Non-contact physiological signal waveform reconstruction method and system based on millimeter wave radar

Abstract

The invention provides a non-contact physiological signal waveform reconstruction method and system based on a millimeter wave radar. The method comprises the following steps: preprocessing an acquired original radar signal into a mixed physiological signal containing a pulse wave signal, a heart rate signal and a respiration signal; performing short-time Fourier transform on the mixed physiological information to obtain an input feature map, and extracting a high-level semantic feature map from the input feature map through an encoder; separating each independent physiological signal feature map from the high-level semantic feature map through a separator; a corresponding physiological signal waveform is reconstructed through an independent decoder corresponding to each physiological signal according to an independent physiological signal feature map and a high-level semantic feature map, a blood pressure decoder adopts a structural design symmetrical to a heart rate decoder and a respiration decoder, and the blood pressure decoder adopts a structural design symmetrical to the heart rate decoder and the respiration decoder; and the blood pressure decoder completes the reconstruction of the blood pressure waveform according to the analysis of the pulse wave signal. The continuous blood pressure waveform, the electrocardiogram waveform and the respiration waveform are jointly reconstructed through the millimeter wave radar.

Description

Technical Field

[0001] This invention relates to the fields of biomedical signal processing and wireless sensing technology, and in particular to a non-contact physiological signal waveform reconstruction method and system based on millimeter-wave radar. Background Technology

[0002] Continuous, non-invasive vital sign monitoring has wide applications in daily life and modern healthcare. By continuously monitoring vital signs such as heart rate variability and respiratory rate, an individual's emotional state can be inferred, thus enabling emotion recognition. Blood pressure and heart rate are crucial indicators of cardiovascular health. Regular, long-term monitoring of blood pressure and heart rate signals allows for the early detection and diagnosis of cardiovascular abnormalities, which is essential for timely intervention and treatment. Among the many key physiological parameters in the biomedical field, electrocardiogram (ECG), respiratory rate (RSP), and blood pressure (BP) signals are some of the most commonly used. The common method for monitoring ECG is using specialized medical ECG monitoring devices. Common methods for monitoring respiratory signals can be broadly divided into two types: one is the chest strap method, which measures changes in strap tension by placing a chest strap on the chest or abdomen to detect respiratory signals; the other is an impedance-based method, which monitors respiratory signals by measuring impedance changes caused by changes in abdominal and chest volume during respiration. Blood pressure signals are typically estimated through analysis of pulse wave waveforms. A pulse wave is a biological signal that reflects changes in blood pressure caused by the heart's contraction and relaxation as they propagate through arteries. Common pulse wave acquisition methods include photoplethysmography (PPG) and millimeter-wave (mmWave) radar. Compared to monitoring heart rate signals via electrocardiogram (ECG), monitoring respiratory signals via chest straps, and monitoring blood pressure signals via PPG, millimeter-wave radar can acquire chest movements caused by respiration and heartbeat, as well as pulse waves reflecting blood pressure signals, without contact by extracting micro-motion signals from the body surface. Recent advances in millimeter-wave radar sensing technology have demonstrated its superior ability to acquire mechanical cardiopulmonary signals and pulse waves non-contactly. The pioneering research of Wu et al. proposed the RSSRnet model, which can separate and reconstruct the waveforms of heart rate (ECG) and respiratory signals (RSP) from radar data.

[0003] Joint monitoring of these three waveforms is crucial for the early detection of potential cardiovascular risks and chronic respiratory diseases. Significant progress has been made in non-invasive monitoring of respiratory and heart rate signals. The main challenge in non-invasive joint monitoring of blood pressure, respiration, and heart rate signals lies in the non-invasive monitoring of blood pressure. Specifically, regarding non-invasive blood pressure monitoring, one approach involves indirect estimation of blood pressure via pulse arrival time (PAT) through complex vascular conduction mechanisms, which introduces significant errors. Another approach is based on millimeter-wave radar-based blood pressure measurement technology, which has also made considerable progress. Existing methods can be broadly categorized into two types: traditional machine learning methods and neural network methods. Traditional machine learning methods typically only obtain blood pressure values ​​and cannot capture the blood pressure waveform containing richer physiological information. With the development of deep learning technology, neural networks can automatically extract deep features from pulse waves. The advantage of neural network methods lies in their ability to convert input pulse wave signals into continuous blood pressure waveforms, thereby obtaining blood pressure variation curves under non-invasive conditions.

[0004] However, a key gap remains: current radar systems are inherently incapable of jointly reconstructing continuous blood pressure, electrocardiogram, and respiratory waveforms. Therefore, there is an urgent need for a non-contact joint reconstruction method that can effectively address the challenges of reconstructing BP, ECG, and RSP signals from millimeter-wave radar. Summary of the Invention

[0005] To address the aforementioned problems in the prior art, this invention provides a non-contact physiological signal waveform reconstruction method and system based on millimeter-wave radar, which can jointly reconstruct continuous blood pressure waveforms, electrocardiogram waveforms, and respiratory waveforms using millimeter-wave radar.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a non-contact physiological signal waveform reconstruction method based on millimeter-wave radar, comprising: Step S1: Preprocess the acquired raw radar signals into a mixed physiological signal containing pulse wave signal, heart rate signal and respiratory signal; Step S2: After the mixed physiological information is subjected to short-time Fourier transform, an input feature map is obtained. The high-level semantic feature map is extracted from the input feature map by an encoder. Step S3: Separate individual physiological signal feature maps from the high-level semantic feature map using a separator; Step S4: The corresponding physiological signal waveform is reconstructed based on the independent physiological signal feature map and high-level semantic feature map through the independent decoder corresponding to each physiological signal. The blood pressure decoder adopts a structural design symmetrical to the heart rate decoder and the respiration decoder, and the blood pressure decoder completes the reconstruction of the blood pressure waveform based on the analysis of the pulse wave signal.

[0007] The beneficial effects of this invention are as follows: the blood pressure waveform reconstruction branch adopts a structure symmetrical to the heart rate and respiratory waveform reconstruction branches. Experimental verification shows that the blood pressure waveform reconstruction branch can effectively model the pulse wave information contained in the radar signal, so as to complete the reconstruction of the blood pressure waveform through the analysis of the pulse wave, thereby jointly reconstructing the continuous blood pressure waveform, electrocardiogram waveform and respiratory waveform through millimeter wave radar.

[0008] Optionally, step S2 includes: By analyzing the frequency domain information of the input feature map in real time, a dynamic weighted convolution kernel is generated. The dynamic weighted convolution kernel is then used to perform grouped convolution operations on the input feature map to obtain the dynamically convolved convolutional feature map.

[0009] Optionally, in step S2, generating a dynamic weighted convolutional kernel by analyzing the frequency domain information of the input feature map in real time includes: Calculate the spectral entropy of each channel on the input feature map in the two-dimensional time-frequency plane, and enhance the spatial correlation of the spectral entropy of each channel through a lightweight CNN network to obtain the enhanced spectral entropy feature; The mean and variance statistics of the enhanced spectral entropy features are extracted along the frequency and time dimensions and then concatenated to form the spectral entropy statistics. The weight generation network, consisting of two linear layers, calculates the weight parameters of the convolution kernel in real time based on the spectral entropy statistics, thereby generating a dynamic weighted convolution kernel.

[0010] As described above, by using dynamic weighted convolution of spectral entropy, the weights of the convolution kernel are adaptively adjusted according to the complexity of the physiological signals of different individuals to enhance feature representation, thereby improving the accuracy of waveform reconstruction.

[0011] Optionally, the formula for calculating the spectral entropy is: ; In the formula, H(F,T) represents the spectral entropy; X(F,T) represents the frequency domain information of the input feature map; F represents the frequency dimension; T represents the time dimension; and ε is a constant.

[0012] Optionally, step S2 includes: The shuffled feature map is obtained by interleaving and recombining the data in each channel of the convolutional feature map.

[0013] Optionally, in step S2, the process of interleaving and recombining the data in each channel of the convolutional feature map to obtain the shuffled feature map includes: The C channels of the convolutional feature map are divided into g groups to obtain a first feature map with shape g×(C / g)×F×T, where F represents the frequency dimension and T represents the time dimension. The tensor of the first feature map is transposed to transform it into a tensor of (C / g)×g×F×T; The tensor transposed to (C / g)×g×F×T is reshaped into a tensor of C×F×T, and the final output is a shuffled feature map with the same number of channels (C) and spatial dimension (F×T).

[0014] As described above, by performing cross-level channel shuffling on the feature maps after each dynamic convolution, multi-scale feature interaction of the feature maps is achieved, breaking the spectral coupling.

[0015] Optionally, step S1 includes: The acquired raw radar signals were subjected to elliptic fitting calibration, phase demodulation, bandpass filtering, and resampling to extract a mixed physiological signal containing pulse wave, heart rate, and respiratory signals.

[0016] Optionally, step S3 includes: By using multiple Transformer encoder blocks in the separator, the high-level semantic feature map is analyzed using the multi-head attention mechanism and feedforward network in the Transformer encoder block, and the feature components belonging to each independent physiological signal are identified and tracked to obtain the feature map of each independent physiological signal.

[0017] Optionally, in step S4, the independent decoder for each physiological signal reconstructs the corresponding physiological signal waveform based on the independent physiological signal feature map and the high-level semantic feature map, including: The independent decoder fuses the received independent physiological signal feature maps and high-level semantic feature maps. After the fused physiological signal feature map is subjected to two-dimensional transposed convolution, normalization, and PRelu activation function processing, the final feature map is obtained. The final feature map is then reconstructed into a time-domain waveform by short-time Fourier transform to obtain the physiological signal waveform.

[0018] In a second aspect, the present invention provides a non-contact physiological signal waveform reconstruction system based on millimeter-wave radar, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the non-contact physiological signal waveform reconstruction method based on millimeter-wave radar of the first aspect.

[0019] The technical effects of the non-contact physiological signal waveform reconstruction system based on millimeter-wave radar provided in the second aspect are described in the relevant description of the non-contact physiological signal waveform reconstruction method based on millimeter-wave radar provided in the first aspect. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the main process of the non-contact physiological signal waveform reconstruction method based on millimeter-wave radar according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall model of the non-contact physiological signal waveform reconstruction method based on millimeter-wave radar according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating step S1 in an embodiment of the present invention; Figure 4 This is a flowchart illustrating step S3 in an embodiment of the present invention; Figure 5 This is a flowchart illustrating step S4 in an embodiment of the present invention; Figure 6 This is a schematic diagram of the two-dimensional dynamic weighted convolutional layer in step S2 of an embodiment of the present invention; Figure 7 This is a schematic diagram of the channel washing layer in step S2 of an embodiment of the present invention; Figure 8 This is a comparison diagram of the reconstructed signal and the true signal involved in the embodiments of the present invention; Figure 9 This is a graph of the cumulative distribution function of the PCC value of the reconstructed signal according to an embodiment of the present invention; Figure 10 This is a graph of the cumulative distribution function of the NRMSE value of the reconstructed signal according to an embodiment of the present invention; Figure 11 This is a schematic diagram of the structure of a non-contact physiological signal waveform reconstruction system based on millimeter-wave radar according to an embodiment of the present invention.

[0021] Explanation of reference numerals in the attached figures: 1: A non-contact physiological signal waveform reconstruction system based on millimeter-wave radar; 2: Processor; 3: Memory. Detailed Implementation

[0022] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.

[0023] Example 1 In this embodiment, Figures 2 to 10 Some of the English definitions are described in the example description; the rest are as follows: 1. Input IQ Signal: The in-phase and quadrature input signals.

[0024] 2. Input: Input.

[0025] 3. Grouping: Grouping.

[0026] 4. Dimension replacement: Dimension transpose.

[0027] 5. Restore: Rebuild.

[0028] 6. Output: Output.

[0029] 7. Groundtruth BP / ECG / RSP: True value signals of blood pressure / ECG / respiratory signals.

[0030] 8. Reconstructed BP / ECG / RSP: Reconstructed signals of blood pressure / ECG / respiratory signals.

[0031] Please refer to Figures 1 to 10 A non-contact physiological signal waveform reconstruction method based on millimeter-wave radar includes the following steps: Step S1: Preprocess the acquired raw radar signals into a mixed physiological signal containing pulse wave signal, heart rate signal and respiratory signal.

[0032] Specifically, the radar signal acquisition employs a 24 GHz continuous wave radar system based on six-port technology. This radar system includes two input ports and four output ports, where the input signals are combined with four preset and fixed phase shifts. The resulting output signals are then down-converted to baseband signals via a diode power detector. Therefore, the obtained baseband signals accurately represent the quadrature-in-phase (I) and quadrature (Q) components in the complex number. The phase change Δφ between the transmitted (Tx) and received (Rx) signals, and the thoracic motion Δx caused by the target's breathing and cardiac activity, can be expressed by the following formula: Δx = 4π(Δφ) / λ; In the formula, λ represents the wavelength of the millimeter-wave radar signal.

[0033] At this point, the acquired raw radar signal is mixed with chest displacement signals from heartbeat and respiration, pulse wave signals, high-frequency noise, and other interference. To obtain a high-quality chest displacement signal with mixed heartbeat, respiration, and pulse wave components, in this embodiment, as follows... Figure 2 As shown, an end-to-end deep learning model (Spectral-Entropy Routed Dynamic Convolution, SERDC) based on millimeter-wave radar signals is proposed.

[0034] like Figure 3 As shown, step S1 includes: The acquired raw radar signals were subjected to elliptic fitting calibration, phase demodulation, bandpass filtering, and resampling to extract a mixed physiological signal containing pulse wave, heart rate, and respiratory signals.

[0035] In actual hardware implementations, IQ signals often exhibit amplitude or phase mismatches. Therefore, the IQ signals must be calibrated before phase demodulation. This embodiment employs an ellipse fitting method to calibrate the IQ signals. Ellipse fitting methods include, but are not limited to, least squares fitting, gradient descent, and the Statistical Random Sample Consensus (RANSAC) algorithm.

[0036] After elliptic fitting of the IQ signals, in order to obtain the micro-motion information of the body surface contained in the phase of the IQ signals, this embodiment uses the differential and cross-multiplication algorithm (DACM) to achieve phase demodulation. The DAM algorithm process can be represented by the following formula: ; In the formula, k represents the index of the discrete time series, which is used to traverse the sampling points of the radar IQ signal.

[0037] Next, a fourth-order Butterworth filter with a bandpass range of 0.1-16 Hz is used to filter the phase signal to remove non-respiratory, non-heartbeat, and non-pulse wave components. For the reference signals, this embodiment only performs simple bandpass filtering (BPF) processing. Specifically, a fourth-order Butterworth filter with a bandpass range of 1-16 Hz is used to eliminate physiological and electrical noise in the ECG and BP signals, and the impedance-respiratory response (Z0) signal is processed using a fourth-order Butterworth filter with a bandpass range of 0.1-0.6 Hz. This frequency range is consistent with the typical human respiratory frequency range. After bandpass filtering, this embodiment can obtain a high-quality, high signal-to-noise ratio reference signal. Finally, all signals are resampled to 100 Hz.

[0038] Step S2: After the mixed physiological information is subjected to short-time Fourier transform, the input feature map is obtained. The high-level semantic feature map is extracted from the input feature map by the encoder.

[0039] In this embodiment, step S2 includes: S21. The mixed physiological information is transformed by a short-time Fourier transform to obtain the input feature map.

[0040] The mixed physiological information, after undergoing a Short-Time Fourier Transform (STFT), yields an input feature map with a shape of B×C×F×T. Here, C is 2, representing the amplitude and phase frequencies after the STFT, respectively; F is the frequency dimension (in this embodiment, only the first 64 frequency bins are used as the input to the encoder); and T is the time dimension.

[0041] like Figure 2 As can be seen, the encoder part of SERDCnet consists of four encoder blocks. Each encoder block consists of a two-dimensional dynamic weighted convolutional layer with spectral entropy routing, a channel shuffling layer, a normalization layer (BatchNorm), a channel attention layer (CA), and a PRelu activation function layer. After passing through each encoder block, the input feature map is upsampled by a factor of 2 in the channel dimension and downsampled by a factor of 2 in the frequency dimension.

[0042] S22. By analyzing the frequency domain information of the input feature map in real time, a dynamic weighted convolution kernel is generated. The dynamic weighted convolution kernel is used to perform grouped convolution operations on the input feature map to obtain the convolved feature map after dynamic convolution.

[0043] Physiological signals exhibit significant non-stationary characteristics, with energy distribution across different frequency bands dynamically changing over time. Traditional static convolutional kernels cannot adapt to this variation, leading to insufficient extraction of key features. This invention proposes a Spectral-Entropy Routed Dynamic 2D Convolution (SERDy 2-DConv) layer, referencing... Figure 6 Therefore, step S22 specifically includes: S221. Calculate the spectral entropy of each channel on the input feature map in the two-dimensional time-frequency plane. Use a lightweight CNN network to enhance the spatial correlation of the spectral entropy of each channel to obtain the enhanced spectral entropy feature.

[0044] The formula for calculating the spectral entropy is as follows: ; In the formula, H(F,T) represents the spectral entropy; X(F,T) represents the frequency domain information of the input feature map; F represents the frequency dimension; T represents the time dimension; and ε is a constant.

[0045] S222. Extract the mean and variance statistics of the enhanced spectral entropy features along the frequency and time dimensions, and then concatenate them into a spectral entropy statistic.

[0046] The mean and variance statistics are concatenated to form the spectral entropy statistic (mip), which serves as the basis for generating the dynamic weight generation network.

[0047] S223. A weight generation network consisting of two linear layers calculates the weight parameters of the convolution kernel in real time based on the spectral entropy statistics to generate a dynamic weight convolution kernel.

[0048] ReLU is used as the activation function between the two linear layers.

[0049] S224. Use dynamic weighted convolution kernels to perform grouped convolution operations on the input feature map to obtain the dynamically convolved convolutional feature map.

[0050] In this embodiment, group convolution (Group Conv) is used to efficiently achieve independent convolution within a batch. Therefore, this embodiment uses dynamic weighted convolution based on spectral entropy to adaptively adjust the convolution kernel weights according to the complexity of the physiological signals of different individuals, thereby enhancing feature representation and improving the accuracy of waveform reconstruction.

[0051] S23. After interleaving and recombining the data in each channel of the convolutional feature map, a shuffled feature map is obtained.

[0052] In this embodiment, physiological signals exhibit inherent coupling in the time-frequency domain. The respiratory sinus arrhythmia (RSA) phenomenon indicates that the respiratory cycle modulates the heart rhythm. This cross-modal dynamic correlation is difficult to model using only convolutional layers. Figure 7 As shown, this embodiment uses a channel shuffle layer to decouple the fixed connections within feature groups and force cross-group interaction of features. The channel shuffle grouping employs an adaptive grouping strategy, setting a lower limit for the number of groups while using the number of channels in the input feature map as the basis for group division. After channel shuffling, the feature map is then input into a lightweight 1×1 convolutional layer to enhance cross-group feature fusion. Specifically, step S23 includes: S231. Divide the C channels of the convolutional feature map into g groups to obtain a first feature map with shape g×(C / g)×F×T, where F represents the frequency dimension and T represents the time dimension.

[0053] S232. Transpose the tensor of the first feature map into a tensor of (C / g)×g×F×T.

[0054] In this step, the original grouping is changed to grouping by channel index unit within a group. Now, the first dimension of the tensor is the "channel index within the group", and the second dimension is the "group number". This is equivalent to extracting channels at the same internal position in different groups and putting them together.

[0055] S233. Reshape the tensor transposed to (C / g)×g×F×T into a tensor of C×F×T, and finally output a shuffled feature map with the same number of channels (C) and spatial dimension (F×T).

[0056] In this approach, channels from different groups are arranged in an alternating manner. For example, the first g positions of the output channel now come from the first channel of the first group, the first channel of the second group, ..., the first channel of the g-th group of the original input, respectively. Ultimately, the output shuffled feature map and its spatial dimensions remain unchanged, but the channel order has been shuffled, achieving thorough mixing of features from different groups. Therefore, this embodiment achieves multi-scale feature interaction of the feature map and breaks spectral coupling by performing cross-level channel shuffling on the feature map after each dynamic convolution.

[0057] S24. After the shuffled feature map is processed sequentially through a normalization layer, a channel attention layer, and a PRelu activation function layer, a high-level semantic feature map is obtained.

[0058] Step S3: Separate individual physiological signal feature maps from the high-level semantic feature map using a separator.

[0059] like Figure 4 As can be seen, the separator in this embodiment is composed of several transformer encoder blocks. The separator receives the output of the encoder and uses a self-attention mechanism to track different components in the feature map belonging to blood pressure, heart rate, and respiratory signals respectively and establish long-term dependencies on them. That is, step S3 includes: By using multiple Transformer encoder blocks in the separator, the high-level semantic feature maps are analyzed using the multi-head attention mechanism and feed-forward network in the Transformer encoder blocks, and the feature components belonging to each independent physiological signal are identified and tracked to obtain the feature maps of each independent physiological signal.

[0060] Step S4: The corresponding physiological signal waveform is reconstructed based on the independent physiological signal feature map and high-level semantic feature map through the independent decoder corresponding to each physiological signal. The blood pressure decoder adopts a structural design that is symmetrical with the heart rate decoder and the respiration decoder, and the blood pressure decoder completes the reconstruction of the blood pressure waveform based on the analysis of the pulse wave signal.

[0061] In this embodiment, the decoder of the SERDCnet model receives independent physiological feature maps processed by the separator and high-level semantic feature maps skipped from the same layer of the encoder via cross-channel attention (CCA). The decoder consists of four decoder blocks, operating in reverse order of the encoder. Specifically, as follows... Figure 5 As shown, in step S4, the independent decoder for each physiological signal reconstructs the corresponding physiological signal waveform based on the independent physiological signal feature map and the high-level semantic feature map, including: The independent decoder fuses the received independent physiological signal feature maps and high-level semantic feature maps. After the fused physiological signal feature maps are processed by 2-D ConvTranspose, normalization, and PRelu activation function, the final feature map is obtained. The final feature map is then reconstructed from the time-domain waveform by short-time Fourier transform to obtain the physiological signal waveform.

[0062] Therefore, the blood pressure waveform reconstruction branch in this embodiment adopts a structure symmetrical to the heart rate and respiratory waveform reconstruction branches. Experimental verification shows that the blood pressure waveform reconstruction branch can effectively model the pulse wave information contained in the radar signal, so as to complete the reconstruction of the blood pressure waveform through the analysis of the pulse wave, thereby jointly reconstructing the continuous blood pressure waveform, electrocardiogram waveform and respiratory waveform through millimeter wave radar.

[0063] To verify the effectiveness of this embodiment, it will be demonstrated on the public dataset provided by Schellenberger et al., and the performance of the improved module of this invention will be evaluated through experiments.

[0064] This dataset was obtained through a carefully controlled data collection process conducted by healthcare professionals in the Department of Palliative Medicine at Erlangen University Hospital. The study recruited 30 participants. During the experiment, subjects lay on a table. A 24 GHz continuous wave radar was mounted approximately 40 cm above the skin of the subjects' chests. The two raw signal components, I and Q, were simultaneously digitized using a 24-bit analog-to-digital converter with a sampling frequency of 2000 Hz.

[0065] In addition to the radar system, a reference system was used to record various contact signals. The sampling rates for ECG, RSP, and BP signals were 1000Hz, 100Hz, and 200Hz, respectively. In this embodiment, well-labeled data without significant errors were selected from 30 subjects in a normal resting setting for the experiment. During the preprocessing stage, the data samples in each data pair were resampled to 100Hz and divided into 8-second segments, with a sampling step size set to 20% of this length. Furthermore, the min-max normalization method was applied to normalize each data pair.

[0066] To evaluate the performance of this embodiment, ablation experiments were conducted for each improvement. In this embodiment, all collected data samples were randomly shuffled, with 80% used for training and validation, and the remaining 20% ​​used for testing. During training, the ratio of training set to validation set samples was 4:1.

[0067] Throughout the training process, this embodiment uses an Adam optimizer with a batch size of 20, a learning rate of 1e-4, β1=0.9, and β2=0.9 to minimize the loss function and update the network parameters.

[0068] To accurately evaluate the model's performance, this embodiment uses PCC and NRMSE as metrics to measure the accuracy of the reconstructed ECG, RSP, and BP signals.

[0069] The Pearson correlation coefficient (PCC) measures the degree of covariance between the reconstructed signal and its corresponding ground truth signal. The coefficient ranges from -1 to 1, where the magnitude represents the strength of the correlation, and the sign determines whether it is positive or negative. The mathematical expression for PCC can be represented by the following formula: ; In the formula, x represents the reconstructed signal, and x' represents the true signal. The Euclidean norm is used... This indicates that the mean function is used... express.

[0070] Another metric used in this embodiment is the Normalized Root Mean Square Error (NRMSE), which measures the difference between the reconstructed signal and the true signal. A smaller NRMSE value indicates a better match between the predicted and actual values. Normalization in the formula ensures that NRMSE is scale-independent. The mathematical expression for NRMSE is as follows: ; In the formula, N represents the sample size, x max and x min These represent the maximum and minimum values ​​of the signal, respectively.

[0071] To fully evaluate the effectiveness of this embodiment, corresponding ablation experiments were conducted on each of the improved modules.

[0072] 1) Baseline: Basic three-branch waveform reconstruction model. The purpose of this experiment is to verify the feasibility of the blood pressure reconstruction branch, evaluate the impact of adding the blood pressure reconstruction branch on ECG / RSP reconstruction, quantify the initial performance of blood pressure waveform reconstruction, and test the architectural compatibility of the three-signal joint reconstruction.

[0073] 2) DynamicConv: A three-branch reconstruction model with only the addition of a spectral entropy dynamic weighted convolutional layer was used to verify the performance improvement of the dynamic weighted convolutional layer for three-branch signal reconstruction.

[0074] 3) Channel-Shuf: A three-branch reconstruction model with only channel shuffling added to evaluate the ability of channel shuffling to model cross-modal correlations and its contribution to improving signal reconstruction performance.

[0075] 4) Proposed: The complete model of the invention in this embodiment, evaluating the performance of the model proposed in this embodiment in the joint reconstruction task of three-branch signals and the synergistic effect of dynamic convolution and channel shuffling in feature extraction.

[0076] Depend on Figure 8 As shown, the baseline three-branch reconstruction model demonstrates the rationality of the extended blood pressure branch waveform reconstruction in this embodiment. However, the baseline model's reconstruction effect on ECG signals deteriorates, indicating that the static two-dimensional convolution in the encoder does not sufficiently extract physiological signal features, resulting in mutual inhibition between multi-branch signals.

[0077] Table 1 shows the performance improvement of each part compared to the baseline model. The results indicate that the model proposed in this embodiment achieves superior overall performance in the three-branch signal reconstruction task. It achieves high PCC values ​​(0.963±0.032 and 0.985±0.052) for both heart rate and respiratory signal reconstruction, and very low NRMSE values ​​(0.033±0.017 and 0.036±0.030), respectively. This demonstrates that the model proposed in this embodiment has small reconstruction errors and is accurate. Although the baseline model achieved the best performance in blood pressure reconstruction with a PCC of 0.982±0.021, its ECG reconstruction performance was the worst, with a reconstructed signal PCC value of only 0.930±0.038.

[0078] As shown in Table 1, each improvement to the model in this embodiment has achieved a comprehensive performance improvement. Although the performance of each improvement is slightly lower than that of the baseline model in terms of blood pressure waveform reconstruction, the overall performance is better than that of the baseline model. Through a comprehensive comparison of different improvement modules, it is determined that the extended blood pressure waveform reconstruction branch, the dynamic weighted convolution of spectral entropy routing, and the channel shuffling in this embodiment are all effective.

[0079] Table 1. Comparison of performance of different models

[0080] To evaluate the performance of the model proposed in this embodiment on the overall test set, the cumulative distribution function curves of PCC and NRMSE were plotted. (Observation) Figure 9 It can be seen that in this embodiment, more than 90% of the PCC values ​​are higher than 0.9. Figure 9 As can be seen, more than 90% of the NRMSE values ​​in this embodiment are below 0.1, indicating that the model proposed in this embodiment has achieved advanced performance in three-branch signal reconstruction.

[0081] Example 2 Please refer to Figure 11 The non-contact physiological signal waveform reconstruction system 1 based on millimeter-wave radar includes a memory 3, a processor 2, and a computer program stored in the memory 3 and run on the processor 2. When the processor 2 executes the computer program, it implements the steps in the above embodiment 1.

[0082] Since the systems / devices described in the above embodiments of the present invention are systems / devices used to implement the methods of the above embodiments of the present invention, those skilled in the art can understand the specific structure and modifications of the systems / devices based on the methods described in the above embodiments of the present invention, and therefore will not be repeated here. All systems / devices used in the methods of the above embodiments of the present invention fall within the scope of protection of the present invention.

[0083] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0084] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions.

[0085] It should be noted that any reference numerals placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In claims that enumerate several means, several of these means may be embodied by the same hardware. The use of the terms first, second, third, etc., is merely for convenience of expression and does not indicate any order. These terms can be understood as part of the component names.

[0086] Furthermore, it should be noted that in the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0087] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the claims should be interpreted to include both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0088] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, then this invention should also include these modifications and variations.

Claims

1. A non-contact physiological signal waveform reconstruction method based on millimeter-wave radar, characterized in that, include: Step S1: Preprocess the acquired raw radar signals into a mixed physiological signal containing pulse wave signal, heart rate signal and respiratory signal; Step S2: After the mixed physiological information is subjected to short-time Fourier transform, an input feature map is obtained. The high-level semantic feature map is extracted from the input feature map by an encoder. Step S3: Separate individual physiological signal feature maps from the high-level semantic feature map using a separator; Step S4: The corresponding physiological signal waveform is reconstructed based on the independent physiological signal feature map and high-level semantic feature map through the independent decoder corresponding to each physiological signal. The blood pressure decoder adopts a structural design symmetrical to the heart rate decoder and the respiration decoder, and the blood pressure decoder completes the reconstruction of the blood pressure waveform based on the analysis of the pulse wave signal.

2. The non-contact physiological signal waveform reconstruction method based on millimeter-wave radar according to claim 1, characterized in that, Step S2 includes: By analyzing the frequency domain information of the input feature map in real time, a dynamic weighted convolution kernel is generated. The dynamic weighted convolution kernel is then used to perform grouped convolution operations on the input feature map to obtain the dynamically convolved convolutional feature map.

3. The non-contact physiological signal waveform reconstruction method based on millimeter-wave radar according to claim 2, characterized in that, In step S2, generating a dynamic weighted convolution kernel by analyzing the frequency domain information of the input feature map in real time includes: Calculate the spectral entropy of each channel on the input feature map in the two-dimensional time-frequency plane, and enhance the spatial correlation of the spectral entropy of each channel through a lightweight CNN network to obtain the enhanced spectral entropy feature; The mean and variance statistics of the enhanced spectral entropy features are extracted along the frequency and time dimensions and then concatenated to form the spectral entropy statistics. The weight generation network, consisting of two linear layers, calculates the weight parameters of the convolution kernel in real time based on the spectral entropy statistics, thereby generating a dynamic weighted convolution kernel.

4. The non-contact physiological signal waveform reconstruction method based on millimeter-wave radar according to claim 3, characterized in that, The formula for calculating the spectral entropy is: ； In the formula, H(F,T) represents the spectral entropy; X(F,T) represents the frequency domain information of the input feature map; F represents the frequency dimension; T represents the time dimension; and ε is a constant.

5. The non-contact physiological signal waveform reconstruction method based on millimeter-wave radar according to claim 1, characterized in that, Step S2 includes: The shuffled feature map is obtained by interleaving and recombining the data in each channel of the convolutional feature map.

6. The non-contact physiological signal waveform reconstruction method based on millimeter-wave radar according to claim 5, characterized in that, In step S2, the data in each channel of the convolutional feature map are interleaved and recombined to obtain the shuffled feature map, which includes: The C channels of the convolutional feature map are divided into g groups to obtain a first feature map with shape g×(C / g)×F×T, where F represents the frequency dimension and T represents the time dimension. The tensor of the first feature map is transposed to transform it into a tensor of (C / g)×g×F×T; The tensor transposed to (C / g)×g×F×T is reshaped into a tensor of C×F×T, and the final output is a shuffled feature map with the same number of channels (C) and spatial dimension (F×T).

7. The non-contact physiological signal waveform reconstruction method based on millimeter-wave radar according to any one of claims 1 to 6, characterized in that, Step S1 includes: The acquired raw radar signals were subjected to elliptic fitting calibration, phase demodulation, bandpass filtering, and resampling to extract a mixed physiological signal containing pulse wave, heart rate, and respiratory signals.

8. The non-contact physiological signal waveform reconstruction method based on millimeter-wave radar according to any one of claims 1 to 6, characterized in that, Step S3 includes: By using multiple Transformer encoder blocks in the separator, the high-level semantic feature map is analyzed using the multi-head attention mechanism and feedforward network in the Transformer encoder block, and the feature components belonging to each independent physiological signal are identified and tracked to obtain the feature map of each independent physiological signal.

9. The non-contact physiological signal waveform reconstruction method based on millimeter-wave radar according to any one of claims 1 to 6, characterized in that, In step S4, the independent decoder for each physiological signal reconstructs the corresponding physiological signal waveform based on the independent physiological signal feature map and the high-level semantic feature map, including: The independent decoder fuses the received independent physiological signal feature maps and high-level semantic feature maps. After the fused physiological signal feature map is subjected to two-dimensional transposed convolution, normalization, and PRelu activation function processing, the final feature map is obtained. The final feature map is then reconstructed into a time-domain waveform by short-time Fourier transform to obtain the physiological signal waveform.

10. A non-contact physiological signal waveform reconstruction system based on millimeter-wave radar, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the non-contact physiological signal waveform reconstruction method based on millimeter-wave radar as described in any one of claims 1 to 9.

Images