0054] In the step of adding self-loop to the original adjacency matrix and performing symmetric normalization to obtain the normalized adjacency matrix, the corresponding process exists the following relationship:

[0055]

[0056] wherein, denotes the normalized adjacency matrix, denotes taking each diagonal element of the degree matrix to the power of times of its value, denotes the result after adding the unit matrix with the dimension of the number of sensors to the original adjacency matrix ;

[0057] It should be noted that the normalization operation is one of the core technologies of the graph convolution network in the embodiment, which aims to prevent the gradient of the node feature from disappearing or exploding in multi-layer propagation, and balance the contribution weight of nodes with different connection degrees when information is aggregated. In the attached Figure 3 , ST-GCN is a spatio-temporal graph convolution network.

[0058] In the step of taking the node feature on the normalized adjacency matrix as the initial node feature and sending it into the stacked two-layer graph convolution network to extract the spatial feature of the noise, the corresponding process exists the following relationship:

[0059]

[0060] wherein, denotes the final result of the calculation of the layer graph convolution network, denotes a nonlinear activation function, denotes the initial input feature of the time step graph convolution network, denotes the weight matrix learnable by the initial layer graph convolution network, denotes the final result of the calculation of the layer graph convolution network, denotes the weight matrix learnable by the next layer graph convolution network, denotes the spatial feature of the noise.

[0061] ​​It should be noted that the output of the first layer graph convolutional network will be taken as the input of the second layer. By stacking two layers of graph convolutional network (GCN), the model can not only aggregate the information of directly adjacent sensors (first-order neighbors), but also perceive the information of "neighbors of neighbors" (second-order neighbors), thereby obtaining a more comprehensive perception of the spatial distribution of noise. After two layers of GCN processing, the final node feature matrix at this moment is obtained.

[0062] In step 3, the spatial features of the noise are aggregated and time-series modeled by the stacked GRU network in the spatio-temporal feature extraction module to obtain the GRU network output.

[0063] In step 3, the spatial features of the noise are aggregated and time-series modeled by the stacked GRU network in the spatio-temporal feature extraction module to obtain the GRU network output, which includes the following steps:

[0064] The spatial features of the noise are averaged and pooled to obtain a feature vector representing the global spatial state.

[0065] The feature vector representing the global spatial state is aggregated into a feature sequence.

[0066] The feature sequence is sent to the stacked GRU network in the spatio-temporal feature extraction module to learn the dynamic evolution law of the noise pattern in the time dimension, thereby obtaining the GRU network output.

[0067] The spatial features of the noise are averaged and pooled to obtain a feature vector representing the global spatial state, and the relationship formula corresponding to the process is as follows:

[0068] ;

[0069] wherein, represents the feature vector of the global spatial state, represents the average pooling processing.

[0070] In the step of aggregating the feature vector representing the global spatial state into a feature sequence, the relationship formula corresponding to the process is as follows:

[0071] ;

[0072] wherein, represents the aggregated feature sequence, represents the global spatial state feature vector at the first time step, represents the global spatial state feature vector at the total time step represents the global spatial state feature vector at the total time step represents the global spatial state feature vector at the total time step and the output feature dimension is a real set.

[0073] It should be noted that, All vectors from the first time step (t=1) to the last time step (t=T) are arranged in order.

[0074] Further, in order to accurately capture the long-term dependence and nonlinear dynamics of motion artifacts in the time dimension, the feature sequence is sent into a stacked double-layer gated recurrent unit (GRU) network. The reason for choosing the GRU network is that it can effectively capture the long-term dependence of the signal while having fewer parameters and higher computational efficiency than the traditional LSTM network through its sophisticated gating mechanism, making it more suitable for real-time noise reduction in resource-constrained scenarios such as wearable devices. Two layers of GRU are stacked. Specifically, the feature sequence is first sent into the first layer of GRU network, and the output hidden state sequence will be used as the input of the second layer of GRU network. This stacked structure allows the network to learn deeper and more abstract temporal dependencies. The first layer network may learn some basic and short-term noise fluctuation patterns, while the second layer network can learn how these basic patterns combine into more complex and long-term evolution rules. After processing by the double-layer GRU network, a hidden state sequence is finally output, which is the final encoding representation of the original noise signal's spatial and temporal features, and it deeply integrates the spatial topology and temporal evolution dynamics of the noise. It will be used as the subsequent input to generate accurate noise prediction.

[0075] Step 4, decoding and predicting the GRU network output through the noise prediction module to obtain the predicted noise signal.

[0076] In step 4, decoding and predicting the GRU network output through the noise prediction module to obtain the predicted noise signal, including:

[0077] The hidden state sequence in the GRU network output is input into the noise prediction module to process each time step in the sequence independently, and the hidden state feature vectors of all time steps are extracted. The hidden state feature vectors of all time steps are input into the fully connected layer in the noise prediction module for feature mapping to obtain the specific amplitude of all time steps on the sensor. The specific amplitude of all time steps on the sensor is constructed as a predicted noise matrix, and the predicted noise matrix is used as the predicted noise signal. The corresponding process exists the following relationship:

[0078] ;

[0079] where, represents the hidden state sequence in the GRU network output, represents the total time step and the feature dimension of the GRU hidden layer a set of real numbers, denotes the predicted noise signal, denotes the number of sensors and the total time steps a set of real numbers.

[0080] It is noted that the weights of this fully connected layer are shared across all T time steps. This means that the model uses the same set of “decoding rules” to process features from different time instants.

[0081] Further, the hidden state sequence output by the GRU network is a highly condensed and abstracted encoding of the noise temporal dynamics. In order to convert these abstract features into usable, dimensionally consistent predictions of the real noise signal, the present invention designs a specialized output module. The core of this module is a time-distributed fully connected layer, which serves to decode the abstract features and reconstruct the final noise prediction matrix.

[0082] Step 5, calculate the difference between the covariance matrix of the predicted noise signal and the covariance matrix of the real noise signal through the training optimization module to construct a spatial structure loss; calculate the point-by-point difference between the predicted noise signal and the real noise signal to construct a point-by-point precision loss; combine the spatial structure loss and the point-by-point precision loss by weighting to form a composite loss function, to guide the optimization of the noise reduction model, and obtain an optimized noise reduction model; input the noisy PPG signal into the optimized noise reduction model, predict and remove the noise to obtain the final reconstructed PPG signal.

[0083] In step 5, the difference between the covariance matrix of the predicted noise signal and the covariance matrix of the real noise signal is calculated through the training optimization module to construct a spatial structure loss, and the corresponding process exists the following relationship:

[0084] ;

[0085] wherein, denotes the spatial structure loss, denotes the covariance matrix calculation, denotes the square of the Frobenius norm, denotes the Frobenius norm, denotes the covariance matrix, denotes the signal the mean along the time dimension;

[0086] It needs to be explained that this is the core innovation of the application, which is used to make up for the defects of point-by-point precision loss and neglecting spatial correlation. It forces the model to learn and reproduce the internal correlation pattern of noise between multiple sensor channels by explicitly matching the covariance matrix of the predicted noise and the real noise. The traditional denoising model generally uses mean square error as the loss function. The essence of MSE is "point-by-point independent error measurement", which regards each data point in the multi-channel signal matrix as isolated, and only focuses on the difference in amplitude between the predicted value and the real value. This method ignores the physical reality and leads to structural distortion. The key of the application is that "reproducing the spatial structure of noise" is as important as "reducing point-by-point error", and a new technical solution is designed for this purpose.

[0087] In the step of calculating the point-by-point difference between the predicted noise signal and the real noise signal to construct the point-by-point precision loss, the corresponding process exists the following relationship:

[0088] ;

[0089] Wherein, represents the point-by-point precision loss;

[0090] In the step of using the spatial structure loss and the point-by-point precision loss to combine into a composite loss function for guiding optimization of the denoising model, the corresponding process exists the following relationship:

[0091] ;

[0092] Wherein, represents the composite loss function, represents the hyperparameter for balancing the spatial structure loss and the point-by-point precision loss.

[0093] It needs to be explained that after obtaining the predicted noise, the training process of the model is supervised and guided by a carefully designed composite loss function. The function aims to ensure that the prediction result approximates the real noise in both point-by-point precision and spatial structure fidelity. In the training process, the gradient of the composite loss function with respect to the parameters of each layer of the network is calculated by the back propagation algorithm, and the Adam optimizer is used to update the gradient iteratively. By minimizing , the model is effectively guided to generate predicted noise that is highly consistent with the real motion artifact in both numerical value and spatial statistical structure.

[0094] The noisy PPG signal is input into the optimized denoising model, the predicted noise is removed to obtain the final reconstructed PPG signal, and the corresponding process exists the following relationship:

[0095] ;

[0096] wherein, denotes the final reconstructed PPG signal, denotes the noisy PPG signal collected in real time in a new scene, denotes the predicted signal of the noise component in the newly collected signal by the optimized denoising model.

[0097] Referring to Figure 5 , the embodiment of the present application proposes a PPG denoising system based on sensor topology and noise structure, which comprises:

[0098] The construction module is configured to:

[0099] The denoising model is constituted based on the signal collection and graph construction module, the spatial-temporal feature extraction module, the noise prediction module, and the training optimization module.

[0100] The feature extraction module is configured to:

[0101] The PPG signal is collected, and the additive noise model is used to calculate the PPG signal to generate a real noise signal; the signal collection and graph construction module is used to construct an undirected graph according to the physical topology structure of the sensor array, and the noisy PPG signal in the PPG signal is represented as a node feature on the undirected graph; a cascaded graph convolution network is used to process the node feature on the undirected graph to obtain the spatial feature of the noise.

[0102] The spatial feature of the noise is subjected to feature aggregation and time series modeling by the stacked GRU network in the spatial-temporal feature extraction module to obtain the GRU network output.

[0103] The prediction module is configured to:

[0104] The GRU network output is input into the noise prediction module for decoding prediction to obtain a predicted noise signal.

[0105] The reconstruction module is configured to:

[0106] The difference between the covariance matrix of the predicted noise signal and the covariance matrix of the real noise signal is calculated and compared by the training optimization module to construct a spatial structure loss; the point-by-point difference between the predicted noise signal and the real noise signal is calculated to construct a point-by-point accuracy loss; the spatial structure loss and the point-by-point accuracy loss are combined by weighting to form a composite loss function, which is used to guide the optimization of the denoising model to obtain an optimized denoising model; the noisy PPG signal is input into the optimized denoising model to predict and remove the noise to obtain the final reconstructed PPG signal.

[0107] It should be appreciated that parts of the present application can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, several steps or methods can be implemented in software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, and as in another embodiment, any of the following technologies, or a combination thereof, can be used: discrete logic circuitry having logic gates for implementing logic functions upon an application of data signals, application specific integrated circuits having appropriate combinational logic gates, programmable gate arrays (PGA), field programmable gate arrays (FPGA), and the like.

[0108] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Also, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0109] The above-described embodiments only express several implementation manners of the present application, which are described in a more specific and detailed manner, but cannot be understood as a limitation on the patent scope of the present application. It should be noted that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present application, which are all within the protection scope of the present application. Therefore, the patent protection scope of the present application should be subject to the appended claims.

Claims

1. A PPG denoising method based on sensor topology and noise structure, characterized in that, The method comprises the following steps: Step 1, based on the signal acquisition and graph construction module, the space-time feature extraction module, the noise prediction module, and the training optimization module to constitute a noise reduction model; Step 2, collect PPG signals, calculate PPG signals using an additive noise model to generate a real noise signal; construct an undirected graph according to the physical topology structure of the sensor array through the signal acquisition and graph construction module, and express the noisy PPG signal in the PPG signal as a node feature on the undirected graph; adopt a cascaded graph convolution network to process the node feature on the undirected graph to obtain the spatial feature of the noise; Step 3, aggregate and time sequence model the spatial feature of the noise through the stacked GRU network in the space-time feature extraction module to obtain the GRU network output; Step 4, decode and predict the GRU network output through the noise prediction module to obtain a predicted noise signal, specifically including: input the hidden state sequence in the GRU network output into the noise prediction module to independently process each time step in the sequence, and extract the hidden state feature vector of all time steps, input the hidden state feature vector of all time steps into the fully connected layer in the noise prediction module for feature mapping to obtain the specific amplitude of all time steps on the sensor, and construct the specific amplitude of all time steps on the sensor into a predicted noise matrix, and the predicted noise matrix is taken as the predicted noise signal, and the relationship between the corresponding processes is as follows: ； wherein, denotes a sequence of hidden states in the GRU network output, denotes the total time steps and the feature dimension of the GRU hidden layer is a real set, denotes a predicted noise signal, denotes the number of sensors and the total time steps is a real set; Step 5, calculate and compare the difference between the covariance matrix of the predicted noise signal and the covariance matrix of the real noise signal through the training optimization module to construct a spatial structure loss; calculate the point-by-point difference between the predicted noise signal and the real noise signal to construct a point-by-point accuracy loss; combine the spatial structure loss and the point-by-point accuracy loss by weighting to form a composite loss function to guide the optimization of the noise reduction model, and obtain an optimized noise reduction model; input the noisy PPG signal into the optimized noise reduction model to predict and remove the noise to obtain the final reconstructed PPG signal; Wherein, the difference between the covariance matrix of the predicted noise signal and the covariance matrix of the real noise signal is calculated and compared through the training optimization module to construct a spatial structure loss, and the relationship between the corresponding processes is as follows: ； wherein, denotes the spatial structure loss, denotes the covariance matrix computation, denotes the square of the Frobenius norm, denotes the Frobenius norm, denotes the covariance matrix, denotes the signal the mean along the time dimension; In the step of calculating the point-by-point difference between the predicted noise signal and the real noise signal to construct a point-by-point accuracy loss, the relationship between the corresponding processes is as follows: ； wherein represents a point-wise loss of accuracy, represents a real noise signal; In the step of combining the spatial structure loss and the point-by-point accuracy loss by weighting to form a composite loss function to guide the optimization of the noise reduction model, the relationship between the corresponding processes is as follows: ； where, represents the composite loss function, represents a hyperparameter balancing the spatial structure loss and the point-wise accuracy loss.

2. The sensor topology and noise structure based PPG denoising method according to claim 1, wherein, In the step 2, collect PPG signals, calculate PPG signals using an additive noise model to generate a real noise signal; construct an undirected graph according to the physical topology structure of the sensor array through the signal acquisition and graph construction module, and express the noisy PPG signal in the PPG signal as a node feature on the undirected graph; adopt a cascaded graph convolution network to process the node feature on the undirected graph to obtain the spatial feature of the noise, specifically including the following steps: Collect pure PPG signals of users in a static state and noisy PPG signals of users in a motion state; Calculate the pure PPG signals from the noisy PPG signals to generate a real noise signal by using an additive noise model; Divide the noisy PPG signals and the real noise signal into fixed-length signal segments by using a sliding window technique; Abstract the physical adjacency relationship of the sensor array into an undirected graph by using a signal acquisition and graph construction module, and construct an original adjacency matrix according to the undirected graph; Add a self-loop to the original adjacency matrix and perform symmetric normalization processing to obtain a normalized adjacency matrix; Correlate the fixed-length signal segments with the normalized adjacency matrix, so that a signal slice at each time step in the signal segment is defined as a node feature on the normalized adjacency matrix; Send the node features on the normalized adjacency matrix as initial node features into a stacked two-layer graph convolution network for spatial feature extraction to obtain spatial features of the noise.

3. The sensor topology and noise structure based PPG denoising method according to claim 2, wherein, Calculate the pure PPG signals from the noisy PPG signals to generate a real noise signal by using an additive noise model, and the corresponding process exists a relationship as follows: ； wherein, denotes the noisy PPG signal, denotes the clean PPG signal; In the step of abstracting the physical adjacency relationship of the sensor array into an undirected graph, the corresponding process exists a relationship as follows: ； wherein, represents an undirected graph, represents a set of nodes, represents a set of edges; In the step of adding a self-loop to the original adjacency matrix and performing symmetric normalization processing to obtain a normalized adjacency matrix, the corresponding process exists a relationship as follows: ； wherein, denotes the normalized adjacency matrix, denotes taking the value of each diagonal element of the degree matrix to the power of denotes the result after adding a unit matrix of dimension number of sensors to the original adjacency matrix ;​​ In the step of sending the node features on the normalized adjacency matrix as initial node features into a stacked two-layer graph convolution network for spatial feature extraction to obtain spatial features of the noise, the corresponding process exists a relationship as follows: ； wherein, represents the final result of the computation of the layer graph convolution network, represents a non-linear activation function, represents the initial input features of the graph convolution network at time step represents the initial input features of the graph convolution network at time step represents a weight matrix learnable by the initial layer graph convolution network, represents the final result of the computation of the layer graph convolution network, represents a weight matrix learnable by the next layer graph convolution network, represents the spatial features of the noise.

4. The sensor topology and noise structure based PPG denoising method according to claim 3, wherein, In the step 3, the spatial features of the noise are aggregated and time-series modeled by a stacked GRU network in the spatio-temporal feature extraction module to obtain a GRU network output, which specifically includes the following steps: Perform average pooling processing on the spatial features of the noise to obtain a feature vector representing a global spatial state; Aggregate the feature vector representing the global spatial state into a feature sequence; Send the feature sequence into the stacked GRU network in the spatio-temporal feature extraction module to learn the dynamic evolution law of the noise pattern in the time dimension to obtain the GRU network output.

5. The sensor topology and noise structure based PPG denoising method according to claim 4, wherein, Perform average pooling processing on the spatial features of the noise to obtain a feature vector representing a global spatial state, and the corresponding process exists a relationship as follows: ； wherein, a feature vector representing a global spatial state, denotes an average pooling process; In the step of aggregating the feature vector representing the global spatial state into a feature sequence, the corresponding process exists a relationship as follows: ； wherein, denotes the characteristic sequence after the aggregation, denotes the global spatial state feature vector at the 1st time step, denotes the global spatial state feature vector at the total time step denotes the global spatial state feature vector at the total time step denotes the global spatial state feature vector at the total time step denotes the real set of the output feature dimension denotes the real set of the output feature dimension 6. The sensor topology and noise structure based PPG denoising method according to claim 1, wherein, Input the noisy PPG signal into the optimized denoising model to predict and remove the noise to obtain a final reconstructed PPG signal, and the corresponding process exists a relationship as follows: ； wherein, represents the final reconstructed PPG signal, represents the noisy PPG signal acquired in real-time in the new scenario, represents the predicted signal of the noise component in the newly acquired signal by the optimized denoising model. 7.A PPG denoising system based on sensor topology and noise structure, characterized in that, The system adopts the PPG denoising method based on the sensor topology and noise structure according to any one of claims 1-6, and the system comprises: The construction module is configured to: The denoising model is composed of the signal acquisition and graph construction module, the spatio-temporal feature extraction module, the noise prediction module, and the training optimization module; The feature extraction module is configured to: The PPG signal is collected, the PPG signal is calculated by using an additive noise model, and a real noise signal is generated; a signal collection and graph construction module is used to construct an undirected graph according to a physical topology structure of a sensor array, and a noisy PPG signal in the PPG signal is represented as a node feature on the undirected graph; a cascaded graph convolution network is used to process the node feature on the undirected graph, so as to obtain a spatial feature of the noise; The spatial feature of the noise is subjected to feature aggregation and time sequence modeling by a stacked GRU network in the space-time feature extraction module, so as to obtain a GRU network output; The prediction module is configured to: input the GRU network output into a noise prediction module for decoding prediction, so as to obtain a predicted noise signal; The reconstruction module is configured to: calculate and compare a difference between a covariance matrix of the predicted noise signal and a covariance matrix of the real noise signal by using the training optimization module, so as to construct a spatial structure loss; calculate a point-by-point difference between the predicted noise signal and the real noise signal, so as to construct a point-by-point precision loss; and combine the spatial structure loss and the point-by-point precision loss by using a weighted combination, so as to obtain a composite loss function, guide optimization of a denoising model, and obtain an optimized denoising model; input the noisy PPG signal into the optimized denoising model, predict and remove the noise, and obtain a final reconstructed PPG signal.

Images